171e3aac0SAndrea Arcangeli /* 271e3aac0SAndrea Arcangeli * Copyright (C) 2009 Red Hat, Inc. 371e3aac0SAndrea Arcangeli * 471e3aac0SAndrea Arcangeli * This work is licensed under the terms of the GNU GPL, version 2. See 571e3aac0SAndrea Arcangeli * the COPYING file in the top-level directory. 671e3aac0SAndrea Arcangeli */ 771e3aac0SAndrea Arcangeli 8ae3a8c1cSAndrew Morton #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 9ae3a8c1cSAndrew Morton 1071e3aac0SAndrea Arcangeli #include <linux/mm.h> 1171e3aac0SAndrea Arcangeli #include <linux/sched.h> 1271e3aac0SAndrea Arcangeli #include <linux/highmem.h> 1371e3aac0SAndrea Arcangeli #include <linux/hugetlb.h> 1471e3aac0SAndrea Arcangeli #include <linux/mmu_notifier.h> 1571e3aac0SAndrea Arcangeli #include <linux/rmap.h> 1671e3aac0SAndrea Arcangeli #include <linux/swap.h> 1797ae1749SKirill A. Shutemov #include <linux/shrinker.h> 18ba76149fSAndrea Arcangeli #include <linux/mm_inline.h> 19ba76149fSAndrea Arcangeli #include <linux/kthread.h> 20ba76149fSAndrea Arcangeli #include <linux/khugepaged.h> 21878aee7dSAndrea Arcangeli #include <linux/freezer.h> 22a664b2d8SAndrea Arcangeli #include <linux/mman.h> 23325adeb5SRalf Baechle #include <linux/pagemap.h> 244daae3b4SMel Gorman #include <linux/migrate.h> 2543b5fbbdSSasha Levin #include <linux/hashtable.h> 2697ae1749SKirill A. Shutemov 2771e3aac0SAndrea Arcangeli #include <asm/tlb.h> 2871e3aac0SAndrea Arcangeli #include <asm/pgalloc.h> 2971e3aac0SAndrea Arcangeli #include "internal.h" 3071e3aac0SAndrea Arcangeli 31ba76149fSAndrea Arcangeli /* 328bfa3f9aSJianguo Wu * By default transparent hugepage support is disabled in order that avoid 338bfa3f9aSJianguo Wu * to risk increase the memory footprint of applications without a guaranteed 348bfa3f9aSJianguo Wu * benefit. When transparent hugepage support is enabled, is for all mappings, 358bfa3f9aSJianguo Wu * and khugepaged scans all mappings. 368bfa3f9aSJianguo Wu * Defrag is invoked by khugepaged hugepage allocations and by page faults 378bfa3f9aSJianguo Wu * for all hugepage allocations. 38ba76149fSAndrea Arcangeli */ 3971e3aac0SAndrea Arcangeli unsigned long transparent_hugepage_flags __read_mostly = 4013ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 41ba76149fSAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_FLAG)| 4213ece886SAndrea Arcangeli #endif 4313ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 4413ece886SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 4513ece886SAndrea Arcangeli #endif 46d39d33c3SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| 4779da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 4879da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 49ba76149fSAndrea Arcangeli 50ba76149fSAndrea Arcangeli /* default scan 8*512 pte (or vmas) every 30 second */ 51ba76149fSAndrea Arcangeli static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; 52ba76149fSAndrea Arcangeli static unsigned int khugepaged_pages_collapsed; 53ba76149fSAndrea Arcangeli static unsigned int khugepaged_full_scans; 54ba76149fSAndrea Arcangeli static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 55ba76149fSAndrea Arcangeli /* during fragmentation poll the hugepage allocator once every minute */ 56ba76149fSAndrea Arcangeli static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 57ba76149fSAndrea Arcangeli static struct task_struct *khugepaged_thread __read_mostly; 58ba76149fSAndrea Arcangeli static DEFINE_MUTEX(khugepaged_mutex); 59ba76149fSAndrea Arcangeli static DEFINE_SPINLOCK(khugepaged_mm_lock); 60ba76149fSAndrea Arcangeli static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 61ba76149fSAndrea Arcangeli /* 62ba76149fSAndrea Arcangeli * default collapse hugepages if there is at least one pte mapped like 63ba76149fSAndrea Arcangeli * it would have happened if the vma was large enough during page 64ba76149fSAndrea Arcangeli * fault. 65ba76149fSAndrea Arcangeli */ 66ba76149fSAndrea Arcangeli static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; 67ba76149fSAndrea Arcangeli 68ba76149fSAndrea Arcangeli static int khugepaged(void *none); 69ba76149fSAndrea Arcangeli static int khugepaged_slab_init(void); 70ba76149fSAndrea Arcangeli 7143b5fbbdSSasha Levin #define MM_SLOTS_HASH_BITS 10 7243b5fbbdSSasha Levin static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 7343b5fbbdSSasha Levin 74ba76149fSAndrea Arcangeli static struct kmem_cache *mm_slot_cache __read_mostly; 75ba76149fSAndrea Arcangeli 76ba76149fSAndrea Arcangeli /** 77ba76149fSAndrea Arcangeli * struct mm_slot - hash lookup from mm to mm_slot 78ba76149fSAndrea Arcangeli * @hash: hash collision list 79ba76149fSAndrea Arcangeli * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 80ba76149fSAndrea Arcangeli * @mm: the mm that this information is valid for 81ba76149fSAndrea Arcangeli */ 82ba76149fSAndrea Arcangeli struct mm_slot { 83ba76149fSAndrea Arcangeli struct hlist_node hash; 84ba76149fSAndrea Arcangeli struct list_head mm_node; 85ba76149fSAndrea Arcangeli struct mm_struct *mm; 86ba76149fSAndrea Arcangeli }; 87ba76149fSAndrea Arcangeli 88ba76149fSAndrea Arcangeli /** 89ba76149fSAndrea Arcangeli * struct khugepaged_scan - cursor for scanning 90ba76149fSAndrea Arcangeli * @mm_head: the head of the mm list to scan 91ba76149fSAndrea Arcangeli * @mm_slot: the current mm_slot we are scanning 92ba76149fSAndrea Arcangeli * @address: the next address inside that to be scanned 93ba76149fSAndrea Arcangeli * 94ba76149fSAndrea Arcangeli * There is only the one khugepaged_scan instance of this cursor structure. 95ba76149fSAndrea Arcangeli */ 96ba76149fSAndrea Arcangeli struct khugepaged_scan { 97ba76149fSAndrea Arcangeli struct list_head mm_head; 98ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 99ba76149fSAndrea Arcangeli unsigned long address; 1002f1da642SH Hartley Sweeten }; 1012f1da642SH Hartley Sweeten static struct khugepaged_scan khugepaged_scan = { 102ba76149fSAndrea Arcangeli .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 103ba76149fSAndrea Arcangeli }; 104ba76149fSAndrea Arcangeli 105f000565aSAndrea Arcangeli 106f000565aSAndrea Arcangeli static int set_recommended_min_free_kbytes(void) 107f000565aSAndrea Arcangeli { 108f000565aSAndrea Arcangeli struct zone *zone; 109f000565aSAndrea Arcangeli int nr_zones = 0; 110f000565aSAndrea Arcangeli unsigned long recommended_min; 111f000565aSAndrea Arcangeli 11217c230afSXiao Guangrong if (!khugepaged_enabled()) 113f000565aSAndrea Arcangeli return 0; 114f000565aSAndrea Arcangeli 115f000565aSAndrea Arcangeli for_each_populated_zone(zone) 116f000565aSAndrea Arcangeli nr_zones++; 117f000565aSAndrea Arcangeli 118f000565aSAndrea Arcangeli /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */ 119f000565aSAndrea Arcangeli recommended_min = pageblock_nr_pages * nr_zones * 2; 120f000565aSAndrea Arcangeli 121f000565aSAndrea Arcangeli /* 122f000565aSAndrea Arcangeli * Make sure that on average at least two pageblocks are almost free 123f000565aSAndrea Arcangeli * of another type, one for a migratetype to fall back to and a 124f000565aSAndrea Arcangeli * second to avoid subsequent fallbacks of other types There are 3 125f000565aSAndrea Arcangeli * MIGRATE_TYPES we care about. 126f000565aSAndrea Arcangeli */ 127f000565aSAndrea Arcangeli recommended_min += pageblock_nr_pages * nr_zones * 128f000565aSAndrea Arcangeli MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 129f000565aSAndrea Arcangeli 130f000565aSAndrea Arcangeli /* don't ever allow to reserve more than 5% of the lowmem */ 131f000565aSAndrea Arcangeli recommended_min = min(recommended_min, 132f000565aSAndrea Arcangeli (unsigned long) nr_free_buffer_pages() / 20); 133f000565aSAndrea Arcangeli recommended_min <<= (PAGE_SHIFT-10); 134f000565aSAndrea Arcangeli 13542aa83cbSHan Pingtian if (recommended_min > min_free_kbytes) { 13642aa83cbSHan Pingtian if (user_min_free_kbytes >= 0) 13742aa83cbSHan Pingtian pr_info("raising min_free_kbytes from %d to %lu " 13842aa83cbSHan Pingtian "to help transparent hugepage allocations\n", 13942aa83cbSHan Pingtian min_free_kbytes, recommended_min); 14042aa83cbSHan Pingtian 141f000565aSAndrea Arcangeli min_free_kbytes = recommended_min; 14242aa83cbSHan Pingtian } 143f000565aSAndrea Arcangeli setup_per_zone_wmarks(); 144f000565aSAndrea Arcangeli return 0; 145f000565aSAndrea Arcangeli } 146f000565aSAndrea Arcangeli late_initcall(set_recommended_min_free_kbytes); 147f000565aSAndrea Arcangeli 148ba76149fSAndrea Arcangeli static int start_khugepaged(void) 149ba76149fSAndrea Arcangeli { 150ba76149fSAndrea Arcangeli int err = 0; 151ba76149fSAndrea Arcangeli if (khugepaged_enabled()) { 152ba76149fSAndrea Arcangeli if (!khugepaged_thread) 153ba76149fSAndrea Arcangeli khugepaged_thread = kthread_run(khugepaged, NULL, 154ba76149fSAndrea Arcangeli "khugepaged"); 155ba76149fSAndrea Arcangeli if (unlikely(IS_ERR(khugepaged_thread))) { 156ae3a8c1cSAndrew Morton pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 157ba76149fSAndrea Arcangeli err = PTR_ERR(khugepaged_thread); 158ba76149fSAndrea Arcangeli khugepaged_thread = NULL; 159ba76149fSAndrea Arcangeli } 160911891afSXiao Guangrong 161911891afSXiao Guangrong if (!list_empty(&khugepaged_scan.mm_head)) 162ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 163f000565aSAndrea Arcangeli 164f000565aSAndrea Arcangeli set_recommended_min_free_kbytes(); 165911891afSXiao Guangrong } else if (khugepaged_thread) { 166911891afSXiao Guangrong kthread_stop(khugepaged_thread); 167911891afSXiao Guangrong khugepaged_thread = NULL; 168911891afSXiao Guangrong } 169637e3a27SXiao Guangrong 170ba76149fSAndrea Arcangeli return err; 171ba76149fSAndrea Arcangeli } 17271e3aac0SAndrea Arcangeli 17397ae1749SKirill A. Shutemov static atomic_t huge_zero_refcount; 17456873f43SWang, Yalin struct page *huge_zero_page __read_mostly; 1754a6c1297SKirill A. Shutemov 1764a6c1297SKirill A. Shutemov static inline bool is_huge_zero_pmd(pmd_t pmd) 1774a6c1297SKirill A. Shutemov { 1785918d10aSKirill A. Shutemov return is_huge_zero_page(pmd_page(pmd)); 1794a6c1297SKirill A. Shutemov } 1804a6c1297SKirill A. Shutemov 1815918d10aSKirill A. Shutemov static struct page *get_huge_zero_page(void) 18297ae1749SKirill A. Shutemov { 18397ae1749SKirill A. Shutemov struct page *zero_page; 18497ae1749SKirill A. Shutemov retry: 18597ae1749SKirill A. Shutemov if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 1865918d10aSKirill A. Shutemov return ACCESS_ONCE(huge_zero_page); 18797ae1749SKirill A. Shutemov 18897ae1749SKirill A. Shutemov zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 18997ae1749SKirill A. Shutemov HPAGE_PMD_ORDER); 190d8a8e1f0SKirill A. Shutemov if (!zero_page) { 191d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 1925918d10aSKirill A. Shutemov return NULL; 193d8a8e1f0SKirill A. Shutemov } 194d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC); 19597ae1749SKirill A. Shutemov preempt_disable(); 1965918d10aSKirill A. Shutemov if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 19797ae1749SKirill A. Shutemov preempt_enable(); 1985ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 19997ae1749SKirill A. Shutemov goto retry; 20097ae1749SKirill A. Shutemov } 20197ae1749SKirill A. Shutemov 20297ae1749SKirill A. Shutemov /* We take additional reference here. It will be put back by shrinker */ 20397ae1749SKirill A. Shutemov atomic_set(&huge_zero_refcount, 2); 20497ae1749SKirill A. Shutemov preempt_enable(); 2055918d10aSKirill A. Shutemov return ACCESS_ONCE(huge_zero_page); 20697ae1749SKirill A. Shutemov } 20797ae1749SKirill A. Shutemov 20897ae1749SKirill A. Shutemov static void put_huge_zero_page(void) 20997ae1749SKirill A. Shutemov { 21097ae1749SKirill A. Shutemov /* 21197ae1749SKirill A. Shutemov * Counter should never go to zero here. Only shrinker can put 21297ae1749SKirill A. Shutemov * last reference. 21397ae1749SKirill A. Shutemov */ 21497ae1749SKirill A. Shutemov BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 21597ae1749SKirill A. Shutemov } 21697ae1749SKirill A. Shutemov 21748896466SGlauber Costa static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 21897ae1749SKirill A. Shutemov struct shrink_control *sc) 21997ae1749SKirill A. Shutemov { 22097ae1749SKirill A. Shutemov /* we can free zero page only if last reference remains */ 22197ae1749SKirill A. Shutemov return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 22248896466SGlauber Costa } 22397ae1749SKirill A. Shutemov 22448896466SGlauber Costa static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 22548896466SGlauber Costa struct shrink_control *sc) 22648896466SGlauber Costa { 22797ae1749SKirill A. Shutemov if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 2285918d10aSKirill A. Shutemov struct page *zero_page = xchg(&huge_zero_page, NULL); 2295918d10aSKirill A. Shutemov BUG_ON(zero_page == NULL); 2305ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 23148896466SGlauber Costa return HPAGE_PMD_NR; 23297ae1749SKirill A. Shutemov } 23397ae1749SKirill A. Shutemov 23497ae1749SKirill A. Shutemov return 0; 23597ae1749SKirill A. Shutemov } 23697ae1749SKirill A. Shutemov 23797ae1749SKirill A. Shutemov static struct shrinker huge_zero_page_shrinker = { 23848896466SGlauber Costa .count_objects = shrink_huge_zero_page_count, 23948896466SGlauber Costa .scan_objects = shrink_huge_zero_page_scan, 24097ae1749SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 24197ae1749SKirill A. Shutemov }; 24297ae1749SKirill A. Shutemov 24371e3aac0SAndrea Arcangeli #ifdef CONFIG_SYSFS 244ba76149fSAndrea Arcangeli 24571e3aac0SAndrea Arcangeli static ssize_t double_flag_show(struct kobject *kobj, 24671e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 24771e3aac0SAndrea Arcangeli enum transparent_hugepage_flag enabled, 24871e3aac0SAndrea Arcangeli enum transparent_hugepage_flag req_madv) 24971e3aac0SAndrea Arcangeli { 25071e3aac0SAndrea Arcangeli if (test_bit(enabled, &transparent_hugepage_flags)) { 25171e3aac0SAndrea Arcangeli VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); 25271e3aac0SAndrea Arcangeli return sprintf(buf, "[always] madvise never\n"); 25371e3aac0SAndrea Arcangeli } else if (test_bit(req_madv, &transparent_hugepage_flags)) 25471e3aac0SAndrea Arcangeli return sprintf(buf, "always [madvise] never\n"); 25571e3aac0SAndrea Arcangeli else 25671e3aac0SAndrea Arcangeli return sprintf(buf, "always madvise [never]\n"); 25771e3aac0SAndrea Arcangeli } 25871e3aac0SAndrea Arcangeli static ssize_t double_flag_store(struct kobject *kobj, 25971e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 26071e3aac0SAndrea Arcangeli const char *buf, size_t count, 26171e3aac0SAndrea Arcangeli enum transparent_hugepage_flag enabled, 26271e3aac0SAndrea Arcangeli enum transparent_hugepage_flag req_madv) 26371e3aac0SAndrea Arcangeli { 26471e3aac0SAndrea Arcangeli if (!memcmp("always", buf, 26571e3aac0SAndrea Arcangeli min(sizeof("always")-1, count))) { 26671e3aac0SAndrea Arcangeli set_bit(enabled, &transparent_hugepage_flags); 26771e3aac0SAndrea Arcangeli clear_bit(req_madv, &transparent_hugepage_flags); 26871e3aac0SAndrea Arcangeli } else if (!memcmp("madvise", buf, 26971e3aac0SAndrea Arcangeli min(sizeof("madvise")-1, count))) { 27071e3aac0SAndrea Arcangeli clear_bit(enabled, &transparent_hugepage_flags); 27171e3aac0SAndrea Arcangeli set_bit(req_madv, &transparent_hugepage_flags); 27271e3aac0SAndrea Arcangeli } else if (!memcmp("never", buf, 27371e3aac0SAndrea Arcangeli min(sizeof("never")-1, count))) { 27471e3aac0SAndrea Arcangeli clear_bit(enabled, &transparent_hugepage_flags); 27571e3aac0SAndrea Arcangeli clear_bit(req_madv, &transparent_hugepage_flags); 27671e3aac0SAndrea Arcangeli } else 27771e3aac0SAndrea Arcangeli return -EINVAL; 27871e3aac0SAndrea Arcangeli 27971e3aac0SAndrea Arcangeli return count; 28071e3aac0SAndrea Arcangeli } 28171e3aac0SAndrea Arcangeli 28271e3aac0SAndrea Arcangeli static ssize_t enabled_show(struct kobject *kobj, 28371e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 28471e3aac0SAndrea Arcangeli { 28571e3aac0SAndrea Arcangeli return double_flag_show(kobj, attr, buf, 28671e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_FLAG, 28771e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 28871e3aac0SAndrea Arcangeli } 28971e3aac0SAndrea Arcangeli static ssize_t enabled_store(struct kobject *kobj, 29071e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 29171e3aac0SAndrea Arcangeli const char *buf, size_t count) 29271e3aac0SAndrea Arcangeli { 293ba76149fSAndrea Arcangeli ssize_t ret; 294ba76149fSAndrea Arcangeli 295ba76149fSAndrea Arcangeli ret = double_flag_store(kobj, attr, buf, count, 29671e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_FLAG, 29771e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 298ba76149fSAndrea Arcangeli 299ba76149fSAndrea Arcangeli if (ret > 0) { 300911891afSXiao Guangrong int err; 301911891afSXiao Guangrong 302911891afSXiao Guangrong mutex_lock(&khugepaged_mutex); 303911891afSXiao Guangrong err = start_khugepaged(); 304911891afSXiao Guangrong mutex_unlock(&khugepaged_mutex); 305911891afSXiao Guangrong 306ba76149fSAndrea Arcangeli if (err) 307ba76149fSAndrea Arcangeli ret = err; 308ba76149fSAndrea Arcangeli } 309ba76149fSAndrea Arcangeli 310ba76149fSAndrea Arcangeli return ret; 31171e3aac0SAndrea Arcangeli } 31271e3aac0SAndrea Arcangeli static struct kobj_attribute enabled_attr = 31371e3aac0SAndrea Arcangeli __ATTR(enabled, 0644, enabled_show, enabled_store); 31471e3aac0SAndrea Arcangeli 31571e3aac0SAndrea Arcangeli static ssize_t single_flag_show(struct kobject *kobj, 31671e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 31771e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 31871e3aac0SAndrea Arcangeli { 319e27e6151SBen Hutchings return sprintf(buf, "%d\n", 320e27e6151SBen Hutchings !!test_bit(flag, &transparent_hugepage_flags)); 32171e3aac0SAndrea Arcangeli } 322e27e6151SBen Hutchings 32371e3aac0SAndrea Arcangeli static ssize_t single_flag_store(struct kobject *kobj, 32471e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 32571e3aac0SAndrea Arcangeli const char *buf, size_t count, 32671e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 32771e3aac0SAndrea Arcangeli { 328e27e6151SBen Hutchings unsigned long value; 329e27e6151SBen Hutchings int ret; 330e27e6151SBen Hutchings 331e27e6151SBen Hutchings ret = kstrtoul(buf, 10, &value); 332e27e6151SBen Hutchings if (ret < 0) 333e27e6151SBen Hutchings return ret; 334e27e6151SBen Hutchings if (value > 1) 33571e3aac0SAndrea Arcangeli return -EINVAL; 33671e3aac0SAndrea Arcangeli 337e27e6151SBen Hutchings if (value) 338e27e6151SBen Hutchings set_bit(flag, &transparent_hugepage_flags); 339e27e6151SBen Hutchings else 340e27e6151SBen Hutchings clear_bit(flag, &transparent_hugepage_flags); 341e27e6151SBen Hutchings 34271e3aac0SAndrea Arcangeli return count; 34371e3aac0SAndrea Arcangeli } 34471e3aac0SAndrea Arcangeli 34571e3aac0SAndrea Arcangeli /* 34671e3aac0SAndrea Arcangeli * Currently defrag only disables __GFP_NOWAIT for allocation. A blind 34771e3aac0SAndrea Arcangeli * __GFP_REPEAT is too aggressive, it's never worth swapping tons of 34871e3aac0SAndrea Arcangeli * memory just to allocate one more hugepage. 