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 2329721fd82SBaolin Wang #ifdef CONFIG_SPARSEMEM 2339721fd82SBaolin Wang /* 2349721fd82SBaolin Wang * If the PFN falls into an offline section, return the start PFN of the 2359721fd82SBaolin Wang * next online section. If the PFN falls into an online section or if 2369721fd82SBaolin Wang * there is no next online section, return 0. 2379721fd82SBaolin Wang */ 2389721fd82SBaolin Wang static unsigned long skip_offline_sections(unsigned long start_pfn) 2399721fd82SBaolin Wang { 2409721fd82SBaolin Wang unsigned long start_nr = pfn_to_section_nr(start_pfn); 2419721fd82SBaolin Wang 2429721fd82SBaolin Wang if (online_section_nr(start_nr)) 2439721fd82SBaolin Wang return 0; 2449721fd82SBaolin Wang 2459721fd82SBaolin Wang while (++start_nr <= __highest_present_section_nr) { 2469721fd82SBaolin Wang if (online_section_nr(start_nr)) 2479721fd82SBaolin Wang return section_nr_to_pfn(start_nr); 2489721fd82SBaolin Wang } 2499721fd82SBaolin Wang 2509721fd82SBaolin Wang return 0; 2519721fd82SBaolin Wang } 252e6e0c767SBaolin Wang 253e6e0c767SBaolin Wang /* 254e6e0c767SBaolin Wang * If the PFN falls into an offline section, return the end PFN of the 255e6e0c767SBaolin Wang * next online section in reverse. If the PFN falls into an online section 256e6e0c767SBaolin Wang * or if there is no next online section in reverse, return 0. 257e6e0c767SBaolin Wang */ 258e6e0c767SBaolin Wang static unsigned long skip_offline_sections_reverse(unsigned long start_pfn) 259e6e0c767SBaolin Wang { 260e6e0c767SBaolin Wang unsigned long start_nr = pfn_to_section_nr(start_pfn); 261e6e0c767SBaolin Wang 262e6e0c767SBaolin Wang if (!start_nr || online_section_nr(start_nr)) 263e6e0c767SBaolin Wang return 0; 264e6e0c767SBaolin Wang 265e6e0c767SBaolin Wang while (start_nr-- > 0) { 266e6e0c767SBaolin Wang if (online_section_nr(start_nr)) 267e6e0c767SBaolin Wang return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION; 268e6e0c767SBaolin Wang } 269e6e0c767SBaolin Wang 270e6e0c767SBaolin Wang return 0; 271e6e0c767SBaolin Wang } 2729721fd82SBaolin Wang #else 2739721fd82SBaolin Wang static unsigned long skip_offline_sections(unsigned long start_pfn) 2749721fd82SBaolin Wang { 2759721fd82SBaolin Wang return 0; 2769721fd82SBaolin Wang } 277e6e0c767SBaolin Wang 278e6e0c767SBaolin Wang static unsigned long skip_offline_sections_reverse(unsigned long start_pfn) 279e6e0c767SBaolin Wang { 280e6e0c767SBaolin Wang return 0; 281e6e0c767SBaolin Wang } 2829721fd82SBaolin Wang #endif 2839721fd82SBaolin Wang 284bb13ffebSMel Gorman /* 2852271b016SHui Su * Compound pages of >= pageblock_order should consistently be skipped until 286b527cfe5SVlastimil Babka * released. It is always pointless to compact pages of such order (if they are 287b527cfe5SVlastimil Babka * migratable), and the pageblocks they occupy cannot contain any free pages. 28821dc7e02SDavid Rientjes */ 289b527cfe5SVlastimil Babka static bool pageblock_skip_persistent(struct page *page) 29021dc7e02SDavid Rientjes { 291b527cfe5SVlastimil Babka if (!PageCompound(page)) 29221dc7e02SDavid Rientjes return false; 293b527cfe5SVlastimil Babka 294b527cfe5SVlastimil Babka page = compound_head(page); 295b527cfe5SVlastimil Babka 296b527cfe5SVlastimil Babka if (compound_order(page) >= pageblock_order) 29721dc7e02SDavid Rientjes return true; 298b527cfe5SVlastimil Babka 299b527cfe5SVlastimil Babka return false; 30021dc7e02SDavid Rientjes } 30121dc7e02SDavid Rientjes 302e332f741SMel Gorman static bool 303e332f741SMel Gorman __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, 304e332f741SMel Gorman bool check_target) 305e332f741SMel Gorman { 306e332f741SMel Gorman struct page *page = pfn_to_online_page(pfn); 3076b0868c8SMel Gorman struct page *block_page; 308e332f741SMel Gorman struct page *end_page; 309e332f741SMel Gorman unsigned long block_pfn; 310e332f741SMel Gorman 311e332f741SMel Gorman if (!page) 312e332f741SMel Gorman return false; 313e332f741SMel Gorman if (zone != page_zone(page)) 314e332f741SMel Gorman return false; 315e332f741SMel Gorman if (pageblock_skip_persistent(page)) 316e332f741SMel Gorman return false; 317e332f741SMel Gorman 318e332f741SMel Gorman /* 319e332f741SMel Gorman * If skip is already cleared do no further checking once the 320e332f741SMel Gorman * restart points have been set. 321e332f741SMel Gorman */ 322e332f741SMel Gorman if (check_source && check_target && !get_pageblock_skip(page)) 323e332f741SMel Gorman return true; 324e332f741SMel Gorman 325e332f741SMel Gorman /* 326e332f741SMel Gorman * If clearing skip for the target scanner, do not select a 327e332f741SMel Gorman * non-movable pageblock as the starting point. 328e332f741SMel Gorman */ 329e332f741SMel Gorman if (!check_source && check_target && 330e332f741SMel Gorman get_pageblock_migratetype(page) != MIGRATE_MOVABLE) 331e332f741SMel Gorman return false; 332e332f741SMel Gorman 3336b0868c8SMel Gorman /* Ensure the start of the pageblock or zone is online and valid */ 3346b0868c8SMel Gorman block_pfn = pageblock_start_pfn(pfn); 335a2e9a5afSVlastimil Babka block_pfn = max(block_pfn, zone->zone_start_pfn); 336a2e9a5afSVlastimil Babka block_page = pfn_to_online_page(block_pfn); 3376b0868c8SMel Gorman if (block_page) { 3386b0868c8SMel Gorman page = block_page; 3396b0868c8SMel Gorman pfn = block_pfn; 3406b0868c8SMel Gorman } 3416b0868c8SMel Gorman 3426b0868c8SMel Gorman /* Ensure the end of the pageblock or zone is online and valid */ 343a2e9a5afSVlastimil Babka block_pfn = pageblock_end_pfn(pfn) - 1; 3446b0868c8SMel Gorman block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); 3456b0868c8SMel Gorman end_page = pfn_to_online_page(block_pfn); 3466b0868c8SMel Gorman if (!end_page) 3476b0868c8SMel Gorman return false; 3486b0868c8SMel Gorman 349e332f741SMel Gorman /* 350e332f741SMel Gorman * Only clear the hint if a sample indicates there is either a 351e332f741SMel Gorman * free page or an LRU page in the block. One or other condition 352e332f741SMel Gorman * is necessary for the block to be a migration source/target. 353e332f741SMel Gorman */ 354e332f741SMel Gorman do { 355e332f741SMel Gorman if (check_source && PageLRU(page)) { 356e332f741SMel Gorman clear_pageblock_skip(page); 357e332f741SMel Gorman return true; 358e332f741SMel Gorman } 359e332f741SMel Gorman 360e332f741SMel Gorman if (check_target && PageBuddy(page)) { 361e332f741SMel Gorman clear_pageblock_skip(page); 362e332f741SMel Gorman return true; 363e332f741SMel Gorman } 364e332f741SMel Gorman 365e332f741SMel Gorman page += (1 << PAGE_ALLOC_COSTLY_ORDER); 366a2e9a5afSVlastimil Babka } while (page <= end_page); 367e332f741SMel Gorman 368e332f741SMel Gorman return false; 369e332f741SMel Gorman } 370e332f741SMel Gorman 37121dc7e02SDavid Rientjes /* 372bb13ffebSMel Gorman * This function is called to clear all cached information on pageblocks that 373bb13ffebSMel Gorman * should be skipped for page isolation when the migrate and free page scanner 374bb13ffebSMel Gorman * meet. 375bb13ffebSMel Gorman */ 37662997027SMel Gorman static void __reset_isolation_suitable(struct zone *zone) 377bb13ffebSMel Gorman { 378e332f741SMel Gorman unsigned long migrate_pfn = zone->zone_start_pfn; 3796b0868c8SMel Gorman unsigned long free_pfn = zone_end_pfn(zone) - 1; 380e332f741SMel Gorman unsigned long reset_migrate = free_pfn; 381e332f741SMel Gorman unsigned long reset_free = migrate_pfn; 382e332f741SMel Gorman bool source_set = false; 383e332f741SMel Gorman bool free_set = false; 384e332f741SMel Gorman 385e332f741SMel Gorman if (!zone->compact_blockskip_flush) 386e332f741SMel Gorman return; 387bb13ffebSMel Gorman 38862997027SMel Gorman zone->compact_blockskip_flush = false; 389bb13ffebSMel Gorman 390e332f741SMel Gorman /* 391e332f741SMel Gorman * Walk the zone and update pageblock skip information. Source looks 392e332f741SMel Gorman * for PageLRU while target looks for PageBuddy. When the scanner 393e332f741SMel Gorman * is found, both PageBuddy and PageLRU are checked as the pageblock 394e332f741SMel Gorman * is suitable as both source and target. 395e332f741SMel Gorman */ 396e332f741SMel Gorman for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, 397e332f741SMel Gorman free_pfn -= pageblock_nr_pages) { 398bb13ffebSMel Gorman cond_resched(); 399bb13ffebSMel Gorman 400e332f741SMel Gorman /* Update the migrate PFN */ 401e332f741SMel Gorman if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && 402e332f741SMel Gorman migrate_pfn < reset_migrate) { 403e332f741SMel Gorman source_set = true; 404e332f741SMel Gorman reset_migrate = migrate_pfn; 405e332f741SMel Gorman zone->compact_init_migrate_pfn = reset_migrate; 406e332f741SMel Gorman zone->compact_cached_migrate_pfn[0] = reset_migrate; 407e332f741SMel Gorman zone->compact_cached_migrate_pfn[1] = reset_migrate; 408bb13ffebSMel Gorman } 40902333641SVlastimil Babka 410e332f741SMel Gorman /* Update the free PFN */ 411e332f741SMel Gorman if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && 412e332f741SMel Gorman free_pfn > reset_free) { 413e332f741SMel Gorman free_set = true; 414e332f741SMel Gorman reset_free = free_pfn; 415e332f741SMel Gorman zone->compact_init_free_pfn = reset_free; 416e332f741SMel Gorman zone->compact_cached_free_pfn = reset_free; 417e332f741SMel Gorman } 418e332f741SMel Gorman } 419e332f741SMel Gorman 420e332f741SMel Gorman /* Leave no distance if no suitable block was reset */ 421e332f741SMel Gorman if (reset_migrate >= reset_free) { 422e332f741SMel Gorman zone->compact_cached_migrate_pfn[0] = migrate_pfn; 423e332f741SMel Gorman zone->compact_cached_migrate_pfn[1] = migrate_pfn; 424e332f741SMel Gorman zone->compact_cached_free_pfn = free_pfn; 425e332f741SMel Gorman } 426bb13ffebSMel Gorman } 427bb13ffebSMel Gorman 42862997027SMel Gorman void reset_isolation_suitable(pg_data_t *pgdat) 42962997027SMel Gorman { 43062997027SMel Gorman int zoneid; 43162997027SMel Gorman 43262997027SMel Gorman for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 43362997027SMel Gorman struct zone *zone = &pgdat->node_zones[zoneid]; 43462997027SMel Gorman if (!populated_zone(zone)) 43562997027SMel Gorman continue; 43662997027SMel Gorman 43762997027SMel Gorman /* Only flush if a full compaction finished recently */ 43862997027SMel Gorman if (zone->compact_blockskip_flush) 43962997027SMel Gorman __reset_isolation_suitable(zone); 44062997027SMel Gorman } 44162997027SMel Gorman } 44262997027SMel Gorman 443bb13ffebSMel Gorman /* 444e380bebeSMel Gorman * Sets the pageblock skip bit if it was clear. Note that this is a hint as 445e380bebeSMel Gorman * locks are not required for read/writers. Returns true if it was already set. 446e380bebeSMel Gorman */ 447590ccea8SMel Gorman static bool test_and_set_skip(struct compact_control *cc, struct page *page) 448e380bebeSMel Gorman { 449e380bebeSMel Gorman bool skip; 450e380bebeSMel Gorman 451590ccea8SMel Gorman /* Do not update if skip hint is being ignored */ 452e380bebeSMel Gorman if (cc->ignore_skip_hint) 453e380bebeSMel Gorman return false; 454e380bebeSMel Gorman 455e380bebeSMel Gorman skip = get_pageblock_skip(page); 456e380bebeSMel Gorman if (!skip && !cc->no_set_skip_hint) 457e380bebeSMel Gorman set_pageblock_skip(page); 458e380bebeSMel Gorman 459e380bebeSMel Gorman return skip; 460e380bebeSMel Gorman } 461e380bebeSMel Gorman 462e380bebeSMel Gorman static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) 463e380bebeSMel Gorman { 464e380bebeSMel Gorman struct zone *zone = cc->zone; 465e380bebeSMel Gorman 466e380bebeSMel Gorman /* Set for isolation rather than compaction */ 467e380bebeSMel Gorman if (cc->no_set_skip_hint) 468e380bebeSMel Gorman return; 469e380bebeSMel Gorman 470*3c099a2bSKemeng Shi pfn = pageblock_end_pfn(pfn); 471*3c099a2bSKemeng Shi 472e380bebeSMel Gorman if (pfn > zone->compact_cached_migrate_pfn[0]) 473e380bebeSMel Gorman zone->compact_cached_migrate_pfn[0] = pfn; 474e380bebeSMel Gorman if (cc->mode != MIGRATE_ASYNC && 475e380bebeSMel Gorman pfn > zone->compact_cached_migrate_pfn[1]) 476e380bebeSMel Gorman zone->compact_cached_migrate_pfn[1] = pfn; 477e380bebeSMel Gorman } 478e380bebeSMel Gorman 479e380bebeSMel Gorman /* 480bb13ffebSMel Gorman * If no pages were isolated then mark this pageblock to be skipped in the 48162997027SMel Gorman * future. The information is later cleared by __reset_isolation_suitable(). 482bb13ffebSMel Gorman */ 483c89511abSMel Gorman static void update_pageblock_skip(struct compact_control *cc, 484d097a6f6SMel Gorman struct page *page, unsigned long pfn) 485bb13ffebSMel Gorman { 486c89511abSMel Gorman struct zone *zone = cc->zone; 4876815bf3fSJoonsoo Kim 4882583d671SVlastimil Babka if (cc->no_set_skip_hint) 4896815bf3fSJoonsoo Kim return; 4906815bf3fSJoonsoo Kim 491bb13ffebSMel Gorman set_pageblock_skip(page); 492c89511abSMel Gorman 49335979ef3SDavid Rientjes /* Update where async and sync compaction should restart */ 49435979ef3SDavid Rientjes if (pfn < zone->compact_cached_free_pfn) 495c89511abSMel Gorman zone->compact_cached_free_pfn = pfn; 496c89511abSMel Gorman } 497bb13ffebSMel Gorman #else 498bb13ffebSMel Gorman static inline bool isolation_suitable(struct compact_control *cc, 499bb13ffebSMel Gorman struct page *page) 500bb13ffebSMel Gorman { 501bb13ffebSMel Gorman return true; 502bb13ffebSMel Gorman } 503bb13ffebSMel Gorman 504b527cfe5SVlastimil Babka static inline bool pageblock_skip_persistent(struct page *page) 50521dc7e02SDavid Rientjes { 50621dc7e02SDavid Rientjes return false; 50721dc7e02SDavid Rientjes } 50821dc7e02SDavid Rientjes 50921dc7e02SDavid Rientjes static inline void update_pageblock_skip(struct compact_control *cc, 510d097a6f6SMel Gorman struct page *page, unsigned long pfn) 511bb13ffebSMel Gorman { 512bb13ffebSMel Gorman } 513e380bebeSMel Gorman 514e380bebeSMel Gorman static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) 515e380bebeSMel Gorman { 516e380bebeSMel Gorman } 517e380bebeSMel Gorman 518590ccea8SMel Gorman static bool test_and_set_skip(struct compact_control *cc, struct page *page) 519e380bebeSMel Gorman { 520e380bebeSMel Gorman return false; 521e380bebeSMel Gorman } 522bb13ffebSMel Gorman #endif /* CONFIG_COMPACTION */ 523bb13ffebSMel Gorman 5241f9efdefSVlastimil Babka /* 5258b44d279SVlastimil Babka * Compaction requires the taking of some coarse locks that are potentially 526cb2dcaf0SMel Gorman * very heavily contended. For async compaction, trylock and record if the 527cb2dcaf0SMel Gorman * lock is contended. The lock will still be acquired but compaction will 528cb2dcaf0SMel Gorman * abort when the current block is finished regardless of success rate. 529cb2dcaf0SMel Gorman * Sync compaction acquires the lock. 5308b44d279SVlastimil Babka * 531cb2dcaf0SMel Gorman * Always returns true which makes it easier to track lock state in callers. 5321f9efdefSVlastimil Babka */ 533cb2dcaf0SMel Gorman static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, 5348b44d279SVlastimil Babka struct compact_control *cc) 53577337edeSJules Irenge __acquires(lock) 5368b44d279SVlastimil Babka { 537cb2dcaf0SMel Gorman /* Track if the lock is contended in async mode */ 538cb2dcaf0SMel Gorman if (cc->mode == MIGRATE_ASYNC && !cc->contended) { 539cb2dcaf0SMel Gorman if (spin_trylock_irqsave(lock, *flags)) 540cb2dcaf0SMel Gorman return true; 541cb2dcaf0SMel Gorman 542c3486f53SVlastimil Babka cc->contended = true; 5438b44d279SVlastimil Babka } 5441f9efdefSVlastimil Babka 545cb2dcaf0SMel Gorman spin_lock_irqsave(lock, *flags); 5468b44d279SVlastimil Babka return true; 5472a1402aaSMel Gorman } 5482a1402aaSMel Gorman 54985aa125fSMichal Nazarewicz /* 550c67fe375SMel Gorman * Compaction requires the taking of some coarse locks that are potentially 5518b44d279SVlastimil Babka * very heavily contended. The lock should be periodically unlocked to avoid 5528b44d279SVlastimil Babka * having disabled IRQs for a long time, even when there is nobody waiting on 5538b44d279SVlastimil Babka * the lock. It might also be that allowing the IRQs will result in 554d56c1584SMiaohe Lin * need_resched() becoming true. If scheduling is needed, compaction schedules. 5558b44d279SVlastimil Babka * Either compaction type will also abort if a fatal signal is pending. 5568b44d279SVlastimil Babka * In either case if the lock was locked, it is dropped and not regained. 557c67fe375SMel Gorman * 558d56c1584SMiaohe Lin * Returns true if compaction should abort due to fatal signal pending. 559d56c1584SMiaohe Lin * Returns false when compaction can continue. 560c67fe375SMel Gorman */ 5618b44d279SVlastimil Babka static bool compact_unlock_should_abort(spinlock_t *lock, 5628b44d279SVlastimil Babka unsigned long flags, bool *locked, struct compact_control *cc) 563c67fe375SMel Gorman { 5648b44d279SVlastimil Babka if (*locked) { 5658b44d279SVlastimil Babka spin_unlock_irqrestore(lock, flags); 5668b44d279SVlastimil Babka *locked = false; 567c67fe375SMel Gorman } 568c67fe375SMel Gorman 5698b44d279SVlastimil Babka if (fatal_signal_pending(current)) { 570c3486f53SVlastimil Babka cc->contended = true; 5718b44d279SVlastimil Babka return true; 5728b44d279SVlastimil Babka } 5738b44d279SVlastimil Babka 574cf66f070SMel Gorman cond_resched(); 575be976572SVlastimil Babka 576be976572SVlastimil Babka return false; 577be976572SVlastimil Babka } 578be976572SVlastimil Babka 579c67fe375SMel Gorman /* 5809e4be470SJerome Marchand * Isolate free pages onto a private freelist. If @strict is true, will abort 5819e4be470SJerome Marchand * returning 0 on any invalid PFNs or non-free pages inside of the pageblock 5829e4be470SJerome Marchand * (even though it may still end up isolating some pages). 58385aa125fSMichal Nazarewicz */ 584f40d1e42SMel Gorman static unsigned long isolate_freepages_block(struct compact_control *cc, 585e14c720eSVlastimil Babka unsigned long *start_pfn, 58685aa125fSMichal Nazarewicz unsigned long end_pfn, 58785aa125fSMichal Nazarewicz struct list_head *freelist, 5884fca9730SMel Gorman unsigned int stride, 58985aa125fSMichal Nazarewicz bool strict) 590748446bbSMel Gorman { 591b7aba698SMel Gorman int nr_scanned = 0, total_isolated = 0; 592d097a6f6SMel Gorman struct page *cursor; 593b8b2d825SXiubo Li unsigned long flags = 0; 594f40d1e42SMel Gorman bool locked = false; 595e14c720eSVlastimil Babka unsigned long blockpfn = *start_pfn; 59666c64223SJoonsoo Kim unsigned int order; 597748446bbSMel Gorman 5984fca9730SMel Gorman /* Strict mode is for isolation, speed is secondary */ 5994fca9730SMel Gorman if (strict) 6004fca9730SMel Gorman stride = 1; 6014fca9730SMel Gorman 602748446bbSMel Gorman cursor = pfn_to_page(blockpfn); 603748446bbSMel Gorman 604f40d1e42SMel Gorman /* Isolate free pages. */ 6054fca9730SMel Gorman for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) { 60666c64223SJoonsoo Kim int isolated; 607748446bbSMel Gorman struct page *page = cursor; 608748446bbSMel Gorman 6098b44d279SVlastimil Babka /* 6108b44d279SVlastimil Babka * Periodically drop the lock (if held) regardless of its 6118b44d279SVlastimil Babka * contention, to give chance to IRQs. Abort if fatal signal 612d56c1584SMiaohe Lin * pending. 6138b44d279SVlastimil Babka */ 614c036ddffSMiaohe Lin if (!