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 871e3aac0SAndrea Arcangeli #include <linux/mm.h> 971e3aac0SAndrea Arcangeli #include <linux/sched.h> 1071e3aac0SAndrea Arcangeli #include <linux/highmem.h> 1171e3aac0SAndrea Arcangeli #include <linux/hugetlb.h> 1271e3aac0SAndrea Arcangeli #include <linux/mmu_notifier.h> 1371e3aac0SAndrea Arcangeli #include <linux/rmap.h> 1471e3aac0SAndrea Arcangeli #include <linux/swap.h> 1597ae1749SKirill A. Shutemov #include <linux/shrinker.h> 16ba76149fSAndrea Arcangeli #include <linux/mm_inline.h> 17ba76149fSAndrea Arcangeli #include <linux/kthread.h> 18ba76149fSAndrea Arcangeli #include <linux/khugepaged.h> 19878aee7dSAndrea Arcangeli #include <linux/freezer.h> 20a664b2d8SAndrea Arcangeli #include <linux/mman.h> 21325adeb5SRalf Baechle #include <linux/pagemap.h> 2297ae1749SKirill A. Shutemov 2371e3aac0SAndrea Arcangeli #include <asm/tlb.h> 2471e3aac0SAndrea Arcangeli #include <asm/pgalloc.h> 2571e3aac0SAndrea Arcangeli #include "internal.h" 2671e3aac0SAndrea Arcangeli 27ba76149fSAndrea Arcangeli /* 28ba76149fSAndrea Arcangeli * By default transparent hugepage support is enabled for all mappings 29ba76149fSAndrea Arcangeli * and khugepaged scans all mappings. Defrag is only invoked by 30ba76149fSAndrea Arcangeli * khugepaged hugepage allocations and by page faults inside 31ba76149fSAndrea Arcangeli * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived 32ba76149fSAndrea Arcangeli * allocations. 33ba76149fSAndrea Arcangeli */ 3471e3aac0SAndrea Arcangeli unsigned long transparent_hugepage_flags __read_mostly = 3513ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 36ba76149fSAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_FLAG)| 3713ece886SAndrea Arcangeli #endif 3813ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 3913ece886SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 4013ece886SAndrea Arcangeli #endif 41d39d33c3SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| 42*79da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 43*79da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 44ba76149fSAndrea Arcangeli 45ba76149fSAndrea Arcangeli /* default scan 8*512 pte (or vmas) every 30 second */ 46ba76149fSAndrea Arcangeli static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; 47ba76149fSAndrea Arcangeli static unsigned int khugepaged_pages_collapsed; 48ba76149fSAndrea Arcangeli static unsigned int khugepaged_full_scans; 49ba76149fSAndrea Arcangeli static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 50ba76149fSAndrea Arcangeli /* during fragmentation poll the hugepage allocator once every minute */ 51ba76149fSAndrea Arcangeli static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 52ba76149fSAndrea Arcangeli static struct task_struct *khugepaged_thread __read_mostly; 53ba76149fSAndrea Arcangeli static DEFINE_MUTEX(khugepaged_mutex); 54ba76149fSAndrea Arcangeli static DEFINE_SPINLOCK(khugepaged_mm_lock); 55ba76149fSAndrea Arcangeli static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 56ba76149fSAndrea Arcangeli /* 57ba76149fSAndrea Arcangeli * default collapse hugepages if there is at least one pte mapped like 58ba76149fSAndrea Arcangeli * it would have happened if the vma was large enough during page 59ba76149fSAndrea Arcangeli * fault. 60ba76149fSAndrea Arcangeli */ 61ba76149fSAndrea Arcangeli static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; 62ba76149fSAndrea Arcangeli 63ba76149fSAndrea Arcangeli static int khugepaged(void *none); 64ba76149fSAndrea Arcangeli static int mm_slots_hash_init(void); 65ba76149fSAndrea Arcangeli static int khugepaged_slab_init(void); 66ba76149fSAndrea Arcangeli static void khugepaged_slab_free(void); 67ba76149fSAndrea Arcangeli 68ba76149fSAndrea Arcangeli #define MM_SLOTS_HASH_HEADS 1024 69ba76149fSAndrea Arcangeli static struct hlist_head *mm_slots_hash __read_mostly; 70ba76149fSAndrea Arcangeli static struct kmem_cache *mm_slot_cache __read_mostly; 71ba76149fSAndrea Arcangeli 72ba76149fSAndrea Arcangeli /** 73ba76149fSAndrea Arcangeli * struct mm_slot - hash lookup from mm to mm_slot 74ba76149fSAndrea Arcangeli * @hash: hash collision list 75ba76149fSAndrea Arcangeli * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 76ba76149fSAndrea Arcangeli * @mm: the mm that this information is valid for 77ba76149fSAndrea Arcangeli */ 78ba76149fSAndrea Arcangeli struct mm_slot { 79ba76149fSAndrea Arcangeli struct hlist_node hash; 80ba76149fSAndrea Arcangeli struct list_head mm_node; 81ba76149fSAndrea Arcangeli struct mm_struct *mm; 82ba76149fSAndrea Arcangeli }; 83ba76149fSAndrea Arcangeli 84ba76149fSAndrea Arcangeli /** 85ba76149fSAndrea Arcangeli * struct khugepaged_scan - cursor for scanning 86ba76149fSAndrea Arcangeli * @mm_head: the head of the mm list to scan 87ba76149fSAndrea Arcangeli * @mm_slot: the current mm_slot we are scanning 88ba76149fSAndrea Arcangeli * @address: the next address inside that to be scanned 89ba76149fSAndrea Arcangeli * 90ba76149fSAndrea Arcangeli * There is only the one khugepaged_scan instance of this cursor structure. 91ba76149fSAndrea Arcangeli */ 92ba76149fSAndrea Arcangeli struct khugepaged_scan { 93ba76149fSAndrea Arcangeli struct list_head mm_head; 94ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 95ba76149fSAndrea Arcangeli unsigned long address; 962f1da642SH Hartley Sweeten }; 972f1da642SH Hartley Sweeten static struct khugepaged_scan khugepaged_scan = { 98ba76149fSAndrea Arcangeli .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 99ba76149fSAndrea Arcangeli }; 100ba76149fSAndrea Arcangeli 101f000565aSAndrea Arcangeli 102f000565aSAndrea Arcangeli static int set_recommended_min_free_kbytes(void) 103f000565aSAndrea Arcangeli { 104f000565aSAndrea Arcangeli struct zone *zone; 105f000565aSAndrea Arcangeli int nr_zones = 0; 106f000565aSAndrea Arcangeli unsigned long recommended_min; 107f000565aSAndrea Arcangeli extern int min_free_kbytes; 108f000565aSAndrea Arcangeli 10917c230afSXiao Guangrong if (!khugepaged_enabled()) 110f000565aSAndrea Arcangeli return 0; 111f000565aSAndrea Arcangeli 112f000565aSAndrea Arcangeli for_each_populated_zone(zone) 113f000565aSAndrea Arcangeli nr_zones++; 114f000565aSAndrea Arcangeli 115f000565aSAndrea Arcangeli /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */ 116f000565aSAndrea Arcangeli recommended_min = pageblock_nr_pages * nr_zones * 2; 117f000565aSAndrea Arcangeli 118f000565aSAndrea Arcangeli /* 119f000565aSAndrea Arcangeli * Make sure that on average at least two pageblocks are almost free 120f000565aSAndrea Arcangeli * of another type, one for a migratetype to fall back to and a 121f000565aSAndrea Arcangeli * second to avoid subsequent fallbacks of other types There are 3 122f000565aSAndrea Arcangeli * MIGRATE_TYPES we care about. 