34971e3aac0SAndrea Arcangeli */ 35071e3aac0SAndrea Arcangeli static ssize_t defrag_show(struct kobject *kobj, 35171e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 35271e3aac0SAndrea Arcangeli { 35371e3aac0SAndrea Arcangeli return double_flag_show(kobj, attr, buf, 35471e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 35571e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 35671e3aac0SAndrea Arcangeli } 35771e3aac0SAndrea Arcangeli static ssize_t defrag_store(struct kobject *kobj, 35871e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 35971e3aac0SAndrea Arcangeli const char *buf, size_t count) 36071e3aac0SAndrea Arcangeli { 36171e3aac0SAndrea Arcangeli return double_flag_store(kobj, attr, buf, count, 36271e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 36371e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 36471e3aac0SAndrea Arcangeli } 36571e3aac0SAndrea Arcangeli static struct kobj_attribute defrag_attr = 36671e3aac0SAndrea Arcangeli __ATTR(defrag, 0644, defrag_show, defrag_store); 36771e3aac0SAndrea Arcangeli 36879da5407SKirill A. Shutemov static ssize_t use_zero_page_show(struct kobject *kobj, 36979da5407SKirill A. Shutemov struct kobj_attribute *attr, char *buf) 37079da5407SKirill A. Shutemov { 37179da5407SKirill A. Shutemov return single_flag_show(kobj, attr, buf, 37279da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 37379da5407SKirill A. Shutemov } 37479da5407SKirill A. Shutemov static ssize_t use_zero_page_store(struct kobject *kobj, 37579da5407SKirill A. Shutemov struct kobj_attribute *attr, const char *buf, size_t count) 37679da5407SKirill A. Shutemov { 37779da5407SKirill A. Shutemov return single_flag_store(kobj, attr, buf, count, 37879da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 37979da5407SKirill A. Shutemov } 38079da5407SKirill A. Shutemov static struct kobj_attribute use_zero_page_attr = 38179da5407SKirill A. Shutemov __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 38271e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 38371e3aac0SAndrea Arcangeli static ssize_t debug_cow_show(struct kobject *kobj, 38471e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 38571e3aac0SAndrea Arcangeli { 38671e3aac0SAndrea Arcangeli return single_flag_show(kobj, attr, buf, 38771e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 38871e3aac0SAndrea Arcangeli } 38971e3aac0SAndrea Arcangeli static ssize_t debug_cow_store(struct kobject *kobj, 39071e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 39171e3aac0SAndrea Arcangeli const char *buf, size_t count) 39271e3aac0SAndrea Arcangeli { 39371e3aac0SAndrea Arcangeli return single_flag_store(kobj, attr, buf, count, 39471e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 39571e3aac0SAndrea Arcangeli } 39671e3aac0SAndrea Arcangeli static struct kobj_attribute debug_cow_attr = 39771e3aac0SAndrea Arcangeli __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 39871e3aac0SAndrea Arcangeli #endif /* CONFIG_DEBUG_VM */ 39971e3aac0SAndrea Arcangeli 40071e3aac0SAndrea Arcangeli static struct attribute *hugepage_attr[] = { 40171e3aac0SAndrea Arcangeli &enabled_attr.attr, 40271e3aac0SAndrea Arcangeli &defrag_attr.attr, 40379da5407SKirill A. Shutemov &use_zero_page_attr.attr, 40471e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 40571e3aac0SAndrea Arcangeli &debug_cow_attr.attr, 40671e3aac0SAndrea Arcangeli #endif 40771e3aac0SAndrea Arcangeli NULL, 40871e3aac0SAndrea Arcangeli }; 40971e3aac0SAndrea Arcangeli 41071e3aac0SAndrea Arcangeli static struct attribute_group hugepage_attr_group = { 41171e3aac0SAndrea Arcangeli .attrs = hugepage_attr, 412ba76149fSAndrea Arcangeli }; 413ba76149fSAndrea Arcangeli 414ba76149fSAndrea Arcangeli static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 415ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 416ba76149fSAndrea Arcangeli char *buf) 417ba76149fSAndrea Arcangeli { 418ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); 419ba76149fSAndrea Arcangeli } 420ba76149fSAndrea Arcangeli 421ba76149fSAndrea Arcangeli static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 422ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 423ba76149fSAndrea Arcangeli const char *buf, size_t count) 424ba76149fSAndrea Arcangeli { 425ba76149fSAndrea Arcangeli unsigned long msecs; 426ba76149fSAndrea Arcangeli int err; 427ba76149fSAndrea Arcangeli 4283dbb95f7SJingoo Han err = kstrtoul(buf, 10, &msecs); 429ba76149fSAndrea Arcangeli if (err || msecs > UINT_MAX) 430ba76149fSAndrea Arcangeli return -EINVAL; 431ba76149fSAndrea Arcangeli 432ba76149fSAndrea Arcangeli khugepaged_scan_sleep_millisecs = msecs; 433ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 434ba76149fSAndrea Arcangeli 435ba76149fSAndrea Arcangeli return count; 436ba76149fSAndrea Arcangeli } 437ba76149fSAndrea Arcangeli static struct kobj_attribute scan_sleep_millisecs_attr = 438ba76149fSAndrea Arcangeli __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 439ba76149fSAndrea Arcangeli scan_sleep_millisecs_store); 440ba76149fSAndrea Arcangeli 441ba76149fSAndrea Arcangeli static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 442ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 443ba76149fSAndrea Arcangeli char *buf) 444ba76149fSAndrea Arcangeli { 445ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 446ba76149fSAndrea Arcangeli } 447ba76149fSAndrea Arcangeli 448ba76149fSAndrea Arcangeli static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 449ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 450ba76149fSAndrea Arcangeli const char *buf, size_t count) 451ba76149fSAndrea Arcangeli { 452ba76149fSAndrea Arcangeli unsigned long msecs; 453ba76149fSAndrea Arcangeli int err; 454ba76149fSAndrea Arcangeli 4553dbb95f7SJingoo Han err = kstrtoul(buf, 10, &msecs); 456ba76149fSAndrea Arcangeli if (err || msecs > UINT_MAX) 457ba76149fSAndrea Arcangeli return -EINVAL; 458ba76149fSAndrea Arcangeli 459ba76149fSAndrea Arcangeli khugepaged_alloc_sleep_millisecs = msecs; 460ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 461ba76149fSAndrea Arcangeli 462ba76149fSAndrea Arcangeli return count; 463ba76149fSAndrea Arcangeli } 464ba76149fSAndrea Arcangeli static struct kobj_attribute alloc_sleep_millisecs_attr = 465ba76149fSAndrea Arcangeli __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 466ba76149fSAndrea Arcangeli alloc_sleep_millisecs_store); 467ba76149fSAndrea Arcangeli 468ba76149fSAndrea Arcangeli static ssize_t pages_to_scan_show(struct kobject *kobj, 469ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 470ba76149fSAndrea Arcangeli char *buf) 471ba76149fSAndrea Arcangeli { 472ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_pages_to_scan); 473ba76149fSAndrea Arcangeli } 474ba76149fSAndrea Arcangeli static ssize_t pages_to_scan_store(struct kobject *kobj, 475ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 476ba76149fSAndrea Arcangeli const char *buf, size_t count) 477ba76149fSAndrea Arcangeli { 478ba76149fSAndrea Arcangeli int err; 479ba76149fSAndrea Arcangeli unsigned long pages; 480ba76149fSAndrea Arcangeli 4813dbb95f7SJingoo Han err = kstrtoul(buf, 10, &pages); 482ba76149fSAndrea Arcangeli if (err || !pages || pages > UINT_MAX) 483ba76149fSAndrea Arcangeli return -EINVAL; 484ba76149fSAndrea Arcangeli 485ba76149fSAndrea Arcangeli khugepaged_pages_to_scan = pages; 486ba76149fSAndrea Arcangeli 487ba76149fSAndrea Arcangeli return count; 488ba76149fSAndrea Arcangeli } 489ba76149fSAndrea Arcangeli static struct kobj_attribute pages_to_scan_attr = 490ba76149fSAndrea Arcangeli __ATTR(pages_to_scan, 0644, pages_to_scan_show, 491ba76149fSAndrea Arcangeli pages_to_scan_store); 492ba76149fSAndrea Arcangeli 493ba76149fSAndrea Arcangeli static ssize_t pages_collapsed_show(struct kobject *kobj, 494ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 495ba76149fSAndrea Arcangeli char *buf) 496ba76149fSAndrea Arcangeli { 497ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_pages_collapsed); 498ba76149fSAndrea Arcangeli } 499ba76149fSAndrea Arcangeli static struct kobj_attribute pages_collapsed_attr = 500ba76149fSAndrea Arcangeli __ATTR_RO(pages_collapsed); 501ba76149fSAndrea Arcangeli 502ba76149fSAndrea Arcangeli static ssize_t full_scans_show(struct kobject *kobj, 503ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 504ba76149fSAndrea Arcangeli char *buf) 505ba76149fSAndrea Arcangeli { 506ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_full_scans); 507ba76149fSAndrea Arcangeli } 508ba76149fSAndrea Arcangeli static struct kobj_attribute full_scans_attr = 509ba76149fSAndrea Arcangeli __ATTR_RO(full_scans); 510ba76149fSAndrea Arcangeli 511ba76149fSAndrea Arcangeli static ssize_t khugepaged_defrag_show(struct kobject *kobj, 512ba76149fSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 513ba76149fSAndrea Arcangeli { 514ba76149fSAndrea Arcangeli return single_flag_show(kobj, attr, buf, 515ba76149fSAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 516ba76149fSAndrea Arcangeli } 517ba76149fSAndrea Arcangeli static ssize_t khugepaged_defrag_store(struct kobject *kobj, 518ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 519ba76149fSAndrea Arcangeli const char *buf, size_t count) 520ba76149fSAndrea Arcangeli { 521ba76149fSAndrea Arcangeli return single_flag_store(kobj, attr, buf, count, 522ba76149fSAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 523ba76149fSAndrea Arcangeli } 524ba76149fSAndrea Arcangeli static struct kobj_attribute khugepaged_defrag_attr = 525ba76149fSAndrea Arcangeli __ATTR(defrag, 0644, khugepaged_defrag_show, 526ba76149fSAndrea Arcangeli khugepaged_defrag_store); 527ba76149fSAndrea Arcangeli 528ba76149fSAndrea Arcangeli /* 529ba76149fSAndrea Arcangeli * max_ptes_none controls if khugepaged should collapse hugepages over 530ba76149fSAndrea Arcangeli * any unmapped ptes in turn potentially increasing the memory 531ba76149fSAndrea Arcangeli * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 532ba76149fSAndrea Arcangeli * reduce the available free memory in the system as it 533ba76149fSAndrea Arcangeli * runs. Increasing max_ptes_none will instead potentially reduce the 534ba76149fSAndrea Arcangeli * free memory in the system during the khugepaged scan. 535ba76149fSAndrea Arcangeli */ 536ba76149fSAndrea Arcangeli static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 537ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 538ba76149fSAndrea Arcangeli char *buf) 539ba76149fSAndrea Arcangeli { 540ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_max_ptes_none); 541ba76149fSAndrea Arcangeli } 542ba76149fSAndrea Arcangeli static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 543ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 544ba76149fSAndrea Arcangeli const char *buf, size_t count) 545ba76149fSAndrea Arcangeli { 546ba76149fSAndrea Arcangeli int err; 547ba76149fSAndrea Arcangeli unsigned long max_ptes_none; 548ba76149fSAndrea Arcangeli 5493dbb95f7SJingoo Han err = kstrtoul(buf, 10, &max_ptes_none); 550ba76149fSAndrea Arcangeli if (err || max_ptes_none > HPAGE_PMD_NR-1) 551ba76149fSAndrea Arcangeli return -EINVAL; 552ba76149fSAndrea Arcangeli 553ba76149fSAndrea Arcangeli khugepaged_max_ptes_none = max_ptes_none; 554ba76149fSAndrea Arcangeli 555ba76149fSAndrea Arcangeli return count; 556ba76149fSAndrea Arcangeli } 557ba76149fSAndrea Arcangeli static struct kobj_attribute khugepaged_max_ptes_none_attr = 558ba76149fSAndrea Arcangeli __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 559ba76149fSAndrea Arcangeli khugepaged_max_ptes_none_store); 560ba76149fSAndrea Arcangeli 561ba76149fSAndrea Arcangeli static struct attribute *khugepaged_attr[] = { 562ba76149fSAndrea Arcangeli &khugepaged_defrag_attr.attr, 563ba76149fSAndrea Arcangeli &khugepaged_max_ptes_none_attr.attr, 564ba76149fSAndrea Arcangeli &pages_to_scan_attr.attr, 565ba76149fSAndrea Arcangeli &pages_collapsed_attr.attr, 566ba76149fSAndrea Arcangeli &full_scans_attr.attr, 567ba76149fSAndrea Arcangeli &scan_sleep_millisecs_attr.attr, 568ba76149fSAndrea Arcangeli &alloc_sleep_millisecs_attr.attr, 569ba76149fSAndrea Arcangeli NULL, 570ba76149fSAndrea Arcangeli }; 571ba76149fSAndrea Arcangeli 572ba76149fSAndrea Arcangeli static struct attribute_group khugepaged_attr_group = { 573ba76149fSAndrea Arcangeli .attrs = khugepaged_attr, 574ba76149fSAndrea Arcangeli .name = "khugepaged", 57571e3aac0SAndrea Arcangeli }; 576569e5590SShaohua Li 577569e5590SShaohua Li static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 578569e5590SShaohua Li { 579569e5590SShaohua Li int err; 580569e5590SShaohua Li 581569e5590SShaohua Li *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 582569e5590SShaohua Li if (unlikely(!*hugepage_kobj)) { 583ae3a8c1cSAndrew Morton pr_err("failed to create transparent hugepage kobject\n"); 584569e5590SShaohua Li return -ENOMEM; 585569e5590SShaohua Li } 586569e5590SShaohua Li 587569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 588569e5590SShaohua Li if (err) { 589ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 590569e5590SShaohua Li goto delete_obj; 591569e5590SShaohua Li } 592569e5590SShaohua Li 593569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 594569e5590SShaohua Li if (err) { 595ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 596569e5590SShaohua Li goto remove_hp_group; 597569e5590SShaohua Li } 598569e5590SShaohua Li 599569e5590SShaohua Li return 0; 600569e5590SShaohua Li 601569e5590SShaohua Li remove_hp_group: 602569e5590SShaohua Li sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 603569e5590SShaohua Li delete_obj: 604569e5590SShaohua Li kobject_put(*hugepage_kobj); 605569e5590SShaohua Li return err; 606569e5590SShaohua Li } 607569e5590SShaohua Li 608569e5590SShaohua Li static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 609569e5590SShaohua Li { 610569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 611569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 612569e5590SShaohua Li kobject_put(hugepage_kobj); 613569e5590SShaohua Li } 614569e5590SShaohua Li #else 615569e5590SShaohua Li static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 616569e5590SShaohua Li { 617569e5590SShaohua Li return 0; 618569e5590SShaohua Li } 619569e5590SShaohua Li 620569e5590SShaohua Li static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 621569e5590SShaohua Li { 622569e5590SShaohua Li } 62371e3aac0SAndrea Arcangeli #endif /* CONFIG_SYSFS */ 62471e3aac0SAndrea Arcangeli 62571e3aac0SAndrea Arcangeli static int __init hugepage_init(void) 62671e3aac0SAndrea Arcangeli { 62771e3aac0SAndrea Arcangeli int err; 628569e5590SShaohua Li struct kobject *hugepage_kobj; 62971e3aac0SAndrea Arcangeli 6304b7167b9SAndrea Arcangeli if (!has_transparent_hugepage()) { 6314b7167b9SAndrea Arcangeli transparent_hugepage_flags = 0; 632569e5590SShaohua Li return -EINVAL; 6334b7167b9SAndrea Arcangeli } 6344b7167b9SAndrea Arcangeli 635569e5590SShaohua Li err = hugepage_init_sysfs(&hugepage_kobj); 636569e5590SShaohua Li if (err) 637569e5590SShaohua Li return err; 638ba76149fSAndrea Arcangeli 639ba76149fSAndrea Arcangeli err = khugepaged_slab_init(); 640ba76149fSAndrea Arcangeli if (err) 641ba76149fSAndrea Arcangeli goto out; 642ba76149fSAndrea Arcangeli 64397ae1749SKirill A. Shutemov register_shrinker(&huge_zero_page_shrinker); 64497ae1749SKirill A. Shutemov 64597562cd2SRik van Riel /* 64697562cd2SRik van Riel * By default disable transparent hugepages on smaller systems, 64797562cd2SRik van Riel * where the extra memory used could hurt more than TLB overhead 64897562cd2SRik van Riel * is likely to save. The admin can still enable it through /sys. 64997562cd2SRik van Riel */ 65097562cd2SRik van Riel if (totalram_pages < (512 << (20 - PAGE_SHIFT))) 65197562cd2SRik van Riel transparent_hugepage_flags = 0; 65297562cd2SRik van Riel 653ba76149fSAndrea Arcangeli start_khugepaged(); 654ba76149fSAndrea Arcangeli 655569e5590SShaohua Li return 0; 656ba76149fSAndrea Arcangeli out: 657569e5590SShaohua Li hugepage_exit_sysfs(hugepage_kobj); 658ba76149fSAndrea Arcangeli return err; 65971e3aac0SAndrea Arcangeli } 660a64fb3cdSPaul Gortmaker subsys_initcall(hugepage_init); 66171e3aac0SAndrea Arcangeli 66271e3aac0SAndrea Arcangeli static int __init setup_transparent_hugepage(char *str) 66371e3aac0SAndrea Arcangeli { 66471e3aac0SAndrea Arcangeli int ret = 0; 66571e3aac0SAndrea Arcangeli if (!str) 66671e3aac0SAndrea Arcangeli goto out; 66771e3aac0SAndrea Arcangeli if (!strcmp(str, "always")) { 66871e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_FLAG, 66971e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 67071e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 67171e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 67271e3aac0SAndrea Arcangeli ret = 1; 67371e3aac0SAndrea Arcangeli } else if (!strcmp(str, "madvise")) { 67471e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 67571e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 67671e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 67771e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 67871e3aac0SAndrea Arcangeli ret = 1; 67971e3aac0SAndrea Arcangeli } else if (!strcmp(str, "never")) { 68071e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 68171e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 68271e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 68371e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 68471e3aac0SAndrea Arcangeli ret = 1; 68571e3aac0SAndrea Arcangeli } 68671e3aac0SAndrea Arcangeli out: 68771e3aac0SAndrea Arcangeli if (!ret) 688ae3a8c1cSAndrew Morton pr_warn("transparent_hugepage= cannot parse, ignored\n"); 68971e3aac0SAndrea Arcangeli return ret; 69071e3aac0SAndrea Arcangeli } 69171e3aac0SAndrea Arcangeli __setup("transparent_hugepage=", setup_transparent_hugepage); 69271e3aac0SAndrea Arcangeli 693b32967ffSMel Gorman pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 69471e3aac0SAndrea Arcangeli { 69571e3aac0SAndrea Arcangeli if (likely(vma->vm_flags & VM_WRITE)) 69671e3aac0SAndrea Arcangeli pmd = pmd_mkwrite(pmd); 69771e3aac0SAndrea Arcangeli return pmd; 69871e3aac0SAndrea Arcangeli } 69971e3aac0SAndrea Arcangeli 7003122359aSKirill A. Shutemov static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) 701b3092b3bSBob Liu { 702b3092b3bSBob Liu pmd_t entry; 7033122359aSKirill A. Shutemov entry = mk_pmd(page, prot); 704b3092b3bSBob Liu entry = pmd_mkhuge(entry); 705b3092b3bSBob Liu return entry; 706b3092b3bSBob Liu } 707b3092b3bSBob Liu 70871e3aac0SAndrea Arcangeli static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, 70971e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 71071e3aac0SAndrea Arcangeli unsigned long haddr, pmd_t *pmd, 71171e3aac0SAndrea Arcangeli struct page *page) 71271e3aac0SAndrea Arcangeli { 71300501b53SJohannes Weiner struct mem_cgroup *memcg; 71471e3aac0SAndrea Arcangeli pgtable_t pgtable; 715c4088ebdSKirill A. Shutemov spinlock_t *ptl; 71671e3aac0SAndrea Arcangeli 717309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page), page); 71800501b53SJohannes Weiner 71900501b53SJohannes Weiner if (mem_cgroup_try_charge(page, mm, GFP_TRANSHUGE, &memcg)) 72071e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 72171e3aac0SAndrea Arcangeli 72200501b53SJohannes Weiner pgtable = pte_alloc_one(mm, haddr); 72300501b53SJohannes Weiner if (unlikely(!pgtable)) { 72400501b53SJohannes Weiner mem_cgroup_cancel_charge(page, memcg); 72500501b53SJohannes Weiner return VM_FAULT_OOM; 72600501b53SJohannes Weiner } 72700501b53SJohannes Weiner 72871e3aac0SAndrea Arcangeli clear_huge_page(page, haddr, HPAGE_PMD_NR); 72952f37629SMinchan Kim /* 73052f37629SMinchan Kim * The memory barrier inside __SetPageUptodate makes sure that 73152f37629SMinchan Kim * clear_huge_page writes become visible before the set_pmd_at() 73252f37629SMinchan Kim * write. 