(blockpfn % COMPACT_CLUSTER_MAX) 6158b44d279SVlastimil Babka && compact_unlock_should_abort(&cc->zone->lock, flags, 6168b44d279SVlastimil Babka &locked, cc)) 6178b44d279SVlastimil Babka break; 6188b44d279SVlastimil Babka 619b7aba698SMel Gorman nr_scanned++; 6202af120bcSLaura Abbott 6219fcd6d2eSVlastimil Babka /* 6229fcd6d2eSVlastimil Babka * For compound pages such as THP and hugetlbfs, we can save 6239fcd6d2eSVlastimil Babka * potentially a lot of iterations if we skip them at once. 6249fcd6d2eSVlastimil Babka * The check is racy, but we can consider only valid values 6259fcd6d2eSVlastimil Babka * and the only danger is skipping too much. 6269fcd6d2eSVlastimil Babka */ 6279fcd6d2eSVlastimil Babka if (PageCompound(page)) { 62821dc7e02SDavid Rientjes const unsigned int order = compound_order(page); 6299fcd6d2eSVlastimil Babka 63023baf831SKirill A. Shutemov if (likely(order <= MAX_ORDER)) { 63121dc7e02SDavid Rientjes blockpfn += (1UL << order) - 1; 63221dc7e02SDavid Rientjes cursor += (1UL << order) - 1; 63356d48d8dSBaolin Wang nr_scanned += (1UL << order) - 1; 6349fcd6d2eSVlastimil Babka } 6359fcd6d2eSVlastimil Babka goto isolate_fail; 6369fcd6d2eSVlastimil Babka } 6379fcd6d2eSVlastimil Babka 638f40d1e42SMel Gorman if (!PageBuddy(page)) 6392af120bcSLaura Abbott goto isolate_fail; 640f40d1e42SMel Gorman 64185f73e6dSMiaohe Lin /* If we already hold the lock, we can skip some rechecking. */ 64269b7189fSVlastimil Babka if (!locked) { 643cb2dcaf0SMel Gorman locked = compact_lock_irqsave(&cc->zone->lock, 6448b44d279SVlastimil Babka &flags, cc); 645f40d1e42SMel Gorman 646f40d1e42SMel Gorman /* Recheck this is a buddy page under lock */ 647f40d1e42SMel Gorman if (!PageBuddy(page)) 6482af120bcSLaura Abbott goto isolate_fail; 64969b7189fSVlastimil Babka } 650748446bbSMel Gorman 65166c64223SJoonsoo Kim /* Found a free page, will break it into order-0 pages */ 652ab130f91SMatthew Wilcox (Oracle) order = buddy_order(page); 65366c64223SJoonsoo Kim isolated = __isolate_free_page(page, order); 654a4f04f2cSDavid Rientjes if (!isolated) 655a4f04f2cSDavid Rientjes break; 65666c64223SJoonsoo Kim set_page_private(page, order); 657a4f04f2cSDavid Rientjes 658b717d6b9SWilliam Lam nr_scanned += isolated - 1; 659748446bbSMel Gorman total_isolated += isolated; 660a4f04f2cSDavid Rientjes cc->nr_freepages += isolated; 66166c64223SJoonsoo Kim list_add_tail(&page->lru, freelist); 66266c64223SJoonsoo Kim 663a4f04f2cSDavid Rientjes if (!strict && cc->nr_migratepages <= cc->nr_freepages) { 664932ff6bbSJoonsoo Kim blockpfn += isolated; 665932ff6bbSJoonsoo Kim break; 666932ff6bbSJoonsoo Kim } 667a4f04f2cSDavid Rientjes /* Advance to the end of split page */ 668748446bbSMel Gorman blockpfn += isolated - 1; 669748446bbSMel Gorman cursor += isolated - 1; 6702af120bcSLaura Abbott continue; 6712af120bcSLaura Abbott 6722af120bcSLaura Abbott isolate_fail: 6732af120bcSLaura Abbott if (strict) 6742af120bcSLaura Abbott break; 6752af120bcSLaura Abbott else 6762af120bcSLaura Abbott continue; 6772af120bcSLaura Abbott 678748446bbSMel Gorman } 679748446bbSMel Gorman 680a4f04f2cSDavid Rientjes if (locked) 681a4f04f2cSDavid Rientjes spin_unlock_irqrestore(&cc->zone->lock, flags); 682a4f04f2cSDavid Rientjes 6839fcd6d2eSVlastimil Babka /* 6849fcd6d2eSVlastimil Babka * There is a tiny chance that we have read bogus compound_order(), 6859fcd6d2eSVlastimil Babka * so be careful to not go outside of the pageblock. 6869fcd6d2eSVlastimil Babka */ 6879fcd6d2eSVlastimil Babka if (unlikely(blockpfn > end_pfn)) 6889fcd6d2eSVlastimil Babka blockpfn = end_pfn; 6899fcd6d2eSVlastimil Babka 690e34d85f0SJoonsoo Kim trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, 691e34d85f0SJoonsoo Kim nr_scanned, total_isolated); 692e34d85f0SJoonsoo Kim 693e14c720eSVlastimil Babka /* Record how far we have got within the block */ 694e14c720eSVlastimil Babka *start_pfn = blockpfn; 695e14c720eSVlastimil Babka 696f40d1e42SMel Gorman /* 697f40d1e42SMel Gorman * If strict isolation is requested by CMA then check that all the 698f40d1e42SMel Gorman * pages requested were isolated. If there were any failures, 0 is 699f40d1e42SMel Gorman * returned and CMA will fail. 700f40d1e42SMel Gorman */ 7012af120bcSLaura Abbott if (strict && blockpfn < end_pfn) 702f40d1e42SMel Gorman total_isolated = 0; 703f40d1e42SMel Gorman 7047f354a54SDavid Rientjes cc->total_free_scanned += nr_scanned; 705397487dbSMel Gorman if (total_isolated) 706010fc29aSMinchan Kim count_compact_events(COMPACTISOLATED, total_isolated); 707748446bbSMel Gorman return total_isolated; 708748446bbSMel Gorman } 709748446bbSMel Gorman 71085aa125fSMichal Nazarewicz /** 71185aa125fSMichal Nazarewicz * isolate_freepages_range() - isolate free pages. 712e8b098fcSMike Rapoport * @cc: Compaction control structure. 71385aa125fSMichal Nazarewicz * @start_pfn: The first PFN to start isolating. 71485aa125fSMichal Nazarewicz * @end_pfn: The one-past-last PFN. 71585aa125fSMichal Nazarewicz * 71685aa125fSMichal Nazarewicz * Non-free pages, invalid PFNs, or zone boundaries within the 71785aa125fSMichal Nazarewicz * [start_pfn, end_pfn) range are considered errors, cause function to 71885aa125fSMichal Nazarewicz * undo its actions and return zero. 71985aa125fSMichal Nazarewicz * 72085aa125fSMichal Nazarewicz * Otherwise, function returns one-past-the-last PFN of isolated page 72185aa125fSMichal Nazarewicz * (which may be greater then end_pfn if end fell in a middle of 72285aa125fSMichal Nazarewicz * a free page). 72385aa125fSMichal Nazarewicz */ 724ff9543fdSMichal Nazarewicz unsigned long 725bb13ffebSMel Gorman isolate_freepages_range(struct compact_control *cc, 726bb13ffebSMel Gorman unsigned long start_pfn, unsigned long end_pfn) 72785aa125fSMichal Nazarewicz { 728e1409c32SJoonsoo Kim unsigned long isolated, pfn, block_start_pfn, block_end_pfn; 72985aa125fSMichal Nazarewicz LIST_HEAD(freelist); 73085aa125fSMichal Nazarewicz 7317d49d886SVlastimil Babka pfn = start_pfn; 73206b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(pfn); 733e1409c32SJoonsoo Kim if (block_start_pfn < cc->zone->zone_start_pfn) 734e1409c32SJoonsoo Kim block_start_pfn = cc->zone->zone_start_pfn; 73506b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(pfn); 7367d49d886SVlastimil Babka 7377d49d886SVlastimil Babka for (; pfn < end_pfn; pfn += isolated, 738e1409c32SJoonsoo Kim block_start_pfn = block_end_pfn, 7397d49d886SVlastimil Babka block_end_pfn += pageblock_nr_pages) { 740e14c720eSVlastimil Babka /* Protect pfn from changing by isolate_freepages_block */ 741e14c720eSVlastimil Babka unsigned long isolate_start_pfn = pfn; 7427d49d886SVlastimil Babka 74385aa125fSMichal Nazarewicz block_end_pfn = min(block_end_pfn, end_pfn); 74485aa125fSMichal Nazarewicz 74558420016SJoonsoo Kim /* 74658420016SJoonsoo Kim * pfn could pass the block_end_pfn if isolated freepage 74758420016SJoonsoo Kim * is more than pageblock order. In this case, we adjust 74858420016SJoonsoo Kim * scanning range to right one. 74958420016SJoonsoo Kim */ 75058420016SJoonsoo Kim if (pfn >= block_end_pfn) { 75106b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(pfn); 75206b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(pfn); 75358420016SJoonsoo Kim block_end_pfn = min(block_end_pfn, end_pfn); 75458420016SJoonsoo Kim } 75558420016SJoonsoo Kim 756e1409c32SJoonsoo Kim if (!pageblock_pfn_to_page(block_start_pfn, 757e1409c32SJoonsoo Kim block_end_pfn, cc->zone)) 7587d49d886SVlastimil Babka break; 7597d49d886SVlastimil Babka 760e14c720eSVlastimil Babka isolated = isolate_freepages_block(cc, &isolate_start_pfn, 7614fca9730SMel Gorman block_end_pfn, &freelist, 0, true); 76285aa125fSMichal Nazarewicz 76385aa125fSMichal Nazarewicz /* 76485aa125fSMichal Nazarewicz * In strict mode, isolate_freepages_block() returns 0 if 76585aa125fSMichal Nazarewicz * there are any holes in the block (ie. invalid PFNs or 76685aa125fSMichal Nazarewicz * non-free pages). 76785aa125fSMichal Nazarewicz */ 76885aa125fSMichal Nazarewicz if (!isolated) 76985aa125fSMichal Nazarewicz break; 77085aa125fSMichal Nazarewicz 77185aa125fSMichal Nazarewicz /* 77285aa125fSMichal Nazarewicz * If we managed to isolate pages, it is always (1 << n) * 77385aa125fSMichal Nazarewicz * pageblock_nr_pages for some non-negative n. (Max order 77485aa125fSMichal Nazarewicz * page may span two pageblocks). 77585aa125fSMichal Nazarewicz */ 77685aa125fSMichal Nazarewicz } 77785aa125fSMichal Nazarewicz 77866c64223SJoonsoo Kim /* __isolate_free_page() does not map the pages */ 7794469ab98SMel Gorman split_map_pages(&freelist); 78085aa125fSMichal Nazarewicz 78185aa125fSMichal Nazarewicz if (pfn < end_pfn) { 78285aa125fSMichal Nazarewicz /* Loop terminated early, cleanup. */ 78385aa125fSMichal Nazarewicz release_freepages(&freelist); 78485aa125fSMichal Nazarewicz return 0; 78585aa125fSMichal Nazarewicz } 78685aa125fSMichal Nazarewicz 78785aa125fSMichal Nazarewicz /* We don't use freelists for anything. */ 78885aa125fSMichal Nazarewicz return pfn; 78985aa125fSMichal Nazarewicz } 79085aa125fSMichal Nazarewicz 791748446bbSMel Gorman /* Similar to reclaim, but different enough that they don't share logic */ 7924fbbb3fdSJohannes Weiner static bool too_many_isolated(struct compact_control *cc) 793748446bbSMel Gorman { 7944fbbb3fdSJohannes Weiner pg_data_t *pgdat = cc->zone->zone_pgdat; 795d818fca1SMel Gorman bool too_many; 796d818fca1SMel Gorman 797bc693045SMinchan Kim unsigned long active, inactive, isolated; 798748446bbSMel Gorman 7995f438eeeSAndrey Ryabinin inactive = node_page_state(pgdat, NR_INACTIVE_FILE) + 8005f438eeeSAndrey Ryabinin node_page_state(pgdat, NR_INACTIVE_ANON); 8015f438eeeSAndrey Ryabinin active = node_page_state(pgdat, NR_ACTIVE_FILE) + 8025f438eeeSAndrey Ryabinin node_page_state(pgdat, NR_ACTIVE_ANON); 8035f438eeeSAndrey Ryabinin isolated = node_page_state(pgdat, NR_ISOLATED_FILE) + 8045f438eeeSAndrey Ryabinin node_page_state(pgdat, NR_ISOLATED_ANON); 805748446bbSMel Gorman 8064fbbb3fdSJohannes Weiner /* 8074fbbb3fdSJohannes Weiner * Allow GFP_NOFS to isolate past the limit set for regular 8084fbbb3fdSJohannes Weiner * compaction runs. This prevents an ABBA deadlock when other 8094fbbb3fdSJohannes Weiner * compactors have already isolated to the limit, but are 8104fbbb3fdSJohannes Weiner * blocked on filesystem locks held by the GFP_NOFS thread. 8114fbbb3fdSJohannes Weiner */ 8124fbbb3fdSJohannes Weiner if (cc->gfp_mask & __GFP_FS) { 8134fbbb3fdSJohannes Weiner inactive >>= 3; 8144fbbb3fdSJohannes Weiner active >>= 3; 8154fbbb3fdSJohannes Weiner } 8164fbbb3fdSJohannes Weiner 817d818fca1SMel Gorman too_many = isolated > (inactive + active) / 2; 818d818fca1SMel Gorman if (!too_many) 819d818fca1SMel Gorman wake_throttle_isolated(pgdat); 820d818fca1SMel Gorman 821d818fca1SMel Gorman return too_many; 822748446bbSMel Gorman } 823748446bbSMel Gorman 8242fe86e00SMichal Nazarewicz /** 825edc2ca61SVlastimil Babka * isolate_migratepages_block() - isolate all migrate-able pages within 826edc2ca61SVlastimil Babka * a single pageblock 8272fe86e00SMichal Nazarewicz * @cc: Compaction control structure. 828edc2ca61SVlastimil Babka * @low_pfn: The first PFN to isolate 829edc2ca61SVlastimil Babka * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock 83089f6c88aSHugh Dickins * @mode: Isolation mode to be used. 8312fe86e00SMichal Nazarewicz * 8322fe86e00SMichal Nazarewicz * Isolate all pages that can be migrated from the range specified by 833edc2ca61SVlastimil Babka * [low_pfn, end_pfn). The range is expected to be within same pageblock. 834c2ad7a1fSOscar Salvador * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion, 835369fa227SOscar Salvador * -ENOMEM in case we could not allocate a page, or 0. 836c2ad7a1fSOscar Salvador * cc->migrate_pfn will contain the next pfn to scan. 8372fe86e00SMichal Nazarewicz * 838edc2ca61SVlastimil Babka * The pages are isolated on cc->migratepages list (not required to be empty), 839c2ad7a1fSOscar Salvador * and cc->nr_migratepages is updated accordingly. 840748446bbSMel Gorman */ 841c2ad7a1fSOscar Salvador static int 842edc2ca61SVlastimil Babka isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, 84389f6c88aSHugh Dickins unsigned long end_pfn, isolate_mode_t mode) 844748446bbSMel Gorman { 8455f438eeeSAndrey Ryabinin pg_data_t *pgdat = cc->zone->zone_pgdat; 846b7aba698SMel Gorman unsigned long nr_scanned = 0, nr_isolated = 0; 847fa9add64SHugh Dickins struct lruvec *lruvec; 848b8b2d825SXiubo Li unsigned long flags = 0; 8496168d0daSAlex Shi struct lruvec *locked = NULL; 85056ae0bb3SKefeng Wang struct folio *folio = NULL; 851bb13ffebSMel Gorman struct page *page = NULL, *valid_page = NULL; 85289f6c88aSHugh Dickins struct address_space *mapping; 853e34d85f0SJoonsoo Kim unsigned long start_pfn = low_pfn; 854fdd048e1SVlastimil Babka bool skip_on_failure = false; 855fdd048e1SVlastimil Babka unsigned long next_skip_pfn = 0; 856e380bebeSMel Gorman bool skip_updated = false; 857c2ad7a1fSOscar Salvador int ret = 0; 858c2ad7a1fSOscar Salvador 859c2ad7a1fSOscar Salvador cc->migrate_pfn = low_pfn; 860748446bbSMel Gorman 861748446bbSMel Gorman /* 862748446bbSMel Gorman * Ensure that there are not too many pages isolated from the LRU 863748446bbSMel Gorman * list by either parallel reclaimers or compaction. If there are, 864748446bbSMel Gorman * delay for some time until fewer pages are isolated 865748446bbSMel Gorman */ 8664fbbb3fdSJohannes Weiner while (unlikely(too_many_isolated(cc))) { 867d20bdd57SZi Yan /* stop isolation if there are still pages not migrated */ 868d20bdd57SZi Yan if (cc->nr_migratepages) 869c2ad7a1fSOscar Salvador return -EAGAIN; 870d20bdd57SZi Yan 871f9e35b3bSMel Gorman /* async migration should just abort */ 872e0b9daebSDavid Rientjes if (cc->mode == MIGRATE_ASYNC) 873c2ad7a1fSOscar Salvador return -EAGAIN; 874f9e35b3bSMel Gorman 875c3f4a9a2SMel Gorman reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); 876748446bbSMel Gorman 877748446bbSMel Gorman if (fatal_signal_pending(current)) 878c2ad7a1fSOscar Salvador return -EINTR; 879748446bbSMel Gorman } 880748446bbSMel Gorman 881cf66f070SMel Gorman cond_resched(); 882aeef4b83SDavid Rientjes 883fdd048e1SVlastimil Babka if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { 884fdd048e1SVlastimil Babka skip_on_failure = true; 885fdd048e1SVlastimil Babka next_skip_pfn = block_end_pfn(low_pfn, cc->order); 886fdd048e1SVlastimil Babka } 887fdd048e1SVlastimil Babka 888748446bbSMel Gorman /* Time to isolate some pages for migration */ 889748446bbSMel Gorman for (; low_pfn < end_pfn; low_pfn++) { 89029c0dde8SVlastimil Babka 891fdd048e1SVlastimil Babka if (skip_on_failure && low_pfn >= next_skip_pfn) { 892fdd048e1SVlastimil Babka /* 893fdd048e1SVlastimil Babka * We have isolated all migration candidates in the 894fdd048e1SVlastimil Babka * previous order-aligned block, and did not skip it due 895fdd048e1SVlastimil Babka * to failure. We should migrate the pages now and 896fdd048e1SVlastimil Babka * hopefully succeed compaction. 897fdd048e1SVlastimil Babka */ 898fdd048e1SVlastimil Babka if (nr_isolated) 899fdd048e1SVlastimil Babka break; 900fdd048e1SVlastimil Babka 901fdd048e1SVlastimil Babka /* 902fdd048e1SVlastimil Babka * We failed to isolate in the previous order-aligned 903fdd048e1SVlastimil Babka * block. Set the new boundary to the end of the 904fdd048e1SVlastimil Babka * current block. Note we can't simply increase 905fdd048e1SVlastimil Babka * next_skip_pfn by 1 << order, as low_pfn might have 906fdd048e1SVlastimil Babka * been incremented by a higher number due to skipping 907fdd048e1SVlastimil Babka * a compound or a high-order buddy page in the 908fdd048e1SVlastimil Babka * previous loop iteration. 909fdd048e1SVlastimil Babka */ 910fdd048e1SVlastimil Babka next_skip_pfn = block_end_pfn(low_pfn, cc->order); 911fdd048e1SVlastimil Babka } 912fdd048e1SVlastimil Babka 9138b44d279SVlastimil Babka /* 9148b44d279SVlastimil Babka * Periodically drop the lock (if held) regardless of its 915670105a2SMel Gorman * contention, to give chance to IRQs. Abort completely if 916670105a2SMel Gorman * a fatal signal is pending. 9178b44d279SVlastimil Babka */ 918c036ddffSMiaohe Lin if (!(low_pfn % COMPACT_CLUSTER_MAX)) { 9196168d0daSAlex Shi if (locked) { 9206168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 9216168d0daSAlex Shi locked = NULL; 9226168d0daSAlex Shi } 9236168d0daSAlex Shi 9246168d0daSAlex Shi if (fatal_signal_pending(current)) { 9256168d0daSAlex Shi cc->contended = true; 926c2ad7a1fSOscar Salvador ret = -EINTR; 9276168d0daSAlex Shi 928670105a2SMel Gorman goto fatal_pending; 929670105a2SMel Gorman } 930b2eef8c0SAndrea Arcangeli 9316168d0daSAlex Shi cond_resched(); 9326168d0daSAlex Shi } 9336168d0daSAlex Shi 934b7aba698SMel Gorman nr_scanned++; 935748446bbSMel Gorman 936748446bbSMel Gorman page = pfn_to_page(low_pfn); 937dc908600SMel Gorman 938e380bebeSMel Gorman /* 939e380bebeSMel Gorman * Check if the pageblock has already been marked skipped. 940e380bebeSMel Gorman * Only the aligned PFN is checked as the caller isolates 941e380bebeSMel Gorman * COMPACT_CLUSTER_MAX at a time so the second call must 942e380bebeSMel Gorman * not falsely conclude that the block should be skipped. 943e380bebeSMel Gorman */ 944ee0913c4SKefeng Wang if (!valid_page && pageblock_aligned(low_pfn)) { 9454af12d04SMiaohe Lin if (!isolation_suitable(cc, page)) { 946e380bebeSMel Gorman low_pfn = end_pfn; 94756ae0bb3SKefeng Wang folio = NULL; 948e380bebeSMel Gorman goto isolate_abort; 949e380bebeSMel Gorman } 950bb13ffebSMel Gorman valid_page = page; 951e380bebeSMel Gorman } 952bb13ffebSMel Gorman 953369fa227SOscar Salvador if (PageHuge(page) && cc->alloc_contig) { 9541c06b6a5SBaolin Wang if (locked) { 9551c06b6a5SBaolin Wang unlock_page_lruvec_irqrestore(locked, flags); 9561c06b6a5SBaolin Wang locked = NULL; 9571c06b6a5SBaolin Wang } 9581c06b6a5SBaolin Wang 959ae37c7ffSOscar Salvador ret = isolate_or_dissolve_huge_page(page, &cc->migratepages); 960369fa227SOscar Salvador 961369fa227SOscar Salvador /* 962369fa227SOscar Salvador * Fail isolation in case isolate_or_dissolve_huge_page() 963369fa227SOscar Salvador * reports an error. In case of -ENOMEM, abort right away. 964369fa227SOscar Salvador */ 965369fa227SOscar Salvador if (ret < 0) { 966369fa227SOscar Salvador /* Do not report -EBUSY down the chain */ 967369fa227SOscar Salvador if (ret == -EBUSY) 968369fa227SOscar Salvador ret = 0; 96966fe1cf7SMiaohe Lin low_pfn += compound_nr(page) - 1; 97056d48d8dSBaolin Wang nr_scanned += compound_nr(page) - 1; 971369fa227SOscar Salvador goto isolate_fail; 972369fa227SOscar Salvador } 973369fa227SOscar Salvador 974ae37c7ffSOscar Salvador if (PageHuge(page)) { 975ae37c7ffSOscar Salvador /* 976ae37c7ffSOscar Salvador * Hugepage was successfully isolated and placed 977ae37c7ffSOscar Salvador * on the cc->migratepages list. 978ae37c7ffSOscar Salvador */ 97956ae0bb3SKefeng Wang folio = page_folio(page); 98056ae0bb3SKefeng Wang low_pfn += folio_nr_pages(folio) - 1; 981ae37c7ffSOscar Salvador goto isolate_success_no_list; 982ae37c7ffSOscar Salvador } 983ae37c7ffSOscar Salvador 984369fa227SOscar Salvador /* 985369fa227SOscar Salvador * Ok, the hugepage was dissolved. Now these pages are 986369fa227SOscar Salvador * Buddy and cannot be re-allocated because they are 987369fa227SOscar Salvador * isolated. Fall-through as the check below handles 988369fa227SOscar Salvador * Buddy pages. 989369fa227SOscar Salvador */ 990369fa227SOscar Salvador } 991369fa227SOscar Salvador 992c122b208SJoonsoo Kim /* 99399c0fd5eSVlastimil Babka * Skip if free. We read page order here without zone lock 99499c0fd5eSVlastimil Babka * which is generally unsafe, but the race window is small and 99599c0fd5eSVlastimil Babka * the worst thing that can happen is that we skip some 99699c0fd5eSVlastimil Babka * potential isolation targets. 9976c14466cSMel Gorman */ 99899c0fd5eSVlastimil Babka if (PageBuddy(page)) { 999ab130f91SMatthew Wilcox (Oracle) unsigned long freepage_order = buddy_order_unsafe(page); 100099c0fd5eSVlastimil Babka 100199c0fd5eSVlastimil Babka /* 100299c0fd5eSVlastimil Babka * Without lock, we cannot be sure that what we got is 100399c0fd5eSVlastimil Babka * a valid page order. Consider only values in the 100499c0fd5eSVlastimil Babka * valid order range to prevent low_pfn overflow. 