123f000565aSAndrea Arcangeli */ 124f000565aSAndrea Arcangeli recommended_min += pageblock_nr_pages * nr_zones * 125f000565aSAndrea Arcangeli MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 126f000565aSAndrea Arcangeli 127f000565aSAndrea Arcangeli /* don't ever allow to reserve more than 5% of the lowmem */ 128f000565aSAndrea Arcangeli recommended_min = min(recommended_min, 129f000565aSAndrea Arcangeli (unsigned long) nr_free_buffer_pages() / 20); 130f000565aSAndrea Arcangeli recommended_min <<= (PAGE_SHIFT-10); 131f000565aSAndrea Arcangeli 132f000565aSAndrea Arcangeli if (recommended_min > min_free_kbytes) 133f000565aSAndrea Arcangeli min_free_kbytes = recommended_min; 134f000565aSAndrea Arcangeli setup_per_zone_wmarks(); 135f000565aSAndrea Arcangeli return 0; 136f000565aSAndrea Arcangeli } 137f000565aSAndrea Arcangeli late_initcall(set_recommended_min_free_kbytes); 138f000565aSAndrea Arcangeli 139ba76149fSAndrea Arcangeli static int start_khugepaged(void) 140ba76149fSAndrea Arcangeli { 141ba76149fSAndrea Arcangeli int err = 0; 142ba76149fSAndrea Arcangeli if (khugepaged_enabled()) { 143ba76149fSAndrea Arcangeli if (!khugepaged_thread) 144ba76149fSAndrea Arcangeli khugepaged_thread = kthread_run(khugepaged, NULL, 145ba76149fSAndrea Arcangeli "khugepaged"); 146ba76149fSAndrea Arcangeli if (unlikely(IS_ERR(khugepaged_thread))) { 147ba76149fSAndrea Arcangeli printk(KERN_ERR 148ba76149fSAndrea Arcangeli "khugepaged: kthread_run(khugepaged) failed\n"); 149ba76149fSAndrea Arcangeli err = PTR_ERR(khugepaged_thread); 150ba76149fSAndrea Arcangeli khugepaged_thread = NULL; 151ba76149fSAndrea Arcangeli } 152911891afSXiao Guangrong 153911891afSXiao Guangrong if (!list_empty(&khugepaged_scan.mm_head)) 154ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 155f000565aSAndrea Arcangeli 156f000565aSAndrea Arcangeli set_recommended_min_free_kbytes(); 157911891afSXiao Guangrong } else if (khugepaged_thread) { 158911891afSXiao Guangrong kthread_stop(khugepaged_thread); 159911891afSXiao Guangrong khugepaged_thread = NULL; 160911891afSXiao Guangrong } 161637e3a27SXiao Guangrong 162ba76149fSAndrea Arcangeli return err; 163ba76149fSAndrea Arcangeli } 16471e3aac0SAndrea Arcangeli 16597ae1749SKirill A. Shutemov static atomic_t huge_zero_refcount; 16697ae1749SKirill A. Shutemov static unsigned long huge_zero_pfn __read_mostly; 1674a6c1297SKirill A. Shutemov 1684a6c1297SKirill A. Shutemov static inline bool is_huge_zero_pfn(unsigned long pfn) 1694a6c1297SKirill A. Shutemov { 17097ae1749SKirill A. Shutemov unsigned long zero_pfn = ACCESS_ONCE(huge_zero_pfn); 17197ae1749SKirill A. Shutemov return zero_pfn && pfn == zero_pfn; 1724a6c1297SKirill A. Shutemov } 1734a6c1297SKirill A. Shutemov 1744a6c1297SKirill A. Shutemov static inline bool is_huge_zero_pmd(pmd_t pmd) 1754a6c1297SKirill A. Shutemov { 1764a6c1297SKirill A. Shutemov return is_huge_zero_pfn(pmd_pfn(pmd)); 1774a6c1297SKirill A. Shutemov } 1784a6c1297SKirill A. Shutemov 17997ae1749SKirill A. Shutemov static unsigned long get_huge_zero_page(void) 18097ae1749SKirill A. Shutemov { 18197ae1749SKirill A. Shutemov struct page *zero_page; 18297ae1749SKirill A. Shutemov retry: 18397ae1749SKirill A. Shutemov if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 18497ae1749SKirill A. Shutemov return ACCESS_ONCE(huge_zero_pfn); 18597ae1749SKirill A. Shutemov 18697ae1749SKirill A. Shutemov zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 18797ae1749SKirill A. Shutemov HPAGE_PMD_ORDER); 188d8a8e1f0SKirill A. Shutemov if (!zero_page) { 189d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 19097ae1749SKirill A. Shutemov return 0; 191d8a8e1f0SKirill A. Shutemov } 192d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC); 19397ae1749SKirill A. Shutemov preempt_disable(); 19497ae1749SKirill A. Shutemov if (cmpxchg(&huge_zero_pfn, 0, page_to_pfn(zero_page))) { 19597ae1749SKirill A. Shutemov preempt_enable(); 19697ae1749SKirill A. Shutemov __free_page(zero_page); 19797ae1749SKirill A. Shutemov goto retry; 19897ae1749SKirill A. Shutemov } 19997ae1749SKirill A. Shutemov 20097ae1749SKirill A. Shutemov /* We take additional reference here. It will be put back by shrinker */ 20197ae1749SKirill A. Shutemov atomic_set(&huge_zero_refcount, 2); 20297ae1749SKirill A. Shutemov preempt_enable(); 20397ae1749SKirill A. Shutemov return ACCESS_ONCE(huge_zero_pfn); 20497ae1749SKirill A. Shutemov } 20597ae1749SKirill A. Shutemov 20697ae1749SKirill A. Shutemov static void put_huge_zero_page(void) 20797ae1749SKirill A. Shutemov { 20897ae1749SKirill A. Shutemov /* 20997ae1749SKirill A. Shutemov * Counter should never go to zero here. Only shrinker can put 21097ae1749SKirill A. Shutemov * last reference. 21197ae1749SKirill A. Shutemov */ 21297ae1749SKirill A. Shutemov BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 21397ae1749SKirill A. Shutemov } 21497ae1749SKirill A. Shutemov 21597ae1749SKirill A. Shutemov static int shrink_huge_zero_page(struct shrinker *shrink, 21697ae1749SKirill A. Shutemov struct shrink_control *sc) 21797ae1749SKirill A. Shutemov { 21897ae1749SKirill A. Shutemov if (!sc->nr_to_scan) 21997ae1749SKirill A. Shutemov /* we can free zero page only if last reference remains */ 22097ae1749SKirill A. Shutemov return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 22197ae1749SKirill A. Shutemov 22297ae1749SKirill A. Shutemov if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 22397ae1749SKirill A. Shutemov unsigned long zero_pfn = xchg(&huge_zero_pfn, 0); 22497ae1749SKirill A. Shutemov BUG_ON(zero_pfn == 0); 22597ae1749SKirill A. Shutemov __free_page(__pfn_to_page(zero_pfn)); 22697ae1749SKirill A. Shutemov } 22797ae1749SKirill A. Shutemov 22897ae1749SKirill