73352f37629SMinchan Kim */ 73471e3aac0SAndrea Arcangeli __SetPageUptodate(page); 73571e3aac0SAndrea Arcangeli 736c4088ebdSKirill A. Shutemov ptl = pmd_lock(mm, pmd); 73771e3aac0SAndrea Arcangeli if (unlikely(!pmd_none(*pmd))) { 738c4088ebdSKirill A. Shutemov spin_unlock(ptl); 73900501b53SJohannes Weiner mem_cgroup_cancel_charge(page, memcg); 74071e3aac0SAndrea Arcangeli put_page(page); 74171e3aac0SAndrea Arcangeli pte_free(mm, pgtable); 74271e3aac0SAndrea Arcangeli } else { 74371e3aac0SAndrea Arcangeli pmd_t entry; 7443122359aSKirill A. Shutemov entry = mk_huge_pmd(page, vma->vm_page_prot); 7453122359aSKirill A. Shutemov entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 74671e3aac0SAndrea Arcangeli page_add_new_anon_rmap(page, vma, haddr); 74700501b53SJohannes Weiner mem_cgroup_commit_charge(page, memcg, false); 74800501b53SJohannes Weiner lru_cache_add_active_or_unevictable(page, vma); 7496b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 75071e3aac0SAndrea Arcangeli set_pmd_at(mm, haddr, pmd, entry); 75171e3aac0SAndrea Arcangeli add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 752e1f56c89SKirill A. Shutemov atomic_long_inc(&mm->nr_ptes); 753c4088ebdSKirill A. Shutemov spin_unlock(ptl); 75471e3aac0SAndrea Arcangeli } 75571e3aac0SAndrea Arcangeli 756aa2e878eSDavid Rientjes return 0; 75771e3aac0SAndrea Arcangeli } 75871e3aac0SAndrea Arcangeli 759cc5d462fSAndi Kleen static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) 7600bbbc0b3SAndrea Arcangeli { 761cc5d462fSAndi Kleen return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp; 7620bbbc0b3SAndrea Arcangeli } 7630bbbc0b3SAndrea Arcangeli 764c4088ebdSKirill A. Shutemov /* Caller must hold page table lock. */ 7653ea41e62SKirill A. Shutemov static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 76697ae1749SKirill A. Shutemov struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 7675918d10aSKirill A. Shutemov struct page *zero_page) 768fc9fe822SKirill A. Shutemov { 769fc9fe822SKirill A. Shutemov pmd_t entry; 7703ea41e62SKirill A. Shutemov if (!pmd_none(*pmd)) 7713ea41e62SKirill A. Shutemov return false; 7725918d10aSKirill A. Shutemov entry = mk_pmd(zero_page, vma->vm_page_prot); 773fc9fe822SKirill A. Shutemov entry = pmd_mkhuge(entry); 7746b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 775fc9fe822SKirill A. Shutemov set_pmd_at(mm, haddr, pmd, entry); 776e1f56c89SKirill A. Shutemov atomic_long_inc(&mm->nr_ptes); 7773ea41e62SKirill A. Shutemov return true; 778fc9fe822SKirill A. Shutemov } 779fc9fe822SKirill A. Shutemov 78071e3aac0SAndrea Arcangeli int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 78171e3aac0SAndrea Arcangeli unsigned long address, pmd_t *pmd, 78271e3aac0SAndrea Arcangeli unsigned int flags) 78371e3aac0SAndrea Arcangeli { 784077fcf11SAneesh Kumar K.V gfp_t gfp; 78571e3aac0SAndrea Arcangeli struct page *page; 78671e3aac0SAndrea Arcangeli unsigned long haddr = address & HPAGE_PMD_MASK; 78771e3aac0SAndrea Arcangeli 788128ec037SKirill A. Shutemov if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 789c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 79071e3aac0SAndrea Arcangeli if (unlikely(anon_vma_prepare(vma))) 79171e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 7926d50e60cSDavid Rientjes if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 793ba76149fSAndrea Arcangeli return VM_FAULT_OOM; 794593befa6SDominik Dingel if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) && 79579da5407SKirill A. Shutemov transparent_hugepage_use_zero_page()) { 796c4088ebdSKirill A. Shutemov spinlock_t *ptl; 79780371957SKirill A. Shutemov pgtable_t pgtable; 7985918d10aSKirill A. Shutemov struct page *zero_page; 7993ea41e62SKirill A. Shutemov bool set; 80080371957SKirill A. Shutemov pgtable = pte_alloc_one(mm, haddr); 80180371957SKirill A. Shutemov if (unlikely(!pgtable)) 80280371957SKirill A. Shutemov return VM_FAULT_OOM; 8035918d10aSKirill A. Shutemov zero_page = get_huge_zero_page(); 8045918d10aSKirill A. Shutemov if (unlikely(!zero_page)) { 80597ae1749SKirill A. Shutemov pte_free(mm, pgtable); 80697ae1749SKirill A. Shutemov count_vm_event(THP_FAULT_FALLBACK); 807c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 80897ae1749SKirill A. Shutemov } 809c4088ebdSKirill A. Shutemov ptl = pmd_lock(mm, pmd); 8103ea41e62SKirill A. Shutemov set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd, 8115918d10aSKirill A. Shutemov zero_page); 812c4088ebdSKirill A. Shutemov spin_unlock(ptl); 8133ea41e62SKirill A. Shutemov if (!set) { 8143ea41e62SKirill A. Shutemov pte_free(mm, pgtable); 8153ea41e62SKirill A. Shutemov put_huge_zero_page(); 8163ea41e62SKirill A. Shutemov } 81780371957SKirill A. Shutemov return 0; 81880371957SKirill A. Shutemov } 819077fcf11SAneesh Kumar K.V gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); 820077fcf11SAneesh Kumar K.V page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 82181ab4201SAndi Kleen if (unlikely(!page)) { 82281ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 823c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 82481ab4201SAndi Kleen } 825128ec037SKirill A. Shutemov if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) { 826edad9d2cSDavid Rientjes put_page(page); 82717766ddeSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK); 828c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 829edad9d2cSDavid Rientjes } 83071e3aac0SAndrea Arcangeli 83117766ddeSDavid Rientjes count_vm_event(THP_FAULT_ALLOC); 832edad9d2cSDavid Rientjes return 0; 83371e3aac0SAndrea Arcangeli } 83471e3aac0SAndrea Arcangeli 83571e3aac0SAndrea Arcangeli int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 83671e3aac0SAndrea Arcangeli pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 83771e3aac0SAndrea Arcangeli struct vm_area_struct *vma) 83871e3aac0SAndrea Arcangeli { 839c4088ebdSKirill A. Shutemov spinlock_t *dst_ptl, *src_ptl; 84071e3aac0SAndrea Arcangeli struct page *src_page; 84171e3aac0SAndrea Arcangeli pmd_t pmd; 84271e3aac0SAndrea Arcangeli pgtable_t pgtable; 84371e3aac0SAndrea Arcangeli int ret; 84471e3aac0SAndrea Arcangeli 84571e3aac0SAndrea Arcangeli ret = -ENOMEM; 84671e3aac0SAndrea Arcangeli pgtable = pte_alloc_one(dst_mm, addr); 84771e3aac0SAndrea Arcangeli if (unlikely(!pgtable)) 84871e3aac0SAndrea Arcangeli goto out; 84971e3aac0SAndrea Arcangeli 850c4088ebdSKirill A. Shutemov dst_ptl = pmd_lock(dst_mm, dst_pmd); 851c4088ebdSKirill A. Shutemov src_ptl = pmd_lockptr(src_mm, src_pmd); 852c4088ebdSKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 85371e3aac0SAndrea Arcangeli 85471e3aac0SAndrea Arcangeli ret = -EAGAIN; 85571e3aac0SAndrea Arcangeli pmd = *src_pmd; 85671e3aac0SAndrea Arcangeli if (unlikely(!pmd_trans_huge(pmd))) { 85771e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 85871e3aac0SAndrea Arcangeli goto out_unlock; 85971e3aac0SAndrea Arcangeli } 860fc9fe822SKirill A. Shutemov /* 861c4088ebdSKirill A. Shutemov * When page table lock is held, the huge zero pmd should not be 862fc9fe822SKirill A. Shutemov * under splitting since we don't split the page itself, only pmd to 863fc9fe822SKirill A. Shutemov * a page table. 864fc9fe822SKirill A. Shutemov */ 865fc9fe822SKirill A. Shutemov if (is_huge_zero_pmd(pmd)) { 8665918d10aSKirill A. Shutemov struct page *zero_page; 8673ea41e62SKirill A. Shutemov bool set; 86897ae1749SKirill A. Shutemov /* 86997ae1749SKirill A. Shutemov * get_huge_zero_page() will never allocate a new page here, 87097ae1749SKirill A. Shutemov * since we already have a zero page to copy. It just takes a 87197ae1749SKirill A. Shutemov * reference. 87297ae1749SKirill A. Shutemov */ 8735918d10aSKirill A. Shutemov zero_page = get_huge_zero_page(); 8743ea41e62SKirill A. Shutemov set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 8755918d10aSKirill A. Shutemov zero_page); 8763ea41e62SKirill A. Shutemov BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */ 877fc9fe822SKirill A. Shutemov ret = 0; 878fc9fe822SKirill A. Shutemov goto out_unlock; 879fc9fe822SKirill A. Shutemov } 880de466bd6SMel Gorman 88171e3aac0SAndrea Arcangeli if (unlikely(pmd_trans_splitting(pmd))) { 88271e3aac0SAndrea Arcangeli /* split huge page running from under us */ 883c4088ebdSKirill A. Shutemov spin_unlock(src_ptl); 884c4088ebdSKirill A. Shutemov spin_unlock(dst_ptl); 88571e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 88671e3aac0SAndrea Arcangeli 88771e3aac0SAndrea Arcangeli wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ 88871e3aac0SAndrea Arcangeli goto out; 88971e3aac0SAndrea Arcangeli } 89071e3aac0SAndrea Arcangeli src_page = pmd_page(pmd); 891309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 89271e3aac0SAndrea Arcangeli get_page(src_page); 89371e3aac0SAndrea Arcangeli page_dup_rmap(src_page); 89471e3aac0SAndrea Arcangeli add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 89571e3aac0SAndrea Arcangeli 89671e3aac0SAndrea Arcangeli pmdp_set_wrprotect(src_mm, addr, src_pmd); 89771e3aac0SAndrea Arcangeli pmd = pmd_mkold(pmd_wrprotect(pmd)); 8986b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 89971e3aac0SAndrea Arcangeli set_pmd_at(dst_mm, addr, dst_pmd, pmd); 900e1f56c89SKirill A. Shutemov atomic_long_inc(&dst_mm->nr_ptes); 90171e3aac0SAndrea Arcangeli 90271e3aac0SAndrea Arcangeli ret = 0; 90371e3aac0SAndrea Arcangeli out_unlock: 904c4088ebdSKirill A. Shutemov spin_unlock(src_ptl); 905c4088ebdSKirill A. Shutemov spin_unlock(dst_ptl); 90671e3aac0SAndrea Arcangeli out: 90771e3aac0SAndrea Arcangeli return ret; 90871e3aac0SAndrea Arcangeli } 90971e3aac0SAndrea Arcangeli 910a1dd450bSWill Deacon void huge_pmd_set_accessed(struct mm_struct *mm, 911a1dd450bSWill Deacon struct vm_area_struct *vma, 912a1dd450bSWill Deacon unsigned long address, 913a1dd450bSWill Deacon pmd_t *pmd, pmd_t orig_pmd, 914a1dd450bSWill Deacon int dirty) 915a1dd450bSWill Deacon { 916c4088ebdSKirill A. Shutemov spinlock_t *ptl; 917a1dd450bSWill Deacon pmd_t entry; 918a1dd450bSWill Deacon unsigned long haddr; 919a1dd450bSWill Deacon 920c4088ebdSKirill A. Shutemov ptl = pmd_lock(mm, pmd); 921a1dd450bSWill Deacon if (unlikely(!pmd_same(*pmd, orig_pmd))) 922a1dd450bSWill Deacon goto unlock; 923a1dd450bSWill Deacon 924a1dd450bSWill Deacon entry = pmd_mkyoung(orig_pmd); 925a1dd450bSWill Deacon haddr = address & HPAGE_PMD_MASK; 926a1dd450bSWill Deacon if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) 927a1dd450bSWill Deacon update_mmu_cache_pmd(vma, address, pmd); 928a1dd450bSWill Deacon 929a1dd450bSWill Deacon unlock: 930c4088ebdSKirill A. Shutemov spin_unlock(ptl); 931a1dd450bSWill Deacon } 932a1dd450bSWill Deacon 9335338a937SHugh Dickins /* 9345338a937SHugh Dickins * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages 9355338a937SHugh Dickins * during copy_user_huge_page()'s copy_page_rep(): in the case when 9365338a937SHugh Dickins * the source page gets split and a tail freed before copy completes. 9375338a937SHugh Dickins * Called under pmd_lock of checked pmd, so safe from splitting itself. 9385338a937SHugh Dickins */ 9395338a937SHugh Dickins static void get_user_huge_page(struct page *page) 9405338a937SHugh Dickins { 9415338a937SHugh Dickins if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { 9425338a937SHugh Dickins struct page *endpage = page + HPAGE_PMD_NR; 9435338a937SHugh Dickins 9445338a937SHugh Dickins atomic_add(HPAGE_PMD_NR, &page->_count); 9455338a937SHugh Dickins while (++page < endpage) 9465338a937SHugh Dickins get_huge_page_tail(page); 9475338a937SHugh Dickins } else { 9485338a937SHugh Dickins get_page(page); 9495338a937SHugh Dickins } 9505338a937SHugh Dickins } 9515338a937SHugh Dickins 9525338a937SHugh Dickins static void put_user_huge_page(struct page *page) 9535338a937SHugh Dickins { 9545338a937SHugh Dickins if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { 9555338a937SHugh Dickins struct page *endpage = page + HPAGE_PMD_NR; 9565338a937SHugh Dickins 9575338a937SHugh Dickins while (page < endpage) 9585338a937SHugh Dickins put_page(page++); 9595338a937SHugh Dickins } else { 9605338a937SHugh Dickins put_page(page); 9615338a937SHugh Dickins } 9625338a937SHugh Dickins } 9635338a937SHugh Dickins 96471e3aac0SAndrea Arcangeli static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, 96571e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 96671e3aac0SAndrea Arcangeli unsigned long address, 96771e3aac0SAndrea Arcangeli pmd_t *pmd, pmd_t orig_pmd, 96871e3aac0SAndrea Arcangeli struct page *page, 96971e3aac0SAndrea Arcangeli unsigned long haddr) 97071e3aac0SAndrea Arcangeli { 97100501b53SJohannes Weiner struct mem_cgroup *memcg; 972c4088ebdSKirill A. Shutemov spinlock_t *ptl; 97371e3aac0SAndrea Arcangeli pgtable_t pgtable; 97471e3aac0SAndrea Arcangeli pmd_t _pmd; 97571e3aac0SAndrea Arcangeli int ret = 0, i; 97671e3aac0SAndrea Arcangeli struct page **pages; 9772ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 9782ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 97971e3aac0SAndrea Arcangeli 98071e3aac0SAndrea Arcangeli pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 98171e3aac0SAndrea Arcangeli GFP_KERNEL); 98271e3aac0SAndrea Arcangeli if (unlikely(!pages)) { 98371e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 98471e3aac0SAndrea Arcangeli goto out; 98571e3aac0SAndrea Arcangeli } 98671e3aac0SAndrea Arcangeli 98771e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 988cc5d462fSAndi Kleen pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | 989cc5d462fSAndi Kleen __GFP_OTHER_NODE, 99019ee151eSAndi Kleen vma, address, page_to_nid(page)); 991b9bbfbe3SAndrea Arcangeli if (unlikely(!pages[i] || 99200501b53SJohannes Weiner mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL, 99300501b53SJohannes Weiner &memcg))) { 994b9bbfbe3SAndrea Arcangeli if (pages[i]) 99571e3aac0SAndrea Arcangeli put_page(pages[i]); 996b9bbfbe3SAndrea Arcangeli while (--i >= 0) { 99700501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 99800501b53SJohannes Weiner set_page_private(pages[i], 0); 99900501b53SJohannes Weiner mem_cgroup_cancel_charge(pages[i], memcg); 1000b9bbfbe3SAndrea Arcangeli put_page(pages[i]); 1001b9bbfbe3SAndrea Arcangeli } 100271e3aac0SAndrea Arcangeli kfree(pages); 100371e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 100471e3aac0SAndrea Arcangeli goto out; 100571e3aac0SAndrea Arcangeli } 100600501b53SJohannes Weiner set_page_private(pages[i], (unsigned long)memcg); 100771e3aac0SAndrea Arcangeli } 100871e3aac0SAndrea Arcangeli 100971e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 101071e3aac0SAndrea Arcangeli copy_user_highpage(pages[i], page + i, 10110089e485SHillf Danton haddr + PAGE_SIZE * i, vma); 101271e3aac0SAndrea Arcangeli __SetPageUptodate(pages[i]); 101371e3aac0SAndrea Arcangeli cond_resched(); 101471e3aac0SAndrea Arcangeli } 101571e3aac0SAndrea Arcangeli 10162ec74c3eSSagi Grimberg mmun_start = haddr; 10172ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 10182ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 10192ec74c3eSSagi Grimberg 1020c4088ebdSKirill A. Shutemov ptl = pmd_lock(mm, pmd); 102171e3aac0SAndrea Arcangeli if (unlikely(!pmd_same(*pmd, orig_pmd))) 102271e3aac0SAndrea Arcangeli goto out_free_pages; 1023309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 102471e3aac0SAndrea Arcangeli 102534ee645eSJoerg Roedel pmdp_clear_flush_notify(vma, haddr, pmd); 102671e3aac0SAndrea Arcangeli /* leave pmd empty until pte is filled */ 102771e3aac0SAndrea Arcangeli 10286b0b50b0SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(mm, pmd); 102971e3aac0SAndrea Arcangeli pmd_populate(mm, &_pmd, pgtable); 103071e3aac0SAndrea Arcangeli 103171e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 103271e3aac0SAndrea Arcangeli pte_t *pte, entry; 103371e3aac0SAndrea Arcangeli entry = mk_pte(pages[i], vma->vm_page_prot); 103471e3aac0SAndrea Arcangeli entry = maybe_mkwrite(pte_mkdirty(entry), vma); 103500501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 103600501b53SJohannes Weiner set_page_private(pages[i], 0); 103771e3aac0SAndrea Arcangeli page_add_new_anon_rmap(pages[i], vma, haddr); 103800501b53SJohannes Weiner mem_cgroup_commit_charge(pages[i], memcg, false); 103900501b53SJohannes Weiner lru_cache_add_active_or_unevictable(pages[i], vma); 104071e3aac0SAndrea Arcangeli pte = pte_offset_map(&_pmd, haddr); 104171e3aac0SAndrea Arcangeli VM_BUG_ON(!pte_none(*pte)); 104271e3aac0SAndrea Arcangeli set_pte_at(mm, haddr, pte, entry); 104371e3aac0SAndrea Arcangeli pte_unmap(pte); 104471e3aac0SAndrea Arcangeli } 104571e3aac0SAndrea Arcangeli kfree(pages); 104671e3aac0SAndrea Arcangeli 104771e3aac0SAndrea Arcangeli smp_wmb(); /* make pte visible before pmd */ 104871e3aac0SAndrea Arcangeli pmd_populate(mm, pmd, pgtable); 104971e3aac0SAndrea Arcangeli page_remove_rmap(page); 1050c4088ebdSKirill A. Shutemov spin_unlock(ptl); 105171e3aac0SAndrea Arcangeli 10522ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 10532ec74c3eSSagi Grimberg 105471e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 105571e3aac0SAndrea Arcangeli put_page(page); 105671e3aac0SAndrea Arcangeli 105771e3aac0SAndrea Arcangeli out: 105871e3aac0SAndrea Arcangeli return ret; 105971e3aac0SAndrea Arcangeli 106071e3aac0SAndrea Arcangeli out_free_pages: 1061c4088ebdSKirill A. Shutemov spin_unlock(ptl); 10622ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1063b9bbfbe3SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 106400501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 106500501b53SJohannes Weiner set_page_private(pages[i], 0); 106600501b53SJohannes Weiner mem_cgroup_cancel_charge(pages[i], memcg); 106771e3aac0SAndrea Arcangeli put_page(pages[i]); 1068b9bbfbe3SAndrea Arcangeli } 106971e3aac0SAndrea Arcangeli kfree(pages); 107071e3aac0SAndrea Arcangeli goto out; 107171e3aac0SAndrea Arcangeli } 107271e3aac0SAndrea Arcangeli 107371e3aac0SAndrea Arcangeli int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 107471e3aac0SAndrea Arcangeli unsigned long address, pmd_t *pmd, pmd_t orig_pmd) 107571e3aac0SAndrea Arcangeli { 1076c4088ebdSKirill A. Shutemov spinlock_t *ptl; 107771e3aac0SAndrea Arcangeli int ret = 0; 107893b4796dSKirill A. Shutemov struct page *page = NULL, *new_page; 107900501b53SJohannes Weiner struct mem_cgroup *memcg; 108071e3aac0SAndrea Arcangeli unsigned long haddr; 10812ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 10822ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 108371e3aac0SAndrea Arcangeli 1084c4088ebdSKirill A. Shutemov ptl = pmd_lockptr(mm, pmd); 108581d1b09cSSasha Levin VM_BUG_ON_VMA(!vma->anon_vma, vma); 108693b4796dSKirill A. Shutemov haddr = address & HPAGE_PMD_MASK; 108793b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 108893b4796dSKirill A. Shutemov goto alloc; 1089c4088ebdSKirill A. Shutemov spin_lock(ptl); 109071e3aac0SAndrea Arcangeli if (unlikely(!pmd_same(*pmd, orig_pmd))) 109171e3aac0SAndrea Arcangeli goto out_unlock; 109271e3aac0SAndrea Arcangeli 109371e3aac0SAndrea Arcangeli page = pmd_page(orig_pmd); 1094309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 109571e3aac0SAndrea Arcangeli if (page_mapcount(page) == 1) { 109671e3aac0SAndrea Arcangeli pmd_t entry; 109771e3aac0SAndrea Arcangeli entry = pmd_mkyoung(orig_pmd); 109871e3aac0SAndrea Arcangeli entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 109971e3aac0SAndrea Arcangeli if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) 1100b113da65SDavid Miller update_mmu_cache_pmd(vma, address, pmd); 110171e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 110271e3aac0SAndrea Arcangeli goto out_unlock; 110371e3aac0SAndrea Arcangeli } 11045338a937SHugh Dickins get_user_huge_page(page); 1105c4088ebdSKirill A. Shutemov spin_unlock(ptl); 110693b4796dSKirill A. Shutemov alloc: 110771e3aac0SAndrea Arcangeli if (transparent_hugepage_enabled(vma) && 1108077fcf11SAneesh Kumar K.V !transparent_hugepage_debug_cow()) { 1109077fcf11SAneesh Kumar K.V gfp_t gfp; 1110077fcf11SAneesh Kumar K.V 1111077fcf11SAneesh Kumar K.V gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); 1112077fcf11SAneesh Kumar K.V new_page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 1113077fcf11SAneesh Kumar K.V } else 111471e3aac0SAndrea Arcangeli new_page = NULL; 111571e3aac0SAndrea Arcangeli 111671e3aac0SAndrea