100599c0fd5eSVlastimil Babka */ 100656d48d8dSBaolin Wang if (freepage_order > 0 && freepage_order <= MAX_ORDER) { 100799c0fd5eSVlastimil Babka low_pfn += (1UL << freepage_order) - 1; 100856d48d8dSBaolin Wang nr_scanned += (1UL << freepage_order) - 1; 100956d48d8dSBaolin Wang } 1010748446bbSMel Gorman continue; 101199c0fd5eSVlastimil Babka } 1012748446bbSMel Gorman 10139927af74SMel Gorman /* 101429c0dde8SVlastimil Babka * Regardless of being on LRU, compound pages such as THP and 10151da2f328SRik van Riel * hugetlbfs are not to be compacted unless we are attempting 10161da2f328SRik van Riel * an allocation much larger than the huge page size (eg CMA). 10171da2f328SRik van Riel * We can potentially save a lot of iterations if we skip them 10181da2f328SRik van Riel * at once. The check is racy, but we can consider only valid 10191da2f328SRik van Riel * values and the only danger is skipping too much. 1020bc835011SAndrea Arcangeli */ 10211da2f328SRik van Riel if (PageCompound(page) && !cc->alloc_contig) { 102221dc7e02SDavid Rientjes const unsigned int order = compound_order(page); 102329c0dde8SVlastimil Babka 102456d48d8dSBaolin Wang if (likely(order <= MAX_ORDER)) { 102521dc7e02SDavid Rientjes low_pfn += (1UL << order) - 1; 102656d48d8dSBaolin Wang nr_scanned += (1UL << order) - 1; 102756d48d8dSBaolin Wang } 1028fdd048e1SVlastimil Babka goto isolate_fail; 10292a1402aaSMel Gorman } 10302a1402aaSMel Gorman 1031bda807d4SMinchan Kim /* 1032bda807d4SMinchan Kim * Check may be lockless but that's ok as we recheck later. 1033bda807d4SMinchan Kim * It's possible to migrate LRU and non-lru movable pages. 1034bda807d4SMinchan Kim * Skip any other type of page 1035bda807d4SMinchan Kim */ 1036bda807d4SMinchan Kim if (!PageLRU(page)) { 1037bda807d4SMinchan Kim /* 1038bda807d4SMinchan Kim * __PageMovable can return false positive so we need 1039bda807d4SMinchan Kim * to verify it under page_lock. 1040bda807d4SMinchan Kim */ 1041bda807d4SMinchan Kim if (unlikely(__PageMovable(page)) && 1042bda807d4SMinchan Kim !PageIsolated(page)) { 1043bda807d4SMinchan Kim if (locked) { 10446168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 10456168d0daSAlex Shi locked = NULL; 1046bda807d4SMinchan Kim } 1047bda807d4SMinchan Kim 104856ae0bb3SKefeng Wang if (isolate_movable_page(page, mode)) { 104956ae0bb3SKefeng Wang folio = page_folio(page); 1050bda807d4SMinchan Kim goto isolate_success; 1051bda807d4SMinchan Kim } 105256ae0bb3SKefeng Wang } 1053bda807d4SMinchan Kim 1054fdd048e1SVlastimil Babka goto isolate_fail; 1055bda807d4SMinchan Kim } 105629c0dde8SVlastimil Babka 1057119d6d59SDavid Rientjes /* 10589df41314SAlex Shi * Be careful not to clear PageLRU until after we're 10599df41314SAlex Shi * sure the page is not being freed elsewhere -- the 10609df41314SAlex Shi * page release code relies on it. 10619df41314SAlex Shi */ 106256ae0bb3SKefeng Wang folio = folio_get_nontail_page(page); 106356ae0bb3SKefeng Wang if (unlikely(!folio)) 10649df41314SAlex Shi goto isolate_fail; 10659df41314SAlex Shi 1066829ae0f8SGavin Shan /* 1067829ae0f8SGavin Shan * Migration will fail if an anonymous page is pinned in memory, 1068829ae0f8SGavin Shan * so avoid taking lru_lock and isolating it unnecessarily in an 1069829ae0f8SGavin Shan * admittedly racy check. 1070829ae0f8SGavin Shan */ 107156ae0bb3SKefeng Wang mapping = folio_mapping(folio); 107256ae0bb3SKefeng Wang if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio)) 1073829ae0f8SGavin Shan goto isolate_fail_put; 1074829ae0f8SGavin Shan 1075829ae0f8SGavin Shan /* 1076829ae0f8SGavin Shan * Only allow to migrate anonymous pages in GFP_NOFS context 1077829ae0f8SGavin Shan * because those do not depend on fs locks. 1078829ae0f8SGavin Shan */ 1079829ae0f8SGavin Shan if (!(cc->gfp_mask & __GFP_FS) && mapping) 1080829ae0f8SGavin Shan goto isolate_fail_put; 1081829ae0f8SGavin Shan 108289f6c88aSHugh Dickins /* Only take pages on LRU: a check now makes later tests safe */ 108356ae0bb3SKefeng Wang if (!folio_test_lru(folio)) 10849df41314SAlex Shi goto isolate_fail_put; 10859df41314SAlex Shi 108689f6c88aSHugh Dickins /* Compaction might skip unevictable pages but CMA takes them */ 108756ae0bb3SKefeng Wang if (!(mode & ISOLATE_UNEVICTABLE) && folio_test_unevictable(folio)) 108889f6c88aSHugh Dickins goto isolate_fail_put; 108989f6c88aSHugh Dickins 109089f6c88aSHugh Dickins /* 109189f6c88aSHugh Dickins * To minimise LRU disruption, the caller can indicate with 109289f6c88aSHugh Dickins * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages 109389f6c88aSHugh Dickins * it will be able to migrate without blocking - clean pages 109489f6c88aSHugh Dickins * for the most part. PageWriteback would require blocking. 109589f6c88aSHugh Dickins */ 109656ae0bb3SKefeng Wang if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio)) 109789f6c88aSHugh Dickins goto isolate_fail_put; 109889f6c88aSHugh Dickins 109956ae0bb3SKefeng Wang if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_dirty(folio)) { 110089f6c88aSHugh Dickins bool migrate_dirty; 110189f6c88aSHugh Dickins 110289f6c88aSHugh Dickins /* 110389f6c88aSHugh Dickins * Only pages without mappings or that have a 11049d0ddc0cSMatthew Wilcox (Oracle) * ->migrate_folio callback are possible to migrate 110589f6c88aSHugh Dickins * without blocking. However, we can be racing with 110689f6c88aSHugh Dickins * truncation so it's necessary to lock the page 110789f6c88aSHugh Dickins * to stabilise the mapping as truncation holds 110889f6c88aSHugh Dickins * the page lock until after the page is removed 110989f6c88aSHugh Dickins * from the page cache. 111089f6c88aSHugh Dickins */ 111156ae0bb3SKefeng Wang if (!folio_trylock(folio)) 111289f6c88aSHugh Dickins goto isolate_fail_put; 111389f6c88aSHugh Dickins 111456ae0bb3SKefeng Wang mapping = folio_mapping(folio); 11155490da4fSMatthew Wilcox (Oracle) migrate_dirty = !mapping || 11169d0ddc0cSMatthew Wilcox (Oracle) mapping->a_ops->migrate_folio; 111756ae0bb3SKefeng Wang folio_unlock(folio); 111889f6c88aSHugh Dickins if (!migrate_dirty) 111989f6c88aSHugh Dickins goto isolate_fail_put; 112089f6c88aSHugh Dickins } 112189f6c88aSHugh Dickins 112256ae0bb3SKefeng Wang /* Try isolate the folio */ 112356ae0bb3SKefeng Wang if (!folio_test_clear_lru(folio)) 11249df41314SAlex Shi goto isolate_fail_put; 11259df41314SAlex Shi 112656ae0bb3SKefeng Wang lruvec = folio_lruvec(folio); 11276168d0daSAlex Shi 112869b7189fSVlastimil Babka /* If we already hold the lock, we can skip some rechecking */ 11296168d0daSAlex Shi if (lruvec != locked) { 11306168d0daSAlex Shi if (locked) 11316168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 11326168d0daSAlex Shi 11336168d0daSAlex Shi compact_lock_irqsave(&lruvec->lru_lock, &flags, cc); 11346168d0daSAlex Shi locked = lruvec; 11356168d0daSAlex Shi 113656ae0bb3SKefeng Wang lruvec_memcg_debug(lruvec, folio); 1137e380bebeSMel Gorman 1138590ccea8SMel Gorman /* 1139590ccea8SMel Gorman * Try get exclusive access under lock. If marked for 1140590ccea8SMel Gorman * skip, the scan is aborted unless the current context 1141590ccea8SMel Gorman * is a rescan to reach the end of the pageblock. 1142590ccea8SMel Gorman */ 1143590ccea8SMel Gorman if (!skip_updated && valid_page) { 1144e380bebeSMel Gorman skip_updated = true; 1145590ccea8SMel Gorman if (test_and_set_skip(cc, valid_page) && 1146590ccea8SMel Gorman !cc->finish_pageblock) { 1147e380bebeSMel Gorman goto isolate_abort; 1148e380bebeSMel Gorman } 1149590ccea8SMel Gorman } 11502a1402aaSMel Gorman 115129c0dde8SVlastimil Babka /* 115256ae0bb3SKefeng Wang * folio become large since the non-locked check, 115356ae0bb3SKefeng Wang * and it's on LRU. 115429c0dde8SVlastimil Babka */ 115556ae0bb3SKefeng Wang if (unlikely(folio_test_large(folio) && !cc->alloc_contig)) { 115656ae0bb3SKefeng Wang low_pfn += folio_nr_pages(folio) - 1; 115756ae0bb3SKefeng Wang nr_scanned += folio_nr_pages(folio) - 1; 115856ae0bb3SKefeng Wang folio_set_lru(folio); 11599df41314SAlex Shi goto isolate_fail_put; 1160bc835011SAndrea Arcangeli } 1161d99fd5feSAlex Shi } 1162fa9add64SHugh Dickins 116356ae0bb3SKefeng Wang /* The folio is taken off the LRU */ 116456ae0bb3SKefeng Wang if (folio_test_large(folio)) 116556ae0bb3SKefeng Wang low_pfn += folio_nr_pages(folio) - 1; 1166bc835011SAndrea Arcangeli 1167748446bbSMel Gorman /* Successfully isolated */ 116856ae0bb3SKefeng Wang lruvec_del_folio(lruvec, folio); 116956ae0bb3SKefeng Wang node_stat_mod_folio(folio, 117056ae0bb3SKefeng Wang NR_ISOLATED_ANON + folio_is_file_lru(folio), 117156ae0bb3SKefeng Wang folio_nr_pages(folio)); 1172b6c75016SJoonsoo Kim 1173b6c75016SJoonsoo Kim isolate_success: 117456ae0bb3SKefeng Wang list_add(&folio->lru, &cc->migratepages); 1175ae37c7ffSOscar Salvador isolate_success_no_list: 117656ae0bb3SKefeng Wang cc->nr_migratepages += folio_nr_pages(folio); 117756ae0bb3SKefeng Wang nr_isolated += folio_nr_pages(folio); 117856ae0bb3SKefeng Wang nr_scanned += folio_nr_pages(folio) - 1; 1179748446bbSMel Gorman 1180804d3121SMel Gorman /* 1181804d3121SMel Gorman * Avoid isolating too much unless this block is being 118248731c84SMel Gorman * fully scanned (e.g. dirty/writeback pages, parallel allocation) 1183cb2dcaf0SMel Gorman * or a lock is contended. For contention, isolate quickly to 1184cb2dcaf0SMel Gorman * potentially remove one source of contention. 1185804d3121SMel Gorman */ 118638935861SZi Yan if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX && 118748731c84SMel Gorman !cc->finish_pageblock && !cc->contended) { 118831b8384aSHillf Danton ++low_pfn; 1189748446bbSMel Gorman break; 1190748446bbSMel Gorman } 1191fdd048e1SVlastimil Babka 1192fdd048e1SVlastimil Babka continue; 11939df41314SAlex Shi 11949df41314SAlex Shi isolate_fail_put: 11959df41314SAlex Shi /* Avoid potential deadlock in freeing page under lru_lock */ 11969df41314SAlex Shi if (locked) { 11976168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 11986168d0daSAlex Shi locked = NULL; 11999df41314SAlex Shi } 120056ae0bb3SKefeng Wang folio_put(folio); 12019df41314SAlex Shi 1202fdd048e1SVlastimil Babka isolate_fail: 1203369fa227SOscar Salvador if (!skip_on_failure && ret != -ENOMEM) 1204fdd048e1SVlastimil Babka continue; 1205fdd048e1SVlastimil Babka 1206fdd048e1SVlastimil Babka /* 1207fdd048e1SVlastimil Babka * We have isolated some pages, but then failed. Release them 1208fdd048e1SVlastimil Babka * instead of migrating, as we cannot form the cc->order buddy 1209fdd048e1SVlastimil Babka * page anyway. 1210fdd048e1SVlastimil Babka */ 1211fdd048e1SVlastimil Babka if (nr_isolated) { 1212fdd048e1SVlastimil Babka if (locked) { 12136168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 12146168d0daSAlex Shi locked = NULL; 1215fdd048e1SVlastimil Babka } 1216fdd048e1SVlastimil Babka putback_movable_pages(&cc->migratepages); 1217fdd048e1SVlastimil Babka cc->nr_migratepages = 0; 1218fdd048e1SVlastimil Babka nr_isolated = 0; 1219fdd048e1SVlastimil Babka } 1220fdd048e1SVlastimil Babka 1221fdd048e1SVlastimil Babka if (low_pfn < next_skip_pfn) { 1222fdd048e1SVlastimil Babka low_pfn = next_skip_pfn - 1; 1223fdd048e1SVlastimil Babka /* 1224fdd048e1SVlastimil Babka * The check near the loop beginning would have updated 1225fdd048e1SVlastimil Babka * next_skip_pfn too, but this is a bit simpler. 1226fdd048e1SVlastimil Babka */ 1227fdd048e1SVlastimil Babka next_skip_pfn += 1UL << cc->order; 1228fdd048e1SVlastimil Babka } 1229369fa227SOscar Salvador 1230369fa227SOscar Salvador if (ret == -ENOMEM) 1231369fa227SOscar Salvador break; 123231b8384aSHillf Danton } 1233748446bbSMel Gorman 123499c0fd5eSVlastimil Babka /* 123599c0fd5eSVlastimil Babka * The PageBuddy() check could have potentially brought us outside 123699c0fd5eSVlastimil Babka * the range to be scanned. 123799c0fd5eSVlastimil Babka */ 123899c0fd5eSVlastimil Babka if (unlikely(low_pfn > end_pfn)) 123999c0fd5eSVlastimil Babka low_pfn = end_pfn; 124099c0fd5eSVlastimil Babka 124156ae0bb3SKefeng Wang folio = NULL; 12429df41314SAlex Shi 1243e380bebeSMel Gorman isolate_abort: 1244c67fe375SMel Gorman if (locked) 12456168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 124656ae0bb3SKefeng Wang if (folio) { 124756ae0bb3SKefeng Wang folio_set_lru(folio); 124856ae0bb3SKefeng Wang folio_put(folio); 12499df41314SAlex Shi } 1250748446bbSMel Gorman 125150b5b094SVlastimil Babka /* 125248731c84SMel Gorman * Update the cached scanner pfn once the pageblock has been scanned. 1253804d3121SMel Gorman * Pages will either be migrated in which case there is no point 1254804d3121SMel Gorman * scanning in the near future or migration failed in which case the 1255804d3121SMel Gorman * failure reason may persist. The block is marked for skipping if 1256804d3121SMel Gorman * there were no pages isolated in the block or if the block is 1257804d3121SMel Gorman * rescanned twice in a row. 125850b5b094SVlastimil Babka */ 125948731c84SMel Gorman if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) { 12608b71b499SBaolin Wang if (!cc->no_set_skip_hint && valid_page && !skip_updated) 1261e380bebeSMel Gorman set_pageblock_skip(valid_page); 1262e380bebeSMel Gorman update_cached_migrate(cc, low_pfn); 1263e380bebeSMel Gorman } 1264bb13ffebSMel Gorman 1265e34d85f0SJoonsoo Kim trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, 1266e34d85f0SJoonsoo Kim nr_scanned, nr_isolated); 1267b7aba698SMel Gorman 1268670105a2SMel Gorman fatal_pending: 12697f354a54SDavid Rientjes cc->total_migrate_scanned += nr_scanned; 1270397487dbSMel Gorman if (nr_isolated) 1271010fc29aSMinchan Kim count_compact_events(COMPACTISOLATED, nr_isolated); 1272397487dbSMel Gorman 1273c2ad7a1fSOscar Salvador cc->migrate_pfn = low_pfn; 1274c2ad7a1fSOscar Salvador 1275c2ad7a1fSOscar Salvador return ret; 12762fe86e00SMichal Nazarewicz } 12772fe86e00SMichal Nazarewicz 1278edc2ca61SVlastimil Babka /** 1279edc2ca61SVlastimil Babka * isolate_migratepages_range() - isolate migrate-able pages in a PFN range 1280edc2ca61SVlastimil Babka * @cc: Compaction control structure. 1281edc2ca61SVlastimil Babka * @start_pfn: The first PFN to start isolating. 1282edc2ca61SVlastimil Babka * @end_pfn: The one-past-last PFN. 1283edc2ca61SVlastimil Babka * 1284369fa227SOscar Salvador * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM 1285369fa227SOscar Salvador * in case we could not allocate a page, or 0. 1286edc2ca61SVlastimil Babka */ 1287c2ad7a1fSOscar Salvador int 1288edc2ca61SVlastimil Babka isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, 1289edc2ca61SVlastimil Babka unsigned long end_pfn) 1290edc2ca61SVlastimil Babka { 1291e1409c32SJoonsoo Kim unsigned long pfn, block_start_pfn, block_end_pfn; 1292c2ad7a1fSOscar Salvador int ret = 0; 1293edc2ca61SVlastimil Babka 1294edc2ca61SVlastimil Babka /* Scan block by block. First and last block may be incomplete */ 1295edc2ca61SVlastimil Babka pfn = start_pfn; 129606b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(pfn); 1297e1409c32SJoonsoo Kim if (block_start_pfn < cc->zone->zone_start_pfn) 1298e1409c32SJoonsoo Kim block_start_pfn = cc->zone->zone_start_pfn; 129906b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(pfn); 1300edc2ca61SVlastimil Babka 1301edc2ca61SVlastimil Babka for (; pfn < end_pfn; pfn = block_end_pfn, 1302e1409c32SJoonsoo Kim block_start_pfn = block_end_pfn, 1303edc2ca61SVlastimil Babka block_end_pfn += pageblock_nr_pages) { 1304edc2ca61SVlastimil Babka 1305edc2ca61SVlastimil Babka block_end_pfn = min(block_end_pfn, end_pfn); 1306edc2ca61SVlastimil Babka 1307e1409c32SJoonsoo Kim if (!pageblock_pfn_to_page(block_start_pfn, 1308e1409c32SJoonsoo Kim block_end_pfn, cc->zone)) 1309edc2ca61SVlastimil Babka continue; 1310edc2ca61SVlastimil Babka 1311c2ad7a1fSOscar Salvador ret = isolate_migratepages_block(cc, pfn, block_end_pfn, 1312edc2ca61SVlastimil Babka ISOLATE_UNEVICTABLE); 1313edc2ca61SVlastimil Babka 1314c2ad7a1fSOscar Salvador if (ret) 1315edc2ca61SVlastimil Babka break; 13166ea41c0cSJoonsoo Kim 131738935861SZi Yan if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX) 13186ea41c0cSJoonsoo Kim break; 1319edc2ca61SVlastimil Babka } 1320edc2ca61SVlastimil Babka 1321c2ad7a1fSOscar Salvador return ret; 1322edc2ca61SVlastimil Babka } 1323edc2ca61SVlastimil Babka 1324ff9543fdSMichal Nazarewicz #endif /* CONFIG_COMPACTION || CONFIG_CMA */ 1325ff9543fdSMichal Nazarewicz #ifdef CONFIG_COMPACTION 1326018e9a49SAndrew Morton 1327b682debdSVlastimil Babka static bool suitable_migration_source(struct compact_control *cc, 1328b682debdSVlastimil Babka struct page *page) 1329b682debdSVlastimil Babka { 1330282722b0SVlastimil Babka int block_mt; 1331282722b0SVlastimil Babka 13329bebefd5SMel Gorman if (pageblock_skip_persistent(page)) 13339bebefd5SMel Gorman return false; 13349bebefd5SMel Gorman 1335282722b0SVlastimil Babka if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) 1336b682debdSVlastimil Babka return true; 1337b682debdSVlastimil Babka 1338282722b0SVlastimil Babka block_mt = get_pageblock_migratetype(page); 1339282722b0SVlastimil Babka 1340282722b0SVlastimil Babka if (cc->migratetype == MIGRATE_MOVABLE) 1341282722b0SVlastimil Babka return is_migrate_movable(block_mt); 1342282722b0SVlastimil Babka else 1343282722b0SVlastimil Babka return block_mt == cc->migratetype; 1344b682debdSVlastimil Babka } 1345b682debdSVlastimil Babka 1346018e9a49SAndrew Morton /* Returns true if the page is within a block suitable for migration to */ 13479f7e3387SVlastimil Babka static bool suitable_migration_target(struct compact_control *cc, 13489f7e3387SVlastimil Babka struct page *page) 1349018e9a49SAndrew Morton { 1350018e9a49SAndrew Morton /* If the page is a large free page, then disallow migration */ 1351018e9a49SAndrew Morton if (PageBuddy(page)) { 1352018e9a49SAndrew Morton /* 1353018e9a49SAndrew Morton * We are checking page_order without zone->lock taken. But 1354018e9a49SAndrew Morton * the only small danger is that we skip a potentially suitable 1355018e9a49SAndrew Morton * pageblock, so it's not worth to check order for valid range. 1356018e9a49SAndrew Morton */ 1357ab130f91SMatthew Wilcox (Oracle) if (buddy_order_unsafe(page) >= pageblock_order) 1358018e9a49SAndrew Morton return false; 1359018e9a49SAndrew Morton } 1360018e9a49SAndrew Morton 13611ef36db2SYisheng Xie if (cc->ignore_block_suitable) 13621ef36db2SYisheng Xie return true; 13631ef36db2SYisheng Xie 1364018e9a49SAndrew Morton /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ 1365b682debdSVlastimil Babka if (is_migrate_movable(get_pageblock_migratetype(page))) 1366018e9a49SAndrew Morton return true; 1367018e9a49SAndrew Morton 1368018e9a49SAndrew Morton /* Otherwise skip the block */ 1369018e9a49SAndrew Morton return false; 1370018e9a49SAndrew Morton } 1371018e9a49SAndrew Morton 137270b44595SMel Gorman static inline unsigned int 137370b44595SMel Gorman freelist_scan_limit(struct compact_control *cc) 137470b44595SMel Gorman { 1375dd7ef7bdSQian Cai unsigned short shift = BITS_PER_LONG - 1; 1376dd7ef7bdSQian Cai 1377dd7ef7bdSQian Cai return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1; 137870b44595SMel Gorman } 137970b44595SMel Gorman 1380ff9543fdSMichal Nazarewicz /* 1381f2849aa0SVlastimil Babka * Test whether the free scanner has reached the same or lower pageblock than 1382f2849aa0SVlastimil Babka * the migration scanner, and compaction should thus terminate. 