A. Shutemov return 0; 22997ae1749SKirill A. Shutemov } 23097ae1749SKirill A. Shutemov 23197ae1749SKirill A. Shutemov static struct shrinker huge_zero_page_shrinker = { 23297ae1749SKirill A. Shutemov .shrink = shrink_huge_zero_page, 23397ae1749SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 23497ae1749SKirill A. Shutemov }; 23597ae1749SKirill A. Shutemov 23671e3aac0SAndrea Arcangeli #ifdef CONFIG_SYSFS 237ba76149fSAndrea Arcangeli 23871e3aac0SAndrea Arcangeli static ssize_t double_flag_show(struct kobject *kobj, 23971e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 24071e3aac0SAndrea Arcangeli enum transparent_hugepage_flag enabled, 24171e3aac0SAndrea Arcangeli enum transparent_hugepage_flag req_madv) 24271e3aac0SAndrea Arcangeli { 24371e3aac0SAndrea Arcangeli if (test_bit(enabled, &transparent_hugepage_flags)) { 24471e3aac0SAndrea Arcangeli VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); 24571e3aac0SAndrea Arcangeli return sprintf(buf, "[always] madvise never\n"); 24671e3aac0SAndrea Arcangeli } else if (test_bit(req_madv, &transparent_hugepage_flags)) 24771e3aac0SAndrea Arcangeli return sprintf(buf, "always [madvise] never\n"); 24871e3aac0SAndrea Arcangeli else 24971e3aac0SAndrea Arcangeli return sprintf(buf, "always madvise [never]\n"); 25071e3aac0SAndrea Arcangeli } 25171e3aac0SAndrea Arcangeli static ssize_t double_flag_store(struct kobject *kobj, 25271e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 25371e3aac0SAndrea Arcangeli const char *buf, size_t count, 25471e3aac0SAndrea Arcangeli enum transparent_hugepage_flag enabled, 25571e3aac0SAndrea Arcangeli enum transparent_hugepage_flag req_madv) 25671e3aac0SAndrea Arcangeli { 25771e3aac0SAndrea Arcangeli if (!memcmp("always", buf, 25871e3aac0SAndrea Arcangeli min(sizeof("always")-1, count))) { 25971e3aac0SAndrea Arcangeli set_bit(enabled, &transparent_hugepage_flags); 26071e3aac0SAndrea Arcangeli clear_bit(req_madv, &transparent_hugepage_flags); 26171e3aac0SAndrea Arcangeli } else if (!memcmp("madvise", buf, 26271e3aac0SAndrea Arcangeli min(sizeof("madvise")-1, count))) { 26371e3aac0SAndrea Arcangeli clear_bit(enabled, &transparent_hugepage_flags); 26471e3aac0SAndrea Arcangeli set_bit(req_madv, &transparent_hugepage_flags); 26571e3aac0SAndrea Arcangeli } else if (!memcmp("never", buf, 26671e3aac0SAndrea Arcangeli min(sizeof("never")-1, count))) { 26771e3aac0SAndrea Arcangeli clear_bit(enabled, &transparent_hugepage_flags); 26871e3aac0SAndrea Arcangeli clear_bit(req_madv, &transparent_hugepage_flags); 26971e3aac0SAndrea Arcangeli } else 27071e3aac0SAndrea Arcangeli return -EINVAL; 27171e3aac0SAndrea Arcangeli 27271e3aac0SAndrea Arcangeli return count; 27371e3aac0SAndrea Arcangeli } 27471e3aac0SAndrea Arcangeli 27571e3aac0SAndrea Arcangeli static ssize_t enabled_show(struct kobject *kobj, 27671e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 27771e3aac0SAndrea Arcangeli { 27871e3aac0SAndrea Arcangeli return double_flag_show(kobj, attr, buf, 27971e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_FLAG, 28071e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 28171e3aac0SAndrea Arcangeli } 28271e3aac0SAndrea Arcangeli static ssize_t enabled_store(struct kobject *kobj, 28371e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 28471e3aac0SAndrea Arcangeli const char *buf, size_t count) 28571e3aac0SAndrea Arcangeli { 286ba76149fSAndrea Arcangeli ssize_t ret; 287ba76149fSAndrea Arcangeli 288ba76149fSAndrea Arcangeli ret = double_flag_store(kobj, attr, buf, count, 28971e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_FLAG, 29071e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 291ba76149fSAndrea Arcangeli 292ba76149fSAndrea Arcangeli if (ret > 0) { 293911891afSXiao Guangrong int err; 294911891afSXiao Guangrong 295911891afSXiao Guangrong mutex_lock(&khugepaged_mutex); 296911891afSXiao Guangrong err = start_khugepaged(); 297911891afSXiao Guangrong mutex_unlock(&khugepaged_mutex); 298911891afSXiao Guangrong 299ba76149fSAndrea Arcangeli if (err) 300ba76149fSAndrea Arcangeli ret = err; 301ba76149fSAndrea Arcangeli } 302ba76149fSAndrea Arcangeli 303ba76149fSAndrea Arcangeli return ret; 30471e3aac0SAndrea Arcangeli } 30571e3aac0SAndrea Arcangeli static struct kobj_attribute enabled_attr = 30671e3aac0SAndrea Arcangeli __ATTR(enabled, 0644, enabled_show, enabled_store); 30771e3aac0SAndrea Arcangeli 30871e3aac0SAndrea Arcangeli static ssize_t single_flag_show(struct kobject *kobj, 30971e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 31071e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 31171e3aac0SAndrea Arcangeli { 312e27e6151SBen Hutchings return sprintf(buf, "%d\n", 313e27e6151SBen Hutchings !!test_bit(flag, &transparent_hugepage_flags)); 31471e3aac0SAndrea Arcangeli } 315e27e6151SBen Hutchings 31671e3aac0SAndrea Arcangeli static ssize_t single_flag_store(struct kobject *kobj, 31771e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 31871e3aac0SAndrea Arcangeli const char *buf, size_t count, 31971e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 32071e3aac0SAndrea Arcangeli { 321e27e6151SBen Hutchings unsigned long value; 322e27e6151SBen Hutchings int ret; 323e27e6151SBen Hutchings 324e27e6151SBen Hutchings ret = kstrtoul(buf, 10, &value); 325e27e6151SBen Hutchings if (ret < 0) 326e27e6151SBen Hutchings return ret; 327e27e6151SBen Hutchings if (value > 1) 32871e3aac0SAndrea Arcangeli return -EINVAL; 32971e3aac0SAndrea Arcangeli 330e27e6151SBen Hutchings if (value) 331e27e6151SBen Hutchings set_bit(flag, &transparent_hugepage_flags); 332e27e6151SBen Hutchings else 333e27e6151SBen Hutchings clear_bit(flag, &transparent_hugepage_flags); 334e27e6151SBen Hutchings 33571e3aac0SAndrea Arcangeli return count; 33671e3aac0SAndrea Arcangeli } 33771e3aac0SAndrea Arcangeli 33871e3aac0SAndrea Arcangeli /* 33971e3aac0SAndrea Arcangeli * Currently defrag only disables __GFP_NOWAIT for allocation. A blind 34071e3aac0SAndrea Arcangeli * __GFP_REPEAT is too aggressive, it's never worth swapping tons of 34171e3aac0SAndrea Arcangeli * memory just to allocate one more hugepage. 