Arcangeli if (unlikely(!new_page)) { 1117eecc1e42SHugh Dickins if (!page) { 1118e9b71ca9SKirill A. Shutemov split_huge_page_pmd(vma, address, pmd); 1119e9b71ca9SKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 112093b4796dSKirill A. Shutemov } else { 112171e3aac0SAndrea Arcangeli ret = do_huge_pmd_wp_page_fallback(mm, vma, address, 112271e3aac0SAndrea Arcangeli pmd, orig_pmd, page, haddr); 11239845cbbdSKirill A. Shutemov if (ret & VM_FAULT_OOM) { 11241f1d06c3SDavid Rientjes split_huge_page(page); 11259845cbbdSKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 11269845cbbdSKirill A. Shutemov } 11275338a937SHugh Dickins put_user_huge_page(page); 112893b4796dSKirill A. Shutemov } 112917766ddeSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK); 113071e3aac0SAndrea Arcangeli goto out; 113171e3aac0SAndrea Arcangeli } 113271e3aac0SAndrea Arcangeli 113300501b53SJohannes Weiner if (unlikely(mem_cgroup_try_charge(new_page, mm, 113400501b53SJohannes Weiner GFP_TRANSHUGE, &memcg))) { 1135b9bbfbe3SAndrea Arcangeli put_page(new_page); 113693b4796dSKirill A. Shutemov if (page) { 11371f1d06c3SDavid Rientjes split_huge_page(page); 11385338a937SHugh Dickins put_user_huge_page(page); 11399845cbbdSKirill A. Shutemov } else 11409845cbbdSKirill A. Shutemov split_huge_page_pmd(vma, address, pmd); 11419845cbbdSKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 114217766ddeSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK); 1143b9bbfbe3SAndrea Arcangeli goto out; 1144b9bbfbe3SAndrea Arcangeli } 1145b9bbfbe3SAndrea Arcangeli 114617766ddeSDavid Rientjes count_vm_event(THP_FAULT_ALLOC); 114717766ddeSDavid Rientjes 1148eecc1e42SHugh Dickins if (!page) 114993b4796dSKirill A. Shutemov clear_huge_page(new_page, haddr, HPAGE_PMD_NR); 115093b4796dSKirill A. Shutemov else 115171e3aac0SAndrea Arcangeli copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 115271e3aac0SAndrea Arcangeli __SetPageUptodate(new_page); 115371e3aac0SAndrea Arcangeli 11542ec74c3eSSagi Grimberg mmun_start = haddr; 11552ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 11562ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 11572ec74c3eSSagi Grimberg 1158c4088ebdSKirill A. Shutemov spin_lock(ptl); 115993b4796dSKirill A. Shutemov if (page) 11605338a937SHugh Dickins put_user_huge_page(page); 1161b9bbfbe3SAndrea Arcangeli if (unlikely(!pmd_same(*pmd, orig_pmd))) { 1162c4088ebdSKirill A. Shutemov spin_unlock(ptl); 116300501b53SJohannes Weiner mem_cgroup_cancel_charge(new_page, memcg); 116471e3aac0SAndrea Arcangeli put_page(new_page); 11652ec74c3eSSagi Grimberg goto out_mn; 1166b9bbfbe3SAndrea Arcangeli } else { 116771e3aac0SAndrea Arcangeli pmd_t entry; 11683122359aSKirill A. Shutemov entry = mk_huge_pmd(new_page, vma->vm_page_prot); 11693122359aSKirill A. Shutemov entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 117034ee645eSJoerg Roedel pmdp_clear_flush_notify(vma, haddr, pmd); 117171e3aac0SAndrea Arcangeli page_add_new_anon_rmap(new_page, vma, haddr); 117200501b53SJohannes Weiner mem_cgroup_commit_charge(new_page, memcg, false); 117300501b53SJohannes Weiner lru_cache_add_active_or_unevictable(new_page, vma); 117471e3aac0SAndrea Arcangeli set_pmd_at(mm, haddr, pmd, entry); 1175b113da65SDavid Miller update_mmu_cache_pmd(vma, address, pmd); 1176eecc1e42SHugh Dickins if (!page) { 117793b4796dSKirill A. Shutemov add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 117897ae1749SKirill A. Shutemov put_huge_zero_page(); 117997ae1749SKirill A. Shutemov } else { 1180309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 118171e3aac0SAndrea Arcangeli page_remove_rmap(page); 118271e3aac0SAndrea Arcangeli put_page(page); 118393b4796dSKirill A. Shutemov } 118471e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 118571e3aac0SAndrea Arcangeli } 1186c4088ebdSKirill A. Shutemov spin_unlock(ptl); 11872ec74c3eSSagi Grimberg out_mn: 11882ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 11892ec74c3eSSagi Grimberg out: 11902ec74c3eSSagi Grimberg return ret; 119171e3aac0SAndrea Arcangeli out_unlock: 1192c4088ebdSKirill A. Shutemov spin_unlock(ptl); 119371e3aac0SAndrea Arcangeli return ret; 119471e3aac0SAndrea Arcangeli } 119571e3aac0SAndrea Arcangeli 1196b676b293SDavid Rientjes struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 119771e3aac0SAndrea Arcangeli unsigned long addr, 119871e3aac0SAndrea Arcangeli pmd_t *pmd, 119971e3aac0SAndrea Arcangeli unsigned int flags) 120071e3aac0SAndrea Arcangeli { 1201b676b293SDavid Rientjes struct mm_struct *mm = vma->vm_mm; 120271e3aac0SAndrea Arcangeli struct page *page = NULL; 120371e3aac0SAndrea Arcangeli 1204c4088ebdSKirill A. Shutemov assert_spin_locked(pmd_lockptr(mm, pmd)); 120571e3aac0SAndrea Arcangeli 120671e3aac0SAndrea Arcangeli if (flags & FOLL_WRITE && !pmd_write(*pmd)) 120771e3aac0SAndrea Arcangeli goto out; 120871e3aac0SAndrea Arcangeli 120985facf25SKirill A. Shutemov /* Avoid dumping huge zero page */ 121085facf25SKirill A. Shutemov if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 121185facf25SKirill A. Shutemov return ERR_PTR(-EFAULT); 121285facf25SKirill A. Shutemov 12132b4847e7SMel Gorman /* Full NUMA hinting faults to serialise migration in fault paths */ 12148a0516edSMel Gorman if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 12152b4847e7SMel Gorman goto out; 12162b4847e7SMel Gorman 121771e3aac0SAndrea Arcangeli page = pmd_page(*pmd); 1218309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 121971e3aac0SAndrea Arcangeli if (flags & FOLL_TOUCH) { 122071e3aac0SAndrea Arcangeli pmd_t _pmd; 122171e3aac0SAndrea Arcangeli /* 122271e3aac0SAndrea Arcangeli * We should set the dirty bit only for FOLL_WRITE but 122371e3aac0SAndrea Arcangeli * for now the dirty bit in the pmd is meaningless. 122471e3aac0SAndrea Arcangeli * And if the dirty bit will become meaningful and 122571e3aac0SAndrea Arcangeli * we'll only set it with FOLL_WRITE, an atomic 122671e3aac0SAndrea Arcangeli * set_bit will be required on the pmd to set the 122771e3aac0SAndrea Arcangeli * young bit, instead of the current set_pmd_at. 122871e3aac0SAndrea Arcangeli */ 122971e3aac0SAndrea Arcangeli _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 12308663890aSAneesh Kumar K.V if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 12318663890aSAneesh Kumar K.V pmd, _pmd, 1)) 12328663890aSAneesh Kumar K.V update_mmu_cache_pmd(vma, addr, pmd); 123371e3aac0SAndrea Arcangeli } 1234b676b293SDavid Rientjes if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1235b676b293SDavid Rientjes if (page->mapping && trylock_page(page)) { 1236b676b293SDavid Rientjes lru_add_drain(); 1237b676b293SDavid Rientjes if (page->mapping) 1238b676b293SDavid Rientjes mlock_vma_page(page); 1239b676b293SDavid Rientjes unlock_page(page); 1240b676b293SDavid Rientjes } 1241b676b293SDavid Rientjes } 124271e3aac0SAndrea Arcangeli page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1243309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page), page); 124471e3aac0SAndrea Arcangeli if (flags & FOLL_GET) 124570b50f94SAndrea Arcangeli get_page_foll(page); 124671e3aac0SAndrea Arcangeli 124771e3aac0SAndrea Arcangeli out: 124871e3aac0SAndrea Arcangeli return page; 124971e3aac0SAndrea Arcangeli } 125071e3aac0SAndrea Arcangeli 1251d10e63f2SMel Gorman /* NUMA hinting page fault entry point for trans huge pmds */ 12524daae3b4SMel Gorman int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, 12534daae3b4SMel Gorman unsigned long addr, pmd_t pmd, pmd_t *pmdp) 1254d10e63f2SMel Gorman { 1255c4088ebdSKirill A. Shutemov spinlock_t *ptl; 1256b8916634SMel Gorman struct anon_vma *anon_vma = NULL; 1257b32967ffSMel Gorman struct page *page; 1258d10e63f2SMel Gorman unsigned long haddr = addr & HPAGE_PMD_MASK; 12598191acbdSMel Gorman int page_nid = -1, this_nid = numa_node_id(); 126090572890SPeter Zijlstra int target_nid, last_cpupid = -1; 12618191acbdSMel Gorman bool page_locked; 12628191acbdSMel Gorman bool migrated = false; 12636688cc05SPeter Zijlstra int flags = 0; 1264d10e63f2SMel Gorman 1265c4088ebdSKirill A. Shutemov ptl = pmd_lock(mm, pmdp); 1266d10e63f2SMel Gorman if (unlikely(!pmd_same(pmd, *pmdp))) 1267d10e63f2SMel Gorman goto out_unlock; 1268d10e63f2SMel Gorman 1269de466bd6SMel Gorman /* 1270de466bd6SMel Gorman * If there are potential migrations, wait for completion and retry 1271de466bd6SMel Gorman * without disrupting NUMA hinting information. Do not relock and 1272de466bd6SMel Gorman * check_same as the page may no longer be mapped. 1273de466bd6SMel Gorman */ 1274de466bd6SMel Gorman if (unlikely(pmd_trans_migrating(*pmdp))) { 12755d833062SMel Gorman page = pmd_page(*pmdp); 1276de466bd6SMel Gorman spin_unlock(ptl); 12775d833062SMel Gorman wait_on_page_locked(page); 1278de466bd6SMel Gorman goto out; 1279de466bd6SMel Gorman } 1280de466bd6SMel Gorman 1281d10e63f2SMel Gorman page = pmd_page(pmd); 1282a1a46184SMel Gorman BUG_ON(is_huge_zero_page(page)); 12838191acbdSMel Gorman page_nid = page_to_nid(page); 128490572890SPeter Zijlstra last_cpupid = page_cpupid_last(page); 128503c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS); 128604bb2f94SRik van Riel if (page_nid == this_nid) { 128703c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 128804bb2f94SRik van Riel flags |= TNF_FAULT_LOCAL; 128904bb2f94SRik van Riel } 12904daae3b4SMel Gorman 1291ff9042b1SMel Gorman /* 12926688cc05SPeter Zijlstra * Avoid grouping on DSO/COW pages in specific and RO pages 12936688cc05SPeter Zijlstra * in general, RO pages shouldn't hurt as much anyway since 12946688cc05SPeter Zijlstra * they can be in shared cache state. 12956688cc05SPeter Zijlstra */ 12966688cc05SPeter Zijlstra if (!pmd_write(pmd)) 12976688cc05SPeter Zijlstra flags |= TNF_NO_GROUP; 12986688cc05SPeter Zijlstra 12996688cc05SPeter Zijlstra /* 1300ff9042b1SMel Gorman * Acquire the page lock to serialise THP migrations but avoid dropping 1301ff9042b1SMel Gorman * page_table_lock if at all possible 1302ff9042b1SMel Gorman */ 1303b8916634SMel Gorman page_locked = trylock_page(page); 1304b8916634SMel Gorman target_nid = mpol_misplaced(page, vma, haddr); 1305b8916634SMel Gorman if (target_nid == -1) { 1306b8916634SMel Gorman /* If the page was locked, there are no parallel migrations */ 1307a54a407fSMel Gorman if (page_locked) 1308b8916634SMel Gorman goto clear_pmdnuma; 13092b4847e7SMel Gorman } 1310cbee9f88SPeter Zijlstra 1311de466bd6SMel Gorman /* Migration could have started since the pmd_trans_migrating check */ 13122b4847e7SMel Gorman if (!page_locked) { 1313c4088ebdSKirill A. Shutemov spin_unlock(ptl); 1314b8916634SMel Gorman wait_on_page_locked(page); 1315a54a407fSMel Gorman page_nid = -1; 1316b8916634SMel Gorman goto out; 1317b8916634SMel Gorman } 1318b8916634SMel Gorman 13192b4847e7SMel Gorman /* 13202b4847e7SMel Gorman * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 13212b4847e7SMel Gorman * to serialises splits 13222b4847e7SMel Gorman */ 1323b8916634SMel Gorman get_page(page); 1324c4088ebdSKirill A. Shutemov spin_unlock(ptl); 1325b8916634SMel Gorman anon_vma = page_lock_anon_vma_read(page); 1326b32967ffSMel Gorman 1327c69307d5SPeter Zijlstra /* Confirm the PMD did not change while page_table_lock was released */ 1328c4088ebdSKirill A. Shutemov spin_lock(ptl); 1329b32967ffSMel Gorman if (unlikely(!pmd_same(pmd, *pmdp))) { 1330b32967ffSMel Gorman unlock_page(page); 1331b32967ffSMel Gorman put_page(page); 1332a54a407fSMel Gorman page_nid = -1; 1333b32967ffSMel Gorman goto out_unlock; 1334b32967ffSMel Gorman } 1335ff9042b1SMel Gorman 1336c3a489caSMel Gorman /* Bail if we fail to protect against THP splits for any reason */ 1337c3a489caSMel Gorman if (unlikely(!anon_vma)) { 1338c3a489caSMel Gorman put_page(page); 1339c3a489caSMel Gorman page_nid = -1; 1340c3a489caSMel Gorman goto clear_pmdnuma; 1341c3a489caSMel Gorman } 1342c3a489caSMel Gorman 1343a54a407fSMel Gorman /* 1344a54a407fSMel Gorman * Migrate the THP to the requested node, returns with page unlocked 13458a0516edSMel Gorman * and access rights restored. 1346a54a407fSMel Gorman */ 1347c4088ebdSKirill A. Shutemov spin_unlock(ptl); 1348b32967ffSMel Gorman migrated = migrate_misplaced_transhuge_page(mm, vma, 1349340ef390SHugh Dickins pmdp, pmd, addr, page, target_nid); 13506688cc05SPeter Zijlstra if (migrated) { 13516688cc05SPeter Zijlstra flags |= TNF_MIGRATED; 13528191acbdSMel Gorman page_nid = target_nid; 13536688cc05SPeter Zijlstra } 1354b32967ffSMel Gorman 13558191acbdSMel Gorman goto out; 13564daae3b4SMel Gorman clear_pmdnuma: 1357a54a407fSMel Gorman BUG_ON(!PageLocked(page)); 13584d942466SMel Gorman pmd = pmd_modify(pmd, vma->vm_page_prot); 1359d10e63f2SMel Gorman set_pmd_at(mm, haddr, pmdp, pmd); 1360d10e63f2SMel Gorman update_mmu_cache_pmd(vma, addr, pmdp); 1361a54a407fSMel Gorman unlock_page(page); 1362d10e63f2SMel Gorman out_unlock: 1363c4088ebdSKirill A. Shutemov spin_unlock(ptl); 1364b8916634SMel Gorman 1365b8916634SMel Gorman out: 1366b8916634SMel Gorman if (anon_vma) 1367b8916634SMel Gorman page_unlock_anon_vma_read(anon_vma); 1368b8916634SMel Gorman 13698191acbdSMel Gorman if (page_nid != -1) 13706688cc05SPeter Zijlstra task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); 13718191acbdSMel Gorman 1372d10e63f2SMel Gorman return 0; 1373d10e63f2SMel Gorman } 1374d10e63f2SMel Gorman 137571e3aac0SAndrea Arcangeli int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1376f21760b1SShaohua Li pmd_t *pmd, unsigned long addr) 137771e3aac0SAndrea Arcangeli { 1378bf929152SKirill A. Shutemov spinlock_t *ptl; 137971e3aac0SAndrea Arcangeli int ret = 0; 138071e3aac0SAndrea Arcangeli 1381bf929152SKirill A. Shutemov if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { 138271e3aac0SAndrea Arcangeli struct page *page; 138371e3aac0SAndrea Arcangeli pgtable_t pgtable; 1384f5c8ad47SDavid Miller pmd_t orig_pmd; 1385a6bf2bb0SAneesh Kumar K.V /* 1386a6bf2bb0SAneesh Kumar K.V * For architectures like ppc64 we look at deposited pgtable 1387a6bf2bb0SAneesh Kumar K.V * when calling pmdp_get_and_clear. So do the 1388a6bf2bb0SAneesh Kumar K.V * pgtable_trans_huge_withdraw after finishing pmdp related 1389a6bf2bb0SAneesh Kumar K.V * operations. 1390a6bf2bb0SAneesh Kumar K.V */ 1391fcbe08d6SMartin Schwidefsky orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd, 1392fcbe08d6SMartin Schwidefsky tlb->fullmm); 1393f21760b1SShaohua Li tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1394a6bf2bb0SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd); 1395479f0abbSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) { 1396e1f56c89SKirill A. Shutemov atomic_long_dec(&tlb->mm->nr_ptes); 1397bf929152SKirill A. Shutemov spin_unlock(ptl); 139897ae1749SKirill A. Shutemov put_huge_zero_page(); 1399479f0abbSKirill A. Shutemov } else { 1400479f0abbSKirill A. Shutemov page = pmd_page(orig_pmd); 140171e3aac0SAndrea Arcangeli page_remove_rmap(page); 1402309381feSSasha Levin VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 140371e3aac0SAndrea Arcangeli add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1404309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1405e1f56c89SKirill A. Shutemov atomic_long_dec(&tlb->mm->nr_ptes); 1406bf929152SKirill A. Shutemov spin_unlock(ptl); 140771e3aac0SAndrea Arcangeli tlb_remove_page(tlb, page); 1408479f0abbSKirill A. Shutemov } 140971e3aac0SAndrea Arcangeli pte_free(tlb->mm, pgtable); 141071e3aac0SAndrea Arcangeli ret = 1; 141171e3aac0SAndrea Arcangeli } 141271e3aac0SAndrea Arcangeli return ret; 141371e3aac0SAndrea Arcangeli } 141471e3aac0SAndrea Arcangeli 141537a1c49aSAndrea Arcangeli int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, 141637a1c49aSAndrea Arcangeli unsigned long old_addr, 141737a1c49aSAndrea Arcangeli unsigned long new_addr, unsigned long old_end, 141837a1c49aSAndrea Arcangeli pmd_t *old_pmd, pmd_t *new_pmd) 141937a1c49aSAndrea Arcangeli { 1420bf929152SKirill A. Shutemov spinlock_t *old_ptl, *new_ptl; 142137a1c49aSAndrea Arcangeli int ret = 0; 142237a1c49aSAndrea Arcangeli pmd_t pmd; 142337a1c49aSAndrea Arcangeli 142437a1c49aSAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 142537a1c49aSAndrea Arcangeli 142637a1c49aSAndrea Arcangeli if ((old_addr & ~HPAGE_PMD_MASK) || 142737a1c49aSAndrea Arcangeli (new_addr & ~HPAGE_PMD_MASK) || 142837a1c49aSAndrea Arcangeli old_end - old_addr < HPAGE_PMD_SIZE || 142937a1c49aSAndrea Arcangeli (new_vma->vm_flags & VM_NOHUGEPAGE)) 143037a1c49aSAndrea Arcangeli goto out; 143137a1c49aSAndrea Arcangeli 143237a1c49aSAndrea Arcangeli /* 143337a1c49aSAndrea Arcangeli * The destination pmd shouldn't be established, free_pgtables() 143437a1c49aSAndrea Arcangeli * should have release it. 143537a1c49aSAndrea Arcangeli */ 143637a1c49aSAndrea Arcangeli if (WARN_ON(!pmd_none(*new_pmd))) { 143737a1c49aSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*new_pmd)); 143837a1c49aSAndrea Arcangeli goto out; 143937a1c49aSAndrea Arcangeli } 144037a1c49aSAndrea Arcangeli 1441bf929152SKirill A. Shutemov /* 1442bf929152SKirill A. Shutemov * We don't have to worry about the ordering of src and dst 1443bf929152SKirill A. Shutemov * ptlocks because exclusive mmap_sem prevents deadlock. 1444bf929152SKirill A. Shutemov */ 1445bf929152SKirill A. Shutemov ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl); 1446025c5b24SNaoya Horiguchi if (ret == 1) { 1447bf929152SKirill A. Shutemov new_ptl = pmd_lockptr(mm, new_pmd); 1448bf929152SKirill A. Shutemov if (new_ptl != old_ptl) 1449bf929152SKirill A. Shutemov spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 145037a1c49aSAndrea Arcangeli pmd = pmdp_get_and_clear(mm, old_addr, old_pmd); 145137a1c49aSAndrea Arcangeli VM_BUG_ON(!pmd_none(*new_pmd)); 14523592806cSKirill A. Shutemov 1453b3084f4dSAneesh Kumar K.V if (pmd_move_must_withdraw(new_ptl, old_ptl)) { 1454b3084f4dSAneesh Kumar K.V pgtable_t pgtable; 14553592806cSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 14563592806cSKirill A. Shutemov pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 14573592806cSKirill A. Shutemov } 1458b3084f4dSAneesh Kumar K.V set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); 1459b3084f4dSAneesh Kumar K.V if (new_ptl != old_ptl) 1460b3084f4dSAneesh Kumar K.V spin_unlock(new_ptl); 1461bf929152SKirill A. Shutemov spin_unlock(old_ptl); 146237a1c49aSAndrea Arcangeli } 146337a1c49aSAndrea Arcangeli out: 146437a1c49aSAndrea Arcangeli return ret; 146537a1c49aSAndrea Arcangeli } 146637a1c49aSAndrea Arcangeli 1467f123d74aSMel Gorman /* 1468f123d74aSMel Gorman * Returns 1469f123d74aSMel Gorman * - 0 if PMD could not be locked 1470f123d74aSMel Gorman * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1471f123d74aSMel Gorman * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1472f123d74aSMel Gorman */ 1473cd7548abSJohannes Weiner int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1474*e944fd67SMel Gorman unsigned long addr, pgprot_t newprot, int prot_numa) 1475cd7548abSJohannes Weiner { 1476cd7548abSJohannes Weiner struct mm_struct *mm = vma->vm_mm; 1477bf929152SKirill A. Shutemov spinlock_t *ptl; 1478cd7548abSJohannes Weiner int ret = 0; 1479cd7548abSJohannes Weiner 1480bf929152SKirill A. Shutemov if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { 1481cd7548abSJohannes Weiner pmd_t entry; 1482*e944fd67SMel Gorman 1483*e944fd67SMel Gorman /* 1484*e944fd67SMel Gorman * Avoid trapping faults against the zero page. The read-only 1485*e944fd67SMel Gorman * data is likely to be read-cached on the local CPU and 1486*e944fd67SMel Gorman * local/remote hits to the zero page are not interesting. 