1383f2849aa0SVlastimil Babka */ 1384f2849aa0SVlastimil Babka static inline bool compact_scanners_met(struct compact_control *cc) 1385f2849aa0SVlastimil Babka { 1386f2849aa0SVlastimil Babka return (cc->free_pfn >> pageblock_order) 1387f2849aa0SVlastimil Babka <= (cc->migrate_pfn >> pageblock_order); 1388f2849aa0SVlastimil Babka } 1389f2849aa0SVlastimil Babka 13905a811889SMel Gorman /* 13915a811889SMel Gorman * Used when scanning for a suitable migration target which scans freelists 13925a811889SMel Gorman * in reverse. Reorders the list such as the unscanned pages are scanned 13935a811889SMel Gorman * first on the next iteration of the free scanner 13945a811889SMel Gorman */ 13955a811889SMel Gorman static void 13965a811889SMel Gorman move_freelist_head(struct list_head *freelist, struct page *freepage) 13975a811889SMel Gorman { 13985a811889SMel Gorman LIST_HEAD(sublist); 13995a811889SMel Gorman 14005a811889SMel Gorman if (!list_is_last(freelist, &freepage->lru)) { 14015a811889SMel Gorman list_cut_before(&sublist, freelist, &freepage->lru); 14025a811889SMel Gorman list_splice_tail(&sublist, freelist); 14035a811889SMel Gorman } 14045a811889SMel Gorman } 14055a811889SMel Gorman 14065a811889SMel Gorman /* 14075a811889SMel Gorman * Similar to move_freelist_head except used by the migration scanner 14085a811889SMel Gorman * when scanning forward. It's possible for these list operations to 14095a811889SMel Gorman * move against each other if they search the free list exactly in 14105a811889SMel Gorman * lockstep. 14115a811889SMel Gorman */ 141270b44595SMel Gorman static void 141370b44595SMel Gorman move_freelist_tail(struct list_head *freelist, struct page *freepage) 141470b44595SMel Gorman { 141570b44595SMel Gorman LIST_HEAD(sublist); 141670b44595SMel Gorman 141770b44595SMel Gorman if (!list_is_first(freelist, &freepage->lru)) { 141870b44595SMel Gorman list_cut_position(&sublist, freelist, &freepage->lru); 141970b44595SMel Gorman list_splice_tail(&sublist, freelist); 142070b44595SMel Gorman } 142170b44595SMel Gorman } 142270b44595SMel Gorman 14235a811889SMel Gorman static void 1424be21b32aSNARIBAYASHI Akira fast_isolate_around(struct compact_control *cc, unsigned long pfn) 14255a811889SMel Gorman { 14265a811889SMel Gorman unsigned long start_pfn, end_pfn; 14276e2b7044SVlastimil Babka struct page *page; 14285a811889SMel Gorman 14295a811889SMel Gorman /* Do not search around if there are enough pages already */ 14305a811889SMel Gorman if (cc->nr_freepages >= cc->nr_migratepages) 14315a811889SMel Gorman return; 14325a811889SMel Gorman 14335a811889SMel Gorman /* Minimise scanning during async compaction */ 14345a811889SMel Gorman if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) 14355a811889SMel Gorman return; 14365a811889SMel Gorman 14375a811889SMel Gorman /* Pageblock boundaries */ 14386e2b7044SVlastimil Babka start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn); 14396e2b7044SVlastimil Babka end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)); 14406e2b7044SVlastimil Babka 14416e2b7044SVlastimil Babka page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone); 14426e2b7044SVlastimil Babka if (!page) 14436e2b7044SVlastimil Babka return; 14445a811889SMel Gorman 14454fca9730SMel Gorman isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false); 14465a811889SMel Gorman 14475a811889SMel Gorman /* Skip this pageblock in the future as it's full or nearly full */ 1448cf650342SBaolin Wang if (start_pfn == end_pfn) 14495a811889SMel Gorman set_pageblock_skip(page); 1450be21b32aSNARIBAYASHI Akira 1451be21b32aSNARIBAYASHI Akira return; 14525a811889SMel Gorman } 14535a811889SMel Gorman 1454dbe2d4e4SMel Gorman /* Search orders in round-robin fashion */ 1455dbe2d4e4SMel Gorman static int next_search_order(struct compact_control *cc, int order) 1456dbe2d4e4SMel Gorman { 1457dbe2d4e4SMel Gorman order--; 1458dbe2d4e4SMel Gorman if (order < 0) 1459dbe2d4e4SMel Gorman order = cc->order - 1; 1460dbe2d4e4SMel Gorman 1461dbe2d4e4SMel Gorman /* Search wrapped around? */ 1462dbe2d4e4SMel Gorman if (order == cc->search_order) { 1463dbe2d4e4SMel Gorman cc->search_order--; 1464dbe2d4e4SMel Gorman if (cc->search_order < 0) 1465dbe2d4e4SMel Gorman cc->search_order = cc->order - 1; 1466dbe2d4e4SMel Gorman return -1; 1467dbe2d4e4SMel Gorman } 1468dbe2d4e4SMel Gorman 1469dbe2d4e4SMel Gorman return order; 1470dbe2d4e4SMel Gorman } 1471dbe2d4e4SMel Gorman 14722dbd9005SBaolin Wang static void fast_isolate_freepages(struct compact_control *cc) 14735a811889SMel Gorman { 1474b55ca526SWonhyuk Yang unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1); 1475447ba886SBaolin Wang unsigned int nr_scanned = 0, total_isolated = 0; 147674e21484SRokudo Yan unsigned long low_pfn, min_pfn, highest = 0; 14775a811889SMel Gorman unsigned long nr_isolated = 0; 14785a811889SMel Gorman unsigned long distance; 14795a811889SMel Gorman struct page *page = NULL; 14805a811889SMel Gorman bool scan_start = false; 14815a811889SMel Gorman int order; 14825a811889SMel Gorman 14835a811889SMel Gorman /* Full compaction passes in a negative order */ 14845a811889SMel Gorman if (cc->order <= 0) 14852dbd9005SBaolin Wang return; 14865a811889SMel Gorman 14875a811889SMel Gorman /* 14885a811889SMel Gorman * If starting the scan, use a deeper search and use the highest 14895a811889SMel Gorman * PFN found if a suitable one is not found. 14905a811889SMel Gorman */ 1491e332f741SMel Gorman if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { 14925a811889SMel Gorman limit = pageblock_nr_pages >> 1; 14935a811889SMel Gorman scan_start = true; 14945a811889SMel Gorman } 14955a811889SMel Gorman 14965a811889SMel Gorman /* 14975a811889SMel Gorman * Preferred point is in the top quarter of the scan space but take 14985a811889SMel Gorman * a pfn from the top half if the search is problematic. 14995a811889SMel Gorman */ 15005a811889SMel Gorman distance = (cc->free_pfn - cc->migrate_pfn); 15015a811889SMel Gorman low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); 15025a811889SMel Gorman min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); 15035a811889SMel Gorman 15045a811889SMel Gorman if (WARN_ON_ONCE(min_pfn > low_pfn)) 15055a811889SMel Gorman low_pfn = min_pfn; 15065a811889SMel Gorman 1507dbe2d4e4SMel Gorman /* 1508dbe2d4e4SMel Gorman * Search starts from the last successful isolation order or the next 1509dbe2d4e4SMel Gorman * order to search after a previous failure 1510dbe2d4e4SMel Gorman */ 1511dbe2d4e4SMel Gorman cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); 1512dbe2d4e4SMel Gorman 1513dbe2d4e4SMel Gorman for (order = cc->search_order; 1514dbe2d4e4SMel Gorman !page && order >= 0; 1515dbe2d4e4SMel Gorman order = next_search_order(cc, order)) { 15165a811889SMel Gorman struct free_area *area = &cc->zone->free_area[order]; 15175a811889SMel Gorman struct list_head *freelist; 15185a811889SMel Gorman struct page *freepage; 15195a811889SMel Gorman unsigned long flags; 15205a811889SMel Gorman unsigned int order_scanned = 0; 152174e21484SRokudo Yan unsigned long high_pfn = 0; 15225a811889SMel Gorman 15235a811889SMel Gorman if (!area->nr_free) 15245a811889SMel Gorman continue; 15255a811889SMel Gorman 15265a811889SMel Gorman spin_lock_irqsave(&cc->zone->lock, flags); 15275a811889SMel Gorman freelist = &area->free_list[MIGRATE_MOVABLE]; 152894ec2003SBaolin Wang list_for_each_entry_reverse(freepage, freelist, buddy_list) { 15295a811889SMel Gorman unsigned long pfn; 15305a811889SMel Gorman 15315a811889SMel Gorman order_scanned++; 15325a811889SMel Gorman nr_scanned++; 15335a811889SMel Gorman pfn = page_to_pfn(freepage); 15345a811889SMel Gorman 15355a811889SMel Gorman if (pfn >= highest) 15366e2b7044SVlastimil Babka highest = max(pageblock_start_pfn(pfn), 15376e2b7044SVlastimil Babka cc->zone->zone_start_pfn); 15385a811889SMel Gorman 15395a811889SMel Gorman if (pfn >= low_pfn) { 15405a811889SMel Gorman cc->fast_search_fail = 0; 1541dbe2d4e4SMel Gorman cc->search_order = order; 15425a811889SMel Gorman page = freepage; 15435a811889SMel Gorman break; 15445a811889SMel Gorman } 15455a811889SMel Gorman 15465a811889SMel Gorman if (pfn >= min_pfn && pfn > high_pfn) { 15475a811889SMel Gorman high_pfn = pfn; 15485a811889SMel Gorman 15495a811889SMel Gorman /* Shorten the scan if a candidate is found */ 15505a811889SMel Gorman limit >>= 1; 15515a811889SMel Gorman } 15525a811889SMel Gorman 15535a811889SMel Gorman if (order_scanned >= limit) 15545a811889SMel Gorman break; 15555a811889SMel Gorman } 15565a811889SMel Gorman 1557e6bd14ecSKemeng Shi /* Use a maximum candidate pfn if a preferred one was not found */ 15585a811889SMel Gorman if (!page && high_pfn) { 15595a811889SMel Gorman page = pfn_to_page(high_pfn); 15605a811889SMel Gorman 15615a811889SMel Gorman /* Update freepage for the list reorder below */ 15625a811889SMel Gorman freepage = page; 15635a811889SMel Gorman } 15645a811889SMel Gorman 15655a811889SMel Gorman /* Reorder to so a future search skips recent pages */ 15665a811889SMel Gorman move_freelist_head(freelist, freepage); 15675a811889SMel Gorman 15685a811889SMel Gorman /* Isolate the page if available */ 15695a811889SMel Gorman if (page) { 15705a811889SMel Gorman if (__isolate_free_page(page, order)) { 15715a811889SMel Gorman set_page_private(page, order); 15725a811889SMel Gorman nr_isolated = 1 << order; 1573b717d6b9SWilliam Lam nr_scanned += nr_isolated - 1; 1574447ba886SBaolin Wang total_isolated += nr_isolated; 15755a811889SMel Gorman cc->nr_freepages += nr_isolated; 15765a811889SMel Gorman list_add_tail(&page->lru, &cc->freepages); 15775a811889SMel Gorman count_compact_events(COMPACTISOLATED, nr_isolated); 15785a811889SMel Gorman } else { 15795a811889SMel Gorman /* If isolation fails, abort the search */ 15805b56d996SQian Cai order = cc->search_order + 1; 15815a811889SMel Gorman page = NULL; 15825a811889SMel Gorman } 15835a811889SMel Gorman } 15845a811889SMel Gorman 15855a811889SMel Gorman spin_unlock_irqrestore(&cc->zone->lock, flags); 15865a811889SMel Gorman 1587a8d13355SBaolin Wang /* Skip fast search if enough freepages isolated */ 1588a8d13355SBaolin Wang if (cc->nr_freepages >= cc->nr_migratepages) 1589a8d13355SBaolin Wang break; 1590a8d13355SBaolin Wang 15915a811889SMel Gorman /* 1592b55ca526SWonhyuk Yang * Smaller scan on next order so the total scan is related 15935a811889SMel Gorman * to freelist_scan_limit. 15945a811889SMel Gorman */ 15955a811889SMel Gorman if (order_scanned >= limit) 1596b55ca526SWonhyuk Yang limit = max(1U, limit >> 1); 15975a811889SMel Gorman } 15985a811889SMel Gorman 1599447ba886SBaolin Wang trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn, 1600447ba886SBaolin Wang nr_scanned, total_isolated); 1601447ba886SBaolin Wang 16025a811889SMel Gorman if (!page) { 16035a811889SMel Gorman cc->fast_search_fail++; 16045a811889SMel Gorman if (scan_start) { 16055a811889SMel Gorman /* 16065a811889SMel Gorman * Use the highest PFN found above min. If one was 1607f3867755SEthon Paul * not found, be pessimistic for direct compaction 16085a811889SMel Gorman * and use the min mark. 16095a811889SMel Gorman */ 1610ca2864e5SMiaohe Lin if (highest >= min_pfn) { 16115a811889SMel Gorman page = pfn_to_page(highest); 16125a811889SMel Gorman cc->free_pfn = highest; 16135a811889SMel Gorman } else { 1614e577c8b6SSuzuki K Poulose if (cc->direct_compaction && pfn_valid(min_pfn)) { 161573a6e474SBaoquan He page = pageblock_pfn_to_page(min_pfn, 16166e2b7044SVlastimil Babka min(pageblock_end_pfn(min_pfn), 16176e2b7044SVlastimil Babka zone_end_pfn(cc->zone)), 161873a6e474SBaoquan He cc->zone); 16195a811889SMel Gorman cc->free_pfn = min_pfn; 16205a811889SMel Gorman } 16215a811889SMel Gorman } 16225a811889SMel Gorman } 16235a811889SMel Gorman } 16245a811889SMel Gorman 1625d097a6f6SMel Gorman if (highest && highest >= cc->zone->compact_cached_free_pfn) { 1626d097a6f6SMel Gorman highest -= pageblock_nr_pages; 16275a811889SMel Gorman cc->zone->compact_cached_free_pfn = highest; 1628d097a6f6SMel Gorman } 16295a811889SMel Gorman 16305a811889SMel Gorman cc->total_free_scanned += nr_scanned; 16315a811889SMel Gorman if (!page) 16322dbd9005SBaolin Wang return; 16335a811889SMel Gorman 16345a811889SMel Gorman low_pfn = page_to_pfn(page); 1635be21b32aSNARIBAYASHI Akira fast_isolate_around(cc, low_pfn); 16365a811889SMel Gorman } 16375a811889SMel Gorman 1638f2849aa0SVlastimil Babka /* 1639ff9543fdSMichal Nazarewicz * Based on information in the current compact_control, find blocks 1640ff9543fdSMichal Nazarewicz * suitable for isolating free pages from and then isolate them. 1641ff9543fdSMichal Nazarewicz */ 1642edc2ca61SVlastimil Babka static void isolate_freepages(struct compact_control *cc) 1643ff9543fdSMichal Nazarewicz { 1644edc2ca61SVlastimil Babka struct zone *zone = cc->zone; 1645ff9543fdSMichal Nazarewicz struct page *page; 1646c96b9e50SVlastimil Babka unsigned long block_start_pfn; /* start of current pageblock */ 1647e14c720eSVlastimil Babka unsigned long isolate_start_pfn; /* exact pfn we start at */ 1648c96b9e50SVlastimil Babka unsigned long block_end_pfn; /* end of current pageblock */ 1649c96b9e50SVlastimil Babka unsigned long low_pfn; /* lowest pfn scanner is able to scan */ 1650ff9543fdSMichal Nazarewicz struct list_head *freelist = &cc->freepages; 16514fca9730SMel Gorman unsigned int stride; 16522fe86e00SMichal Nazarewicz 16535a811889SMel Gorman /* Try a small search of the free lists for a candidate */ 165400bc102fSMiaohe Lin fast_isolate_freepages(cc); 16555a811889SMel Gorman if (cc->nr_freepages) 16565a811889SMel Gorman goto splitmap; 16575a811889SMel Gorman 1658ff9543fdSMichal Nazarewicz /* 1659ff9543fdSMichal Nazarewicz * Initialise the free scanner. The starting point is where we last 166049e068f0SVlastimil Babka * successfully isolated from, zone-cached value, or the end of the 1661e14c720eSVlastimil Babka * zone when isolating for the first time. For looping we also need 1662e14c720eSVlastimil Babka * this pfn aligned down to the pageblock boundary, because we do 1663c96b9e50SVlastimil Babka * block_start_pfn -= pageblock_nr_pages in the for loop. 1664c96b9e50SVlastimil Babka * For ending point, take care when isolating in last pageblock of a 1665a1c1dbebSRandy Dunlap * zone which ends in the middle of a pageblock. 166649e068f0SVlastimil Babka * The low boundary is the end of the pageblock the migration scanner 166749e068f0SVlastimil Babka * is using. 1668ff9543fdSMichal Nazarewicz */ 1669e14c720eSVlastimil Babka isolate_start_pfn = cc->free_pfn; 16705a811889SMel Gorman block_start_pfn = pageblock_start_pfn(isolate_start_pfn); 1671c96b9e50SVlastimil Babka block_end_pfn = min(block_start_pfn + pageblock_nr_pages, 1672c96b9e50SVlastimil Babka zone_end_pfn(zone)); 167306b6640aSVlastimil Babka low_pfn = pageblock_end_pfn(cc->migrate_pfn); 16744fca9730SMel Gorman stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; 16752fe86e00SMichal Nazarewicz 1676ff9543fdSMichal Nazarewicz /* 1677ff9543fdSMichal Nazarewicz * Isolate free pages until enough are available to migrate the 1678ff9543fdSMichal Nazarewicz * pages on cc->migratepages. We stop searching if the migrate 1679ff9543fdSMichal Nazarewicz * and free page scanners meet or enough free pages are isolated. 1680ff9543fdSMichal Nazarewicz */ 1681f5f61a32SVlastimil Babka for (; block_start_pfn >= low_pfn; 1682c96b9e50SVlastimil Babka block_end_pfn = block_start_pfn, 1683e14c720eSVlastimil Babka block_start_pfn -= pageblock_nr_pages, 1684e14c720eSVlastimil Babka isolate_start_pfn = block_start_pfn) { 16854fca9730SMel Gorman unsigned long nr_isolated; 16864fca9730SMel Gorman 1687f6ea3adbSDavid Rientjes /* 1688f6ea3adbSDavid Rientjes * This can iterate a massively long zone without finding any 1689cb810ad2SMel Gorman * suitable migration targets, so periodically check resched. 1690f6ea3adbSDavid Rientjes */ 1691c036ddffSMiaohe Lin if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) 1692cf66f070SMel Gorman cond_resched(); 1693f6ea3adbSDavid Rientjes 16947d49d886SVlastimil Babka page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, 16957d49d886SVlastimil Babka zone); 1696e6e0c767SBaolin Wang if (!page) { 1697e6e0c767SBaolin Wang unsigned long next_pfn; 1698e6e0c767SBaolin Wang 1699e6e0c767SBaolin Wang next_pfn = skip_offline_sections_reverse(block_start_pfn); 1700e6e0c767SBaolin Wang if (next_pfn) 1701e6e0c767SBaolin Wang block_start_pfn = max(next_pfn, low_pfn); 1702e6e0c767SBaolin Wang 1703ff9543fdSMichal Nazarewicz continue; 1704e6e0c767SBaolin Wang } 1705ff9543fdSMichal Nazarewicz 1706ff9543fdSMichal Nazarewicz /* Check the block is suitable for migration */ 17079f7e3387SVlastimil Babka if (!suitable_migration_target(cc, page)) 1708ff9543fdSMichal Nazarewicz continue; 170968e3e926SLinus Torvalds 1710bb13ffebSMel Gorman /* If isolation recently failed, do not retry */ 1711bb13ffebSMel Gorman if (!isolation_suitable(cc, page)) 1712bb13ffebSMel Gorman continue; 1713bb13ffebSMel Gorman 1714e14c720eSVlastimil Babka /* Found a block suitable for isolating free pages from. */ 17154fca9730SMel Gorman nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, 17164fca9730SMel Gorman block_end_pfn, freelist, stride, false); 1717ff9543fdSMichal Nazarewicz 1718d097a6f6SMel Gorman /* Update the skip hint if the full pageblock was scanned */ 1719d097a6f6SMel Gorman if (isolate_start_pfn == block_end_pfn) 1720d097a6f6SMel Gorman update_pageblock_skip(cc, page, block_start_pfn); 1721d097a6f6SMel Gorman 1722cb2dcaf0SMel Gorman /* Are enough freepages isolated? */ 1723cb2dcaf0SMel Gorman if (cc->nr_freepages >= cc->nr_migratepages) { 1724a46cbf3bSDavid Rientjes if (isolate_start_pfn >= block_end_pfn) { 1725a46cbf3bSDavid Rientjes /* 1726a46cbf3bSDavid Rientjes * Restart at previous pageblock if more 1727a46cbf3bSDavid Rientjes * freepages can be isolated next time. 1728a46cbf3bSDavid Rientjes */ 1729f5f61a32SVlastimil Babka isolate_start_pfn = 1730e14c720eSVlastimil Babka block_start_pfn - pageblock_nr_pages; 1731a46cbf3bSDavid Rientjes } 1732be976572SVlastimil Babka break; 1733a46cbf3bSDavid Rientjes } else if (isolate_start_pfn < block_end_pfn) { 1734f5f61a32SVlastimil Babka /* 1735a46cbf3bSDavid Rientjes * If isolation failed early, do not continue 1736a46cbf3bSDavid Rientjes * needlessly. 