34271e3aac0SAndrea Arcangeli */ 34371e3aac0SAndrea Arcangeli static ssize_t defrag_show(struct kobject *kobj, 34471e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 34571e3aac0SAndrea Arcangeli { 34671e3aac0SAndrea Arcangeli return double_flag_show(kobj, attr, buf, 34771e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 34871e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 34971e3aac0SAndrea Arcangeli } 35071e3aac0SAndrea Arcangeli static ssize_t defrag_store(struct kobject *kobj, 35171e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 35271e3aac0SAndrea Arcangeli const char *buf, size_t count) 35371e3aac0SAndrea Arcangeli { 35471e3aac0SAndrea Arcangeli return double_flag_store(kobj, attr, buf, count, 35571e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 35671e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 35771e3aac0SAndrea Arcangeli } 35871e3aac0SAndrea Arcangeli static struct kobj_attribute defrag_attr = 35971e3aac0SAndrea Arcangeli __ATTR(defrag, 0644, defrag_show, defrag_store); 36071e3aac0SAndrea Arcangeli 361*79da5407SKirill A. Shutemov static ssize_t use_zero_page_show(struct kobject *kobj, 362*79da5407SKirill A. Shutemov struct kobj_attribute *attr, char *buf) 363*79da5407SKirill A. Shutemov { 364*79da5407SKirill A. Shutemov return single_flag_show(kobj, attr, buf, 365*79da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 366*79da5407SKirill A. Shutemov } 367*79da5407SKirill A. Shutemov static ssize_t use_zero_page_store(struct kobject *kobj, 368*79da5407SKirill A. Shutemov struct kobj_attribute *attr, const char *buf, size_t count) 369*79da5407SKirill A. Shutemov { 370*79da5407SKirill A. Shutemov return single_flag_store(kobj, attr, buf, count, 371*79da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 372*79da5407SKirill A. Shutemov } 373*79da5407SKirill A. Shutemov static struct kobj_attribute use_zero_page_attr = 374*79da5407SKirill A. Shutemov __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 37571e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 37671e3aac0SAndrea Arcangeli static ssize_t debug_cow_show(struct kobject *kobj, 37771e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 37871e3aac0SAndrea Arcangeli { 37971e3aac0SAndrea Arcangeli return single_flag_show(kobj, attr, buf, 38071e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 38171e3aac0SAndrea Arcangeli } 38271e3aac0SAndrea Arcangeli static ssize_t debug_cow_store(struct kobject *kobj, 38371e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 38471e3aac0SAndrea Arcangeli const char *buf, size_t count) 38571e3aac0SAndrea Arcangeli { 38671e3aac0SAndrea Arcangeli return single_flag_store(kobj, attr, buf, count, 38771e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 38871e3aac0SAndrea Arcangeli } 38971e3aac0SAndrea Arcangeli static struct kobj_attribute debug_cow_attr = 39071e3aac0SAndrea Arcangeli __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 39171e3aac0SAndrea Arcangeli #endif /* CONFIG_DEBUG_VM */ 39271e3aac0SAndrea Arcangeli 39371e3aac0SAndrea Arcangeli static struct attribute *hugepage_attr[] = { 39471e3aac0SAndrea Arcangeli &enabled_attr.attr, 39571e3aac0SAndrea Arcangeli &defrag_attr.attr, 396*79da5407SKirill A. Shutemov &use_zero_page_attr.attr, 39771e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 39871e3aac0SAndrea Arcangeli &debug_cow_attr.attr, 39971e3aac0SAndrea Arcangeli #endif 40071e3aac0SAndrea Arcangeli NULL, 40171e3aac0SAndrea Arcangeli }; 40271e3aac0SAndrea Arcangeli 40371e3aac0SAndrea Arcangeli static struct attribute_group hugepage_attr_group = { 40471e3aac0SAndrea Arcangeli .attrs = hugepage_attr, 405ba76149fSAndrea Arcangeli }; 406ba76149fSAndrea Arcangeli 407ba76149fSAndrea Arcangeli static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 408ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 409ba76149fSAndrea Arcangeli char *buf) 410ba76149fSAndrea Arcangeli { 411ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); 412ba76149fSAndrea Arcangeli } 413ba76149fSAndrea Arcangeli 414ba76149fSAndrea Arcangeli static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 415ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 416ba76149fSAndrea Arcangeli const char *buf, size_t count) 417ba76149fSAndrea Arcangeli { 418ba76149fSAndrea Arcangeli unsigned long msecs; 419ba76149fSAndrea Arcangeli int err; 420ba76149fSAndrea Arcangeli 421ba76149fSAndrea Arcangeli err = strict_strtoul(buf, 10, &msecs); 422ba76149fSAndrea Arcangeli if (err || msecs > UINT_MAX) 423ba76149fSAndrea Arcangeli return -EINVAL; 424ba76149fSAndrea Arcangeli 425ba76149fSAndrea Arcangeli khugepaged_scan_sleep_millisecs = msecs; 426ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 427ba76149fSAndrea Arcangeli 428ba76149fSAndrea Arcangeli return count; 429ba76149fSAndrea Arcangeli } 430ba76149fSAndrea Arcangeli static struct kobj_attribute scan_sleep_millisecs_attr = 431ba76149fSAndrea Arcangeli __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 432ba76149fSAndrea Arcangeli scan_sleep_millisecs_store); 433ba76149fSAndrea Arcangeli 434ba76149fSAndrea Arcangeli static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 435ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 436ba76149fSAndrea Arcangeli char *buf) 437ba76149fSAndrea Arcangeli { 438ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 439ba76149fSAndrea Arcangeli } 440ba76149fSAndrea Arcangeli 441ba76149fSAndrea Arcangeli static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 442ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 443ba76149fSAndrea Arcangeli const char *buf, size_t count) 444ba76149fSAndrea Arcangeli { 445ba76149fSAndrea Arcangeli unsigned long msecs; 446ba76149fSAndrea Arcangeli int err; 447ba76149fSAndrea Arcangeli 448ba76149fSAndrea Arcangeli err = strict_strtoul(buf, 10, &msecs); 449ba76149fSAndrea Arcangeli if (err || msecs > UINT_MAX) 450ba76149fSAndrea Arcangeli return -EINVAL; 451ba76149fSAndrea Arcangeli 452ba76149fSAndrea Arcangeli khugepaged_alloc_sleep_millisecs = msecs; 453ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 454ba76149fSAndrea Arcangeli 455ba76149fSAndrea Arcangeli return count; 456ba76149fSAndrea Arcangeli } 457ba76149fSAndrea Arcangeli static struct kobj_attribute alloc_sleep_millisecs_attr = 458ba76149fSAndrea Arcangeli __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 459ba76149fSAndrea Arcangeli