1487*e944fd67SMel Gorman */ 1488*e944fd67SMel Gorman if (prot_numa && is_huge_zero_pmd(*pmd)) { 1489*e944fd67SMel Gorman spin_unlock(ptl); 1490*e944fd67SMel Gorman return 0; 1491*e944fd67SMel Gorman } 1492*e944fd67SMel Gorman 1493f123d74aSMel Gorman ret = 1; 149434ee645eSJoerg Roedel entry = pmdp_get_and_clear_notify(mm, addr, pmd); 1495cd7548abSJohannes Weiner entry = pmd_modify(entry, newprot); 1496f123d74aSMel Gorman ret = HPAGE_PMD_NR; 149756eecdb9SAneesh Kumar K.V set_pmd_at(mm, addr, pmd, entry); 1498a4f1de17SHugh Dickins BUG_ON(pmd_write(entry)); 1499bf929152SKirill A. Shutemov spin_unlock(ptl); 1500cd7548abSJohannes Weiner } 1501cd7548abSJohannes Weiner 1502cd7548abSJohannes Weiner return ret; 1503cd7548abSJohannes Weiner } 1504cd7548abSJohannes Weiner 1505025c5b24SNaoya Horiguchi /* 1506025c5b24SNaoya Horiguchi * Returns 1 if a given pmd maps a stable (not under splitting) thp. 1507025c5b24SNaoya Horiguchi * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. 1508025c5b24SNaoya Horiguchi * 1509025c5b24SNaoya Horiguchi * Note that if it returns 1, this routine returns without unlocking page 1510025c5b24SNaoya Horiguchi * table locks. So callers must unlock them. 1511025c5b24SNaoya Horiguchi */ 1512bf929152SKirill A. Shutemov int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, 1513bf929152SKirill A. Shutemov spinlock_t **ptl) 1514025c5b24SNaoya Horiguchi { 1515bf929152SKirill A. Shutemov *ptl = pmd_lock(vma->vm_mm, pmd); 1516025c5b24SNaoya Horiguchi if (likely(pmd_trans_huge(*pmd))) { 1517025c5b24SNaoya Horiguchi if (unlikely(pmd_trans_splitting(*pmd))) { 1518bf929152SKirill A. Shutemov spin_unlock(*ptl); 1519025c5b24SNaoya Horiguchi wait_split_huge_page(vma->anon_vma, pmd); 1520025c5b24SNaoya Horiguchi return -1; 1521025c5b24SNaoya Horiguchi } else { 1522025c5b24SNaoya Horiguchi /* Thp mapped by 'pmd' is stable, so we can 1523025c5b24SNaoya Horiguchi * handle it as it is. */ 1524025c5b24SNaoya Horiguchi return 1; 1525025c5b24SNaoya Horiguchi } 1526025c5b24SNaoya Horiguchi } 1527bf929152SKirill A. Shutemov spin_unlock(*ptl); 1528025c5b24SNaoya Horiguchi return 0; 1529025c5b24SNaoya Horiguchi } 1530025c5b24SNaoya Horiguchi 1531117b0791SKirill A. Shutemov /* 1532117b0791SKirill A. Shutemov * This function returns whether a given @page is mapped onto the @address 1533117b0791SKirill A. Shutemov * in the virtual space of @mm. 1534117b0791SKirill A. Shutemov * 1535117b0791SKirill A. Shutemov * When it's true, this function returns *pmd with holding the page table lock 1536117b0791SKirill A. Shutemov * and passing it back to the caller via @ptl. 1537117b0791SKirill A. Shutemov * If it's false, returns NULL without holding the page table lock. 1538117b0791SKirill A. Shutemov */ 153971e3aac0SAndrea Arcangeli pmd_t *page_check_address_pmd(struct page *page, 154071e3aac0SAndrea Arcangeli struct mm_struct *mm, 154171e3aac0SAndrea Arcangeli unsigned long address, 1542117b0791SKirill A. Shutemov enum page_check_address_pmd_flag flag, 1543117b0791SKirill A. Shutemov spinlock_t **ptl) 154471e3aac0SAndrea Arcangeli { 1545b5a8cad3SKirill A. Shutemov pgd_t *pgd; 1546b5a8cad3SKirill A. Shutemov pud_t *pud; 1547117b0791SKirill A. Shutemov pmd_t *pmd; 154871e3aac0SAndrea Arcangeli 154971e3aac0SAndrea Arcangeli if (address & ~HPAGE_PMD_MASK) 1550117b0791SKirill A. Shutemov return NULL; 155171e3aac0SAndrea Arcangeli 1552b5a8cad3SKirill A. Shutemov pgd = pgd_offset(mm, address); 1553b5a8cad3SKirill A. Shutemov if (!pgd_present(*pgd)) 1554117b0791SKirill A. Shutemov return NULL; 1555b5a8cad3SKirill A. Shutemov pud = pud_offset(pgd, address); 1556b5a8cad3SKirill A. Shutemov if (!pud_present(*pud)) 1557b5a8cad3SKirill A. Shutemov return NULL; 1558b5a8cad3SKirill A. Shutemov pmd = pmd_offset(pud, address); 1559b5a8cad3SKirill A. Shutemov 1560117b0791SKirill A. Shutemov *ptl = pmd_lock(mm, pmd); 1561b5a8cad3SKirill A. Shutemov if (!pmd_present(*pmd)) 1562117b0791SKirill A. Shutemov goto unlock; 156371e3aac0SAndrea Arcangeli if (pmd_page(*pmd) != page) 1564117b0791SKirill A. Shutemov goto unlock; 156594fcc585SAndrea Arcangeli /* 156694fcc585SAndrea Arcangeli * split_vma() may create temporary aliased mappings. There is 156794fcc585SAndrea Arcangeli * no risk as long as all huge pmd are found and have their 156894fcc585SAndrea Arcangeli * splitting bit set before __split_huge_page_refcount 156994fcc585SAndrea Arcangeli * runs. Finding the same huge pmd more than once during the 157094fcc585SAndrea Arcangeli * same rmap walk is not a problem. 157194fcc585SAndrea Arcangeli */ 157294fcc585SAndrea Arcangeli if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && 157394fcc585SAndrea Arcangeli pmd_trans_splitting(*pmd)) 1574117b0791SKirill A. Shutemov goto unlock; 157571e3aac0SAndrea Arcangeli if (pmd_trans_huge(*pmd)) { 157671e3aac0SAndrea Arcangeli VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && 157771e3aac0SAndrea Arcangeli !pmd_trans_splitting(*pmd)); 1578117b0791SKirill A. Shutemov return pmd; 157971e3aac0SAndrea Arcangeli } 1580117b0791SKirill A. Shutemov unlock: 1581117b0791SKirill A. Shutemov spin_unlock(*ptl); 1582117b0791SKirill A. Shutemov return NULL; 158371e3aac0SAndrea Arcangeli } 158471e3aac0SAndrea Arcangeli 158571e3aac0SAndrea Arcangeli static int __split_huge_page_splitting(struct page *page, 158671e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 158771e3aac0SAndrea Arcangeli unsigned long address) 158871e3aac0SAndrea Arcangeli { 158971e3aac0SAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 1590117b0791SKirill A. Shutemov spinlock_t *ptl; 159171e3aac0SAndrea Arcangeli pmd_t *pmd; 159271e3aac0SAndrea Arcangeli int ret = 0; 15932ec74c3eSSagi Grimberg /* For mmu_notifiers */ 15942ec74c3eSSagi Grimberg const unsigned long mmun_start = address; 15952ec74c3eSSagi Grimberg const unsigned long mmun_end = address + HPAGE_PMD_SIZE; 159671e3aac0SAndrea Arcangeli 15972ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 159871e3aac0SAndrea Arcangeli pmd = page_check_address_pmd(page, mm, address, 1599117b0791SKirill A. Shutemov PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl); 160071e3aac0SAndrea Arcangeli if (pmd) { 160171e3aac0SAndrea Arcangeli /* 160271e3aac0SAndrea Arcangeli * We can't temporarily set the pmd to null in order 160371e3aac0SAndrea Arcangeli * to split it, the pmd must remain marked huge at all 160471e3aac0SAndrea Arcangeli * times or the VM won't take the pmd_trans_huge paths 16055a505085SIngo Molnar * and it won't wait on the anon_vma->root->rwsem to 160671e3aac0SAndrea Arcangeli * serialize against split_huge_page*. 160771e3aac0SAndrea Arcangeli */ 16082ec74c3eSSagi Grimberg pmdp_splitting_flush(vma, address, pmd); 160934ee645eSJoerg Roedel 161071e3aac0SAndrea Arcangeli ret = 1; 1611117b0791SKirill A. Shutemov spin_unlock(ptl); 161271e3aac0SAndrea Arcangeli } 16132ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 161471e3aac0SAndrea Arcangeli 161571e3aac0SAndrea Arcangeli return ret; 161671e3aac0SAndrea Arcangeli } 161771e3aac0SAndrea Arcangeli 16185bc7b8acSShaohua Li static void __split_huge_page_refcount(struct page *page, 16195bc7b8acSShaohua Li struct list_head *list) 162071e3aac0SAndrea Arcangeli { 162171e3aac0SAndrea Arcangeli int i; 162271e3aac0SAndrea Arcangeli struct zone *zone = page_zone(page); 1623fa9add64SHugh Dickins struct lruvec *lruvec; 162470b50f94SAndrea Arcangeli int tail_count = 0; 162571e3aac0SAndrea Arcangeli 162671e3aac0SAndrea Arcangeli /* prevent PageLRU to go away from under us, and freeze lru stats */ 162771e3aac0SAndrea Arcangeli spin_lock_irq(&zone->lru_lock); 1628fa9add64SHugh Dickins lruvec = mem_cgroup_page_lruvec(page, zone); 1629fa9add64SHugh Dickins 163071e3aac0SAndrea Arcangeli compound_lock(page); 1631e94c8a9cSKAMEZAWA Hiroyuki /* complete memcg works before add pages to LRU */ 1632e94c8a9cSKAMEZAWA Hiroyuki mem_cgroup_split_huge_fixup(page); 163371e3aac0SAndrea Arcangeli 163445676885SShaohua Li for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 163571e3aac0SAndrea Arcangeli struct page *page_tail = page + i; 163671e3aac0SAndrea Arcangeli 163770b50f94SAndrea Arcangeli /* tail_page->_mapcount cannot change */ 163870b50f94SAndrea Arcangeli BUG_ON(page_mapcount(page_tail) < 0); 163970b50f94SAndrea Arcangeli tail_count += page_mapcount(page_tail); 164070b50f94SAndrea Arcangeli /* check for overflow */ 164170b50f94SAndrea Arcangeli BUG_ON(tail_count < 0); 164270b50f94SAndrea Arcangeli BUG_ON(atomic_read(&page_tail->_count) != 0); 164370b50f94SAndrea Arcangeli /* 164470b50f94SAndrea Arcangeli * tail_page->_count is zero and not changing from 164570b50f94SAndrea Arcangeli * under us. But get_page_unless_zero() may be running 164670b50f94SAndrea Arcangeli * from under us on the tail_page. If we used 164770b50f94SAndrea Arcangeli * atomic_set() below instead of atomic_add(), we 164870b50f94SAndrea Arcangeli * would then run atomic_set() concurrently with 164970b50f94SAndrea Arcangeli * get_page_unless_zero(), and atomic_set() is 165070b50f94SAndrea Arcangeli * implemented in C not using locked ops. spin_unlock 165170b50f94SAndrea Arcangeli * on x86 sometime uses locked ops because of PPro 165270b50f94SAndrea Arcangeli * errata 66, 92, so unless somebody can guarantee 165370b50f94SAndrea Arcangeli * atomic_set() here would be safe on all archs (and 165470b50f94SAndrea Arcangeli * not only on x86), it's safer to use atomic_add(). 165570b50f94SAndrea Arcangeli */ 165670b50f94SAndrea Arcangeli atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, 165770b50f94SAndrea Arcangeli &page_tail->_count); 165871e3aac0SAndrea Arcangeli 165971e3aac0SAndrea Arcangeli /* after clearing PageTail the gup refcount can be released */ 16603a79d52aSWaiman Long smp_mb__after_atomic(); 166171e3aac0SAndrea Arcangeli 1662a6d30dddSJin Dongming /* 1663a6d30dddSJin Dongming * retain hwpoison flag of the poisoned tail page: 1664a6d30dddSJin Dongming * fix for the unsuitable process killed on Guest Machine(KVM) 1665a6d30dddSJin Dongming * by the memory-failure. 1666a6d30dddSJin Dongming */ 1667a6d30dddSJin Dongming page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON; 166871e3aac0SAndrea Arcangeli page_tail->flags |= (page->flags & 166971e3aac0SAndrea Arcangeli ((1L << PG_referenced) | 167071e3aac0SAndrea Arcangeli (1L << PG_swapbacked) | 167171e3aac0SAndrea Arcangeli (1L << PG_mlocked) | 1672e180cf80SKirill A. Shutemov (1L << PG_uptodate) | 1673e180cf80SKirill A. Shutemov (1L << PG_active) | 1674e180cf80SKirill A. Shutemov (1L << PG_unevictable))); 167571e3aac0SAndrea Arcangeli page_tail->flags |= (1L << PG_dirty); 167671e3aac0SAndrea Arcangeli 167770b50f94SAndrea Arcangeli /* clear PageTail before overwriting first_page */ 167871e3aac0SAndrea Arcangeli smp_wmb(); 167971e3aac0SAndrea Arcangeli 168071e3aac0SAndrea Arcangeli /* 168171e3aac0SAndrea Arcangeli * __split_huge_page_splitting() already set the 168271e3aac0SAndrea Arcangeli * splitting bit in all pmd that could map this 168371e3aac0SAndrea Arcangeli * hugepage, that will ensure no CPU can alter the 168471e3aac0SAndrea Arcangeli * mapcount on the head page. The mapcount is only 168571e3aac0SAndrea Arcangeli * accounted in the head page and it has to be 168671e3aac0SAndrea Arcangeli * transferred to all tail pages in the below code. So 168771e3aac0SAndrea Arcangeli * for this code to be safe, the split the mapcount 168871e3aac0SAndrea Arcangeli * can't change. But that doesn't mean userland can't 168971e3aac0SAndrea Arcangeli * keep changing and reading the page contents while 169071e3aac0SAndrea Arcangeli * we transfer the mapcount, so the pmd splitting 169171e3aac0SAndrea Arcangeli * status is achieved setting a reserved bit in the 169271e3aac0SAndrea Arcangeli * pmd, not by clearing the present bit. 169371e3aac0SAndrea Arcangeli */ 169471e3aac0SAndrea Arcangeli page_tail->_mapcount = page->_mapcount; 169571e3aac0SAndrea Arcangeli 169671e3aac0SAndrea Arcangeli BUG_ON(page_tail->mapping); 169771e3aac0SAndrea Arcangeli page_tail->mapping = page->mapping; 169871e3aac0SAndrea Arcangeli 169945676885SShaohua Li page_tail->index = page->index + i; 170090572890SPeter Zijlstra page_cpupid_xchg_last(page_tail, page_cpupid_last(page)); 170171e3aac0SAndrea Arcangeli 170271e3aac0SAndrea Arcangeli BUG_ON(!PageAnon(page_tail)); 170371e3aac0SAndrea Arcangeli BUG_ON(!PageUptodate(page_tail)); 170471e3aac0SAndrea Arcangeli BUG_ON(!PageDirty(page_tail)); 170571e3aac0SAndrea Arcangeli BUG_ON(!PageSwapBacked(page_tail)); 170671e3aac0SAndrea Arcangeli 17075bc7b8acSShaohua Li lru_add_page_tail(page, page_tail, lruvec, list); 170871e3aac0SAndrea Arcangeli } 170970b50f94SAndrea Arcangeli atomic_sub(tail_count, &page->_count); 171070b50f94SAndrea Arcangeli BUG_ON(atomic_read(&page->_count) <= 0); 171171e3aac0SAndrea Arcangeli 1712fa9add64SHugh Dickins __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); 171379134171SAndrea Arcangeli 171471e3aac0SAndrea Arcangeli ClearPageCompound(page); 171571e3aac0SAndrea Arcangeli compound_unlock(page); 171671e3aac0SAndrea Arcangeli spin_unlock_irq(&zone->lru_lock); 171771e3aac0SAndrea Arcangeli 171871e3aac0SAndrea Arcangeli for (i = 1; i < HPAGE_PMD_NR; i++) { 171971e3aac0SAndrea Arcangeli struct page *page_tail = page + i; 172071e3aac0SAndrea Arcangeli BUG_ON(page_count(page_tail) <= 0); 172171e3aac0SAndrea Arcangeli /* 172271e3aac0SAndrea Arcangeli * Tail pages may be freed if there wasn't any mapping 172371e3aac0SAndrea Arcangeli * like if add_to_swap() is running on a lru page that 172471e3aac0SAndrea Arcangeli * had its mapping zapped. And freeing these pages 172571e3aac0SAndrea Arcangeli * requires taking the lru_lock so we do the put_page 172671e3aac0SAndrea Arcangeli * of the tail pages after the split is complete. 172771e3aac0SAndrea Arcangeli */ 172871e3aac0SAndrea Arcangeli put_page(page_tail); 172971e3aac0SAndrea Arcangeli } 173071e3aac0SAndrea Arcangeli 173171e3aac0SAndrea Arcangeli /* 173271e3aac0SAndrea Arcangeli * Only the head page (now become a regular page) is required 173371e3aac0SAndrea Arcangeli * to be pinned by the caller. 173471e3aac0SAndrea Arcangeli */ 173571e3aac0SAndrea Arcangeli BUG_ON(page_count(page) <= 0); 173671e3aac0SAndrea Arcangeli } 173771e3aac0SAndrea Arcangeli 173871e3aac0SAndrea Arcangeli static int __split_huge_page_map(struct page *page, 173971e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 174071e3aac0SAndrea Arcangeli unsigned long address) 174171e3aac0SAndrea Arcangeli { 174271e3aac0SAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 1743117b0791SKirill A. Shutemov spinlock_t *ptl; 174471e3aac0SAndrea Arcangeli pmd_t *pmd, _pmd; 174571e3aac0SAndrea Arcangeli int ret = 0, i; 174671e3aac0SAndrea Arcangeli pgtable_t pgtable; 174771e3aac0SAndrea Arcangeli unsigned long haddr; 174871e3aac0SAndrea Arcangeli 174971e3aac0SAndrea Arcangeli pmd = page_check_address_pmd(page, mm, address, 1750117b0791SKirill A. Shutemov PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl); 175171e3aac0SAndrea Arcangeli if (pmd) { 17526b0b50b0SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(mm, pmd); 175371e3aac0SAndrea Arcangeli pmd_populate(mm, &_pmd, pgtable); 1754f8303c25SWaiman Long if (pmd_write(*pmd)) 1755f8303c25SWaiman Long BUG_ON(page_mapcount(page) != 1); 175671e3aac0SAndrea Arcangeli 1757e3ebcf64SGerald Schaefer haddr = address; 1758e3ebcf64SGerald Schaefer for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 175971e3aac0SAndrea Arcangeli pte_t *pte, entry; 176071e3aac0SAndrea Arcangeli BUG_ON(PageCompound(page+i)); 1761abc40bd2SMel Gorman /* 17628a0516edSMel Gorman * Note that NUMA hinting access restrictions are not 17638a0516edSMel Gorman * transferred to avoid any possibility of altering 17648a0516edSMel Gorman * permissions across VMAs. 1765abc40bd2SMel Gorman */ 176671e3aac0SAndrea Arcangeli entry = mk_pte(page + i, vma->vm_page_prot); 176771e3aac0SAndrea Arcangeli entry = maybe_mkwrite(pte_mkdirty(entry), vma); 176871e3aac0SAndrea Arcangeli if (!pmd_write(*pmd)) 176971e3aac0SAndrea Arcangeli entry = pte_wrprotect(entry); 177071e3aac0SAndrea Arcangeli if (!pmd_young(*pmd)) 177171e3aac0SAndrea Arcangeli entry = pte_mkold(entry); 177271e3aac0SAndrea Arcangeli pte = pte_offset_map(&_pmd, haddr); 177371e3aac0SAndrea Arcangeli BUG_ON(!pte_none(*pte)); 177471e3aac0SAndrea Arcangeli set_pte_at(mm, haddr, pte, entry); 177571e3aac0SAndrea Arcangeli pte_unmap(pte); 177671e3aac0SAndrea Arcangeli } 177771e3aac0SAndrea Arcangeli 177871e3aac0SAndrea Arcangeli smp_wmb(); /* make pte visible before pmd */ 177971e3aac0SAndrea Arcangeli /* 178071e3aac0SAndrea Arcangeli * Up to this point the pmd is present and huge and 178171e3aac0SAndrea Arcangeli * userland has the whole access to the hugepage 178271e3aac0SAndrea Arcangeli * during the split (which happens in place). If we 178371e3aac0SAndrea Arcangeli * overwrite the pmd with the not-huge version 178471e3aac0SAndrea Arcangeli * pointing to the pte here (which of course we could 178571e3aac0SAndrea Arcangeli * if all CPUs were bug free), userland could trigger 178671e3aac0SAndrea Arcangeli * a small page size TLB miss on the small sized TLB 178771e3aac0SAndrea Arcangeli * while the hugepage TLB entry is still established 178871e3aac0SAndrea Arcangeli * in the huge TLB. Some CPU doesn't like that. See 178971e3aac0SAndrea Arcangeli * http://support.amd.com/us/Processor_TechDocs/41322.pdf, 179071e3aac0SAndrea Arcangeli * Erratum 383 on page 93. Intel should be safe but is 179171e3aac0SAndrea Arcangeli * also warns that it's only safe if the permission 179271e3aac0SAndrea Arcangeli * and cache attributes of the two entries loaded in 179371e3aac0SAndrea Arcangeli * the two TLB is identical (which should be the case 179471e3aac0SAndrea Arcangeli * here). But it is generally safer to never allow 179571e3aac0SAndrea Arcangeli * small and huge TLB entries for the same virtual 179671e3aac0SAndrea Arcangeli * address to be loaded simultaneously. So instead of 179771e3aac0SAndrea Arcangeli * doing "pmd_populate(); flush_tlb_range();" we first 179871e3aac0SAndrea Arcangeli * mark the current pmd notpresent (atomically because 179971e3aac0SAndrea Arcangeli * here the pmd_trans_huge and pmd_trans_splitting 180071e3aac0SAndrea Arcangeli * must remain set at all times on the pmd until the 180171e3aac0SAndrea Arcangeli * split is complete for this pmd), then we flush the 180271e3aac0SAndrea Arcangeli * SMP TLB and finally we write the non-huge version 180371e3aac0SAndrea Arcangeli * of the pmd entry with pmd_populate. 