1737f5f61a32SVlastimil Babka */ 1738a46cbf3bSDavid Rientjes break; 1739f5f61a32SVlastimil Babka } 17404fca9730SMel Gorman 17414fca9730SMel Gorman /* Adjust stride depending on isolation */ 17424fca9730SMel Gorman if (nr_isolated) { 17434fca9730SMel Gorman stride = 1; 17444fca9730SMel Gorman continue; 17454fca9730SMel Gorman } 17464fca9730SMel Gorman stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); 1747c89511abSMel Gorman } 1748ff9543fdSMichal Nazarewicz 17497ed695e0SVlastimil Babka /* 1750f5f61a32SVlastimil Babka * Record where the free scanner will restart next time. Either we 1751f5f61a32SVlastimil Babka * broke from the loop and set isolate_start_pfn based on the last 1752f5f61a32SVlastimil Babka * call to isolate_freepages_block(), or we met the migration scanner 1753f5f61a32SVlastimil Babka * and the loop terminated due to isolate_start_pfn < low_pfn 17547ed695e0SVlastimil Babka */ 1755f5f61a32SVlastimil Babka cc->free_pfn = isolate_start_pfn; 17565a811889SMel Gorman 17575a811889SMel Gorman splitmap: 17585a811889SMel Gorman /* __isolate_free_page() does not map the pages */ 17595a811889SMel Gorman split_map_pages(freelist); 1760748446bbSMel Gorman } 1761748446bbSMel Gorman 1762748446bbSMel Gorman /* 1763748446bbSMel Gorman * This is a migrate-callback that "allocates" freepages by taking pages 1764748446bbSMel Gorman * from the isolated freelists in the block we are migrating to. 1765748446bbSMel Gorman */ 17664e096ae1SMatthew Wilcox (Oracle) static struct folio *compaction_alloc(struct folio *src, unsigned long data) 1767748446bbSMel Gorman { 1768748446bbSMel Gorman struct compact_control *cc = (struct compact_control *)data; 17694e096ae1SMatthew Wilcox (Oracle) struct folio *dst; 1770748446bbSMel Gorman 1771748446bbSMel Gorman if (list_empty(&cc->freepages)) { 1772edc2ca61SVlastimil Babka isolate_freepages(cc); 1773748446bbSMel Gorman 1774748446bbSMel Gorman if (list_empty(&cc->freepages)) 1775748446bbSMel Gorman return NULL; 1776748446bbSMel Gorman } 1777748446bbSMel Gorman 17784e096ae1SMatthew Wilcox (Oracle) dst = list_entry(cc->freepages.next, struct folio, lru); 17794e096ae1SMatthew Wilcox (Oracle) list_del(&dst->lru); 1780748446bbSMel Gorman cc->nr_freepages--; 1781748446bbSMel Gorman 17824e096ae1SMatthew Wilcox (Oracle) return dst; 1783748446bbSMel Gorman } 1784748446bbSMel Gorman 1785748446bbSMel Gorman /* 1786d53aea3dSDavid Rientjes * This is a migrate-callback that "frees" freepages back to the isolated 1787d53aea3dSDavid Rientjes * freelist. All pages on the freelist are from the same zone, so there is no 1788d53aea3dSDavid Rientjes * special handling needed for NUMA. 1789d53aea3dSDavid Rientjes */ 17904e096ae1SMatthew Wilcox (Oracle) static void compaction_free(struct folio *dst, unsigned long data) 1791d53aea3dSDavid Rientjes { 1792d53aea3dSDavid Rientjes struct compact_control *cc = (struct compact_control *)data; 1793d53aea3dSDavid Rientjes 17944e096ae1SMatthew Wilcox (Oracle) list_add(&dst->lru, &cc->freepages); 1795d53aea3dSDavid Rientjes cc->nr_freepages++; 1796d53aea3dSDavid Rientjes } 1797d53aea3dSDavid Rientjes 1798ff9543fdSMichal Nazarewicz /* possible outcome of isolate_migratepages */ 1799ff9543fdSMichal Nazarewicz typedef enum { 1800ff9543fdSMichal Nazarewicz ISOLATE_ABORT, /* Abort compaction now */ 1801ff9543fdSMichal Nazarewicz ISOLATE_NONE, /* No pages isolated, continue scanning */ 1802ff9543fdSMichal Nazarewicz ISOLATE_SUCCESS, /* Pages isolated, migrate */ 1803ff9543fdSMichal Nazarewicz } isolate_migrate_t; 1804ff9543fdSMichal Nazarewicz 1805ff9543fdSMichal Nazarewicz /* 18065bbe3547SEric B Munson * Allow userspace to control policy on scanning the unevictable LRU for 18075bbe3547SEric B Munson * compactable pages. 18085bbe3547SEric B Munson */ 180948fe8ab8SMinghao Chi static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT; 181048fe8ab8SMinghao Chi /* 181148fe8ab8SMinghao Chi * Tunable for proactive compaction. It determines how 181248fe8ab8SMinghao Chi * aggressively the kernel should compact memory in the 181348fe8ab8SMinghao Chi * background. It takes values in the range [0, 100]. 181448fe8ab8SMinghao Chi */ 181548fe8ab8SMinghao Chi static unsigned int __read_mostly sysctl_compaction_proactiveness = 20; 181648fe8ab8SMinghao Chi static int sysctl_extfrag_threshold = 500; 18178b9167cdSWen Yang static int __read_mostly sysctl_compact_memory; 18185bbe3547SEric B Munson 181970b44595SMel Gorman static inline void 182070b44595SMel Gorman update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) 182170b44595SMel Gorman { 182270b44595SMel Gorman if (cc->fast_start_pfn == ULONG_MAX) 182370b44595SMel Gorman return; 182470b44595SMel Gorman 182570b44595SMel Gorman if (!cc->fast_start_pfn) 182670b44595SMel Gorman cc->fast_start_pfn = pfn; 182770b44595SMel Gorman 182870b44595SMel Gorman cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); 182970b44595SMel Gorman } 183070b44595SMel Gorman 183170b44595SMel Gorman static inline unsigned long 183270b44595SMel Gorman reinit_migrate_pfn(struct compact_control *cc) 183370b44595SMel Gorman { 183470b44595SMel Gorman if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) 183570b44595SMel Gorman return cc->migrate_pfn; 183670b44595SMel Gorman 183770b44595SMel Gorman cc->migrate_pfn = cc->fast_start_pfn; 183870b44595SMel Gorman cc->fast_start_pfn = ULONG_MAX; 183970b44595SMel Gorman 184070b44595SMel Gorman return cc->migrate_pfn; 184170b44595SMel Gorman } 184270b44595SMel Gorman 184370b44595SMel Gorman /* 184470b44595SMel Gorman * Briefly search the free lists for a migration source that already has 184570b44595SMel Gorman * some free pages to reduce the number of pages that need migration 184670b44595SMel Gorman * before a pageblock is free. 184770b44595SMel Gorman */ 184870b44595SMel Gorman static unsigned long fast_find_migrateblock(struct compact_control *cc) 184970b44595SMel Gorman { 185070b44595SMel Gorman unsigned int limit = freelist_scan_limit(cc); 185170b44595SMel Gorman unsigned int nr_scanned = 0; 185270b44595SMel Gorman unsigned long distance; 185370b44595SMel Gorman unsigned long pfn = cc->migrate_pfn; 185470b44595SMel Gorman unsigned long high_pfn; 185570b44595SMel Gorman int order; 185615d28d0dSWonhyuk Yang bool found_block = false; 185770b44595SMel Gorman 185870b44595SMel Gorman /* Skip hints are relied on to avoid repeats on the fast search */ 185970b44595SMel Gorman if (cc->ignore_skip_hint) 186070b44595SMel Gorman return pfn; 186170b44595SMel Gorman 186270b44595SMel Gorman /* 1863f9d7fc1aSMel Gorman * If the pageblock should be finished then do not select a different 1864f9d7fc1aSMel Gorman * pageblock. 1865f9d7fc1aSMel Gorman */ 1866f9d7fc1aSMel Gorman if (cc->finish_pageblock) 1867f9d7fc1aSMel Gorman return pfn; 1868f9d7fc1aSMel Gorman 1869f9d7fc1aSMel Gorman /* 187070b44595SMel Gorman * If the migrate_pfn is not at the start of a zone or the start 187170b44595SMel Gorman * of a pageblock then assume this is a continuation of a previous 187270b44595SMel Gorman * scan restarted due to COMPACT_CLUSTER_MAX. 187370b44595SMel Gorman */ 187470b44595SMel Gorman if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) 187570b44595SMel Gorman return pfn; 187670b44595SMel Gorman 187770b44595SMel Gorman /* 187870b44595SMel Gorman * For smaller orders, just linearly scan as the number of pages 187970b44595SMel Gorman * to migrate should be relatively small and does not necessarily 188070b44595SMel Gorman * justify freeing up a large block for a small allocation. 188170b44595SMel Gorman */ 188270b44595SMel Gorman if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) 188370b44595SMel Gorman return pfn; 188470b44595SMel Gorman 188570b44595SMel Gorman /* 188670b44595SMel Gorman * Only allow kcompactd and direct requests for movable pages to 188770b44595SMel Gorman * quickly clear out a MOVABLE pageblock for allocation. This 188870b44595SMel Gorman * reduces the risk that a large movable pageblock is freed for 188970b44595SMel Gorman * an unmovable/reclaimable small allocation. 189070b44595SMel Gorman */ 189170b44595SMel Gorman if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) 189270b44595SMel Gorman return pfn; 189370b44595SMel Gorman 189470b44595SMel Gorman /* 189570b44595SMel Gorman * When starting the migration scanner, pick any pageblock within the 189670b44595SMel Gorman * first half of the search space. Otherwise try and pick a pageblock 189770b44595SMel Gorman * within the first eighth to reduce the chances that a migration 189870b44595SMel Gorman * target later becomes a source. 189970b44595SMel Gorman */ 190070b44595SMel Gorman distance = (cc->free_pfn - cc->migrate_pfn) >> 1; 190170b44595SMel Gorman if (cc->migrate_pfn != cc->zone->zone_start_pfn) 190270b44595SMel Gorman distance >>= 2; 190370b44595SMel Gorman high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); 190470b44595SMel Gorman 190570b44595SMel Gorman for (order = cc->order - 1; 190615d28d0dSWonhyuk Yang order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; 190770b44595SMel Gorman order--) { 190870b44595SMel Gorman struct free_area *area = &cc->zone->free_area[order]; 190970b44595SMel Gorman struct list_head *freelist; 191070b44595SMel Gorman unsigned long flags; 191170b44595SMel Gorman struct page *freepage; 191270b44595SMel Gorman 191370b44595SMel Gorman if (!area->nr_free) 191470b44595SMel Gorman continue; 191570b44595SMel Gorman 191670b44595SMel Gorman spin_lock_irqsave(&cc->zone->lock, flags); 191770b44595SMel Gorman freelist = &area->free_list[MIGRATE_MOVABLE]; 191894ec2003SBaolin Wang list_for_each_entry(freepage, freelist, buddy_list) { 191970b44595SMel Gorman unsigned long free_pfn; 192070b44595SMel Gorman 192115d28d0dSWonhyuk Yang if (nr_scanned++ >= limit) { 192215d28d0dSWonhyuk Yang move_freelist_tail(freelist, freepage); 192315d28d0dSWonhyuk Yang break; 192415d28d0dSWonhyuk Yang } 192515d28d0dSWonhyuk Yang 192670b44595SMel Gorman free_pfn = page_to_pfn(freepage); 192770b44595SMel Gorman if (free_pfn < high_pfn) { 192870b44595SMel Gorman /* 192970b44595SMel Gorman * Avoid if skipped recently. Ideally it would 193070b44595SMel Gorman * move to the tail but even safe iteration of 193170b44595SMel Gorman * the list assumes an entry is deleted, not 193270b44595SMel Gorman * reordered. 193370b44595SMel Gorman */ 193415d28d0dSWonhyuk Yang if (get_pageblock_skip(freepage)) 193570b44595SMel Gorman continue; 193670b44595SMel Gorman 193770b44595SMel Gorman /* Reorder to so a future search skips recent pages */ 193870b44595SMel Gorman move_freelist_tail(freelist, freepage); 193970b44595SMel Gorman 1940e380bebeSMel Gorman update_fast_start_pfn(cc, free_pfn); 194170b44595SMel Gorman pfn = pageblock_start_pfn(free_pfn); 1942bbe832b9SRei Yamamoto if (pfn < cc->zone->zone_start_pfn) 1943bbe832b9SRei Yamamoto pfn = cc->zone->zone_start_pfn; 194470b44595SMel Gorman cc->fast_search_fail = 0; 194515d28d0dSWonhyuk Yang found_block = true; 194670b44595SMel Gorman break; 194770b44595SMel Gorman } 194870b44595SMel Gorman } 194970b44595SMel Gorman spin_unlock_irqrestore(&cc->zone->lock, flags); 195070b44595SMel Gorman } 195170b44595SMel Gorman 195270b44595SMel Gorman cc->total_migrate_scanned += nr_scanned; 195370b44595SMel Gorman 195470b44595SMel Gorman /* 195570b44595SMel Gorman * If fast scanning failed then use a cached entry for a page block 195670b44595SMel Gorman * that had free pages as the basis for starting a linear scan. 195770b44595SMel Gorman */ 195815d28d0dSWonhyuk Yang if (!found_block) { 195915d28d0dSWonhyuk Yang cc->fast_search_fail++; 196070b44595SMel Gorman pfn = reinit_migrate_pfn(cc); 196115d28d0dSWonhyuk Yang } 196270b44595SMel Gorman return pfn; 196370b44595SMel Gorman } 196470b44595SMel Gorman 19655bbe3547SEric B Munson /* 1966edc2ca61SVlastimil Babka * Isolate all pages that can be migrated from the first suitable block, 1967edc2ca61SVlastimil Babka * starting at the block pointed to by the migrate scanner pfn within 1968edc2ca61SVlastimil Babka * compact_control. 1969ff9543fdSMichal Nazarewicz */ 197032aaf055SPengfei Li static isolate_migrate_t isolate_migratepages(struct compact_control *cc) 1971ff9543fdSMichal Nazarewicz { 1972e1409c32SJoonsoo Kim unsigned long block_start_pfn; 1973e1409c32SJoonsoo Kim unsigned long block_end_pfn; 1974e1409c32SJoonsoo Kim unsigned long low_pfn; 1975edc2ca61SVlastimil Babka struct page *page; 1976edc2ca61SVlastimil Babka const isolate_mode_t isolate_mode = 19775bbe3547SEric B Munson (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | 19781d2047feSHugh Dickins (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); 197970b44595SMel Gorman bool fast_find_block; 1980ff9543fdSMichal Nazarewicz 1981edc2ca61SVlastimil Babka /* 1982edc2ca61SVlastimil Babka * Start at where we last stopped, or beginning of the zone as 198370b44595SMel Gorman * initialized by compact_zone(). The first failure will use 198470b44595SMel Gorman * the lowest PFN as the starting point for linear scanning. 1985edc2ca61SVlastimil Babka */ 198670b44595SMel Gorman low_pfn = fast_find_migrateblock(cc); 198706b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(low_pfn); 198832aaf055SPengfei Li if (block_start_pfn < cc->zone->zone_start_pfn) 198932aaf055SPengfei Li block_start_pfn = cc->zone->zone_start_pfn; 1990ff9543fdSMichal Nazarewicz 199170b44595SMel Gorman /* 199270b44595SMel Gorman * fast_find_migrateblock marks a pageblock skipped so to avoid 199370b44595SMel Gorman * the isolation_suitable check below, check whether the fast 199470b44595SMel Gorman * search was successful. 199570b44595SMel Gorman */ 199670b44595SMel Gorman fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; 199770b44595SMel Gorman 1998ff9543fdSMichal Nazarewicz /* Only scan within a pageblock boundary */ 199906b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(low_pfn); 2000ff9543fdSMichal Nazarewicz 2001edc2ca61SVlastimil Babka /* 2002edc2ca61SVlastimil Babka * Iterate over whole pageblocks until we find the first suitable. 2003edc2ca61SVlastimil Babka * Do not cross the free scanner. 2004edc2ca61SVlastimil Babka */ 2005e1409c32SJoonsoo Kim for (; block_end_pfn <= cc->free_pfn; 200670b44595SMel Gorman fast_find_block = false, 2007c2ad7a1fSOscar Salvador cc->migrate_pfn = low_pfn = block_end_pfn, 2008e1409c32SJoonsoo Kim block_start_pfn = block_end_pfn, 2009e1409c32SJoonsoo Kim block_end_pfn += pageblock_nr_pages) { 2010edc2ca61SVlastimil Babka 2011edc2ca61SVlastimil Babka /* 2012edc2ca61SVlastimil Babka * This can potentially iterate a massively long zone with 2013edc2ca61SVlastimil Babka * many pageblocks unsuitable, so periodically check if we 2014cb810ad2SMel Gorman * need to schedule. 2015edc2ca61SVlastimil Babka */ 2016c036ddffSMiaohe Lin if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) 2017cf66f070SMel Gorman cond_resched(); 2018edc2ca61SVlastimil Babka 201932aaf055SPengfei Li page = pageblock_pfn_to_page(block_start_pfn, 202032aaf055SPengfei Li block_end_pfn, cc->zone); 20219721fd82SBaolin Wang if (!page) { 20229721fd82SBaolin Wang unsigned long next_pfn; 20239721fd82SBaolin Wang 20249721fd82SBaolin Wang next_pfn = skip_offline_sections(block_start_pfn); 20259721fd82SBaolin Wang if (next_pfn) 20269721fd82SBaolin Wang block_end_pfn = min(next_pfn, cc->free_pfn); 2027edc2ca61SVlastimil Babka continue; 20289721fd82SBaolin Wang } 2029edc2ca61SVlastimil Babka 2030e380bebeSMel Gorman /* 2031e380bebeSMel Gorman * If isolation recently failed, do not retry. Only check the 2032e380bebeSMel Gorman * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock 2033e380bebeSMel Gorman * to be visited multiple times. Assume skip was checked 2034e380bebeSMel Gorman * before making it "skip" so other compaction instances do 2035e380bebeSMel Gorman * not scan the same block. 2036e380bebeSMel Gorman */ 2037ee0913c4SKefeng Wang if (pageblock_aligned(low_pfn) && 2038e380bebeSMel Gorman !fast_find_block && !isolation_suitable(cc, page)) 2039edc2ca61SVlastimil Babka continue; 2040edc2ca61SVlastimil Babka 2041edc2ca61SVlastimil Babka /* 2042556162bfSMiaohe Lin * For async direct compaction, only scan the pageblocks of the 2043556162bfSMiaohe Lin * same migratetype without huge pages. Async direct compaction 2044556162bfSMiaohe Lin * is optimistic to see if the minimum amount of work satisfies 2045556162bfSMiaohe Lin * the allocation. The cached PFN is updated as it's possible 2046556162bfSMiaohe Lin * that all remaining blocks between source and target are 2047556162bfSMiaohe Lin * unsuitable and the compaction scanners fail to meet. 2048edc2ca61SVlastimil Babka */ 20499bebefd5SMel Gorman if (!suitable_migration_source(cc, page)) { 20509bebefd5SMel Gorman update_cached_migrate(cc, block_end_pfn); 2051edc2ca61SVlastimil Babka continue; 20529bebefd5SMel Gorman } 2053ff9543fdSMichal Nazarewicz 2054ff9543fdSMichal Nazarewicz /* Perform the isolation */ 2055c2ad7a1fSOscar Salvador if (isolate_migratepages_block(cc, low_pfn, block_end_pfn, 2056c2ad7a1fSOscar Salvador isolate_mode)) 2057ff9543fdSMichal Nazarewicz return ISOLATE_ABORT; 2058ff9543fdSMichal Nazarewicz 2059edc2ca61SVlastimil Babka /* 2060edc2ca61SVlastimil Babka * Either we isolated something and proceed with migration. Or 2061edc2ca61SVlastimil Babka * we failed and compact_zone should decide if we should 2062edc2ca61SVlastimil Babka * continue or not. 2063edc2ca61SVlastimil Babka */ 2064edc2ca61SVlastimil Babka break; 2065edc2ca61SVlastimil Babka } 2066edc2ca61SVlastimil Babka 2067edc2ca61SVlastimil Babka return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; 2068ff9543fdSMichal Nazarewicz } 2069ff9543fdSMichal Nazarewicz 207021c527a3SYaowei Bai /* 207121c527a3SYaowei Bai * order == -1 is expected when compacting via 207221c527a3SYaowei Bai * /proc/sys/vm/compact_memory 207321c527a3SYaowei Bai */ 207421c527a3SYaowei Bai static inline bool is_via_compact_memory(int order) 207521c527a3SYaowei Bai { 207621c527a3SYaowei Bai return order == -1; 207721c527a3SYaowei Bai } 207821c527a3SYaowei Bai 2079b4a0215eSKefeng Wang /* 2080b4a0215eSKefeng Wang * Determine whether kswapd is (or recently was!) running on this node. 2081b4a0215eSKefeng Wang * 2082b4a0215eSKefeng Wang * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't 2083b4a0215eSKefeng Wang * zero it. 2084b4a0215eSKefeng Wang */ 2085facdaa91SNitin Gupta static bool kswapd_is_running(pg_data_t *pgdat) 2086facdaa91SNitin Gupta { 2087b4a0215eSKefeng Wang bool running; 2088b4a0215eSKefeng Wang 2089b4a0215eSKefeng Wang pgdat_kswapd_lock(pgdat); 2090b4a0215eSKefeng Wang running = pgdat->kswapd && task_is_running(pgdat->kswapd); 2091b4a0215eSKefeng Wang pgdat_kswapd_unlock(pgdat); 2092b4a0215eSKefeng Wang 2093b4a0215eSKefeng Wang return running; 2094facdaa91SNitin Gupta } 2095facdaa91SNitin Gupta 2096facdaa91SNitin Gupta /* 2097facdaa91SNitin Gupta * A zone's fragmentation score is the external fragmentation wrt to the 209840d7e203SCharan Teja Reddy * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100]. 209940d7e203SCharan Teja Reddy */ 210040d7e203SCharan Teja Reddy static unsigned int fragmentation_score_zone(struct zone *zone) 210140d7e203SCharan Teja Reddy { 210240d7e203SCharan Teja Reddy return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER); 210340d7e203SCharan Teja Reddy } 210440d7e203SCharan Teja Reddy 210540d7e203SCharan Teja Reddy /* 210640d7e203SCharan Teja Reddy * A weighted zone's fragmentation score is the external fragmentation 210740d7e203SCharan Teja Reddy * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It 210840d7e203SCharan Teja Reddy * returns a value in the range [0, 100]. 2109facdaa91SNitin Gupta * 2110facdaa91SNitin Gupta * The scaling factor ensures that proactive compaction focuses on larger 2111facdaa91SNitin Gupta * zones like ZONE_NORMAL, rather than smaller, specialized zones like 2112facdaa91SNitin Gupta * ZONE_DMA32. For smaller zones, the score value remains close to zero, 2113facdaa91SNitin Gupta * and thus never exceeds the high threshold for proactive compaction. 2114facdaa91SNitin Gupta */ 211540d7e203SCharan Teja Reddy static unsigned int fragmentation_score_zone_weighted(struct zone *zone) 2116facdaa91SNitin Gupta { 2117facdaa91SNitin Gupta unsigned long score; 2118facdaa91SNitin Gupta 211940d7e203SCharan Teja Reddy score = zone->present_pages * fragmentation_score_zone(zone); 2120facdaa91SNitin Gupta return div64_ul(score, zone->zone_pgdat->node_present_pages + 1); 2121facdaa91SNitin Gupta } 2122facdaa91SNitin Gupta 2123facdaa91SNitin Gupta /* 2124facdaa91SNitin Gupta * The per-node proactive (background) compaction process is started by its 2125facdaa91SNitin Gupta * corresponding kcompactd thread when the node's fragmentation score 2126facdaa91SNitin Gupta * exceeds the high threshold. The compaction process remains active till 2127facdaa91SNitin Gupta * the node's score falls below the low threshold, or one of the back-off 2128facdaa91SNitin Gupta * conditions is met. 2129facdaa91SNitin Gupta */ 2130d34c0a75SNitin Gupta static unsigned int fragmentation_score_node(pg_data_t *pgdat) 2131facdaa91SNitin Gupta { 2132d34c0a75SNitin Gupta unsigned int score = 0; 2133facdaa91SNitin Gupta int zoneid; 2134facdaa91SNitin Gupta 2135facdaa91SNitin Gupta for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 2136facdaa91SNitin Gupta struct zone *zone; 2137facdaa91SNitin Gupta 2138facdaa91SNitin Gupta zone = &pgdat->node_zones[zoneid]; 21399e552271SBaolin Wang if (!populated_zone(zone)) 21409e552271SBaolin Wang continue; 214140d7e203SCharan Teja Reddy score += fragmentation_score_zone_weighted(zone); 2142facdaa91SNitin Gupta } 2143facdaa91SNitin Gupta 2144facdaa91SNitin Gupta return score; 2145facdaa91SNitin Gupta } 2146facdaa91SNitin Gupta 2147d34c0a75SNitin Gupta static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low) 2148facdaa91SNitin Gupta { 2149d34c0a75SNitin Gupta unsigned int wmark_low; 2150facdaa91SNitin Gupta 2151facdaa91SNitin Gupta /* 2152f0953a1bSIngo Molnar * Cap the low watermark to avoid excessive compaction 2153f0953a1bSIngo Molnar * activity in case a user sets the proactiveness tunable 2154facdaa91SNitin Gupta * close to 100 (maximum). 