alloc_sleep_millisecs_store); 460ba76149fSAndrea Arcangeli 461ba76149fSAndrea Arcangeli static ssize_t pages_to_scan_show(struct kobject *kobj, 462ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 463ba76149fSAndrea Arcangeli char *buf) 464ba76149fSAndrea Arcangeli { 465ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_pages_to_scan); 466ba76149fSAndrea Arcangeli } 467ba76149fSAndrea Arcangeli static ssize_t pages_to_scan_store(struct kobject *kobj, 468ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 469ba76149fSAndrea Arcangeli const char *buf, size_t count) 470ba76149fSAndrea Arcangeli { 471ba76149fSAndrea Arcangeli int err; 472ba76149fSAndrea Arcangeli unsigned long pages; 473ba76149fSAndrea Arcangeli 474ba76149fSAndrea Arcangeli err = strict_strtoul(buf, 10, &pages); 475ba76149fSAndrea Arcangeli if (err || !pages || pages > UINT_MAX) 476ba76149fSAndrea Arcangeli return -EINVAL; 477ba76149fSAndrea Arcangeli 478ba76149fSAndrea Arcangeli khugepaged_pages_to_scan = pages; 479ba76149fSAndrea Arcangeli 480ba76149fSAndrea Arcangeli return count; 481ba76149fSAndrea Arcangeli } 482ba76149fSAndrea Arcangeli static struct kobj_attribute pages_to_scan_attr = 483ba76149fSAndrea Arcangeli __ATTR(pages_to_scan, 0644, pages_to_scan_show, 484ba76149fSAndrea Arcangeli pages_to_scan_store); 485ba76149fSAndrea Arcangeli 486ba76149fSAndrea Arcangeli static ssize_t pages_collapsed_show(struct kobject *kobj, 487ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 488ba76149fSAndrea Arcangeli char *buf) 489ba76149fSAndrea Arcangeli { 490ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_pages_collapsed); 491ba76149fSAndrea Arcangeli } 492ba76149fSAndrea Arcangeli static struct kobj_attribute pages_collapsed_attr = 493ba76149fSAndrea Arcangeli __ATTR_RO(pages_collapsed); 494ba76149fSAndrea Arcangeli 495ba76149fSAndrea Arcangeli static ssize_t full_scans_show(struct kobject *kobj, 496ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 497ba76149fSAndrea Arcangeli char *buf) 498ba76149fSAndrea Arcangeli { 499ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_full_scans); 500ba76149fSAndrea Arcangeli } 501ba76149fSAndrea Arcangeli static struct kobj_attribute full_scans_attr = 502ba76149fSAndrea Arcangeli __ATTR_RO(full_scans); 503ba76149fSAndrea Arcangeli 504ba76149fSAndrea Arcangeli static ssize_t khugepaged_defrag_show(struct kobject *kobj, 505ba76149fSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 506ba76149fSAndrea Arcangeli { 507ba76149fSAndrea Arcangeli return single_flag_show(kobj, attr, buf, 508ba76149fSAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 509ba76149fSAndrea Arcangeli } 510ba76149fSAndrea Arcangeli static ssize_t khugepaged_defrag_store(struct kobject *kobj, 511ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 512ba76149fSAndrea Arcangeli const char *buf, size_t count) 513ba76149fSAndrea Arcangeli { 514ba76149fSAndrea Arcangeli return single_flag_store(kobj, attr, buf, count, 515ba76149fSAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 516ba76149fSAndrea Arcangeli } 517ba76149fSAndrea Arcangeli static struct kobj_attribute khugepaged_defrag_attr = 518ba76149fSAndrea Arcangeli __ATTR(defrag, 0644, khugepaged_defrag_show, 519ba76149fSAndrea Arcangeli khugepaged_defrag_store); 520ba76149fSAndrea Arcangeli 521ba76149fSAndrea Arcangeli /* 522ba76149fSAndrea Arcangeli * max_ptes_none controls if khugepaged should collapse hugepages over 523ba76149fSAndrea Arcangeli * any unmapped ptes in turn potentially increasing the memory 524ba76149fSAndrea Arcangeli * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 525ba76149fSAndrea Arcangeli * reduce the available free memory in the system as it 526ba76149fSAndrea Arcangeli * runs. Increasing max_ptes_none will instead potentially reduce the 527ba76149fSAndrea Arcangeli * free memory in the system during the khugepaged scan. 528ba76149fSAndrea Arcangeli */ 529ba76149fSAndrea Arcangeli static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 530ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 531ba76149fSAndrea Arcangeli char *buf) 532ba76149fSAndrea Arcangeli { 533ba76149fSAndrea Arcangeli return sprintf(buf, "%u\n", khugepaged_max_ptes_none); 534ba76149fSAndrea Arcangeli } 535ba76149fSAndrea Arcangeli static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 536ba76149fSAndrea Arcangeli struct kobj_attribute *attr, 537ba76149fSAndrea Arcangeli const char *buf, size_t count) 538ba76149fSAndrea Arcangeli { 539ba76149fSAndrea Arcangeli int err; 540ba76149fSAndrea Arcangeli unsigned long max_ptes_none; 541ba76149fSAndrea Arcangeli 542ba76149fSAndrea Arcangeli err = strict_strtoul(buf, 10, &max_ptes_none); 543ba76149fSAndrea Arcangeli if (err || max_ptes_none > HPAGE_PMD_NR-1) 544ba76149fSAndrea Arcangeli return -EINVAL; 545ba76149fSAndrea Arcangeli 546ba76149fSAndrea Arcangeli khugepaged_max_ptes_none = max_ptes_none; 547ba76149fSAndrea Arcangeli 548ba76149fSAndrea Arcangeli return count; 549ba76149fSAndrea Arcangeli } 550ba76149fSAndrea Arcangeli static struct kobj_attribute khugepaged_max_ptes_none_attr = 551ba76149fSAndrea Arcangeli __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 552ba76149fSAndrea Arcangeli khugepaged_max_ptes_none_store); 553ba76149fSAndrea Arcangeli 554ba76149fSAndrea Arcangeli static struct attribute *khugepaged_attr[] = { 555ba76149fSAndrea Arcangeli &khugepaged_defrag_attr.attr, 556ba76149fSAndrea Arcangeli &khugepaged_max_ptes_none_attr.attr, 557ba76149fSAndrea Arcangeli &pages_to_scan_attr.attr, 558ba76149fSAndrea Arcangeli &pages_collapsed_attr.attr, 559ba76149fSAndrea Arcangeli &full_scans_attr.attr, 560ba76149fSAndrea Arcangeli &scan_sleep_millisecs_attr.attr, 561ba76149fSAndrea Arcangeli &alloc_sleep_millisecs_attr.attr, 562ba76149fSAndrea Arcangeli NULL, 563ba76149fSAndrea Arcangeli }; 564ba76149fSAndrea Arcangeli 565ba76149fSAndrea Arcangeli static struct attribute_group khugepaged_attr_group = { 566ba76149fSAndrea Arcangeli .attrs = khugepaged_attr, 567ba76149fSAndrea Arcangeli .name = "khugepaged", 56871e3aac0SAndrea Arcangeli }; 569569e5590SShaohua Li 570569e5590SShaohua Li static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 571569e5590SShaohua Li { 572569e5590SShaohua Li int err; 573569e5590SShaohua Li 574569e5590SShaohua Li *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 575569e5590SShaohua Li if (unlikely(!