180471e3aac0SAndrea Arcangeli */ 180546dcde73SGerald Schaefer pmdp_invalidate(vma, address, pmd); 180671e3aac0SAndrea Arcangeli pmd_populate(mm, pmd, pgtable); 180771e3aac0SAndrea Arcangeli ret = 1; 1808117b0791SKirill A. Shutemov spin_unlock(ptl); 180971e3aac0SAndrea Arcangeli } 181071e3aac0SAndrea Arcangeli 181171e3aac0SAndrea Arcangeli return ret; 181271e3aac0SAndrea Arcangeli } 181371e3aac0SAndrea Arcangeli 18145a505085SIngo Molnar /* must be called with anon_vma->root->rwsem held */ 181571e3aac0SAndrea Arcangeli static void __split_huge_page(struct page *page, 18165bc7b8acSShaohua Li struct anon_vma *anon_vma, 18175bc7b8acSShaohua Li struct list_head *list) 181871e3aac0SAndrea Arcangeli { 181971e3aac0SAndrea Arcangeli int mapcount, mapcount2; 1820bf181b9fSMichel Lespinasse pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 182171e3aac0SAndrea Arcangeli struct anon_vma_chain *avc; 182271e3aac0SAndrea Arcangeli 182371e3aac0SAndrea Arcangeli BUG_ON(!PageHead(page)); 182471e3aac0SAndrea Arcangeli BUG_ON(PageTail(page)); 182571e3aac0SAndrea Arcangeli 182671e3aac0SAndrea Arcangeli mapcount = 0; 1827bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 182871e3aac0SAndrea Arcangeli struct vm_area_struct *vma = avc->vma; 182971e3aac0SAndrea Arcangeli unsigned long addr = vma_address(page, vma); 183071e3aac0SAndrea Arcangeli BUG_ON(is_vma_temporary_stack(vma)); 183171e3aac0SAndrea Arcangeli mapcount += __split_huge_page_splitting(page, vma, addr); 183271e3aac0SAndrea Arcangeli } 183305759d38SAndrea Arcangeli /* 183405759d38SAndrea Arcangeli * It is critical that new vmas are added to the tail of the 183505759d38SAndrea Arcangeli * anon_vma list. This guarantes that if copy_huge_pmd() runs 183605759d38SAndrea Arcangeli * and establishes a child pmd before 183705759d38SAndrea Arcangeli * __split_huge_page_splitting() freezes the parent pmd (so if 183805759d38SAndrea Arcangeli * we fail to prevent copy_huge_pmd() from running until the 183905759d38SAndrea Arcangeli * whole __split_huge_page() is complete), we will still see 184005759d38SAndrea Arcangeli * the newly established pmd of the child later during the 184105759d38SAndrea Arcangeli * walk, to be able to set it as pmd_trans_splitting too. 184205759d38SAndrea Arcangeli */ 1843ff9e43ebSKirill A. Shutemov if (mapcount != page_mapcount(page)) { 1844ae3a8c1cSAndrew Morton pr_err("mapcount %d page_mapcount %d\n", 184505759d38SAndrea Arcangeli mapcount, page_mapcount(page)); 1846ff9e43ebSKirill A. Shutemov BUG(); 1847ff9e43ebSKirill A. Shutemov } 184871e3aac0SAndrea Arcangeli 18495bc7b8acSShaohua Li __split_huge_page_refcount(page, list); 185071e3aac0SAndrea Arcangeli 185171e3aac0SAndrea Arcangeli mapcount2 = 0; 1852bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 185371e3aac0SAndrea Arcangeli struct vm_area_struct *vma = avc->vma; 185471e3aac0SAndrea Arcangeli unsigned long addr = vma_address(page, vma); 185571e3aac0SAndrea Arcangeli BUG_ON(is_vma_temporary_stack(vma)); 185671e3aac0SAndrea Arcangeli mapcount2 += __split_huge_page_map(page, vma, addr); 185771e3aac0SAndrea Arcangeli } 1858ff9e43ebSKirill A. Shutemov if (mapcount != mapcount2) { 1859ae3a8c1cSAndrew Morton pr_err("mapcount %d mapcount2 %d page_mapcount %d\n", 186005759d38SAndrea Arcangeli mapcount, mapcount2, page_mapcount(page)); 1861ff9e43ebSKirill A. Shutemov BUG(); 1862ff9e43ebSKirill A. Shutemov } 186371e3aac0SAndrea Arcangeli } 186471e3aac0SAndrea Arcangeli 18655bc7b8acSShaohua Li /* 18665bc7b8acSShaohua Li * Split a hugepage into normal pages. This doesn't change the position of head 18675bc7b8acSShaohua Li * page. If @list is null, tail pages will be added to LRU list, otherwise, to 18685bc7b8acSShaohua Li * @list. Both head page and tail pages will inherit mapping, flags, and so on 18695bc7b8acSShaohua Li * from the hugepage. 18705bc7b8acSShaohua Li * Return 0 if the hugepage is split successfully otherwise return 1. 18715bc7b8acSShaohua Li */ 18725bc7b8acSShaohua Li int split_huge_page_to_list(struct page *page, struct list_head *list) 187371e3aac0SAndrea Arcangeli { 187471e3aac0SAndrea Arcangeli struct anon_vma *anon_vma; 187571e3aac0SAndrea Arcangeli int ret = 1; 187671e3aac0SAndrea Arcangeli 18775918d10aSKirill A. Shutemov BUG_ON(is_huge_zero_page(page)); 187871e3aac0SAndrea Arcangeli BUG_ON(!PageAnon(page)); 1879062f1af2SMel Gorman 1880062f1af2SMel Gorman /* 1881062f1af2SMel Gorman * The caller does not necessarily hold an mmap_sem that would prevent 1882062f1af2SMel Gorman * the anon_vma disappearing so we first we take a reference to it 1883062f1af2SMel Gorman * and then lock the anon_vma for write. This is similar to 1884062f1af2SMel Gorman * page_lock_anon_vma_read except the write lock is taken to serialise 1885062f1af2SMel Gorman * against parallel split or collapse operations. 1886062f1af2SMel Gorman */ 1887062f1af2SMel Gorman anon_vma = page_get_anon_vma(page); 188871e3aac0SAndrea Arcangeli if (!anon_vma) 188971e3aac0SAndrea Arcangeli goto out; 1890062f1af2SMel Gorman anon_vma_lock_write(anon_vma); 1891062f1af2SMel Gorman 189271e3aac0SAndrea Arcangeli ret = 0; 189371e3aac0SAndrea Arcangeli if (!PageCompound(page)) 189471e3aac0SAndrea Arcangeli goto out_unlock; 189571e3aac0SAndrea Arcangeli 189671e3aac0SAndrea Arcangeli BUG_ON(!PageSwapBacked(page)); 18975bc7b8acSShaohua Li __split_huge_page(page, anon_vma, list); 189881ab4201SAndi Kleen count_vm_event(THP_SPLIT); 189971e3aac0SAndrea Arcangeli 190071e3aac0SAndrea Arcangeli BUG_ON(PageCompound(page)); 190171e3aac0SAndrea Arcangeli out_unlock: 190208b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 1903062f1af2SMel Gorman put_anon_vma(anon_vma); 190471e3aac0SAndrea Arcangeli out: 190571e3aac0SAndrea Arcangeli return ret; 190671e3aac0SAndrea Arcangeli } 190771e3aac0SAndrea Arcangeli 19089050d7ebSVlastimil Babka #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) 190978f11a25SAndrea Arcangeli 191060ab3244SAndrea Arcangeli int hugepage_madvise(struct vm_area_struct *vma, 191160ab3244SAndrea Arcangeli unsigned long *vm_flags, int advice) 19120af4e98bSAndrea Arcangeli { 1913a664b2d8SAndrea Arcangeli switch (advice) { 1914a664b2d8SAndrea Arcangeli case MADV_HUGEPAGE: 19151e1836e8SAlex Thorlton #ifdef CONFIG_S390 19161e1836e8SAlex Thorlton /* 19171e1836e8SAlex Thorlton * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 19181e1836e8SAlex Thorlton * can't handle this properly after s390_enable_sie, so we simply 19191e1836e8SAlex Thorlton * ignore the madvise to prevent qemu from causing a SIGSEGV. 19201e1836e8SAlex Thorlton */ 19211e1836e8SAlex Thorlton if (mm_has_pgste(vma->vm_mm)) 19221e1836e8SAlex Thorlton return 0; 19231e1836e8SAlex Thorlton #endif 19240af4e98bSAndrea Arcangeli /* 19250af4e98bSAndrea Arcangeli * Be somewhat over-protective like KSM for now! 19260af4e98bSAndrea Arcangeli */ 192778f11a25SAndrea Arcangeli if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP)) 19280af4e98bSAndrea Arcangeli return -EINVAL; 1929a664b2d8SAndrea Arcangeli *vm_flags &= ~VM_NOHUGEPAGE; 19300af4e98bSAndrea Arcangeli *vm_flags |= VM_HUGEPAGE; 193160ab3244SAndrea Arcangeli /* 193260ab3244SAndrea Arcangeli * If the vma become good for khugepaged to scan, 193360ab3244SAndrea Arcangeli * register it here without waiting a page fault that 193460ab3244SAndrea Arcangeli * may not happen any time soon. 193560ab3244SAndrea Arcangeli */ 19366d50e60cSDavid Rientjes if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) 193760ab3244SAndrea Arcangeli return -ENOMEM; 1938a664b2d8SAndrea Arcangeli break; 1939a664b2d8SAndrea Arcangeli case MADV_NOHUGEPAGE: 1940a664b2d8SAndrea Arcangeli /* 1941a664b2d8SAndrea Arcangeli * Be somewhat over-protective like KSM for now! 1942a664b2d8SAndrea Arcangeli */ 194378f11a25SAndrea Arcangeli if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP)) 1944a664b2d8SAndrea Arcangeli return -EINVAL; 1945a664b2d8SAndrea Arcangeli *vm_flags &= ~VM_HUGEPAGE; 1946a664b2d8SAndrea Arcangeli *vm_flags |= VM_NOHUGEPAGE; 194760ab3244SAndrea Arcangeli /* 194860ab3244SAndrea Arcangeli * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 194960ab3244SAndrea Arcangeli * this vma even if we leave the mm registered in khugepaged if 195060ab3244SAndrea Arcangeli * it got registered before VM_NOHUGEPAGE was set. 195160ab3244SAndrea Arcangeli */ 1952a664b2d8SAndrea Arcangeli break; 1953a664b2d8SAndrea Arcangeli } 19540af4e98bSAndrea Arcangeli 19550af4e98bSAndrea Arcangeli return 0; 19560af4e98bSAndrea Arcangeli } 19570af4e98bSAndrea Arcangeli 1958ba76149fSAndrea Arcangeli static int __init khugepaged_slab_init(void) 1959ba76149fSAndrea Arcangeli { 1960ba76149fSAndrea Arcangeli mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 1961ba76149fSAndrea Arcangeli sizeof(struct mm_slot), 1962ba76149fSAndrea Arcangeli __alignof__(struct mm_slot), 0, NULL); 1963ba76149fSAndrea Arcangeli if (!mm_slot_cache) 1964ba76149fSAndrea Arcangeli return -ENOMEM; 1965ba76149fSAndrea Arcangeli 1966ba76149fSAndrea Arcangeli return 0; 1967ba76149fSAndrea Arcangeli } 1968ba76149fSAndrea Arcangeli 1969ba76149fSAndrea Arcangeli static inline struct mm_slot *alloc_mm_slot(void) 1970ba76149fSAndrea Arcangeli { 1971ba76149fSAndrea Arcangeli if (!mm_slot_cache) /* initialization failed */ 1972ba76149fSAndrea Arcangeli return NULL; 1973ba76149fSAndrea Arcangeli return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 1974ba76149fSAndrea Arcangeli } 1975ba76149fSAndrea Arcangeli 1976ba76149fSAndrea Arcangeli static inline void free_mm_slot(struct mm_slot *mm_slot) 1977ba76149fSAndrea Arcangeli { 1978ba76149fSAndrea Arcangeli kmem_cache_free(mm_slot_cache, mm_slot); 1979ba76149fSAndrea Arcangeli } 1980ba76149fSAndrea Arcangeli 1981ba76149fSAndrea Arcangeli static struct mm_slot *get_mm_slot(struct mm_struct *mm) 1982ba76149fSAndrea Arcangeli { 1983ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 1984ba76149fSAndrea Arcangeli 1985b67bfe0dSSasha Levin hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 1986ba76149fSAndrea Arcangeli if (mm == mm_slot->mm) 1987ba76149fSAndrea Arcangeli return mm_slot; 198843b5fbbdSSasha Levin 1989ba76149fSAndrea Arcangeli return NULL; 1990ba76149fSAndrea Arcangeli } 1991ba76149fSAndrea Arcangeli 1992ba76149fSAndrea Arcangeli static void insert_to_mm_slots_hash(struct mm_struct *mm, 1993ba76149fSAndrea Arcangeli struct mm_slot *mm_slot) 1994ba76149fSAndrea Arcangeli { 1995ba76149fSAndrea Arcangeli mm_slot->mm = mm; 199643b5fbbdSSasha Levin hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 1997ba76149fSAndrea Arcangeli } 1998ba76149fSAndrea Arcangeli 1999ba76149fSAndrea Arcangeli static inline int khugepaged_test_exit(struct mm_struct *mm) 2000ba76149fSAndrea Arcangeli { 2001ba76149fSAndrea Arcangeli return atomic_read(&mm->mm_users) == 0; 2002ba76149fSAndrea Arcangeli } 2003ba76149fSAndrea Arcangeli 2004ba76149fSAndrea Arcangeli int __khugepaged_enter(struct mm_struct *mm) 2005ba76149fSAndrea Arcangeli { 2006ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 2007ba76149fSAndrea Arcangeli int wakeup; 2008ba76149fSAndrea Arcangeli 2009ba76149fSAndrea Arcangeli mm_slot = alloc_mm_slot(); 2010ba76149fSAndrea Arcangeli if (!mm_slot) 2011ba76149fSAndrea Arcangeli return -ENOMEM; 2012ba76149fSAndrea Arcangeli 2013ba76149fSAndrea Arcangeli /* __khugepaged_exit() must not run from under us */ 201496dad67fSSasha Levin VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); 2015ba76149fSAndrea Arcangeli if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 2016ba76149fSAndrea Arcangeli free_mm_slot(mm_slot); 2017ba76149fSAndrea Arcangeli return 0; 2018ba76149fSAndrea Arcangeli } 2019ba76149fSAndrea Arcangeli 2020ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2021ba76149fSAndrea Arcangeli insert_to_mm_slots_hash(mm, mm_slot); 2022ba76149fSAndrea Arcangeli /* 2023ba76149fSAndrea Arcangeli * Insert just behind the scanning cursor, to let the area settle 2024ba76149fSAndrea Arcangeli * down a little. 2025ba76149fSAndrea Arcangeli */ 2026ba76149fSAndrea Arcangeli wakeup = list_empty(&khugepaged_scan.mm_head); 2027ba76149fSAndrea Arcangeli list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 2028ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 2029ba76149fSAndrea Arcangeli 2030ba76149fSAndrea Arcangeli atomic_inc(&mm->mm_count); 2031ba76149fSAndrea Arcangeli if (wakeup) 2032ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 2033ba76149fSAndrea Arcangeli 2034ba76149fSAndrea Arcangeli return 0; 2035ba76149fSAndrea Arcangeli } 2036ba76149fSAndrea Arcangeli 20376d50e60cSDavid Rientjes int khugepaged_enter_vma_merge(struct vm_area_struct *vma, 20386d50e60cSDavid Rientjes unsigned long vm_flags) 2039ba76149fSAndrea Arcangeli { 2040ba76149fSAndrea Arcangeli unsigned long hstart, hend; 2041ba76149fSAndrea Arcangeli if (!vma->anon_vma) 2042ba76149fSAndrea Arcangeli /* 2043ba76149fSAndrea Arcangeli * Not yet faulted in so we will register later in the 2044ba76149fSAndrea Arcangeli * page fault if needed. 2045ba76149fSAndrea Arcangeli */ 2046ba76149fSAndrea Arcangeli return 0; 204778f11a25SAndrea Arcangeli if (vma->vm_ops) 2048ba76149fSAndrea Arcangeli /* khugepaged not yet working on file or special mappings */ 2049ba76149fSAndrea Arcangeli return 0; 20506d50e60cSDavid Rientjes VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma); 2051ba76149fSAndrea Arcangeli hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2052ba76149fSAndrea Arcangeli hend = vma->vm_end & HPAGE_PMD_MASK; 2053ba76149fSAndrea Arcangeli if (hstart < hend) 20546d50e60cSDavid Rientjes return khugepaged_enter(vma, vm_flags); 2055ba76149fSAndrea Arcangeli return 0; 2056ba76149fSAndrea Arcangeli } 2057ba76149fSAndrea Arcangeli 2058ba76149fSAndrea Arcangeli void __khugepaged_exit(struct mm_struct *mm) 2059ba76149fSAndrea Arcangeli { 2060ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 2061ba76149fSAndrea Arcangeli int free = 0; 2062ba76149fSAndrea Arcangeli 2063ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2064ba76149fSAndrea Arcangeli mm_slot = get_mm_slot(mm); 2065ba76149fSAndrea Arcangeli if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 206643b5fbbdSSasha Levin hash_del(&mm_slot->hash); 2067ba76149fSAndrea Arcangeli list_del(&mm_slot->mm_node); 2068ba76149fSAndrea Arcangeli free = 1; 2069ba76149fSAndrea Arcangeli } 2070d788e80aSChris Wright spin_unlock(&khugepaged_mm_lock); 2071ba76149fSAndrea Arcangeli 2072ba76149fSAndrea Arcangeli if (free) { 2073ba76149fSAndrea Arcangeli clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2074ba76149fSAndrea Arcangeli free_mm_slot(mm_slot); 2075ba76149fSAndrea Arcangeli mmdrop(mm); 2076ba76149fSAndrea Arcangeli } else if (mm_slot) { 2077ba76149fSAndrea Arcangeli /* 2078ba76149fSAndrea Arcangeli * This is required to serialize against 2079ba76149fSAndrea Arcangeli * khugepaged_test_exit() (which is guaranteed to run 2080ba76149fSAndrea Arcangeli * under mmap sem read mode). Stop here (after we 2081ba76149fSAndrea Arcangeli * return all pagetables will be destroyed) until 2082ba76149fSAndrea Arcangeli * khugepaged has finished working on the pagetables 2083ba76149fSAndrea Arcangeli * under the mmap_sem. 2084ba76149fSAndrea Arcangeli */ 2085ba76149fSAndrea Arcangeli down_write(&mm->mmap_sem); 2086ba76149fSAndrea Arcangeli up_write(&mm->mmap_sem); 2087d788e80aSChris Wright } 2088ba76149fSAndrea Arcangeli } 2089ba76149fSAndrea Arcangeli 2090ba76149fSAndrea Arcangeli static void release_pte_page(struct page *page) 2091ba76149fSAndrea Arcangeli { 2092ba76149fSAndrea Arcangeli /* 0 stands for page_is_file_cache(page) == false */ 2093ba76149fSAndrea Arcangeli dec_zone_page_state(page, NR_ISOLATED_ANON + 0); 2094ba76149fSAndrea Arcangeli unlock_page(page); 2095ba76149fSAndrea Arcangeli putback_lru_page(page); 2096ba76149fSAndrea Arcangeli } 2097ba76149fSAndrea Arcangeli 2098ba76149fSAndrea Arcangeli static void release_pte_pages(pte_t *pte, pte_t *_pte) 2099ba76149fSAndrea Arcangeli { 2100ba76149fSAndrea Arcangeli while (--_pte >= pte) { 2101ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 2102ba76149fSAndrea Arcangeli if (!pte_none(pteval)) 2103ba76149fSAndrea Arcangeli release_pte_page(pte_page(pteval)); 2104ba76149fSAndrea Arcangeli } 2105ba76149fSAndrea Arcangeli } 2106ba76149fSAndrea Arcangeli 2107ba76149fSAndrea Arcangeli static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 2108ba76149fSAndrea Arcangeli unsigned long address, 2109ba76149fSAndrea Arcangeli pte_t *pte) 2110ba76149fSAndrea Arcangeli { 2111ba76149fSAndrea Arcangeli struct page *page; 2112ba76149fSAndrea Arcangeli pte_t *_pte; 211310359213SEbru Akagunduz int none = 0; 211410359213SEbru Akagunduz bool referenced = false, writable = false; 2115ba76149fSAndrea Arcangeli for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 2116ba76149fSAndrea Arcangeli _pte++, address += PAGE_SIZE) { 2117ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 2118ba76149fSAndrea Arcangeli if (pte_none(pteval)) { 2119ba76149fSAndrea Arcangeli if (++none <= khugepaged_max_ptes_none) 2120ba76149fSAndrea Arcangeli continue; 2121344aa35cSBob Liu else 2122ba76149fSAndrea Arcangeli goto out; 2123ba76149fSAndrea Arcangeli } 212410359213SEbru Akagunduz if (!pte_present(pteval)) 2125ba76149fSAndrea Arcangeli goto out; 2126ba76149fSAndrea Arcangeli page = vm_normal_page(vma, address, pteval); 2127344aa35cSBob Liu if (unlikely(!page)) 2128ba76149fSAndrea Arcangeli goto out; 2129344aa35cSBob Liu 2130309381feSSasha Levin VM_BUG_ON_PAGE(PageCompound(page), page); 2131309381feSSasha Levin VM_BUG_ON_PAGE(!PageAnon(page), page); 2132309381feSSasha Levin VM_BUG_ON_PAGE(!PageSwapBacked(page), page); 2133ba76149fSAndrea Arcangeli 2134ba76149fSAndrea Arcangeli /* 2135ba76149fSAndrea Arcangeli * We can do it before isolate_lru_page because the 2136ba76149fSAndrea Arcangeli * page can't be freed from under us. NOTE: PG_lock 2137ba76149fSAndrea Arcangeli * is needed to serialize against split_huge_page 2138ba76149fSAndrea Arcangeli * when invoked from the VM. 2139ba76149fSAndrea Arcangeli */ 2140344aa35cSBob Liu if (!trylock_page(page)) 2141ba76149fSAndrea Arcangeli goto out; 214210359213SEbru Akagunduz 214310359213SEbru Akagunduz /* 214410359213SEbru Akagunduz * cannot use mapcount: can't collapse if there's a gup pin. 214510359213SEbru Akagunduz * The page must only be referenced by the scanned process 214610359213SEbru Akagunduz * and page swap cache. 214710359213SEbru Akagunduz */ 214810359213SEbru Akagunduz if (page_count(page) != 1 + !!PageSwapCache(page)) { 214910359213SEbru Akagunduz unlock_page(page); 215010359213SEbru Akagunduz goto out; 215110359213SEbru Akagunduz } 215210359213SEbru Akagunduz if (pte_write(pteval)) { 215310359213SEbru Akagunduz writable = true; 215410359213SEbru Akagunduz } else { 215510359213SEbru Akagunduz if (PageSwapCache(page) && !reuse_swap_page(page)) { 215610359213SEbru Akagunduz unlock_page(page); 215710359213SEbru Akagunduz goto out; 215810359213SEbru Akagunduz } 215910359213SEbru Akagunduz /* 216010359213SEbru Akagunduz * Page is not in the swap cache. It can be collapsed 216110359213SEbru Akagunduz * into a THP. 216210359213SEbru Akagunduz */ 216310359213SEbru Akagunduz } 216410359213SEbru Akagunduz 2165ba76149fSAndrea Arcangeli /* 2166ba76149fSAndrea Arcangeli * Isolate the page to avoid collapsing an hugepage 2167ba76149fSAndrea Arcangeli * currently in use by the VM. 2168ba76149fSAndrea Arcangeli */ 2169ba76149fSAndrea Arcangeli if (isolate_lru_page(page)) { 2170ba76149fSAndrea Arcangeli unlock_page(page); 2171ba76149fSAndrea Arcangeli goto out; 2172ba76149fSAndrea Arcangeli } 2173ba76149fSAndrea Arcangeli /* 0 stands for page_is_file_cache(page) == false */ 2174ba76149fSAndrea Arcangeli inc_zone_page_state(page, NR_ISOLATED_ANON + 0); 2175309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 2176309381feSSasha Levin VM_BUG_ON_PAGE(PageLRU(page), page); 2177ba76149fSAndrea Arcangeli 2178ba76149fSAndrea Arcangeli /* If there is no mapped pte young don't collapse the page */ 21798ee53820SAndrea Arcangeli if (pte_young(pteval) || PageReferenced(page) || 21808ee53820SAndrea Arcangeli mmu_notifier_test_young(vma->vm_mm, address)) 218110359213SEbru Akagunduz referenced = true; 2182ba76149fSAndrea Arcangeli } 218310359213SEbru Akagunduz if (likely(referenced && writable)) 2184344aa35cSBob Liu return 1; 2185ba76149fSAndrea Arcangeli out: 2186344aa35cSBob Liu release_pte_pages(pte, _pte); 2187344aa35cSBob Liu return 0; 2188ba76149fSAndrea Arcangeli } 2189ba76149fSAndrea Arcangeli 2190ba76149fSAndrea Arcangeli static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 2191ba76149fSAndrea Arcangeli struct vm_area_struct *vma, 2192ba76149fSAndrea Arcangeli unsigned long address, 2193ba76149fSAndrea Arcangeli spinlock_t *ptl) 2194ba76149fSAndrea Arcangeli { 2195ba76149fSAndrea Arcangeli pte_t *_pte; 2196ba76149fSAndrea Arcangeli for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { 2197ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 2198ba76149fSAndrea Arcangeli struct page *src_page; 2199ba76149fSAndrea Arcangeli 2200ba76149fSAndrea Arcangeli if (pte_none(pteval)) { 2201ba76149fSAndrea Arcangeli clear_user_highpage(page, address); 2202ba76149fSAndrea Arcangeli add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 2203ba76149fSAndrea Arcangeli } else { 2204ba76149fSAndrea Arcangeli src_page = pte_page(pteval); 2205ba76149fSAndrea Arcangeli copy_user_highpage(page, src_page, address, vma); 2206309381feSSasha Levin VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); 2207ba76149fSAndrea Arcangeli release_pte_page(src_page); 2208ba76149fSAndrea Arcangeli /* 2209ba76149fSAndrea Arcangeli * ptl mostly unnecessary, but preempt has to 2210ba76149fSAndrea Arcangeli * be disabled to update the per-cpu stats 2211ba76149fSAndrea Arcangeli * inside page_remove_rmap(). 