2155facdaa91SNitin Gupta */ 2156d34c0a75SNitin Gupta wmark_low = max(100U - sysctl_compaction_proactiveness, 5U); 2157d34c0a75SNitin Gupta return low ? wmark_low : min(wmark_low + 10, 100U); 2158facdaa91SNitin Gupta } 2159facdaa91SNitin Gupta 2160facdaa91SNitin Gupta static bool should_proactive_compact_node(pg_data_t *pgdat) 2161facdaa91SNitin Gupta { 2162facdaa91SNitin Gupta int wmark_high; 2163facdaa91SNitin Gupta 2164facdaa91SNitin Gupta if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat)) 2165facdaa91SNitin Gupta return false; 2166facdaa91SNitin Gupta 2167facdaa91SNitin Gupta wmark_high = fragmentation_score_wmark(pgdat, false); 2168facdaa91SNitin Gupta return fragmentation_score_node(pgdat) > wmark_high; 2169facdaa91SNitin Gupta } 2170facdaa91SNitin Gupta 217140cacbcbSMel Gorman static enum compact_result __compact_finished(struct compact_control *cc) 2172748446bbSMel Gorman { 21738fb74b9fSMel Gorman unsigned int order; 2174d39773a0SVlastimil Babka const int migratetype = cc->migratetype; 2175cb2dcaf0SMel Gorman int ret; 2176748446bbSMel Gorman 2177753341a4SMel Gorman /* Compaction run completes if the migrate and free scanner meet */ 2178f2849aa0SVlastimil Babka if (compact_scanners_met(cc)) { 217955b7c4c9SVlastimil Babka /* Let the next compaction start anew. */ 218040cacbcbSMel Gorman reset_cached_positions(cc->zone); 218155b7c4c9SVlastimil Babka 218262997027SMel Gorman /* 218362997027SMel Gorman * Mark that the PG_migrate_skip information should be cleared 2184accf6242SVlastimil Babka * by kswapd when it goes to sleep. kcompactd does not set the 218562997027SMel Gorman * flag itself as the decision to be clear should be directly 218662997027SMel Gorman * based on an allocation request. 218762997027SMel Gorman */ 2188accf6242SVlastimil Babka if (cc->direct_compaction) 218940cacbcbSMel Gorman cc->zone->compact_blockskip_flush = true; 219062997027SMel Gorman 2191c8f7de0bSMichal Hocko if (cc->whole_zone) 2192748446bbSMel Gorman return COMPACT_COMPLETE; 2193c8f7de0bSMichal Hocko else 2194c8f7de0bSMichal Hocko return COMPACT_PARTIAL_SKIPPED; 2195bb13ffebSMel Gorman } 2196748446bbSMel Gorman 2197facdaa91SNitin Gupta if (cc->proactive_compaction) { 2198facdaa91SNitin Gupta int score, wmark_low; 2199facdaa91SNitin Gupta pg_data_t *pgdat; 2200facdaa91SNitin Gupta 2201facdaa91SNitin Gupta pgdat = cc->zone->zone_pgdat; 2202facdaa91SNitin Gupta if (kswapd_is_running(pgdat)) 2203facdaa91SNitin Gupta return COMPACT_PARTIAL_SKIPPED; 2204facdaa91SNitin Gupta 2205facdaa91SNitin Gupta score = fragmentation_score_zone(cc->zone); 2206facdaa91SNitin Gupta wmark_low = fragmentation_score_wmark(pgdat, true); 2207facdaa91SNitin Gupta 2208facdaa91SNitin Gupta if (score > wmark_low) 2209facdaa91SNitin Gupta ret = COMPACT_CONTINUE; 2210facdaa91SNitin Gupta else 2211facdaa91SNitin Gupta ret = COMPACT_SUCCESS; 2212facdaa91SNitin Gupta 2213facdaa91SNitin Gupta goto out; 2214facdaa91SNitin Gupta } 2215facdaa91SNitin Gupta 221621c527a3SYaowei Bai if (is_via_compact_memory(cc->order)) 221756de7263SMel Gorman return COMPACT_CONTINUE; 221856de7263SMel Gorman 2219baf6a9a1SVlastimil Babka /* 2220efe771c7SMel Gorman * Always finish scanning a pageblock to reduce the possibility of 2221efe771c7SMel Gorman * fallbacks in the future. This is particularly important when 2222efe771c7SMel Gorman * migration source is unmovable/reclaimable but it's not worth 2223efe771c7SMel Gorman * special casing. 2224baf6a9a1SVlastimil Babka */ 2225ee0913c4SKefeng Wang if (!pageblock_aligned(cc->migrate_pfn)) 2226baf6a9a1SVlastimil Babka return COMPACT_CONTINUE; 2227baf6a9a1SVlastimil Babka 222856de7263SMel Gorman /* Direct compactor: Is a suitable page free? */ 2229cb2dcaf0SMel Gorman ret = COMPACT_NO_SUITABLE_PAGE; 223023baf831SKirill A. Shutemov for (order = cc->order; order <= MAX_ORDER; order++) { 223140cacbcbSMel Gorman struct free_area *area = &cc->zone->free_area[order]; 22322149cdaeSJoonsoo Kim bool can_steal; 22338fb74b9fSMel Gorman 223456de7263SMel Gorman /* Job done if page is free of the right migratetype */ 2235b03641afSDan Williams if (!free_area_empty(area, migratetype)) 2236cf378319SVlastimil Babka return COMPACT_SUCCESS; 223756de7263SMel Gorman 22382149cdaeSJoonsoo Kim #ifdef CONFIG_CMA 22392149cdaeSJoonsoo Kim /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ 22402149cdaeSJoonsoo Kim if (migratetype == MIGRATE_MOVABLE && 2241b03641afSDan Williams !free_area_empty(area, MIGRATE_CMA)) 2242cf378319SVlastimil Babka return COMPACT_SUCCESS; 22432149cdaeSJoonsoo Kim #endif 22442149cdaeSJoonsoo Kim /* 22452149cdaeSJoonsoo Kim * Job done if allocation would steal freepages from 22462149cdaeSJoonsoo Kim * other migratetype buddy lists. 22472149cdaeSJoonsoo Kim */ 22482149cdaeSJoonsoo Kim if (find_suitable_fallback(area, order, migratetype, 2249fa599c44SMiaohe Lin true, &can_steal) != -1) 2250baf6a9a1SVlastimil Babka /* 2251fa599c44SMiaohe Lin * Movable pages are OK in any pageblock. If we are 2252fa599c44SMiaohe Lin * stealing for a non-movable allocation, make sure 2253fa599c44SMiaohe Lin * we finish compacting the current pageblock first 2254fa599c44SMiaohe Lin * (which is assured by the above migrate_pfn align 2255fa599c44SMiaohe Lin * check) so it is as free as possible and we won't 2256fa599c44SMiaohe Lin * have to steal another one soon. 2257baf6a9a1SVlastimil Babka */ 2258baf6a9a1SVlastimil Babka return COMPACT_SUCCESS; 2259baf6a9a1SVlastimil Babka } 2260baf6a9a1SVlastimil Babka 2261facdaa91SNitin Gupta out: 2262cb2dcaf0SMel Gorman if (cc->contended || fatal_signal_pending(current)) 2263cb2dcaf0SMel Gorman ret = COMPACT_CONTENDED; 2264cb2dcaf0SMel Gorman 2265cb2dcaf0SMel Gorman return ret; 2266837d026dSJoonsoo Kim } 2267837d026dSJoonsoo Kim 226840cacbcbSMel Gorman static enum compact_result compact_finished(struct compact_control *cc) 2269837d026dSJoonsoo Kim { 2270837d026dSJoonsoo Kim int ret; 2271837d026dSJoonsoo Kim 227240cacbcbSMel Gorman ret = __compact_finished(cc); 227340cacbcbSMel Gorman trace_mm_compaction_finished(cc->zone, cc->order, ret); 2274837d026dSJoonsoo Kim if (ret == COMPACT_NO_SUITABLE_PAGE) 2275837d026dSJoonsoo Kim ret = COMPACT_CONTINUE; 2276837d026dSJoonsoo Kim 2277837d026dSJoonsoo Kim return ret; 2278748446bbSMel Gorman } 2279748446bbSMel Gorman 22803cf04937SJohannes Weiner static bool __compaction_suitable(struct zone *zone, int order, 228197a225e6SJoonsoo Kim int highest_zoneidx, 228286a294a8SMichal Hocko unsigned long wmark_target) 22833e7d3449SMel Gorman { 22843e7d3449SMel Gorman unsigned long watermark; 22853957c776SMichal Hocko /* 22869861a62cSVlastimil Babka * Watermarks for order-0 must be met for compaction to be able to 2287984fdba6SVlastimil Babka * isolate free pages for migration targets. This means that the 2288984fdba6SVlastimil Babka * watermark and alloc_flags have to match, or be more pessimistic than 2289984fdba6SVlastimil Babka * the check in __isolate_free_page(). We don't use the direct 2290984fdba6SVlastimil Babka * compactor's alloc_flags, as they are not relevant for freepage 229197a225e6SJoonsoo Kim * isolation. We however do use the direct compactor's highest_zoneidx 229297a225e6SJoonsoo Kim * to skip over zones where lowmem reserves would prevent allocation 229397a225e6SJoonsoo Kim * even if compaction succeeds. 22948348faf9SVlastimil Babka * For costly orders, we require low watermark instead of min for 22958348faf9SVlastimil Babka * compaction to proceed to increase its chances. 2296d883c6cfSJoonsoo Kim * ALLOC_CMA is used, as pages in CMA pageblocks are considered 2297d883c6cfSJoonsoo Kim * suitable migration targets 22983e7d3449SMel Gorman */ 22998348faf9SVlastimil Babka watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? 23008348faf9SVlastimil Babka low_wmark_pages(zone) : min_wmark_pages(zone); 23018348faf9SVlastimil Babka watermark += compact_gap(order); 23023cf04937SJohannes Weiner return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx, 23033cf04937SJohannes Weiner ALLOC_CMA, wmark_target); 2304cc5c9f09SVlastimil Babka } 2305cc5c9f09SVlastimil Babka 23062b1a20c3SHui Su /* 23072b1a20c3SHui Su * compaction_suitable: Is this suitable to run compaction on this zone now? 23082b1a20c3SHui Su */ 23093cf04937SJohannes Weiner bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx) 2310cc5c9f09SVlastimil Babka { 23113cf04937SJohannes Weiner enum compact_result compact_result; 23123cf04937SJohannes Weiner bool suitable; 2313cc5c9f09SVlastimil Babka 23143cf04937SJohannes Weiner suitable = __compaction_suitable(zone, order, highest_zoneidx, 2315cc5c9f09SVlastimil Babka zone_page_state(zone, NR_FREE_PAGES)); 23163e7d3449SMel Gorman /* 23173e7d3449SMel Gorman * fragmentation index determines if allocation failures are due to 23183e7d3449SMel Gorman * low memory or external fragmentation 23193e7d3449SMel Gorman * 2320ebff3980SVlastimil Babka * index of -1000 would imply allocations might succeed depending on 2321ebff3980SVlastimil Babka * watermarks, but we already failed the high-order watermark check 23223e7d3449SMel Gorman * index towards 0 implies failure is due to lack of memory 23233e7d3449SMel Gorman * index towards 1000 implies failure is due to fragmentation 23243e7d3449SMel Gorman * 232520311420SVlastimil Babka * Only compact if a failure would be due to fragmentation. Also 232620311420SVlastimil Babka * ignore fragindex for non-costly orders where the alternative to 232720311420SVlastimil Babka * a successful reclaim/compaction is OOM. Fragindex and the 232820311420SVlastimil Babka * vm.extfrag_threshold sysctl is meant as a heuristic to prevent 232920311420SVlastimil Babka * excessive compaction for costly orders, but it should not be at the 233020311420SVlastimil Babka * expense of system stability. 23313e7d3449SMel Gorman */ 23323cf04937SJohannes Weiner if (suitable) { 23333cf04937SJohannes Weiner compact_result = COMPACT_CONTINUE; 23343cf04937SJohannes Weiner if (order > PAGE_ALLOC_COSTLY_ORDER) { 23353cf04937SJohannes Weiner int fragindex = fragmentation_index(zone, order); 23363cf04937SJohannes Weiner 23373cf04937SJohannes Weiner if (fragindex >= 0 && 23383cf04937SJohannes Weiner fragindex <= sysctl_extfrag_threshold) { 23393cf04937SJohannes Weiner suitable = false; 23403cf04937SJohannes Weiner compact_result = COMPACT_NOT_SUITABLE_ZONE; 23413cf04937SJohannes Weiner } 23423cf04937SJohannes Weiner } 23433cf04937SJohannes Weiner } else { 23443cf04937SJohannes Weiner compact_result = COMPACT_SKIPPED; 23453e7d3449SMel Gorman } 23463e7d3449SMel Gorman 23473cf04937SJohannes Weiner trace_mm_compaction_suitable(zone, order, compact_result); 2348837d026dSJoonsoo Kim 23493cf04937SJohannes Weiner return suitable; 2350837d026dSJoonsoo Kim } 2351837d026dSJoonsoo Kim 235286a294a8SMichal Hocko bool compaction_zonelist_suitable(struct alloc_context *ac, int order, 235386a294a8SMichal Hocko int alloc_flags) 235486a294a8SMichal Hocko { 235586a294a8SMichal Hocko struct zone *zone; 235686a294a8SMichal Hocko struct zoneref *z; 235786a294a8SMichal Hocko 235886a294a8SMichal Hocko /* 235986a294a8SMichal Hocko * Make sure at least one zone would pass __compaction_suitable if we continue 236086a294a8SMichal Hocko * retrying the reclaim. 236186a294a8SMichal Hocko */ 236297a225e6SJoonsoo Kim for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, 236397a225e6SJoonsoo Kim ac->highest_zoneidx, ac->nodemask) { 236486a294a8SMichal Hocko unsigned long available; 236586a294a8SMichal Hocko 236686a294a8SMichal Hocko /* 236786a294a8SMichal Hocko * Do not consider all the reclaimable memory because we do not 236886a294a8SMichal Hocko * want to trash just for a single high order allocation which 236986a294a8SMichal Hocko * is even not guaranteed to appear even if __compaction_suitable 237086a294a8SMichal Hocko * is happy about the watermark check. 237186a294a8SMichal Hocko */ 23725a1c84b4SMel Gorman available = zone_reclaimable_pages(zone) / order; 237386a294a8SMichal Hocko available += zone_page_state_snapshot(zone, NR_FREE_PAGES); 2374e8606320SJohannes Weiner if (__compaction_suitable(zone, order, ac->highest_zoneidx, 23753cf04937SJohannes Weiner available)) 237686a294a8SMichal Hocko return true; 237786a294a8SMichal Hocko } 237886a294a8SMichal Hocko 237986a294a8SMichal Hocko return false; 238086a294a8SMichal Hocko } 238186a294a8SMichal Hocko 23825e1f0f09SMel Gorman static enum compact_result 23835e1f0f09SMel Gorman compact_zone(struct compact_control *cc, struct capture_control *capc) 2384748446bbSMel Gorman { 2385ea7ab982SMichal Hocko enum compact_result ret; 238640cacbcbSMel Gorman unsigned long start_pfn = cc->zone->zone_start_pfn; 238740cacbcbSMel Gorman unsigned long end_pfn = zone_end_pfn(cc->zone); 2388566e54e1SMel Gorman unsigned long last_migrated_pfn; 2389e0b9daebSDavid Rientjes const bool sync = cc->mode != MIGRATE_ASYNC; 23908854c55fSMel Gorman bool update_cached; 239184b328aaSBaolin Wang unsigned int nr_succeeded = 0; 2392748446bbSMel Gorman 2393a94b5252SYafang Shao /* 2394a94b5252SYafang Shao * These counters track activities during zone compaction. Initialize 2395a94b5252SYafang Shao * them before compacting a new zone. 2396a94b5252SYafang Shao */ 2397a94b5252SYafang Shao cc->total_migrate_scanned = 0; 2398a94b5252SYafang Shao cc->total_free_scanned = 0; 2399a94b5252SYafang Shao cc->nr_migratepages = 0; 2400a94b5252SYafang Shao cc->nr_freepages = 0; 2401a94b5252SYafang Shao INIT_LIST_HEAD(&cc->freepages); 2402a94b5252SYafang Shao INIT_LIST_HEAD(&cc->migratepages); 2403a94b5252SYafang Shao 240401c0bfe0SWei Yang cc->migratetype = gfp_migratetype(cc->gfp_mask); 2405e8606320SJohannes Weiner 2406e8606320SJohannes Weiner if (!is_via_compact_memory(cc->order)) { 2407e8606320SJohannes Weiner unsigned long watermark; 2408e8606320SJohannes Weiner 2409e8606320SJohannes Weiner /* Allocation can already succeed, nothing to do */ 2410e8606320SJohannes Weiner watermark = wmark_pages(cc->zone, 2411e8606320SJohannes Weiner cc->alloc_flags & ALLOC_WMARK_MASK); 2412e8606320SJohannes Weiner if (zone_watermark_ok(cc->zone, cc->order, watermark, 2413e8606320SJohannes Weiner cc->highest_zoneidx, cc->alloc_flags)) 2414e8606320SJohannes Weiner return COMPACT_SUCCESS; 2415e8606320SJohannes Weiner 24163e7d3449SMel Gorman /* Compaction is likely to fail */ 24173cf04937SJohannes Weiner if (!compaction_suitable(cc->zone, cc->order, 24183cf04937SJohannes Weiner cc->highest_zoneidx)) 24193cf04937SJohannes Weiner return COMPACT_SKIPPED; 2420e8606320SJohannes Weiner } 2421c46649deSMichal Hocko 2422c89511abSMel Gorman /* 2423d3132e4bSVlastimil Babka * Clear pageblock skip if there were failures recently and compaction 2424accf6242SVlastimil Babka * is about to be retried after being deferred. 2425d3132e4bSVlastimil Babka */ 242640cacbcbSMel Gorman if (compaction_restarting(cc->zone, cc->order)) 242740cacbcbSMel Gorman __reset_isolation_suitable(cc->zone); 2428d3132e4bSVlastimil Babka 2429d3132e4bSVlastimil Babka /* 2430c89511abSMel Gorman * Setup to move all movable pages to the end of the zone. Used cached 243106ed2998SVlastimil Babka * information on where the scanners should start (unless we explicitly 243206ed2998SVlastimil Babka * want to compact the whole zone), but check that it is initialised 243306ed2998SVlastimil Babka * by ensuring the values are within zone boundaries. 2434c89511abSMel Gorman */ 243570b44595SMel Gorman cc->fast_start_pfn = 0; 243606ed2998SVlastimil Babka if (cc->whole_zone) { 243706ed2998SVlastimil Babka cc->migrate_pfn = start_pfn; 243806ed2998SVlastimil Babka cc->free_pfn = pageblock_start_pfn(end_pfn - 1); 243906ed2998SVlastimil Babka } else { 244040cacbcbSMel Gorman cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; 244140cacbcbSMel Gorman cc->free_pfn = cc->zone->compact_cached_free_pfn; 2442623446e4SJoonsoo Kim if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { 244306b6640aSVlastimil Babka cc->free_pfn = pageblock_start_pfn(end_pfn - 1); 244440cacbcbSMel Gorman cc->zone->compact_cached_free_pfn = cc->free_pfn; 2445c89511abSMel Gorman } 2446623446e4SJoonsoo Kim if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { 2447c89511abSMel Gorman cc->migrate_pfn = start_pfn; 244840cacbcbSMel Gorman cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; 244940cacbcbSMel Gorman cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; 2450c89511abSMel Gorman } 2451c8f7de0bSMichal Hocko 2452e332f741SMel Gorman if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) 2453c8f7de0bSMichal Hocko cc->whole_zone = true; 245406ed2998SVlastimil Babka } 2455c8f7de0bSMichal Hocko 2456566e54e1SMel Gorman last_migrated_pfn = 0; 2457748446bbSMel Gorman 24588854c55fSMel Gorman /* 24598854c55fSMel Gorman * Migrate has separate cached PFNs for ASYNC and SYNC* migration on 24608854c55fSMel Gorman * the basis that some migrations will fail in ASYNC mode. However, 24618854c55fSMel Gorman * if the cached PFNs match and pageblocks are skipped due to having 24628854c55fSMel Gorman * no isolation candidates, then the sync state does not matter. 24638854c55fSMel Gorman * Until a pageblock with isolation candidates is found, keep the 24648854c55fSMel Gorman * cached PFNs in sync to avoid revisiting the same blocks. 24658854c55fSMel Gorman */ 24668854c55fSMel Gorman update_cached = !sync && 24678854c55fSMel Gorman cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; 24688854c55fSMel Gorman 2469abd4349fSBaolin Wang trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync); 24700eb927c0SMel Gorman 2471361a2a22SMinchan Kim /* lru_add_drain_all could be expensive with involving other CPUs */ 2472361a2a22SMinchan Kim lru_add_drain(); 2473748446bbSMel Gorman 247440cacbcbSMel Gorman while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { 24759d502c1cSMinchan Kim int err; 247619d3cf9dSYanfei Xu unsigned long iteration_start_pfn = cc->migrate_pfn; 2477748446bbSMel Gorman 2478804d3121SMel Gorman /* 247948731c84SMel Gorman * Avoid multiple rescans of the same pageblock which can 248048731c84SMel Gorman * happen if a page cannot be isolated (dirty/writeback in 248148731c84SMel Gorman * async mode) or if the migrated pages are being allocated 248248731c84SMel Gorman * before the pageblock is cleared. The first rescan will 248348731c84SMel Gorman * capture the entire pageblock for migration. If it fails, 248448731c84SMel Gorman * it'll be marked skip and scanning will proceed as normal. 