*hugepage_kobj)) { 576569e5590SShaohua Li printk(KERN_ERR "hugepage: failed kobject create\n"); 577569e5590SShaohua Li return -ENOMEM; 578569e5590SShaohua Li } 579569e5590SShaohua Li 580569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 581569e5590SShaohua Li if (err) { 582569e5590SShaohua Li printk(KERN_ERR "hugepage: failed register hugeage group\n"); 583569e5590SShaohua Li goto delete_obj; 584569e5590SShaohua Li } 585569e5590SShaohua Li 586569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 587569e5590SShaohua Li if (err) { 588569e5590SShaohua Li printk(KERN_ERR "hugepage: failed register hugeage group\n"); 589569e5590SShaohua Li goto remove_hp_group; 590569e5590SShaohua Li } 591569e5590SShaohua Li 592569e5590SShaohua Li return 0; 593569e5590SShaohua Li 594569e5590SShaohua Li remove_hp_group: 595569e5590SShaohua Li sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 596569e5590SShaohua Li delete_obj: 597569e5590SShaohua Li kobject_put(*hugepage_kobj); 598569e5590SShaohua Li return err; 599569e5590SShaohua Li } 600569e5590SShaohua Li 601569e5590SShaohua Li static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 602569e5590SShaohua Li { 603569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 604569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 605569e5590SShaohua Li kobject_put(hugepage_kobj); 606569e5590SShaohua Li } 607569e5590SShaohua Li #else 608569e5590SShaohua Li static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 609569e5590SShaohua Li { 610569e5590SShaohua Li return 0; 611569e5590SShaohua Li } 612569e5590SShaohua Li 613569e5590SShaohua Li static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 614569e5590SShaohua Li { 615569e5590SShaohua Li } 61671e3aac0SAndrea Arcangeli #endif /* CONFIG_SYSFS */ 61771e3aac0SAndrea Arcangeli 61871e3aac0SAndrea Arcangeli static int __init hugepage_init(void) 61971e3aac0SAndrea Arcangeli { 62071e3aac0SAndrea Arcangeli int err; 621569e5590SShaohua Li struct kobject *hugepage_kobj; 62271e3aac0SAndrea Arcangeli 6234b7167b9SAndrea Arcangeli if (!has_transparent_hugepage()) { 6244b7167b9SAndrea Arcangeli transparent_hugepage_flags = 0; 625569e5590SShaohua Li return -EINVAL; 6264b7167b9SAndrea Arcangeli } 6274b7167b9SAndrea Arcangeli 628569e5590SShaohua Li err = hugepage_init_sysfs(&hugepage_kobj); 629569e5590SShaohua Li if (err) 630569e5590SShaohua Li return err; 631ba76149fSAndrea Arcangeli 632ba76149fSAndrea Arcangeli err = khugepaged_slab_init(); 633ba76149fSAndrea Arcangeli if (err) 634ba76149fSAndrea Arcangeli goto out; 635ba76149fSAndrea Arcangeli 636ba76149fSAndrea Arcangeli err = mm_slots_hash_init(); 637ba76149fSAndrea Arcangeli if (err) { 638ba76149fSAndrea Arcangeli khugepaged_slab_free(); 639ba76149fSAndrea Arcangeli goto out; 640ba76149fSAndrea Arcangeli } 641ba76149fSAndrea Arcangeli 64297ae1749SKirill A. Shutemov register_shrinker(&huge_zero_page_shrinker); 64397ae1749SKirill A. Shutemov 64497562cd2SRik van Riel /* 64597562cd2SRik van Riel * By default disable transparent hugepages on smaller systems, 64697562cd2SRik van Riel * where the extra memory used could hurt more than TLB overhead 64797562cd2SRik van Riel * is likely to save. The admin can still enable it through /sys. 64897562cd2SRik van Riel */ 64997562cd2SRik van Riel if (totalram_pages < (512 << (20 - PAGE_SHIFT))) 65097562cd2SRik van Riel transparent_hugepage_flags = 0; 65197562cd2SRik van Riel 652ba76149fSAndrea Arcangeli start_khugepaged(); 653ba76149fSAndrea Arcangeli 654569e5590SShaohua Li return 0; 655ba76149fSAndrea Arcangeli out: 656569e5590SShaohua Li hugepage_exit_sysfs(hugepage_kobj); 657ba76149fSAndrea Arcangeli return err; 65871e3aac0SAndrea Arcangeli } 65971e3aac0SAndrea Arcangeli module_init(hugepage_init) 66071e3aac0SAndrea Arcangeli 66171e3aac0SAndrea Arcangeli static int __init setup_transparent_hugepage(char *str) 66271e3aac0SAndrea Arcangeli { 66371e3aac0SAndrea Arcangeli int ret = 0; 66471e3aac0SAndrea Arcangeli if (!str) 66571e3aac0SAndrea Arcangeli goto out; 66671e3aac0SAndrea Arcangeli if (!strcmp(str, "always")) { 66771e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_FLAG, 66871e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 66971e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 67071e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 67171e3aac0SAndrea Arcangeli ret = 1; 67271e3aac0SAndrea Arcangeli } else if (!strcmp(str, "madvise")) { 67371e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 67471e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 67571e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 67671e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 67771e3aac0SAndrea Arcangeli ret = 1; 67871e3aac0SAndrea Arcangeli } else if (!strcmp(str, "never")) { 67971e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 68071e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 68171e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 68271e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 68371e3aac0SAndrea Arcangeli ret = 1; 68471e3aac0SAndrea Arcangeli } 68571e3aac0SAndrea Arcangeli out: 68671e3aac0SAndrea Arcangeli if (!ret) 68771e3aac0SAndrea Arcangeli printk(KERN_WARNING 68871e3aac0SAndrea Arcangeli "transparent_hugepage= cannot parse, ignored\n"); 68971e3aac0SAndrea Arcangeli return ret; 69071e3aac0SAndrea Arcangeli } 69171e3aac0SAndrea Arcangeli __setup("transparent_hugepage=", setup_transparent_hugepage); 69271e3aac0SAndrea Arcangeli 69371e3aac0SAndrea Arcangeli static inline 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 700b3092b3bSBob Liu static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma) 701b3092b3bSBob Liu { 702b3092b3bSBob Liu pmd_t entry; 703b3092b3bSBob Liu entry = mk_pmd(page, vma->vm_page_prot); 704b3092b3bSBob Liu entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 705b3092b3bSBob Liu entry = pmd_mkhuge(entry); 706b3092b3bSBob Liu return entry; 707b3092b3bSBob Liu } 708b3092b3bSBob Liu 70971e3aac0SAndrea Arcangeli static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, 71071e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 