2212ba76149fSAndrea Arcangeli */ 2213ba76149fSAndrea Arcangeli spin_lock(ptl); 2214ba76149fSAndrea Arcangeli /* 2215ba76149fSAndrea Arcangeli * paravirt calls inside pte_clear here are 2216ba76149fSAndrea Arcangeli * superfluous. 2217ba76149fSAndrea Arcangeli */ 2218ba76149fSAndrea Arcangeli pte_clear(vma->vm_mm, address, _pte); 2219ba76149fSAndrea Arcangeli page_remove_rmap(src_page); 2220ba76149fSAndrea Arcangeli spin_unlock(ptl); 2221ba76149fSAndrea Arcangeli free_page_and_swap_cache(src_page); 2222ba76149fSAndrea Arcangeli } 2223ba76149fSAndrea Arcangeli 2224ba76149fSAndrea Arcangeli address += PAGE_SIZE; 2225ba76149fSAndrea Arcangeli page++; 2226ba76149fSAndrea Arcangeli } 2227ba76149fSAndrea Arcangeli } 2228ba76149fSAndrea Arcangeli 222926234f36SXiao Guangrong static void khugepaged_alloc_sleep(void) 223026234f36SXiao Guangrong { 223126234f36SXiao Guangrong wait_event_freezable_timeout(khugepaged_wait, false, 223226234f36SXiao Guangrong msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 223326234f36SXiao Guangrong } 223426234f36SXiao Guangrong 22359f1b868aSBob Liu static int khugepaged_node_load[MAX_NUMNODES]; 22369f1b868aSBob Liu 223714a4e214SDavid Rientjes static bool khugepaged_scan_abort(int nid) 223814a4e214SDavid Rientjes { 223914a4e214SDavid Rientjes int i; 224014a4e214SDavid Rientjes 224114a4e214SDavid Rientjes /* 224214a4e214SDavid Rientjes * If zone_reclaim_mode is disabled, then no extra effort is made to 224314a4e214SDavid Rientjes * allocate memory locally. 224414a4e214SDavid Rientjes */ 224514a4e214SDavid Rientjes if (!zone_reclaim_mode) 224614a4e214SDavid Rientjes return false; 224714a4e214SDavid Rientjes 224814a4e214SDavid Rientjes /* If there is a count for this node already, it must be acceptable */ 224914a4e214SDavid Rientjes if (khugepaged_node_load[nid]) 225014a4e214SDavid Rientjes return false; 225114a4e214SDavid Rientjes 225214a4e214SDavid Rientjes for (i = 0; i < MAX_NUMNODES; i++) { 225314a4e214SDavid Rientjes if (!khugepaged_node_load[i]) 225414a4e214SDavid Rientjes continue; 225514a4e214SDavid Rientjes if (node_distance(nid, i) > RECLAIM_DISTANCE) 225614a4e214SDavid Rientjes return true; 225714a4e214SDavid Rientjes } 225814a4e214SDavid Rientjes return false; 225914a4e214SDavid Rientjes } 226014a4e214SDavid Rientjes 226126234f36SXiao Guangrong #ifdef CONFIG_NUMA 22629f1b868aSBob Liu static int khugepaged_find_target_node(void) 22639f1b868aSBob Liu { 22649f1b868aSBob Liu static int last_khugepaged_target_node = NUMA_NO_NODE; 22659f1b868aSBob Liu int nid, target_node = 0, max_value = 0; 22669f1b868aSBob Liu 22679f1b868aSBob Liu /* find first node with max normal pages hit */ 22689f1b868aSBob Liu for (nid = 0; nid < MAX_NUMNODES; nid++) 22699f1b868aSBob Liu if (khugepaged_node_load[nid] > max_value) { 22709f1b868aSBob Liu max_value = khugepaged_node_load[nid]; 22719f1b868aSBob Liu target_node = nid; 22729f1b868aSBob Liu } 22739f1b868aSBob Liu 22749f1b868aSBob Liu /* do some balance if several nodes have the same hit record */ 22759f1b868aSBob Liu if (target_node <= last_khugepaged_target_node) 22769f1b868aSBob Liu for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; 22779f1b868aSBob Liu nid++) 22789f1b868aSBob Liu if (max_value == khugepaged_node_load[nid]) { 22799f1b868aSBob Liu target_node = nid; 22809f1b868aSBob Liu break; 22819f1b868aSBob Liu } 22829f1b868aSBob Liu 22839f1b868aSBob Liu last_khugepaged_target_node = target_node; 22849f1b868aSBob Liu return target_node; 22859f1b868aSBob Liu } 22869f1b868aSBob Liu 228726234f36SXiao Guangrong static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 228826234f36SXiao Guangrong { 228926234f36SXiao Guangrong if (IS_ERR(*hpage)) { 229026234f36SXiao Guangrong if (!*wait) 229126234f36SXiao Guangrong return false; 229226234f36SXiao Guangrong 229326234f36SXiao Guangrong *wait = false; 2294e3b4126cSXiao Guangrong *hpage = NULL; 229526234f36SXiao Guangrong khugepaged_alloc_sleep(); 229626234f36SXiao Guangrong } else if (*hpage) { 229726234f36SXiao Guangrong put_page(*hpage); 229826234f36SXiao Guangrong *hpage = NULL; 229926234f36SXiao Guangrong } 230026234f36SXiao Guangrong 230126234f36SXiao Guangrong return true; 230226234f36SXiao Guangrong } 230326234f36SXiao Guangrong 230426234f36SXiao Guangrong static struct page 230526234f36SXiao Guangrong *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 230626234f36SXiao Guangrong struct vm_area_struct *vma, unsigned long address, 230726234f36SXiao Guangrong int node) 230826234f36SXiao Guangrong { 2309309381feSSasha Levin VM_BUG_ON_PAGE(*hpage, *hpage); 23108b164568SVlastimil Babka 231126234f36SXiao Guangrong /* 23128b164568SVlastimil Babka * Before allocating the hugepage, release the mmap_sem read lock. 23138b164568SVlastimil Babka * The allocation can take potentially a long time if it involves 23148b164568SVlastimil Babka * sync compaction, and we do not need to hold the mmap_sem during 23158b164568SVlastimil Babka * that. We will recheck the vma after taking it again in write mode. 231626234f36SXiao Guangrong */ 231726234f36SXiao Guangrong up_read(&mm->mmap_sem); 23188b164568SVlastimil Babka 23198b164568SVlastimil Babka *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask( 23208b164568SVlastimil Babka khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER); 232126234f36SXiao Guangrong if (unlikely(!*hpage)) { 232226234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 232326234f36SXiao Guangrong *hpage = ERR_PTR(-ENOMEM); 232426234f36SXiao Guangrong return NULL; 232526234f36SXiao Guangrong } 232626234f36SXiao Guangrong 232726234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC); 232826234f36SXiao Guangrong return *hpage; 232926234f36SXiao Guangrong } 233026234f36SXiao Guangrong #else 23319f1b868aSBob Liu static int khugepaged_find_target_node(void) 23329f1b868aSBob Liu { 23339f1b868aSBob Liu return 0; 23349f1b868aSBob Liu } 23359f1b868aSBob Liu 233610dc4155SBob Liu static inline struct page *alloc_hugepage(int defrag) 233710dc4155SBob Liu { 233810dc4155SBob Liu return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), 233910dc4155SBob Liu HPAGE_PMD_ORDER); 234010dc4155SBob Liu } 234110dc4155SBob Liu 234226234f36SXiao Guangrong static struct page *khugepaged_alloc_hugepage(bool *wait) 234326234f36SXiao Guangrong { 234426234f36SXiao Guangrong struct page *hpage; 234526234f36SXiao Guangrong 234626234f36SXiao Guangrong do { 234726234f36SXiao Guangrong hpage = alloc_hugepage(khugepaged_defrag()); 234826234f36SXiao Guangrong if (!hpage) { 234926234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 235026234f36SXiao Guangrong if (!*wait) 235126234f36SXiao Guangrong return NULL; 235226234f36SXiao Guangrong 235326234f36SXiao Guangrong *wait = false; 235426234f36SXiao Guangrong khugepaged_alloc_sleep(); 235526234f36SXiao Guangrong } else 235626234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC); 235726234f36SXiao Guangrong } while (unlikely(!hpage) && likely(khugepaged_enabled())); 235826234f36SXiao Guangrong 235926234f36SXiao Guangrong return hpage; 236026234f36SXiao Guangrong } 236126234f36SXiao Guangrong 236226234f36SXiao Guangrong static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 236326234f36SXiao Guangrong { 236426234f36SXiao Guangrong if (!*hpage) 236526234f36SXiao Guangrong *hpage = khugepaged_alloc_hugepage(wait); 236626234f36SXiao Guangrong 236726234f36SXiao Guangrong if (unlikely(!*hpage)) 236826234f36SXiao Guangrong return false; 236926234f36SXiao Guangrong 237026234f36SXiao Guangrong return true; 237126234f36SXiao Guangrong } 237226234f36SXiao Guangrong 237326234f36SXiao Guangrong static struct page 237426234f36SXiao Guangrong *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 237526234f36SXiao Guangrong struct vm_area_struct *vma, unsigned long address, 237626234f36SXiao Guangrong int node) 237726234f36SXiao Guangrong { 237826234f36SXiao Guangrong up_read(&mm->mmap_sem); 237926234f36SXiao Guangrong VM_BUG_ON(!*hpage); 238026234f36SXiao Guangrong return *hpage; 238126234f36SXiao Guangrong } 238226234f36SXiao Guangrong #endif 238326234f36SXiao Guangrong 2384fa475e51SBob Liu static bool hugepage_vma_check(struct vm_area_struct *vma) 2385fa475e51SBob Liu { 2386fa475e51SBob Liu if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || 2387fa475e51SBob Liu (vma->vm_flags & VM_NOHUGEPAGE)) 2388fa475e51SBob Liu return false; 2389fa475e51SBob Liu 2390fa475e51SBob Liu if (!vma->anon_vma || vma->vm_ops) 2391fa475e51SBob Liu return false; 2392fa475e51SBob Liu if (is_vma_temporary_stack(vma)) 2393fa475e51SBob Liu return false; 239481d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma); 2395fa475e51SBob Liu return true; 2396fa475e51SBob Liu } 2397fa475e51SBob Liu 2398ba76149fSAndrea Arcangeli static void collapse_huge_page(struct mm_struct *mm, 2399ba76149fSAndrea Arcangeli unsigned long address, 2400ce83d217SAndrea Arcangeli struct page **hpage, 24015c4b4be3SAndi Kleen struct vm_area_struct *vma, 24025c4b4be3SAndi Kleen int node) 2403ba76149fSAndrea Arcangeli { 2404ba76149fSAndrea Arcangeli pmd_t *pmd, _pmd; 2405ba76149fSAndrea Arcangeli pte_t *pte; 2406ba76149fSAndrea Arcangeli pgtable_t pgtable; 2407ba76149fSAndrea Arcangeli struct page *new_page; 2408c4088ebdSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 2409ba76149fSAndrea Arcangeli int isolated; 2410ba76149fSAndrea Arcangeli unsigned long hstart, hend; 241100501b53SJohannes Weiner struct mem_cgroup *memcg; 24122ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 24132ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 2414ba76149fSAndrea Arcangeli 2415ba76149fSAndrea Arcangeli VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2416692e0b35SAndrea Arcangeli 241726234f36SXiao Guangrong /* release the mmap_sem read lock. */ 241826234f36SXiao Guangrong new_page = khugepaged_alloc_page(hpage, mm, vma, address, node); 241926234f36SXiao Guangrong if (!new_page) 2420ce83d217SAndrea Arcangeli return; 2421ce83d217SAndrea Arcangeli 242200501b53SJohannes Weiner if (unlikely(mem_cgroup_try_charge(new_page, mm, 242300501b53SJohannes Weiner GFP_TRANSHUGE, &memcg))) 2424692e0b35SAndrea Arcangeli return; 2425ba76149fSAndrea Arcangeli 2426ba76149fSAndrea Arcangeli /* 2427ba76149fSAndrea Arcangeli * Prevent all access to pagetables with the exception of 2428ba76149fSAndrea Arcangeli * gup_fast later hanlded by the ptep_clear_flush and the VM 2429ba76149fSAndrea Arcangeli * handled by the anon_vma lock + PG_lock. 2430ba76149fSAndrea Arcangeli */ 2431ba76149fSAndrea Arcangeli down_write(&mm->mmap_sem); 2432ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) 2433ba76149fSAndrea Arcangeli goto out; 2434ba76149fSAndrea Arcangeli 2435ba76149fSAndrea Arcangeli vma = find_vma(mm, address); 2436a8f531ebSLibin if (!vma) 2437a8f531ebSLibin goto out; 2438ba76149fSAndrea Arcangeli hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2439ba76149fSAndrea Arcangeli hend = vma->vm_end & HPAGE_PMD_MASK; 2440ba76149fSAndrea Arcangeli if (address < hstart || address + HPAGE_PMD_SIZE > hend) 2441ba76149fSAndrea Arcangeli goto out; 2442fa475e51SBob Liu if (!hugepage_vma_check(vma)) 2443ba76149fSAndrea Arcangeli goto out; 24446219049aSBob Liu pmd = mm_find_pmd(mm, address); 24456219049aSBob Liu if (!pmd) 2446ba76149fSAndrea Arcangeli goto out; 2447ba76149fSAndrea Arcangeli 24484fc3f1d6SIngo Molnar anon_vma_lock_write(vma->anon_vma); 2449ba76149fSAndrea Arcangeli 2450ba76149fSAndrea Arcangeli pte = pte_offset_map(pmd, address); 2451c4088ebdSKirill A. Shutemov pte_ptl = pte_lockptr(mm, pmd); 2452ba76149fSAndrea Arcangeli 24532ec74c3eSSagi Grimberg mmun_start = address; 24542ec74c3eSSagi Grimberg mmun_end = address + HPAGE_PMD_SIZE; 24552ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2456c4088ebdSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 2457ba76149fSAndrea Arcangeli /* 2458ba76149fSAndrea Arcangeli * After this gup_fast can't run anymore. This also removes 2459ba76149fSAndrea Arcangeli * any huge TLB entry from the CPU so we won't allow 2460ba76149fSAndrea Arcangeli * huge and small TLB entries for the same virtual address 2461ba76149fSAndrea Arcangeli * to avoid the risk of CPU bugs in that area. 2462ba76149fSAndrea Arcangeli */ 24632ec74c3eSSagi Grimberg _pmd = pmdp_clear_flush(vma, address, pmd); 2464c4088ebdSKirill A. Shutemov spin_unlock(pmd_ptl); 24652ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2466ba76149fSAndrea Arcangeli 2467c4088ebdSKirill A. Shutemov spin_lock(pte_ptl); 2468ba76149fSAndrea Arcangeli isolated = __collapse_huge_page_isolate(vma, address, pte); 2469c4088ebdSKirill A. Shutemov spin_unlock(pte_ptl); 2470ba76149fSAndrea Arcangeli 2471ba76149fSAndrea Arcangeli if (unlikely(!isolated)) { 2472453c7192SJohannes Weiner pte_unmap(pte); 2473c4088ebdSKirill A. Shutemov spin_lock(pmd_ptl); 2474ba76149fSAndrea Arcangeli BUG_ON(!pmd_none(*pmd)); 24757c342512SAneesh Kumar K.V /* 24767c342512SAneesh Kumar K.V * We can only use set_pmd_at when establishing 24777c342512SAneesh Kumar K.V * hugepmds and never for establishing regular pmds that 24787c342512SAneesh Kumar K.V * points to regular pagetables. Use pmd_populate for that 24797c342512SAneesh Kumar K.V */ 24807c342512SAneesh Kumar K.V pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 2481c4088ebdSKirill A. Shutemov spin_unlock(pmd_ptl); 248208b52706SKonstantin Khlebnikov anon_vma_unlock_write(vma->anon_vma); 2483ce83d217SAndrea Arcangeli goto out; 2484ba76149fSAndrea Arcangeli } 2485ba76149fSAndrea Arcangeli 2486ba76149fSAndrea Arcangeli /* 2487ba76149fSAndrea Arcangeli * All pages are isolated and locked so anon_vma rmap 2488ba76149fSAndrea Arcangeli * can't run anymore. 2489ba76149fSAndrea Arcangeli */ 249008b52706SKonstantin Khlebnikov anon_vma_unlock_write(vma->anon_vma); 2491ba76149fSAndrea Arcangeli 2492c4088ebdSKirill A. Shutemov __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); 2493453c7192SJohannes Weiner pte_unmap(pte); 2494ba76149fSAndrea Arcangeli __SetPageUptodate(new_page); 2495ba76149fSAndrea Arcangeli pgtable = pmd_pgtable(_pmd); 2496ba76149fSAndrea Arcangeli 24973122359aSKirill A. Shutemov _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); 24983122359aSKirill A. Shutemov _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 2499ba76149fSAndrea Arcangeli 2500ba76149fSAndrea Arcangeli /* 2501ba76149fSAndrea Arcangeli * spin_lock() below is not the equivalent of smp_wmb(), so 2502ba76149fSAndrea Arcangeli * this is needed to avoid the copy_huge_page writes to become 2503ba76149fSAndrea Arcangeli * visible after the set_pmd_at() write. 2504ba76149fSAndrea Arcangeli */ 2505ba76149fSAndrea Arcangeli smp_wmb(); 2506ba76149fSAndrea Arcangeli 2507c4088ebdSKirill A. Shutemov spin_lock(pmd_ptl); 2508ba76149fSAndrea Arcangeli BUG_ON(!pmd_none(*pmd)); 2509ba76149fSAndrea Arcangeli page_add_new_anon_rmap(new_page, vma, address); 251000501b53SJohannes Weiner mem_cgroup_commit_charge(new_page, memcg, false); 251100501b53SJohannes Weiner lru_cache_add_active_or_unevictable(new_page, vma); 2512fce144b4SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 2513ba76149fSAndrea Arcangeli set_pmd_at(mm, address, pmd, _pmd); 2514b113da65SDavid Miller update_mmu_cache_pmd(vma, address, pmd); 2515c4088ebdSKirill A. Shutemov spin_unlock(pmd_ptl); 2516ba76149fSAndrea Arcangeli 2517ba76149fSAndrea Arcangeli *hpage = NULL; 2518420256efSXiao Guangrong 2519ba76149fSAndrea Arcangeli khugepaged_pages_collapsed++; 2520ce83d217SAndrea Arcangeli out_up_write: 2521ba76149fSAndrea Arcangeli up_write(&mm->mmap_sem); 25220bbbc0b3SAndrea Arcangeli return; 25230bbbc0b3SAndrea Arcangeli 2524ce83d217SAndrea Arcangeli out: 252500501b53SJohannes Weiner mem_cgroup_cancel_charge(new_page, memcg); 2526ce83d217SAndrea Arcangeli goto out_up_write; 2527ba76149fSAndrea Arcangeli } 2528ba76149fSAndrea Arcangeli 2529ba76149fSAndrea Arcangeli static int khugepaged_scan_pmd(struct mm_struct *mm, 2530ba76149fSAndrea Arcangeli struct vm_area_struct *vma, 2531ba76149fSAndrea Arcangeli unsigned long address, 2532ba76149fSAndrea Arcangeli struct page **hpage) 2533ba76149fSAndrea Arcangeli { 2534ba76149fSAndrea Arcangeli pmd_t *pmd; 2535ba76149fSAndrea Arcangeli pte_t *pte, *_pte; 253610359213SEbru Akagunduz int ret = 0, none = 0; 2537ba76149fSAndrea Arcangeli struct page *page; 2538ba76149fSAndrea Arcangeli unsigned long _address; 2539ba76149fSAndrea Arcangeli spinlock_t *ptl; 254000ef2d2fSDavid Rientjes int node = NUMA_NO_NODE; 254110359213SEbru Akagunduz bool writable = false, referenced = false; 2542ba76149fSAndrea Arcangeli 2543ba76149fSAndrea Arcangeli VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2544ba76149fSAndrea Arcangeli 25456219049aSBob Liu pmd = mm_find_pmd(mm, address); 25466219049aSBob Liu if (!pmd) 2547ba76149fSAndrea Arcangeli goto out; 2548ba76149fSAndrea Arcangeli 25499f1b868aSBob Liu memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 2550ba76149fSAndrea Arcangeli pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2551ba76149fSAndrea Arcangeli for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 2552ba76149fSAndrea Arcangeli _pte++, _address += PAGE_SIZE) { 2553ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 2554ba76149fSAndrea Arcangeli if (pte_none(pteval)) { 2555ba76149fSAndrea Arcangeli if (++none <= khugepaged_max_ptes_none) 2556ba76149fSAndrea Arcangeli continue; 2557ba76149fSAndrea Arcangeli else 2558ba76149fSAndrea Arcangeli goto out_unmap; 2559ba76149fSAndrea Arcangeli } 256010359213SEbru Akagunduz if (!pte_present(pteval)) 2561ba76149fSAndrea Arcangeli goto out_unmap; 256210359213SEbru Akagunduz if (pte_write(pteval)) 256310359213SEbru Akagunduz writable = true; 256410359213SEbru Akagunduz 2565ba76149fSAndrea Arcangeli page = vm_normal_page(vma, _address, pteval); 2566ba76149fSAndrea Arcangeli if (unlikely(!page)) 2567ba76149fSAndrea Arcangeli goto out_unmap; 25685c4b4be3SAndi Kleen /* 25699f1b868aSBob Liu * Record which node the original page is from and save this 25709f1b868aSBob Liu * information to khugepaged_node_load[]. 25719f1b868aSBob Liu * Khupaged will allocate hugepage from the node has the max 25729f1b868aSBob Liu * hit record. 25735c4b4be3SAndi Kleen */ 25745c4b4be3SAndi Kleen node = page_to_nid(page); 257514a4e214SDavid Rientjes if (khugepaged_scan_abort(node)) 257614a4e214SDavid Rientjes goto out_unmap; 25779f1b868aSBob Liu khugepaged_node_load[node]++; 2578309381feSSasha Levin VM_BUG_ON_PAGE(PageCompound(page), page); 2579ba76149fSAndrea Arcangeli if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) 2580ba76149fSAndrea Arcangeli goto out_unmap; 258110359213SEbru Akagunduz /* 258210359213SEbru Akagunduz * cannot use mapcount: can't collapse if there's a gup pin. 258310359213SEbru Akagunduz * The page must only be referenced by the scanned process 258410359213SEbru Akagunduz * and page swap cache. 