2485804d3121SMel Gorman */ 248648731c84SMel Gorman cc->finish_pageblock = false; 2487804d3121SMel Gorman if (pageblock_start_pfn(last_migrated_pfn) == 248819d3cf9dSYanfei Xu pageblock_start_pfn(iteration_start_pfn)) { 248948731c84SMel Gorman cc->finish_pageblock = true; 2490804d3121SMel Gorman } 2491804d3121SMel Gorman 2492cfccd2e6SMel Gorman rescan: 249332aaf055SPengfei Li switch (isolate_migratepages(cc)) { 2494f9e35b3bSMel Gorman case ISOLATE_ABORT: 24952d1e1041SVlastimil Babka ret = COMPACT_CONTENDED; 24965733c7d1SRafael Aquini putback_movable_pages(&cc->migratepages); 2497e64c5237SShaohua Li cc->nr_migratepages = 0; 2498f9e35b3bSMel Gorman goto out; 2499f9e35b3bSMel Gorman case ISOLATE_NONE: 25008854c55fSMel Gorman if (update_cached) { 25018854c55fSMel Gorman cc->zone->compact_cached_migrate_pfn[1] = 25028854c55fSMel Gorman cc->zone->compact_cached_migrate_pfn[0]; 25038854c55fSMel Gorman } 25048854c55fSMel Gorman 2505fdaf7f5cSVlastimil Babka /* 2506fdaf7f5cSVlastimil Babka * We haven't isolated and migrated anything, but 2507fdaf7f5cSVlastimil Babka * there might still be unflushed migrations from 2508fdaf7f5cSVlastimil Babka * previous cc->order aligned block. 2509fdaf7f5cSVlastimil Babka */ 2510fdaf7f5cSVlastimil Babka goto check_drain; 2511f9e35b3bSMel Gorman case ISOLATE_SUCCESS: 25128854c55fSMel Gorman update_cached = false; 251319d3cf9dSYanfei Xu last_migrated_pfn = iteration_start_pfn; 2514f9e35b3bSMel Gorman } 2515748446bbSMel Gorman 2516d53aea3dSDavid Rientjes err = migrate_pages(&cc->migratepages, compaction_alloc, 2517e0b9daebSDavid Rientjes compaction_free, (unsigned long)cc, cc->mode, 251884b328aaSBaolin Wang MR_COMPACTION, &nr_succeeded); 2519748446bbSMel Gorman 2520abd4349fSBaolin Wang trace_mm_compaction_migratepages(cc, nr_succeeded); 2521748446bbSMel Gorman 2522f8c9301fSVlastimil Babka /* All pages were either migrated or will be released */ 2523f8c9301fSVlastimil Babka cc->nr_migratepages = 0; 25249d502c1cSMinchan Kim if (err) { 25255733c7d1SRafael Aquini putback_movable_pages(&cc->migratepages); 25267ed695e0SVlastimil Babka /* 25277ed695e0SVlastimil Babka * migrate_pages() may return -ENOMEM when scanners meet 25287ed695e0SVlastimil Babka * and we want compact_finished() to detect it 25297ed695e0SVlastimil Babka */ 2530f2849aa0SVlastimil Babka if (err == -ENOMEM && !compact_scanners_met(cc)) { 25312d1e1041SVlastimil Babka ret = COMPACT_CONTENDED; 25324bf2bba3SDavid Rientjes goto out; 2533748446bbSMel Gorman } 2534fdd048e1SVlastimil Babka /* 2535cfccd2e6SMel Gorman * If an ASYNC or SYNC_LIGHT fails to migrate a page 25369ecc5fc5SMel Gorman * within the current order-aligned block and 25379ecc5fc5SMel Gorman * fast_find_migrateblock may be used then scan the 2538cfccd2e6SMel Gorman * remainder of the pageblock. This will mark the 2539cfccd2e6SMel Gorman * pageblock "skip" to avoid rescanning in the near 2540cfccd2e6SMel Gorman * future. This will isolate more pages than necessary 2541cfccd2e6SMel Gorman * for the request but avoid loops due to 2542cfccd2e6SMel Gorman * fast_find_migrateblock revisiting blocks that were 2543cfccd2e6SMel Gorman * recently partially scanned. 2544fdd048e1SVlastimil Babka */ 2545539aa041SMel Gorman if (!pageblock_aligned(cc->migrate_pfn) && 25469ecc5fc5SMel Gorman !cc->ignore_skip_hint && !cc->finish_pageblock && 2547cfccd2e6SMel Gorman (cc->mode < MIGRATE_SYNC)) { 2548cfccd2e6SMel Gorman cc->finish_pageblock = true; 2549cfccd2e6SMel Gorman 2550cfccd2e6SMel Gorman /* 2551cfccd2e6SMel Gorman * Draining pcplists does not help THP if 2552cfccd2e6SMel Gorman * any page failed to migrate. Even after 2553cfccd2e6SMel Gorman * drain, the pageblock will not be free. 2554cfccd2e6SMel Gorman */ 2555cfccd2e6SMel Gorman if (cc->order == COMPACTION_HPAGE_ORDER) 2556566e54e1SMel Gorman last_migrated_pfn = 0; 2557cfccd2e6SMel Gorman 2558cfccd2e6SMel Gorman goto rescan; 2559fdd048e1SVlastimil Babka } 25604bf2bba3SDavid Rientjes } 2561fdaf7f5cSVlastimil Babka 256216b3be40SMel Gorman /* Stop if a page has been captured */ 256316b3be40SMel Gorman if (capc && capc->page) { 256416b3be40SMel Gorman ret = COMPACT_SUCCESS; 256516b3be40SMel Gorman break; 256616b3be40SMel Gorman } 256716b3be40SMel Gorman 2568fdaf7f5cSVlastimil Babka check_drain: 2569fdaf7f5cSVlastimil Babka /* 2570fdaf7f5cSVlastimil Babka * Has the migration scanner moved away from the previous 2571fdaf7f5cSVlastimil Babka * cc->order aligned block where we migrated from? If yes, 2572fdaf7f5cSVlastimil Babka * flush the pages that were freed, so that they can merge and 2573fdaf7f5cSVlastimil Babka * compact_finished() can detect immediately if allocation 2574fdaf7f5cSVlastimil Babka * would succeed. 2575fdaf7f5cSVlastimil Babka */ 2576566e54e1SMel Gorman if (cc->order > 0 && last_migrated_pfn) { 2577fdaf7f5cSVlastimil Babka unsigned long current_block_start = 257806b6640aSVlastimil Babka block_start_pfn(cc->migrate_pfn, cc->order); 2579fdaf7f5cSVlastimil Babka 2580566e54e1SMel Gorman if (last_migrated_pfn < current_block_start) { 2581b01b2141SIngo Molnar lru_add_drain_cpu_zone(cc->zone); 2582fdaf7f5cSVlastimil Babka /* No more flushing until we migrate again */ 2583566e54e1SMel Gorman last_migrated_pfn = 0; 2584fdaf7f5cSVlastimil Babka } 2585fdaf7f5cSVlastimil Babka } 2586748446bbSMel Gorman } 2587748446bbSMel Gorman 2588f9e35b3bSMel Gorman out: 25896bace090SVlastimil Babka /* 25906bace090SVlastimil Babka * Release free pages and update where the free scanner should restart, 25916bace090SVlastimil Babka * so we don't leave any returned pages behind in the next attempt. 25926bace090SVlastimil Babka */ 25936bace090SVlastimil Babka if (cc->nr_freepages > 0) { 25946bace090SVlastimil Babka unsigned long free_pfn = release_freepages(&cc->freepages); 25956bace090SVlastimil Babka 25966bace090SVlastimil Babka cc->nr_freepages = 0; 25976bace090SVlastimil Babka VM_BUG_ON(free_pfn == 0); 25986bace090SVlastimil Babka /* The cached pfn is always the first in a pageblock */ 259906b6640aSVlastimil Babka free_pfn = pageblock_start_pfn(free_pfn); 26006bace090SVlastimil Babka /* 26016bace090SVlastimil Babka * Only go back, not forward. The cached pfn might have been 26026bace090SVlastimil Babka * already reset to zone end in compact_finished() 26036bace090SVlastimil Babka */ 260440cacbcbSMel Gorman if (free_pfn > cc->zone->compact_cached_free_pfn) 260540cacbcbSMel Gorman cc->zone->compact_cached_free_pfn = free_pfn; 26066bace090SVlastimil Babka } 2607748446bbSMel Gorman 26087f354a54SDavid Rientjes count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); 26097f354a54SDavid Rientjes count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); 26107f354a54SDavid Rientjes 2611abd4349fSBaolin Wang trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret); 26120eb927c0SMel Gorman 2613753ec50dSBaolin Wang VM_BUG_ON(!list_empty(&cc->freepages)); 2614753ec50dSBaolin Wang VM_BUG_ON(!list_empty(&cc->migratepages)); 2615753ec50dSBaolin Wang 2616748446bbSMel Gorman return ret; 2617748446bbSMel Gorman } 261876ab0f53SMel Gorman 2619ea7ab982SMichal Hocko static enum compact_result compact_zone_order(struct zone *zone, int order, 2620c3486f53SVlastimil Babka gfp_t gfp_mask, enum compact_priority prio, 262197a225e6SJoonsoo Kim unsigned int alloc_flags, int highest_zoneidx, 26225e1f0f09SMel Gorman struct page **capture) 262356de7263SMel Gorman { 2624ea7ab982SMichal Hocko enum compact_result ret; 262556de7263SMel Gorman struct compact_control cc = { 262656de7263SMel Gorman .order = order, 2627dbe2d4e4SMel Gorman .search_order = order, 26286d7ce559SDavid Rientjes .gfp_mask = gfp_mask, 262956de7263SMel Gorman .zone = zone, 2630a5508cd8SVlastimil Babka .mode = (prio == COMPACT_PRIO_ASYNC) ? 2631a5508cd8SVlastimil Babka MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, 2632ebff3980SVlastimil Babka .alloc_flags = alloc_flags, 263397a225e6SJoonsoo Kim .highest_zoneidx = highest_zoneidx, 2634accf6242SVlastimil Babka .direct_compaction = true, 2635a8e025e5SVlastimil Babka .whole_zone = (prio == MIN_COMPACT_PRIORITY), 26369f7e3387SVlastimil Babka .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), 26379f7e3387SVlastimil Babka .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) 263856de7263SMel Gorman }; 26395e1f0f09SMel Gorman struct capture_control capc = { 26405e1f0f09SMel Gorman .cc = &cc, 26415e1f0f09SMel Gorman .page = NULL, 26425e1f0f09SMel Gorman }; 26435e1f0f09SMel Gorman 2644b9e20f0dSVlastimil Babka /* 2645b9e20f0dSVlastimil Babka * Make sure the structs are really initialized before we expose the 2646b9e20f0dSVlastimil Babka * capture control, in case we are interrupted and the interrupt handler 2647b9e20f0dSVlastimil Babka * frees a page. 2648b9e20f0dSVlastimil Babka */ 2649b9e20f0dSVlastimil Babka barrier(); 2650b9e20f0dSVlastimil Babka WRITE_ONCE(current->capture_control, &capc); 265156de7263SMel Gorman 26525e1f0f09SMel Gorman ret = compact_zone(&cc, &capc); 2653e64c5237SShaohua Li 2654b9e20f0dSVlastimil Babka /* 2655b9e20f0dSVlastimil Babka * Make sure we hide capture control first before we read the captured 2656b9e20f0dSVlastimil Babka * page pointer, otherwise an interrupt could free and capture a page 2657b9e20f0dSVlastimil Babka * and we would leak it. 2658b9e20f0dSVlastimil Babka */ 2659b9e20f0dSVlastimil Babka WRITE_ONCE(current->capture_control, NULL); 2660b9e20f0dSVlastimil Babka *capture = READ_ONCE(capc.page); 266106dac2f4SCharan Teja Reddy /* 266206dac2f4SCharan Teja Reddy * Technically, it is also possible that compaction is skipped but 266306dac2f4SCharan Teja Reddy * the page is still captured out of luck(IRQ came and freed the page). 266406dac2f4SCharan Teja Reddy * Returning COMPACT_SUCCESS in such cases helps in properly accounting 266506dac2f4SCharan Teja Reddy * the COMPACT[STALL|FAIL] when compaction is skipped. 266606dac2f4SCharan Teja Reddy */ 266706dac2f4SCharan Teja Reddy if (*capture) 266806dac2f4SCharan Teja Reddy ret = COMPACT_SUCCESS; 26695e1f0f09SMel Gorman 2670e64c5237SShaohua Li return ret; 267156de7263SMel Gorman } 267256de7263SMel Gorman 267356de7263SMel Gorman /** 267456de7263SMel Gorman * try_to_compact_pages - Direct compact to satisfy a high-order allocation 267556de7263SMel Gorman * @gfp_mask: The GFP mask of the current allocation 26761a6d53a1SVlastimil Babka * @order: The order of the current allocation 26771a6d53a1SVlastimil Babka * @alloc_flags: The allocation flags of the current allocation 26781a6d53a1SVlastimil Babka * @ac: The context of current allocation 2679112d2d29SYang Shi * @prio: Determines how hard direct compaction should try to succeed 26806467552cSVlastimil Babka * @capture: Pointer to free page created by compaction will be stored here 268156de7263SMel Gorman * 268256de7263SMel Gorman * This is the main entry point for direct page compaction. 268356de7263SMel Gorman */ 2684ea7ab982SMichal Hocko enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, 2685c603844bSMel Gorman unsigned int alloc_flags, const struct alloc_context *ac, 26865e1f0f09SMel Gorman enum compact_priority prio, struct page **capture) 268756de7263SMel Gorman { 2688fe573327SVasily Averin int may_perform_io = (__force int)(gfp_mask & __GFP_IO); 268956de7263SMel Gorman struct zoneref *z; 269056de7263SMel Gorman struct zone *zone; 26911d4746d3SMichal Hocko enum compact_result rc = COMPACT_SKIPPED; 269256de7263SMel Gorman 269373e64c51SMichal Hocko /* 269473e64c51SMichal Hocko * Check if the GFP flags allow compaction - GFP_NOIO is really 269573e64c51SMichal Hocko * tricky context because the migration might require IO 269673e64c51SMichal Hocko */ 269773e64c51SMichal Hocko if (!may_perform_io) 269853853e2dSVlastimil Babka return COMPACT_SKIPPED; 269956de7263SMel Gorman 2700a5508cd8SVlastimil Babka trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); 2701837d026dSJoonsoo Kim 270256de7263SMel Gorman /* Compact each zone in the list */ 270397a225e6SJoonsoo Kim for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, 270497a225e6SJoonsoo Kim ac->highest_zoneidx, ac->nodemask) { 2705ea7ab982SMichal Hocko enum compact_result status; 270656de7263SMel Gorman 2707a8e025e5SVlastimil Babka if (prio > MIN_COMPACT_PRIORITY 2708a8e025e5SVlastimil Babka && compaction_deferred(zone, order)) { 27091d4746d3SMichal Hocko rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); 271053853e2dSVlastimil Babka continue; 27111d4746d3SMichal Hocko } 271253853e2dSVlastimil Babka 2713a5508cd8SVlastimil Babka status = compact_zone_order(zone, order, gfp_mask, prio, 271497a225e6SJoonsoo Kim alloc_flags, ac->highest_zoneidx, capture); 271556de7263SMel Gorman rc = max(status, rc); 271656de7263SMel Gorman 27177ceb009aSVlastimil Babka /* The allocation should succeed, stop compacting */ 27187ceb009aSVlastimil Babka if (status == COMPACT_SUCCESS) { 271953853e2dSVlastimil Babka /* 272053853e2dSVlastimil Babka * We think the allocation will succeed in this zone, 272153853e2dSVlastimil Babka * but it is not certain, hence the false. The caller 272253853e2dSVlastimil Babka * will repeat this with true if allocation indeed 272353853e2dSVlastimil Babka * succeeds in this zone. 272453853e2dSVlastimil Babka */ 272553853e2dSVlastimil Babka compaction_defer_reset(zone, order, false); 27261f9efdefSVlastimil Babka 2727c3486f53SVlastimil Babka break; 27281f9efdefSVlastimil Babka } 27291f9efdefSVlastimil Babka 2730a5508cd8SVlastimil Babka if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || 2731c3486f53SVlastimil Babka status == COMPACT_PARTIAL_SKIPPED)) 273253853e2dSVlastimil Babka /* 273353853e2dSVlastimil Babka * We think that allocation won't succeed in this zone 273453853e2dSVlastimil Babka * so we defer compaction there. If it ends up 273553853e2dSVlastimil Babka * succeeding after all, it will be reset. 273653853e2dSVlastimil Babka */ 273753853e2dSVlastimil Babka defer_compaction(zone, order); 27381f9efdefSVlastimil Babka 27391f9efdefSVlastimil Babka /* 27401f9efdefSVlastimil Babka * We might have stopped compacting due to need_resched() in 27411f9efdefSVlastimil Babka * async compaction, or due to a fatal signal detected. In that 2742c3486f53SVlastimil Babka * case do not try further zones 27431f9efdefSVlastimil Babka */ 2744c3486f53SVlastimil Babka if ((prio == COMPACT_PRIO_ASYNC && need_resched()) 2745c3486f53SVlastimil Babka || fatal_signal_pending(current)) 27461f9efdefSVlastimil Babka break; 27471f9efdefSVlastimil Babka } 27481f9efdefSVlastimil Babka 274956de7263SMel Gorman return rc; 275056de7263SMel Gorman } 275156de7263SMel Gorman 2752facdaa91SNitin Gupta /* 2753facdaa91SNitin Gupta * Compact all zones within a node till each zone's fragmentation score 2754facdaa91SNitin Gupta * reaches within proactive compaction thresholds (as determined by the 2755facdaa91SNitin Gupta * proactiveness tunable). 2756facdaa91SNitin Gupta * 2757facdaa91SNitin Gupta * It is possible that the function returns before reaching score targets 2758facdaa91SNitin Gupta * due to various back-off conditions, such as, contention on per-node or 2759facdaa91SNitin Gupta * per-zone locks. 2760facdaa91SNitin Gupta */ 2761facdaa91SNitin Gupta static void proactive_compact_node(pg_data_t *pgdat) 2762facdaa91SNitin Gupta { 2763facdaa91SNitin Gupta int zoneid; 2764facdaa91SNitin Gupta struct zone *zone; 2765facdaa91SNitin Gupta struct compact_control cc = { 2766facdaa91SNitin Gupta .order = -1, 2767facdaa91SNitin Gupta .mode = MIGRATE_SYNC_LIGHT, 2768facdaa91SNitin Gupta .ignore_skip_hint = true, 2769facdaa91SNitin Gupta .whole_zone = true, 2770facdaa91SNitin Gupta .gfp_mask = GFP_KERNEL, 2771facdaa91SNitin Gupta .proactive_compaction = true, 2772facdaa91SNitin Gupta }; 2773facdaa91SNitin Gupta 2774facdaa91SNitin Gupta for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 2775facdaa91SNitin Gupta zone = &pgdat->node_zones[zoneid]; 2776facdaa91SNitin Gupta if (!populated_zone(zone)) 2777facdaa91SNitin Gupta continue; 2778facdaa91SNitin Gupta 2779facdaa91SNitin Gupta cc.zone = zone; 2780facdaa91SNitin Gupta 2781facdaa91SNitin Gupta compact_zone(&cc, NULL); 2782facdaa91SNitin Gupta 27831bfb7684SBaolin Wang count_compact_events(KCOMPACTD_MIGRATE_SCANNED, 27841bfb7684SBaolin Wang cc.total_migrate_scanned); 27851bfb7684SBaolin Wang count_compact_events(KCOMPACTD_FREE_SCANNED, 27861bfb7684SBaolin Wang cc.total_free_scanned); 2787facdaa91SNitin Gupta } 2788facdaa91SNitin Gupta } 278956de7263SMel Gorman 279076ab0f53SMel Gorman /* Compact all zones within a node */ 27917103f16dSAndrew Morton static void compact_node(int nid) 27927be62de9SRik van Riel { 2793791cae96SVlastimil Babka pg_data_t *pgdat = NODE_DATA(nid); 2794791cae96SVlastimil Babka int zoneid; 2795791cae96SVlastimil Babka struct zone *zone; 27967be62de9SRik van Riel struct compact_control cc = { 27977be62de9SRik van Riel .order = -1, 2798e0b9daebSDavid Rientjes .mode = MIGRATE_SYNC, 279991ca9186SDavid Rientjes .ignore_skip_hint = true, 280006ed2998SVlastimil Babka .whole_zone = true, 280173e64c51SMichal Hocko .gfp_mask = GFP_KERNEL, 28027be62de9SRik van Riel }; 28037be62de9SRik van Riel 2804791cae96SVlastimil Babka 2805791cae96SVlastimil Babka for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 2806791cae96SVlastimil Babka 2807791cae96SVlastimil Babka zone = &pgdat->node_zones[zoneid]; 2808791cae96SVlastimil Babka if (!populated_zone(zone)) 2809791cae96SVlastimil Babka continue; 2810791cae96SVlastimil Babka 2811791cae96SVlastimil Babka cc.zone = zone; 2812791cae96SVlastimil Babka 28135e1f0f09SMel Gorman compact_zone(&cc, NULL); 2814791cae96SVlastimil Babka } 28157be62de9SRik van Riel } 28167be62de9SRik van Riel 281776ab0f53SMel Gorman /* Compact all nodes in the system */ 28187964c06dSJason Liu static void compact_nodes(void) 281976ab0f53SMel Gorman { 282076ab0f53SMel Gorman int nid; 282176ab0f53SMel Gorman 28228575ec29SHugh Dickins /* Flush pending updates to the LRU lists */ 28238575ec29SHugh Dickins lru_add_drain_all(); 28248575ec29SHugh Dickins 282576ab0f53SMel Gorman for_each_online_node(nid) 282676ab0f53SMel Gorman compact_node(nid); 282776ab0f53SMel Gorman } 282876ab0f53SMel Gorman 282948fe8ab8SMinghao Chi static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write, 283065d759c8SCharan Teja Reddy void *buffer, size_t *length, loff_t *ppos) 283165d759c8SCharan Teja Reddy { 283265d759c8SCharan Teja Reddy int rc, nid; 283365d759c8SCharan Teja Reddy 283465d759c8SCharan Teja Reddy rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 283565d759c8SCharan Teja Reddy if (rc) 283665d759c8SCharan Teja Reddy return rc; 283765d759c8SCharan Teja Reddy 283865d759c8SCharan Teja Reddy if (write && sysctl_compaction_proactiveness) { 283965d759c8SCharan Teja Reddy for_each_online_node(nid) { 284065d759c8SCharan Teja Reddy pg_data_t *pgdat = NODE_DATA(nid); 284165d759c8SCharan Teja Reddy 284265d759c8SCharan Teja Reddy if (pgdat->proactive_compact_trigger) 284365d759c8SCharan Teja Reddy continue; 284465d759c8SCharan Teja Reddy 284565d759c8SCharan Teja Reddy pgdat->proactive_compact_trigger = true; 28468fff8b6fSBaolin Wang trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1, 28478fff8b6fSBaolin Wang pgdat->nr_zones - 1); 284865d759c8SCharan Teja Reddy wake_up_interruptible(&pgdat->kcompactd_wait); 284965d759c8SCharan Teja Reddy } 285065d759c8SCharan Teja Reddy } 285165d759c8SCharan Teja Reddy 285265d759c8SCharan Teja Reddy return 0; 285365d759c8SCharan Teja Reddy } 285465d759c8SCharan Teja Reddy 2855facdaa91SNitin Gupta /* 2856fec4eb2cSYaowei Bai * This is the entry point for compacting all nodes via 2857fec4eb2cSYaowei Bai * /proc/sys/vm/compact_memory 2858fec4eb2cSYaowei Bai */ 285948fe8ab8SMinghao Chi static int sysctl_compaction_handler(struct ctl_table *table, int write, 286032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 286176ab0f53SMel Gorman { 28628b9167cdSWen Yang int ret; 28638b9167cdSWen Yang 28648b9167cdSWen Yang ret = proc_dointvec(table, write, buffer, length, ppos); 28658b9167cdSWen Yang if (ret) 28668b9167cdSWen Yang return ret; 28678b9167cdSWen Yang 28688b9167cdSWen Yang if (sysctl_compact_memory != 1) 28698b9167cdSWen Yang return -EINVAL; 28708b9167cdSWen Yang 287176ab0f53SMel