71171e3aac0SAndrea Arcangeli unsigned long haddr, pmd_t *pmd, 71271e3aac0SAndrea Arcangeli struct page *page) 71371e3aac0SAndrea Arcangeli { 71471e3aac0SAndrea Arcangeli pgtable_t pgtable; 71571e3aac0SAndrea Arcangeli 71671e3aac0SAndrea Arcangeli VM_BUG_ON(!PageCompound(page)); 71771e3aac0SAndrea Arcangeli pgtable = pte_alloc_one(mm, haddr); 718edad9d2cSDavid Rientjes if (unlikely(!pgtable)) 71971e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 72071e3aac0SAndrea Arcangeli 72171e3aac0SAndrea Arcangeli clear_huge_page(page, haddr, HPAGE_PMD_NR); 72271e3aac0SAndrea Arcangeli __SetPageUptodate(page); 72371e3aac0SAndrea Arcangeli 72471e3aac0SAndrea Arcangeli spin_lock(&mm->page_table_lock); 72571e3aac0SAndrea Arcangeli if (unlikely(!pmd_none(*pmd))) { 72671e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 727b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_page(page); 72871e3aac0SAndrea Arcangeli put_page(page); 72971e3aac0SAndrea Arcangeli pte_free(mm, pgtable); 73071e3aac0SAndrea Arcangeli } else { 73171e3aac0SAndrea Arcangeli pmd_t entry; 732b3092b3bSBob Liu entry = mk_huge_pmd(page, vma); 73371e3aac0SAndrea Arcangeli /* 73471e3aac0SAndrea Arcangeli * The spinlocking to take the lru_lock inside 73571e3aac0SAndrea Arcangeli * page_add_new_anon_rmap() acts as a full memory 73671e3aac0SAndrea Arcangeli * barrier to be sure clear_huge_page writes become 73771e3aac0SAndrea Arcangeli * visible after the set_pmd_at() write. 73871e3aac0SAndrea Arcangeli */ 73971e3aac0SAndrea Arcangeli page_add_new_anon_rmap(page, vma, haddr); 74071e3aac0SAndrea Arcangeli set_pmd_at(mm, haddr, pmd, entry); 741e3ebcf64SGerald Schaefer pgtable_trans_huge_deposit(mm, pgtable); 74271e3aac0SAndrea Arcangeli add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 7431c641e84SAndrea Arcangeli mm->nr_ptes++; 74471e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 74571e3aac0SAndrea Arcangeli } 74671e3aac0SAndrea Arcangeli 747aa2e878eSDavid Rientjes return 0; 74871e3aac0SAndrea Arcangeli } 74971e3aac0SAndrea Arcangeli 750cc5d462fSAndi Kleen static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) 7510bbbc0b3SAndrea Arcangeli { 752cc5d462fSAndi Kleen return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp; 7530bbbc0b3SAndrea Arcangeli } 7540bbbc0b3SAndrea Arcangeli 7550bbbc0b3SAndrea Arcangeli static inline struct page *alloc_hugepage_vma(int defrag, 7560bbbc0b3SAndrea Arcangeli struct vm_area_struct *vma, 757cc5d462fSAndi Kleen unsigned long haddr, int nd, 758cc5d462fSAndi Kleen gfp_t extra_gfp) 7590bbbc0b3SAndrea Arcangeli { 760cc5d462fSAndi Kleen return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp), 7615c4b4be3SAndi Kleen HPAGE_PMD_ORDER, vma, haddr, nd); 7620bbbc0b3SAndrea Arcangeli } 7630bbbc0b3SAndrea Arcangeli 7640bbbc0b3SAndrea Arcangeli #ifndef CONFIG_NUMA 76571e3aac0SAndrea Arcangeli static inline struct page *alloc_hugepage(int defrag) 76671e3aac0SAndrea Arcangeli { 767cc5d462fSAndi Kleen return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), 76871e3aac0SAndrea Arcangeli HPAGE_PMD_ORDER); 76971e3aac0SAndrea Arcangeli } 7700bbbc0b3SAndrea Arcangeli #endif 77171e3aac0SAndrea Arcangeli 772fc9fe822SKirill A. Shutemov static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 77397ae1749SKirill A. Shutemov struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 77497ae1749SKirill A. Shutemov unsigned long zero_pfn) 775fc9fe822SKirill A. Shutemov { 776fc9fe822SKirill A. Shutemov pmd_t entry; 77797ae1749SKirill A. Shutemov entry = pfn_pmd(zero_pfn, vma->vm_page_prot); 778fc9fe822SKirill A. Shutemov entry = pmd_wrprotect(entry); 779fc9fe822SKirill A. Shutemov entry = pmd_mkhuge(entry); 780fc9fe822SKirill A. Shutemov set_pmd_at(mm, haddr, pmd, entry); 781fc9fe822SKirill A. Shutemov pgtable_trans_huge_deposit(mm, pgtable); 782fc9fe822SKirill A. Shutemov mm->nr_ptes++; 783fc9fe822SKirill A. Shutemov } 784fc9fe822SKirill A. Shutemov 78571e3aac0SAndrea Arcangeli int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 78671e3aac0SAndrea Arcangeli unsigned long address, pmd_t *pmd, 78771e3aac0SAndrea Arcangeli unsigned int flags) 78871e3aac0SAndrea Arcangeli { 78971e3aac0SAndrea Arcangeli struct page *page; 79071e3aac0SAndrea Arcangeli unsigned long haddr = address & HPAGE_PMD_MASK; 79171e3aac0SAndrea Arcangeli pte_t *pte; 79271e3aac0SAndrea Arcangeli 79371e3aac0SAndrea Arcangeli if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) { 79471e3aac0SAndrea Arcangeli if (unlikely(anon_vma_prepare(vma))) 79571e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 796ba76149fSAndrea Arcangeli if (unlikely(khugepaged_enter(vma))) 797ba76149fSAndrea Arcangeli return VM_FAULT_OOM; 798*79da5407SKirill A. Shutemov if (!(flags & FAULT_FLAG_WRITE) && 799*79da5407SKirill A. Shutemov transparent_hugepage_use_zero_page()) { 80080371957SKirill A. Shutemov pgtable_t pgtable; 80197ae1749SKirill A. Shutemov unsigned long zero_pfn; 80280371957SKirill A. Shutemov pgtable = pte_alloc_one(mm, haddr); 80380371957SKirill A. Shutemov if (unlikely(!pgtable)) 80480371957SKirill A. Shutemov return VM_FAULT_OOM; 80597ae1749SKirill A. Shutemov zero_pfn = get_huge_zero_page(); 80697ae1749SKirill A. Shutemov if (unlikely(!zero_pfn)) { 80797ae1749SKirill A. Shutemov pte_free(mm, pgtable); 80897ae1749SKirill A. Shutemov count_vm_event(THP_FAULT_FALLBACK); 80997ae1749SKirill A. Shutemov goto out; 81097ae1749SKirill A. Shutemov } 81180371957SKirill A. Shutemov spin_lock(&mm->page_table_lock); 81297ae1749SKirill A. Shutemov set_huge_zero_page(pgtable, mm, vma, haddr, pmd, 81397ae1749SKirill A. Shutemov zero_pfn); 81480371957SKirill A. Shutemov spin_unlock(&mm->page_table_lock); 81580371957SKirill A. Shutemov return 0; 81680371957SKirill A. Shutemov } 8170bbbc0b3SAndrea Arcangeli page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), 818cc5d462fSAndi Kleen vma, haddr, numa_node_id(), 0); 81981ab4201SAndi Kleen if (unlikely(!page)) { 82081ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 82171e3aac0SAndrea Arcangeli goto out; 82281ab4201SAndi Kleen } 82381ab4201SAndi Kleen count_vm_event(THP_FAULT_ALLOC); 824b9bbfbe3SAndrea Arcangeli if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) { 825b9bbfbe3SAndrea Arcangeli put_page(page); 826b9bbfbe3SAndrea Arcangeli goto out; 827b9bbfbe3SAndrea Arcangeli } 828edad9d2cSDavid Rientjes if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, 829edad9d2cSDavid Rientjes page))) { 830edad9d2cSDavid Rientjes mem_cgroup_uncharge_page(page); 831edad9d2cSDavid Rientjes put_page(page); 832edad9d2cSDavid Rientjes goto out; 833edad9d2cSDavid Rientjes } 83471e3aac0SAndrea Arcangeli 835edad9d2cSDavid Rientjes return 0; 83671e3aac0SAndrea Arcangeli } 83771e3aac0SAndrea Arcangeli out: 83871e3aac0SAndrea Arcangeli /* 83971e3aac0SAndrea Arcangeli * Use __pte_alloc instead of pte_alloc_map, because we can't 84071e3aac0SAndrea Arcangeli * run pte_offset_map on the pmd, if an huge pmd could 84171e3aac0SAndrea Arcangeli * materialize from under us from a different thread. 