258510359213SEbru Akagunduz */ 258610359213SEbru Akagunduz if (page_count(page) != 1 + !!PageSwapCache(page)) 2587ba76149fSAndrea Arcangeli goto out_unmap; 25888ee53820SAndrea Arcangeli if (pte_young(pteval) || PageReferenced(page) || 25898ee53820SAndrea Arcangeli mmu_notifier_test_young(vma->vm_mm, address)) 259010359213SEbru Akagunduz referenced = true; 2591ba76149fSAndrea Arcangeli } 259210359213SEbru Akagunduz if (referenced && writable) 2593ba76149fSAndrea Arcangeli ret = 1; 2594ba76149fSAndrea Arcangeli out_unmap: 2595ba76149fSAndrea Arcangeli pte_unmap_unlock(pte, ptl); 25969f1b868aSBob Liu if (ret) { 25979f1b868aSBob Liu node = khugepaged_find_target_node(); 2598ce83d217SAndrea Arcangeli /* collapse_huge_page will return with the mmap_sem released */ 25995c4b4be3SAndi Kleen collapse_huge_page(mm, address, hpage, vma, node); 26009f1b868aSBob Liu } 2601ba76149fSAndrea Arcangeli out: 2602ba76149fSAndrea Arcangeli return ret; 2603ba76149fSAndrea Arcangeli } 2604ba76149fSAndrea Arcangeli 2605ba76149fSAndrea Arcangeli static void collect_mm_slot(struct mm_slot *mm_slot) 2606ba76149fSAndrea Arcangeli { 2607ba76149fSAndrea Arcangeli struct mm_struct *mm = mm_slot->mm; 2608ba76149fSAndrea Arcangeli 2609b9980cdcSHugh Dickins VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2610ba76149fSAndrea Arcangeli 2611ba76149fSAndrea Arcangeli if (khugepaged_test_exit(mm)) { 2612ba76149fSAndrea Arcangeli /* free mm_slot */ 261343b5fbbdSSasha Levin hash_del(&mm_slot->hash); 2614ba76149fSAndrea Arcangeli list_del(&mm_slot->mm_node); 2615ba76149fSAndrea Arcangeli 2616ba76149fSAndrea Arcangeli /* 2617ba76149fSAndrea Arcangeli * Not strictly needed because the mm exited already. 2618ba76149fSAndrea Arcangeli * 2619ba76149fSAndrea Arcangeli * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2620ba76149fSAndrea Arcangeli */ 2621ba76149fSAndrea Arcangeli 2622ba76149fSAndrea Arcangeli /* khugepaged_mm_lock actually not necessary for the below */ 2623ba76149fSAndrea Arcangeli free_mm_slot(mm_slot); 2624ba76149fSAndrea Arcangeli mmdrop(mm); 2625ba76149fSAndrea Arcangeli } 2626ba76149fSAndrea Arcangeli } 2627ba76149fSAndrea Arcangeli 2628ba76149fSAndrea Arcangeli static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 2629ba76149fSAndrea Arcangeli struct page **hpage) 26302f1da642SH Hartley Sweeten __releases(&khugepaged_mm_lock) 26312f1da642SH Hartley Sweeten __acquires(&khugepaged_mm_lock) 2632ba76149fSAndrea Arcangeli { 2633ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 2634ba76149fSAndrea Arcangeli struct mm_struct *mm; 2635ba76149fSAndrea Arcangeli struct vm_area_struct *vma; 2636ba76149fSAndrea Arcangeli int progress = 0; 2637ba76149fSAndrea Arcangeli 2638ba76149fSAndrea Arcangeli VM_BUG_ON(!pages); 2639b9980cdcSHugh Dickins VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2640ba76149fSAndrea Arcangeli 2641ba76149fSAndrea Arcangeli if (khugepaged_scan.mm_slot) 2642ba76149fSAndrea Arcangeli mm_slot = khugepaged_scan.mm_slot; 2643ba76149fSAndrea Arcangeli else { 2644ba76149fSAndrea Arcangeli mm_slot = list_entry(khugepaged_scan.mm_head.next, 2645ba76149fSAndrea Arcangeli struct mm_slot, mm_node); 2646ba76149fSAndrea Arcangeli khugepaged_scan.address = 0; 2647ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = mm_slot; 2648ba76149fSAndrea Arcangeli } 2649ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 2650ba76149fSAndrea Arcangeli 2651ba76149fSAndrea Arcangeli mm = mm_slot->mm; 2652ba76149fSAndrea Arcangeli down_read(&mm->mmap_sem); 2653ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) 2654ba76149fSAndrea Arcangeli vma = NULL; 2655ba76149fSAndrea Arcangeli else 2656ba76149fSAndrea Arcangeli vma = find_vma(mm, khugepaged_scan.address); 2657ba76149fSAndrea Arcangeli 2658ba76149fSAndrea Arcangeli progress++; 2659ba76149fSAndrea Arcangeli for (; vma; vma = vma->vm_next) { 2660ba76149fSAndrea Arcangeli unsigned long hstart, hend; 2661ba76149fSAndrea Arcangeli 2662ba76149fSAndrea Arcangeli cond_resched(); 2663ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) { 2664ba76149fSAndrea Arcangeli progress++; 2665ba76149fSAndrea Arcangeli break; 2666ba76149fSAndrea Arcangeli } 2667fa475e51SBob Liu if (!hugepage_vma_check(vma)) { 2668a7d6e4ecSAndrea Arcangeli skip: 2669ba76149fSAndrea Arcangeli progress++; 2670ba76149fSAndrea Arcangeli continue; 2671ba76149fSAndrea Arcangeli } 2672ba76149fSAndrea Arcangeli hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2673ba76149fSAndrea Arcangeli hend = vma->vm_end & HPAGE_PMD_MASK; 2674a7d6e4ecSAndrea Arcangeli if (hstart >= hend) 2675a7d6e4ecSAndrea Arcangeli goto skip; 2676a7d6e4ecSAndrea Arcangeli if (khugepaged_scan.address > hend) 2677a7d6e4ecSAndrea Arcangeli goto skip; 2678ba76149fSAndrea Arcangeli if (khugepaged_scan.address < hstart) 2679ba76149fSAndrea Arcangeli khugepaged_scan.address = hstart; 2680a7d6e4ecSAndrea Arcangeli VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2681ba76149fSAndrea Arcangeli 2682ba76149fSAndrea Arcangeli while (khugepaged_scan.address < hend) { 2683ba76149fSAndrea Arcangeli int ret; 2684ba76149fSAndrea Arcangeli cond_resched(); 2685ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) 2686ba76149fSAndrea Arcangeli goto breakouterloop; 2687ba76149fSAndrea Arcangeli 2688ba76149fSAndrea Arcangeli VM_BUG_ON(khugepaged_scan.address < hstart || 2689ba76149fSAndrea Arcangeli khugepaged_scan.address + HPAGE_PMD_SIZE > 2690ba76149fSAndrea Arcangeli hend); 2691ba76149fSAndrea Arcangeli ret = khugepaged_scan_pmd(mm, vma, 2692ba76149fSAndrea Arcangeli khugepaged_scan.address, 2693ba76149fSAndrea Arcangeli hpage); 2694ba76149fSAndrea Arcangeli /* move to next address */ 2695ba76149fSAndrea Arcangeli khugepaged_scan.address += HPAGE_PMD_SIZE; 2696ba76149fSAndrea Arcangeli progress += HPAGE_PMD_NR; 2697ba76149fSAndrea Arcangeli if (ret) 2698ba76149fSAndrea Arcangeli /* we released mmap_sem so break loop */ 2699ba76149fSAndrea Arcangeli goto breakouterloop_mmap_sem; 2700ba76149fSAndrea Arcangeli if (progress >= pages) 2701ba76149fSAndrea Arcangeli goto breakouterloop; 2702ba76149fSAndrea Arcangeli } 2703ba76149fSAndrea Arcangeli } 2704ba76149fSAndrea Arcangeli breakouterloop: 2705ba76149fSAndrea Arcangeli up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ 2706ba76149fSAndrea Arcangeli breakouterloop_mmap_sem: 2707ba76149fSAndrea Arcangeli 2708ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2709a7d6e4ecSAndrea Arcangeli VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2710ba76149fSAndrea Arcangeli /* 2711ba76149fSAndrea Arcangeli * Release the current mm_slot if this mm is about to die, or 2712ba76149fSAndrea Arcangeli * if we scanned all vmas of this mm. 2713ba76149fSAndrea Arcangeli */ 2714ba76149fSAndrea Arcangeli if (khugepaged_test_exit(mm) || !vma) { 2715ba76149fSAndrea Arcangeli /* 2716ba76149fSAndrea Arcangeli * Make sure that if mm_users is reaching zero while 2717ba76149fSAndrea Arcangeli * khugepaged runs here, khugepaged_exit will find 2718ba76149fSAndrea Arcangeli * mm_slot not pointing to the exiting mm. 2719ba76149fSAndrea Arcangeli */ 2720ba76149fSAndrea Arcangeli if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2721ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = list_entry( 2722ba76149fSAndrea Arcangeli mm_slot->mm_node.next, 2723ba76149fSAndrea Arcangeli struct mm_slot, mm_node); 2724ba76149fSAndrea Arcangeli khugepaged_scan.address = 0; 2725ba76149fSAndrea Arcangeli } else { 2726ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = NULL; 2727ba76149fSAndrea Arcangeli khugepaged_full_scans++; 2728ba76149fSAndrea Arcangeli } 2729ba76149fSAndrea Arcangeli 2730ba76149fSAndrea Arcangeli collect_mm_slot(mm_slot); 2731ba76149fSAndrea Arcangeli } 2732ba76149fSAndrea Arcangeli 2733ba76149fSAndrea Arcangeli return progress; 2734ba76149fSAndrea Arcangeli } 2735ba76149fSAndrea Arcangeli 2736ba76149fSAndrea Arcangeli static int khugepaged_has_work(void) 2737ba76149fSAndrea Arcangeli { 2738ba76149fSAndrea Arcangeli return !list_empty(&khugepaged_scan.mm_head) && 2739ba76149fSAndrea Arcangeli khugepaged_enabled(); 2740ba76149fSAndrea Arcangeli } 2741ba76149fSAndrea Arcangeli 2742ba76149fSAndrea Arcangeli static int khugepaged_wait_event(void) 2743ba76149fSAndrea Arcangeli { 2744ba76149fSAndrea Arcangeli return !list_empty(&khugepaged_scan.mm_head) || 27452017c0bfSXiao Guangrong kthread_should_stop(); 2746ba76149fSAndrea Arcangeli } 2747ba76149fSAndrea Arcangeli 2748d516904bSXiao Guangrong static void khugepaged_do_scan(void) 2749d516904bSXiao Guangrong { 2750d516904bSXiao Guangrong struct page *hpage = NULL; 2751ba76149fSAndrea Arcangeli unsigned int progress = 0, pass_through_head = 0; 2752ba76149fSAndrea Arcangeli unsigned int pages = khugepaged_pages_to_scan; 2753d516904bSXiao Guangrong bool wait = true; 2754ba76149fSAndrea Arcangeli 2755ba76149fSAndrea Arcangeli barrier(); /* write khugepaged_pages_to_scan to local stack */ 2756ba76149fSAndrea Arcangeli 2757ba76149fSAndrea Arcangeli while (progress < pages) { 275826234f36SXiao Guangrong if (!khugepaged_prealloc_page(&hpage, &wait)) 275926234f36SXiao Guangrong break; 2760d516904bSXiao Guangrong 2761420256efSXiao Guangrong cond_resched(); 2762ba76149fSAndrea Arcangeli 2763878aee7dSAndrea Arcangeli if (unlikely(kthread_should_stop() || freezing(current))) 2764878aee7dSAndrea Arcangeli break; 2765878aee7dSAndrea Arcangeli 2766ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2767ba76149fSAndrea Arcangeli if (!khugepaged_scan.mm_slot) 2768ba76149fSAndrea Arcangeli pass_through_head++; 2769ba76149fSAndrea Arcangeli if (khugepaged_has_work() && 2770ba76149fSAndrea Arcangeli pass_through_head < 2) 2771ba76149fSAndrea Arcangeli progress += khugepaged_scan_mm_slot(pages - progress, 2772d516904bSXiao Guangrong &hpage); 2773ba76149fSAndrea Arcangeli else 2774ba76149fSAndrea Arcangeli progress = pages; 2775ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 2776ba76149fSAndrea Arcangeli } 2777ba76149fSAndrea Arcangeli 2778d516904bSXiao Guangrong if (!IS_ERR_OR_NULL(hpage)) 2779d516904bSXiao Guangrong put_page(hpage); 2780ba76149fSAndrea Arcangeli } 27810bbbc0b3SAndrea Arcangeli 27822017c0bfSXiao Guangrong static void khugepaged_wait_work(void) 27832017c0bfSXiao Guangrong { 27842017c0bfSXiao Guangrong try_to_freeze(); 27852017c0bfSXiao Guangrong 27862017c0bfSXiao Guangrong if (khugepaged_has_work()) { 27872017c0bfSXiao Guangrong if (!khugepaged_scan_sleep_millisecs) 27882017c0bfSXiao Guangrong return; 27892017c0bfSXiao Guangrong 27902017c0bfSXiao Guangrong wait_event_freezable_timeout(khugepaged_wait, 27912017c0bfSXiao Guangrong kthread_should_stop(), 27922017c0bfSXiao Guangrong msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); 27932017c0bfSXiao Guangrong return; 27942017c0bfSXiao Guangrong } 27952017c0bfSXiao Guangrong 27962017c0bfSXiao Guangrong if (khugepaged_enabled()) 27972017c0bfSXiao Guangrong wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 27982017c0bfSXiao Guangrong } 27992017c0bfSXiao Guangrong 2800ba76149fSAndrea Arcangeli static int khugepaged(void *none) 2801ba76149fSAndrea Arcangeli { 2802ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 2803ba76149fSAndrea Arcangeli 2804878aee7dSAndrea Arcangeli set_freezable(); 28058698a745SDongsheng Yang set_user_nice(current, MAX_NICE); 2806ba76149fSAndrea Arcangeli 2807b7231789SXiao Guangrong while (!kthread_should_stop()) { 2808b7231789SXiao Guangrong khugepaged_do_scan(); 2809b7231789SXiao Guangrong khugepaged_wait_work(); 2810b7231789SXiao Guangrong } 2811ba76149fSAndrea Arcangeli 2812ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2813ba76149fSAndrea Arcangeli mm_slot = khugepaged_scan.mm_slot; 2814ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = NULL; 2815ba76149fSAndrea Arcangeli if (mm_slot) 2816ba76149fSAndrea Arcangeli collect_mm_slot(mm_slot); 2817ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 2818ba76149fSAndrea Arcangeli return 0; 2819ba76149fSAndrea Arcangeli } 2820ba76149fSAndrea Arcangeli 2821c5a647d0SKirill A. Shutemov static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2822c5a647d0SKirill A. Shutemov unsigned long haddr, pmd_t *pmd) 2823c5a647d0SKirill A. Shutemov { 2824c5a647d0SKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2825c5a647d0SKirill A. Shutemov pgtable_t pgtable; 2826c5a647d0SKirill A. Shutemov pmd_t _pmd; 2827c5a647d0SKirill A. Shutemov int i; 2828c5a647d0SKirill A. Shutemov 282934ee645eSJoerg Roedel pmdp_clear_flush_notify(vma, haddr, pmd); 2830c5a647d0SKirill A. Shutemov /* leave pmd empty until pte is filled */ 2831c5a647d0SKirill A. Shutemov 28326b0b50b0SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2833c5a647d0SKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2834c5a647d0SKirill A. Shutemov 2835c5a647d0SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2836c5a647d0SKirill A. Shutemov pte_t *pte, entry; 2837c5a647d0SKirill A. Shutemov entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2838c5a647d0SKirill A. Shutemov entry = pte_mkspecial(entry); 2839c5a647d0SKirill A. Shutemov pte = pte_offset_map(&_pmd, haddr); 2840c5a647d0SKirill A. Shutemov VM_BUG_ON(!pte_none(*pte)); 2841c5a647d0SKirill A. Shutemov set_pte_at(mm, haddr, pte, entry); 2842c5a647d0SKirill A. Shutemov pte_unmap(pte); 2843c5a647d0SKirill A. Shutemov } 2844c5a647d0SKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2845c5a647d0SKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 284697ae1749SKirill A. Shutemov put_huge_zero_page(); 2847c5a647d0SKirill A. Shutemov } 2848c5a647d0SKirill A. Shutemov 2849e180377fSKirill A. Shutemov void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, 2850e180377fSKirill A. Shutemov pmd_t *pmd) 285171e3aac0SAndrea Arcangeli { 2852c4088ebdSKirill A. Shutemov spinlock_t *ptl; 285371e3aac0SAndrea Arcangeli struct page *page; 2854e180377fSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2855c5a647d0SKirill A. Shutemov unsigned long haddr = address & HPAGE_PMD_MASK; 2856c5a647d0SKirill A. Shutemov unsigned long mmun_start; /* For mmu_notifiers */ 2857c5a647d0SKirill A. Shutemov unsigned long mmun_end; /* For mmu_notifiers */ 2858e180377fSKirill A. Shutemov 2859e180377fSKirill A. Shutemov BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); 286071e3aac0SAndrea Arcangeli 2861c5a647d0SKirill A. Shutemov mmun_start = haddr; 2862c5a647d0SKirill A. Shutemov mmun_end = haddr + HPAGE_PMD_SIZE; 2863750e8165SHugh Dickins again: 2864c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2865c4088ebdSKirill A. Shutemov ptl = pmd_lock(mm, pmd); 286671e3aac0SAndrea Arcangeli if (unlikely(!pmd_trans_huge(*pmd))) { 2867c4088ebdSKirill A. Shutemov spin_unlock(ptl); 2868c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2869c5a647d0SKirill A. Shutemov return; 2870c5a647d0SKirill A. Shutemov } 2871c5a647d0SKirill A. Shutemov if (is_huge_zero_pmd(*pmd)) { 2872c5a647d0SKirill A. Shutemov __split_huge_zero_page_pmd(vma, haddr, pmd); 2873c4088ebdSKirill A. Shutemov spin_unlock(ptl); 2874c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 287571e3aac0SAndrea Arcangeli return; 287671e3aac0SAndrea Arcangeli } 287771e3aac0SAndrea Arcangeli page = pmd_page(*pmd); 2878309381feSSasha Levin VM_BUG_ON_PAGE(!page_count(page), page); 287971e3aac0SAndrea Arcangeli get_page(page); 2880c4088ebdSKirill A. Shutemov spin_unlock(ptl); 2881c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 288271e3aac0SAndrea Arcangeli 288371e3aac0SAndrea Arcangeli split_huge_page(page); 288471e3aac0SAndrea Arcangeli 288571e3aac0SAndrea Arcangeli put_page(page); 2886750e8165SHugh Dickins 2887750e8165SHugh Dickins /* 2888750e8165SHugh Dickins * We don't always have down_write of mmap_sem here: a racing 2889750e8165SHugh Dickins * do_huge_pmd_wp_page() might have copied-on-write to another 2890750e8165SHugh Dickins * huge page before our split_huge_page() got the anon_vma lock. 2891750e8165SHugh Dickins */ 2892750e8165SHugh Dickins if (unlikely(pmd_trans_huge(*pmd))) 2893750e8165SHugh Dickins goto again; 289471e3aac0SAndrea Arcangeli } 289594fcc585SAndrea Arcangeli 2896e180377fSKirill A. Shutemov void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address, 2897e180377fSKirill A. Shutemov pmd_t *pmd) 2898e180377fSKirill A. Shutemov { 2899e180377fSKirill A. Shutemov struct vm_area_struct *vma; 2900e180377fSKirill A. Shutemov 2901e180377fSKirill A. Shutemov vma = find_vma(mm, address); 2902e180377fSKirill A. Shutemov BUG_ON(vma == NULL); 2903e180377fSKirill A. Shutemov split_huge_page_pmd(vma, address, pmd); 2904e180377fSKirill A. Shutemov } 2905e180377fSKirill A. Shutemov 290694fcc585SAndrea Arcangeli static void split_huge_page_address(struct mm_struct *mm, 290794fcc585SAndrea Arcangeli unsigned long address) 290894fcc585SAndrea Arcangeli { 2909f72e7dcdSHugh Dickins pgd_t *pgd; 2910f72e7dcdSHugh Dickins pud_t *pud; 291194fcc585SAndrea Arcangeli pmd_t *pmd; 291294fcc585SAndrea Arcangeli 291394fcc585SAndrea Arcangeli VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); 291494fcc585SAndrea Arcangeli 2915f72e7dcdSHugh Dickins pgd = pgd_offset(mm, address); 2916f72e7dcdSHugh Dickins if (!pgd_present(*pgd)) 2917f72e7dcdSHugh Dickins return; 2918f72e7dcdSHugh Dickins 2919f72e7dcdSHugh Dickins pud = pud_offset(pgd, address); 2920f72e7dcdSHugh Dickins if (!pud_present(*pud)) 2921f72e7dcdSHugh Dickins return; 2922f72e7dcdSHugh Dickins 2923f72e7dcdSHugh Dickins pmd = pmd_offset(pud, address); 2924f72e7dcdSHugh Dickins if (!pmd_present(*pmd)) 292594fcc585SAndrea Arcangeli return; 292694fcc585SAndrea Arcangeli /* 292794fcc585SAndrea Arcangeli * Caller holds the mmap_sem write mode, so a huge pmd cannot 292894fcc585SAndrea Arcangeli * materialize from under us. 292994fcc585SAndrea Arcangeli */ 2930e180377fSKirill A. Shutemov split_huge_page_pmd_mm(mm, address, pmd); 293194fcc585SAndrea Arcangeli } 293294fcc585SAndrea Arcangeli 293394fcc585SAndrea Arcangeli void __vma_adjust_trans_huge(struct vm_area_struct *vma, 293494fcc585SAndrea Arcangeli unsigned long start, 293594fcc585SAndrea Arcangeli unsigned long end, 293694fcc585SAndrea Arcangeli long adjust_next) 293794fcc585SAndrea Arcangeli { 293894fcc585SAndrea Arcangeli /* 293994fcc585SAndrea Arcangeli * If the new start address isn't hpage aligned and it could 294094fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 294194fcc585SAndrea Arcangeli * an huge pmd. 294294fcc585SAndrea Arcangeli */ 294394fcc585SAndrea Arcangeli if (start & ~HPAGE_PMD_MASK && 294494fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) >= vma->vm_start && 294594fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 294694fcc585SAndrea Arcangeli split_huge_page_address(vma->vm_mm, start); 294794fcc585SAndrea Arcangeli 294894fcc585SAndrea Arcangeli /* 294994fcc585SAndrea Arcangeli * If the new end address isn't hpage aligned and it could 295094fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 295194fcc585SAndrea Arcangeli * an huge pmd. 295294fcc585SAndrea Arcangeli */ 295394fcc585SAndrea Arcangeli if (end & ~HPAGE_PMD_MASK && 295494fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) >= vma->vm_start && 295594fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 295694fcc585SAndrea Arcangeli split_huge_page_address(vma->vm_mm, end); 295794fcc585SAndrea Arcangeli 295894fcc585SAndrea Arcangeli /* 295994fcc585SAndrea Arcangeli * If we're also updating the vma->vm_next->vm_start, if the new 296094fcc585SAndrea Arcangeli * vm_next->vm_start isn't page aligned and it could previously 296194fcc585SAndrea Arcangeli * contain an hugepage: check if we need to split an huge pmd. 296294fcc585SAndrea Arcangeli */ 296394fcc585SAndrea Arcangeli if (adjust_next > 0) { 296494fcc585SAndrea Arcangeli struct vm_area_struct *next = vma->vm_next; 296594fcc585SAndrea Arcangeli unsigned long nstart = next->vm_start; 296694fcc585SAndrea Arcangeli nstart += adjust_next << PAGE_SHIFT; 296794fcc585SAndrea Arcangeli if (nstart & ~HPAGE_PMD_MASK && 296894fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) >= next->vm_start && 296994fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 297094fcc585SAndrea Arcangeli split_huge_page_address(next->vm_mm, nstart); 297194fcc585SAndrea Arcangeli } 297294fcc585SAndrea Arcangeli } 2973