Gorman if (write) 28727964c06dSJason Liu compact_nodes(); 287376ab0f53SMel Gorman 287476ab0f53SMel Gorman return 0; 287576ab0f53SMel Gorman } 2876ed4a6d7fSMel Gorman 2877ed4a6d7fSMel Gorman #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) 287817adb230SYueHaibing static ssize_t compact_store(struct device *dev, 287910fbcf4cSKay Sievers struct device_attribute *attr, 2880ed4a6d7fSMel Gorman const char *buf, size_t count) 2881ed4a6d7fSMel Gorman { 28828575ec29SHugh Dickins int nid = dev->id; 28838575ec29SHugh Dickins 28848575ec29SHugh Dickins if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { 28858575ec29SHugh Dickins /* Flush pending updates to the LRU lists */ 28868575ec29SHugh Dickins lru_add_drain_all(); 28878575ec29SHugh Dickins 28888575ec29SHugh Dickins compact_node(nid); 28898575ec29SHugh Dickins } 2890ed4a6d7fSMel Gorman 2891ed4a6d7fSMel Gorman return count; 2892ed4a6d7fSMel Gorman } 289317adb230SYueHaibing static DEVICE_ATTR_WO(compact); 2894ed4a6d7fSMel Gorman 2895ed4a6d7fSMel Gorman int compaction_register_node(struct node *node) 2896ed4a6d7fSMel Gorman { 289710fbcf4cSKay Sievers return device_create_file(&node->dev, &dev_attr_compact); 2898ed4a6d7fSMel Gorman } 2899ed4a6d7fSMel Gorman 2900ed4a6d7fSMel Gorman void compaction_unregister_node(struct node *node) 2901ed4a6d7fSMel Gorman { 290210fbcf4cSKay Sievers return device_remove_file(&node->dev, &dev_attr_compact); 2903ed4a6d7fSMel Gorman } 2904ed4a6d7fSMel Gorman #endif /* CONFIG_SYSFS && CONFIG_NUMA */ 2905ff9543fdSMichal Nazarewicz 2906698b1b30SVlastimil Babka static inline bool kcompactd_work_requested(pg_data_t *pgdat) 2907698b1b30SVlastimil Babka { 290865d759c8SCharan Teja Reddy return pgdat->kcompactd_max_order > 0 || kthread_should_stop() || 290965d759c8SCharan Teja Reddy pgdat->proactive_compact_trigger; 2910698b1b30SVlastimil Babka } 2911698b1b30SVlastimil Babka 2912698b1b30SVlastimil Babka static bool kcompactd_node_suitable(pg_data_t *pgdat) 2913698b1b30SVlastimil Babka { 2914698b1b30SVlastimil Babka int zoneid; 2915698b1b30SVlastimil Babka struct zone *zone; 291697a225e6SJoonsoo Kim enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx; 2917698b1b30SVlastimil Babka 291897a225e6SJoonsoo Kim for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) { 2919698b1b30SVlastimil Babka zone = &pgdat->node_zones[zoneid]; 2920698b1b30SVlastimil Babka 2921698b1b30SVlastimil Babka if (!populated_zone(zone)) 2922698b1b30SVlastimil Babka continue; 2923698b1b30SVlastimil Babka 2924e8606320SJohannes Weiner /* Allocation can already succeed, check other zones */ 2925e8606320SJohannes Weiner if (zone_watermark_ok(zone, pgdat->kcompactd_max_order, 2926e8606320SJohannes Weiner min_wmark_pages(zone), 2927e8606320SJohannes Weiner highest_zoneidx, 0)) 2928e8606320SJohannes Weiner continue; 2929e8606320SJohannes Weiner 2930e8606320SJohannes Weiner if (compaction_suitable(zone, pgdat->kcompactd_max_order, 29313cf04937SJohannes Weiner highest_zoneidx)) 2932698b1b30SVlastimil Babka return true; 2933698b1b30SVlastimil Babka } 2934698b1b30SVlastimil Babka 2935698b1b30SVlastimil Babka return false; 2936698b1b30SVlastimil Babka } 2937698b1b30SVlastimil Babka 2938698b1b30SVlastimil Babka static void kcompactd_do_work(pg_data_t *pgdat) 2939698b1b30SVlastimil Babka { 2940698b1b30SVlastimil Babka /* 2941698b1b30SVlastimil Babka * With no special task, compact all zones so that a page of requested 2942698b1b30SVlastimil Babka * order is allocatable. 2943698b1b30SVlastimil Babka */ 2944698b1b30SVlastimil Babka int zoneid; 2945698b1b30SVlastimil Babka struct zone *zone; 2946698b1b30SVlastimil Babka struct compact_control cc = { 2947698b1b30SVlastimil Babka .order = pgdat->kcompactd_max_order, 2948dbe2d4e4SMel Gorman .search_order = pgdat->kcompactd_max_order, 294997a225e6SJoonsoo Kim .highest_zoneidx = pgdat->kcompactd_highest_zoneidx, 2950698b1b30SVlastimil Babka .mode = MIGRATE_SYNC_LIGHT, 2951a0647dc9SDavid Rientjes .ignore_skip_hint = false, 295273e64c51SMichal Hocko .gfp_mask = GFP_KERNEL, 2953698b1b30SVlastimil Babka }; 2954698b1b30SVlastimil Babka trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, 295597a225e6SJoonsoo Kim cc.highest_zoneidx); 29567f354a54SDavid Rientjes count_compact_event(KCOMPACTD_WAKE); 2957698b1b30SVlastimil Babka 295897a225e6SJoonsoo Kim for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) { 2959698b1b30SVlastimil Babka int status; 2960698b1b30SVlastimil Babka 2961698b1b30SVlastimil Babka zone = &pgdat->node_zones[zoneid]; 2962698b1b30SVlastimil Babka if (!populated_zone(zone)) 2963698b1b30SVlastimil Babka continue; 2964698b1b30SVlastimil Babka 2965698b1b30SVlastimil Babka if (compaction_deferred(zone, cc.order)) 2966698b1b30SVlastimil Babka continue; 2967698b1b30SVlastimil Babka 2968e8606320SJohannes Weiner /* Allocation can already succeed, nothing to do */ 2969e8606320SJohannes Weiner if (zone_watermark_ok(zone, cc.order, 2970e8606320SJohannes Weiner min_wmark_pages(zone), zoneid, 0)) 2971698b1b30SVlastimil Babka continue; 2972698b1b30SVlastimil Babka 29733cf04937SJohannes Weiner if (!compaction_suitable(zone, cc.order, zoneid)) 2974e8606320SJohannes Weiner continue; 2975f98a497eSJohannes Weiner 2976172400c6SVlastimil Babka if (kthread_should_stop()) 2977172400c6SVlastimil Babka return; 2978a94b5252SYafang Shao 2979a94b5252SYafang Shao cc.zone = zone; 29805e1f0f09SMel Gorman status = compact_zone(&cc, NULL); 2981698b1b30SVlastimil Babka 29827ceb009aSVlastimil Babka if (status == COMPACT_SUCCESS) { 2983698b1b30SVlastimil Babka compaction_defer_reset(zone, cc.order, false); 2984c8f7de0bSMichal Hocko } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { 2985698b1b30SVlastimil Babka /* 2986bc3106b2SDavid Rientjes * Buddy pages may become stranded on pcps that could 2987bc3106b2SDavid Rientjes * otherwise coalesce on the zone's free area for 2988bc3106b2SDavid Rientjes * order >= cc.order. This is ratelimited by the 2989bc3106b2SDavid Rientjes * upcoming deferral. 2990bc3106b2SDavid Rientjes */ 2991bc3106b2SDavid Rientjes drain_all_pages(zone); 2992bc3106b2SDavid Rientjes 2993bc3106b2SDavid Rientjes /* 2994698b1b30SVlastimil Babka * We use sync migration mode here, so we defer like 2995698b1b30SVlastimil Babka * sync direct compaction does. 2996698b1b30SVlastimil Babka */ 2997698b1b30SVlastimil Babka defer_compaction(zone, cc.order); 2998698b1b30SVlastimil Babka } 2999698b1b30SVlastimil Babka 30007f354a54SDavid Rientjes count_compact_events(KCOMPACTD_MIGRATE_SCANNED, 30017f354a54SDavid Rientjes cc.total_migrate_scanned); 30027f354a54SDavid Rientjes count_compact_events(KCOMPACTD_FREE_SCANNED, 30037f354a54SDavid Rientjes cc.total_free_scanned); 3004698b1b30SVlastimil Babka } 3005698b1b30SVlastimil Babka 3006698b1b30SVlastimil Babka /* 3007698b1b30SVlastimil Babka * Regardless of success, we are done until woken up next. But remember 300897a225e6SJoonsoo Kim * the requested order/highest_zoneidx in case it was higher/tighter 300997a225e6SJoonsoo Kim * than our current ones 3010698b1b30SVlastimil Babka */ 3011698b1b30SVlastimil Babka if (pgdat->kcompactd_max_order <= cc.order) 3012698b1b30SVlastimil Babka pgdat->kcompactd_max_order = 0; 301397a225e6SJoonsoo Kim if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx) 301497a225e6SJoonsoo Kim pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; 3015698b1b30SVlastimil Babka } 3016698b1b30SVlastimil Babka 301797a225e6SJoonsoo Kim void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx) 3018698b1b30SVlastimil Babka { 3019698b1b30SVlastimil Babka if (!order) 3020698b1b30SVlastimil Babka return; 3021698b1b30SVlastimil Babka 3022698b1b30SVlastimil Babka if (pgdat->kcompactd_max_order < order) 3023698b1b30SVlastimil Babka pgdat->kcompactd_max_order = order; 3024698b1b30SVlastimil Babka 302597a225e6SJoonsoo Kim if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx) 302697a225e6SJoonsoo Kim pgdat->kcompactd_highest_zoneidx = highest_zoneidx; 3027698b1b30SVlastimil Babka 30286818600fSDavidlohr Bueso /* 30296818600fSDavidlohr Bueso * Pairs with implicit barrier in wait_event_freezable() 30306818600fSDavidlohr Bueso * such that wakeups are not missed. 30316818600fSDavidlohr Bueso */ 30326818600fSDavidlohr Bueso if (!wq_has_sleeper(&pgdat->kcompactd_wait)) 3033698b1b30SVlastimil Babka return; 3034698b1b30SVlastimil Babka 3035698b1b30SVlastimil Babka if (!kcompactd_node_suitable(pgdat)) 3036698b1b30SVlastimil Babka return; 3037698b1b30SVlastimil Babka 3038698b1b30SVlastimil Babka trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, 303997a225e6SJoonsoo Kim highest_zoneidx); 3040698b1b30SVlastimil Babka wake_up_interruptible(&pgdat->kcompactd_wait); 3041698b1b30SVlastimil Babka } 3042698b1b30SVlastimil Babka 3043698b1b30SVlastimil Babka /* 3044698b1b30SVlastimil Babka * The background compaction daemon, started as a kernel thread 3045698b1b30SVlastimil Babka * from the init process. 3046698b1b30SVlastimil Babka */ 3047698b1b30SVlastimil Babka static int kcompactd(void *p) 3048698b1b30SVlastimil Babka { 3049698b1b30SVlastimil Babka pg_data_t *pgdat = (pg_data_t *)p; 3050698b1b30SVlastimil Babka struct task_struct *tsk = current; 3051e1e92bfaSCharan Teja Reddy long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC); 3052e1e92bfaSCharan Teja Reddy long timeout = default_timeout; 3053698b1b30SVlastimil Babka 3054698b1b30SVlastimil Babka const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 3055698b1b30SVlastimil Babka 3056698b1b30SVlastimil Babka if (!cpumask_empty(cpumask)) 3057698b1b30SVlastimil Babka set_cpus_allowed_ptr(tsk, cpumask); 3058698b1b30SVlastimil Babka 3059698b1b30SVlastimil Babka set_freezable(); 3060698b1b30SVlastimil Babka 3061698b1b30SVlastimil Babka pgdat->kcompactd_max_order = 0; 306297a225e6SJoonsoo Kim pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; 3063698b1b30SVlastimil Babka 3064698b1b30SVlastimil Babka while (!kthread_should_stop()) { 3065eb414681SJohannes Weiner unsigned long pflags; 3066eb414681SJohannes Weiner 306765d759c8SCharan Teja Reddy /* 306865d759c8SCharan Teja Reddy * Avoid the unnecessary wakeup for proactive compaction 306965d759c8SCharan Teja Reddy * when it is disabled. 307065d759c8SCharan Teja Reddy */ 307165d759c8SCharan Teja Reddy if (!sysctl_compaction_proactiveness) 307265d759c8SCharan Teja Reddy timeout = MAX_SCHEDULE_TIMEOUT; 3073698b1b30SVlastimil Babka trace_mm_compaction_kcompactd_sleep(pgdat->node_id); 3074facdaa91SNitin Gupta if (wait_event_freezable_timeout(pgdat->kcompactd_wait, 307565d759c8SCharan Teja Reddy kcompactd_work_requested(pgdat), timeout) && 307665d759c8SCharan Teja Reddy !pgdat->proactive_compact_trigger) { 3077698b1b30SVlastimil Babka 3078eb414681SJohannes Weiner psi_memstall_enter(&pflags); 3079698b1b30SVlastimil Babka kcompactd_do_work(pgdat); 3080eb414681SJohannes Weiner psi_memstall_leave(&pflags); 3081e1e92bfaSCharan Teja Reddy /* 3082e1e92bfaSCharan Teja Reddy * Reset the timeout value. The defer timeout from 3083e1e92bfaSCharan Teja Reddy * proactive compaction is lost here but that is fine 3084e1e92bfaSCharan Teja Reddy * as the condition of the zone changing substantionally 3085e1e92bfaSCharan Teja Reddy * then carrying on with the previous defer interval is 3086e1e92bfaSCharan Teja Reddy * not useful. 3087e1e92bfaSCharan Teja Reddy */ 3088e1e92bfaSCharan Teja Reddy timeout = default_timeout; 3089facdaa91SNitin Gupta continue; 3090facdaa91SNitin Gupta } 3091facdaa91SNitin Gupta 3092e1e92bfaSCharan Teja Reddy /* 3093e1e92bfaSCharan Teja Reddy * Start the proactive work with default timeout. Based 3094e1e92bfaSCharan Teja Reddy * on the fragmentation score, this timeout is updated. 3095e1e92bfaSCharan Teja Reddy */ 3096e1e92bfaSCharan Teja Reddy timeout = default_timeout; 3097facdaa91SNitin Gupta if (should_proactive_compact_node(pgdat)) { 3098facdaa91SNitin Gupta unsigned int prev_score, score; 3099facdaa91SNitin Gupta 3100facdaa91SNitin Gupta prev_score = fragmentation_score_node(pgdat); 3101facdaa91SNitin Gupta proactive_compact_node(pgdat); 3102facdaa91SNitin Gupta score = fragmentation_score_node(pgdat); 3103facdaa91SNitin Gupta /* 3104facdaa91SNitin Gupta * Defer proactive compaction if the fragmentation 3105facdaa91SNitin Gupta * score did not go down i.e. no progress made. 3106facdaa91SNitin Gupta */ 3107e1e92bfaSCharan Teja Reddy if (unlikely(score >= prev_score)) 3108e1e92bfaSCharan Teja Reddy timeout = 3109e1e92bfaSCharan Teja Reddy default_timeout << COMPACT_MAX_DEFER_SHIFT; 3110facdaa91SNitin Gupta } 311165d759c8SCharan Teja Reddy if (unlikely(pgdat->proactive_compact_trigger)) 311265d759c8SCharan Teja Reddy pgdat->proactive_compact_trigger = false; 3113698b1b30SVlastimil Babka } 3114698b1b30SVlastimil Babka 3115698b1b30SVlastimil Babka return 0; 3116698b1b30SVlastimil Babka } 3117698b1b30SVlastimil Babka 3118698b1b30SVlastimil Babka /* 3119698b1b30SVlastimil Babka * This kcompactd start function will be called by init and node-hot-add. 3120698b1b30SVlastimil Babka * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. 3121698b1b30SVlastimil Babka */ 3122833dfc00SMiaohe Lin void __meminit kcompactd_run(int nid) 3123698b1b30SVlastimil Babka { 3124698b1b30SVlastimil Babka pg_data_t *pgdat = NODE_DATA(nid); 3125698b1b30SVlastimil Babka 3126698b1b30SVlastimil Babka if (pgdat->kcompactd) 3127024c61eaSMiaohe Lin return; 3128698b1b30SVlastimil Babka 3129698b1b30SVlastimil Babka pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); 3130698b1b30SVlastimil Babka if (IS_ERR(pgdat->kcompactd)) { 3131698b1b30SVlastimil Babka pr_err("Failed to start kcompactd on node %d\n", nid); 3132698b1b30SVlastimil Babka pgdat->kcompactd = NULL; 3133698b1b30SVlastimil Babka } 3134698b1b30SVlastimil Babka } 3135698b1b30SVlastimil Babka 3136698b1b30SVlastimil Babka /* 3137698b1b30SVlastimil Babka * Called by memory hotplug when all memory in a node is offlined. Caller must 3138e8da368aSYun-Ze Li * be holding mem_hotplug_begin/done(). 3139698b1b30SVlastimil Babka */ 3140833dfc00SMiaohe Lin void __meminit kcompactd_stop(int nid) 3141698b1b30SVlastimil Babka { 3142698b1b30SVlastimil Babka struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; 3143698b1b30SVlastimil Babka 3144698b1b30SVlastimil Babka if (kcompactd) { 3145698b1b30SVlastimil Babka kthread_stop(kcompactd); 3146698b1b30SVlastimil Babka NODE_DATA(nid)->kcompactd = NULL; 3147698b1b30SVlastimil Babka } 3148698b1b30SVlastimil Babka } 3149698b1b30SVlastimil Babka 3150698b1b30SVlastimil Babka /* 3151698b1b30SVlastimil Babka * It's optimal to keep kcompactd on the same CPUs as their memory, but 3152698b1b30SVlastimil Babka * not required for correctness. So if the last cpu in a node goes 3153698b1b30SVlastimil Babka * away, we get changed to run anywhere: as the first one comes back, 3154698b1b30SVlastimil Babka * restore their cpu bindings. 3155698b1b30SVlastimil Babka */ 3156e46b1db2SAnna-Maria Gleixner static int kcompactd_cpu_online(unsigned int cpu) 3157698b1b30SVlastimil Babka { 3158698b1b30SVlastimil Babka int nid; 3159698b1b30SVlastimil Babka 3160698b1b30SVlastimil Babka for_each_node_state(nid, N_MEMORY) { 3161698b1b30SVlastimil Babka pg_data_t *pgdat = NODE_DATA(nid); 3162698b1b30SVlastimil Babka const struct cpumask *mask; 3163698b1b30SVlastimil Babka 3164698b1b30SVlastimil Babka mask = cpumask_of_node(pgdat->node_id); 3165698b1b30SVlastimil Babka 3166698b1b30SVlastimil Babka if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 3167698b1b30SVlastimil Babka /* One of our CPUs online: restore mask */ 31683109de30SMiaohe Lin if (pgdat->kcompactd) 3169698b1b30SVlastimil Babka set_cpus_allowed_ptr(pgdat->kcompactd, mask); 3170698b1b30SVlastimil Babka } 3171e46b1db2SAnna-Maria Gleixner return 0; 3172698b1b30SVlastimil Babka } 3173698b1b30SVlastimil Babka 317448fe8ab8SMinghao Chi static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table, 317548fe8ab8SMinghao Chi int write, void *buffer, size_t *lenp, loff_t *ppos) 317648fe8ab8SMinghao Chi { 317748fe8ab8SMinghao Chi int ret, old; 317848fe8ab8SMinghao Chi 317948fe8ab8SMinghao Chi if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write) 318048fe8ab8SMinghao Chi return proc_dointvec_minmax(table, write, buffer, lenp, ppos); 318148fe8ab8SMinghao Chi 318248fe8ab8SMinghao Chi old = *(int *)table->data; 318348fe8ab8SMinghao Chi ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 318448fe8ab8SMinghao Chi if (ret) 318548fe8ab8SMinghao Chi return ret; 318648fe8ab8SMinghao Chi if (old != *(int *)table->data) 318748fe8ab8SMinghao Chi pr_warn_once("sysctl attribute %s changed by %s[%d]\n", 318848fe8ab8SMinghao Chi table->procname, current->comm, 318948fe8ab8SMinghao Chi task_pid_nr(current)); 319048fe8ab8SMinghao Chi return ret; 319148fe8ab8SMinghao Chi } 319248fe8ab8SMinghao Chi 319348fe8ab8SMinghao Chi static struct ctl_table vm_compaction[] = { 319448fe8ab8SMinghao Chi { 319548fe8ab8SMinghao Chi .procname = "compact_memory", 31968b9167cdSWen Yang .data = &sysctl_compact_memory, 319748fe8ab8SMinghao Chi .maxlen = sizeof(int), 319848fe8ab8SMinghao Chi .mode = 0200, 319948fe8ab8SMinghao Chi .proc_handler = sysctl_compaction_handler, 320048fe8ab8SMinghao Chi }, 320148fe8ab8SMinghao Chi { 320248fe8ab8SMinghao Chi .procname = "compaction_proactiveness", 320348fe8ab8SMinghao Chi .data = &sysctl_compaction_proactiveness, 320448fe8ab8SMinghao Chi .maxlen = sizeof(sysctl_compaction_proactiveness), 320548fe8ab8SMinghao Chi .mode = 0644, 320648fe8ab8SMinghao Chi .proc_handler = compaction_proactiveness_sysctl_handler, 320748fe8ab8SMinghao Chi .extra1 = SYSCTL_ZERO, 320848fe8ab8SMinghao Chi .extra2 = SYSCTL_ONE_HUNDRED, 320948fe8ab8SMinghao Chi }, 321048fe8ab8SMinghao Chi { 321148fe8ab8SMinghao Chi .procname = "extfrag_threshold", 321248fe8ab8SMinghao Chi .data = &sysctl_extfrag_threshold, 321348fe8ab8SMinghao Chi .maxlen = sizeof(int), 321448fe8ab8SMinghao Chi .mode = 0644, 321548fe8ab8SMinghao Chi .proc_handler = proc_dointvec_minmax, 321648fe8ab8SMinghao Chi .extra1 = SYSCTL_ZERO, 321748fe8ab8SMinghao Chi .extra2 = SYSCTL_ONE_THOUSAND, 321848fe8ab8SMinghao Chi }, 321948fe8ab8SMinghao Chi { 322048fe8ab8SMinghao Chi .procname = "compact_unevictable_allowed", 322148fe8ab8SMinghao Chi .data = &sysctl_compact_unevictable_allowed, 322248fe8ab8SMinghao Chi .maxlen = sizeof(int), 322348fe8ab8SMinghao Chi .mode = 0644, 322448fe8ab8SMinghao Chi .proc_handler = proc_dointvec_minmax_warn_RT_change, 322548fe8ab8SMinghao Chi .extra1 = SYSCTL_ZERO, 322648fe8ab8SMinghao Chi .extra2 = SYSCTL_ONE, 322748fe8ab8SMinghao Chi }, 322848fe8ab8SMinghao Chi { } 322948fe8ab8SMinghao Chi }; 323048fe8ab8SMinghao Chi 3231698b1b30SVlastimil Babka static int __init kcompactd_init(void) 3232698b1b30SVlastimil Babka { 3233698b1b30SVlastimil Babka int nid; 3234e46b1db2SAnna-Maria Gleixner int ret; 3235e46b1db2SAnna-Maria Gleixner 3236e46b1db2SAnna-Maria Gleixner ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 3237e46b1db2SAnna-Maria Gleixner "mm/compaction:online", 3238e46b1db2SAnna-Maria Gleixner kcompactd_cpu_online, NULL); 3239e46b1db2SAnna-Maria Gleixner if (ret < 0) { 3240e46b1db2SAnna-Maria Gleixner pr_err("kcompactd: failed to register hotplug callbacks.\n"); 3241e46b1db2SAnna-Maria Gleixner return ret; 3242e46b1db2SAnna-Maria Gleixner } 3243698b1b30SVlastimil Babka 3244698b1b30SVlastimil Babka for_each_node_state(nid, N_MEMORY) 3245698b1b30SVlastimil Babka kcompactd_run(nid); 324648fe8ab8SMinghao Chi register_sysctl_init("vm", vm_compaction); 3247698b1b30SVlastimil Babka return 0; 3248698b1b30SVlastimil Babka } 3249698b1b30SVlastimil Babka subsys_initcall(kcompactd_init) 3250698b1b30SVlastimil Babka 3251ff9543fdSMichal Nazarewicz #endif /* CONFIG_COMPACTION */ 3252