84271e3aac0SAndrea Arcangeli */ 84371e3aac0SAndrea Arcangeli if (unlikely(__pte_alloc(mm, vma, pmd, address))) 84471e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 84571e3aac0SAndrea Arcangeli /* if an huge pmd materialized from under us just retry later */ 84671e3aac0SAndrea Arcangeli if (unlikely(pmd_trans_huge(*pmd))) 84771e3aac0SAndrea Arcangeli return 0; 84871e3aac0SAndrea Arcangeli /* 84971e3aac0SAndrea Arcangeli * A regular pmd is established and it can't morph into a huge pmd 85071e3aac0SAndrea Arcangeli * from under us anymore at this point because we hold the mmap_sem 85171e3aac0SAndrea Arcangeli * read mode and khugepaged takes it in write mode. So now it's 85271e3aac0SAndrea Arcangeli * safe to run pte_offset_map(). 85371e3aac0SAndrea Arcangeli */ 85471e3aac0SAndrea Arcangeli pte = pte_offset_map(pmd, address); 85571e3aac0SAndrea Arcangeli return handle_pte_fault(mm, vma, address, pte, pmd, flags); 85671e3aac0SAndrea Arcangeli } 85771e3aac0SAndrea Arcangeli 85871e3aac0SAndrea Arcangeli int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 85971e3aac0SAndrea Arcangeli pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 86071e3aac0SAndrea Arcangeli struct vm_area_struct *vma) 86171e3aac0SAndrea Arcangeli { 86271e3aac0SAndrea Arcangeli struct page *src_page; 86371e3aac0SAndrea Arcangeli pmd_t pmd; 86471e3aac0SAndrea Arcangeli pgtable_t pgtable; 86571e3aac0SAndrea Arcangeli int ret; 86671e3aac0SAndrea Arcangeli 86771e3aac0SAndrea Arcangeli ret = -ENOMEM; 86871e3aac0SAndrea Arcangeli pgtable = pte_alloc_one(dst_mm, addr); 86971e3aac0SAndrea Arcangeli if (unlikely(!pgtable)) 87071e3aac0SAndrea Arcangeli goto out; 87171e3aac0SAndrea Arcangeli 87271e3aac0SAndrea Arcangeli spin_lock(&dst_mm->page_table_lock); 87371e3aac0SAndrea Arcangeli spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING); 87471e3aac0SAndrea Arcangeli 87571e3aac0SAndrea Arcangeli ret = -EAGAIN; 87671e3aac0SAndrea Arcangeli pmd = *src_pmd; 87771e3aac0SAndrea Arcangeli if (unlikely(!pmd_trans_huge(pmd))) { 87871e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 87971e3aac0SAndrea Arcangeli goto out_unlock; 88071e3aac0SAndrea Arcangeli } 881fc9fe822SKirill A. Shutemov /* 882fc9fe822SKirill A. Shutemov * mm->page_table_lock is enough to be sure that huge zero pmd is not 883fc9fe822SKirill A. Shutemov * under splitting since we don't split the page itself, only pmd to 884fc9fe822SKirill A. Shutemov * a page table. 885fc9fe822SKirill A. Shutemov */ 886fc9fe822SKirill A. Shutemov if (is_huge_zero_pmd(pmd)) { 88797ae1749SKirill A. Shutemov unsigned long zero_pfn; 88897ae1749SKirill A. Shutemov /* 88997ae1749SKirill A. Shutemov * get_huge_zero_page() will never allocate a new page here, 89097ae1749SKirill A. Shutemov * since we already have a zero page to copy. It just takes a 89197ae1749SKirill A. Shutemov * reference. 89297ae1749SKirill A. Shutemov */ 89397ae1749SKirill A. Shutemov zero_pfn = get_huge_zero_page(); 89497ae1749SKirill A. Shutemov set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 89597ae1749SKirill A. Shutemov zero_pfn); 896fc9fe822SKirill A. Shutemov ret = 0; 897fc9fe822SKirill A. Shutemov goto out_unlock; 898fc9fe822SKirill A. Shutemov } 89971e3aac0SAndrea Arcangeli if (unlikely(pmd_trans_splitting(pmd))) { 90071e3aac0SAndrea Arcangeli /* split huge page running from under us */ 90171e3aac0SAndrea Arcangeli spin_unlock(&src_mm->page_table_lock); 90271e3aac0SAndrea Arcangeli spin_unlock(&dst_mm->page_table_lock); 90371e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 90471e3aac0SAndrea Arcangeli 90571e3aac0SAndrea Arcangeli wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ 90671e3aac0SAndrea Arcangeli goto out; 90771e3aac0SAndrea Arcangeli } 90871e3aac0SAndrea Arcangeli src_page = pmd_page(pmd); 90971e3aac0SAndrea Arcangeli VM_BUG_ON(!PageHead(src_page)); 91071e3aac0SAndrea Arcangeli get_page(src_page); 91171e3aac0SAndrea Arcangeli page_dup_rmap(src_page); 91271e3aac0SAndrea Arcangeli add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 91371e3aac0SAndrea Arcangeli 91471e3aac0SAndrea Arcangeli pmdp_set_wrprotect(src_mm, addr, src_pmd); 91571e3aac0SAndrea Arcangeli pmd = pmd_mkold(pmd_wrprotect(pmd)); 91671e3aac0SAndrea Arcangeli set_pmd_at(dst_mm, addr, dst_pmd, pmd); 917e3ebcf64SGerald Schaefer pgtable_trans_huge_deposit(dst_mm, pgtable); 9181c641e84SAndrea Arcangeli dst_mm->nr_ptes++; 91971e3aac0SAndrea Arcangeli 92071e3aac0SAndrea Arcangeli ret = 0; 92171e3aac0SAndrea Arcangeli out_unlock: 92271e3aac0SAndrea Arcangeli spin_unlock(&src_mm->page_table_lock); 92371e3aac0SAndrea Arcangeli spin_unlock(&dst_mm->page_table_lock); 92471e3aac0SAndrea Arcangeli out: 92571e3aac0SAndrea Arcangeli return ret; 92671e3aac0SAndrea Arcangeli } 92771e3aac0SAndrea Arcangeli 928a1dd450bSWill Deacon void huge_pmd_set_accessed(struct mm_struct *mm, 929a1dd450bSWill Deacon struct vm_area_struct *vma, 930a1dd450bSWill Deacon unsigned long address, 931a1dd450bSWill Deacon pmd_t *pmd, pmd_t orig_pmd, 932a1dd450bSWill Deacon int dirty) 933a1dd450bSWill Deacon { 934a1dd450bSWill Deacon pmd_t entry; 935a1dd450bSWill Deacon unsigned long haddr; 936a1dd450bSWill Deacon 937a1dd450bSWill Deacon spin_lock(&mm->page_table_lock); 938a1dd450bSWill Deacon if (unlikely(!pmd_same(*pmd, orig_pmd))) 939a1dd450bSWill Deacon goto unlock; 940a1dd450bSWill Deacon 941a1dd450bSWill Deacon entry = pmd_mkyoung(orig_pmd); 942a1dd450bSWill Deacon haddr = address & HPAGE_PMD_MASK; 943a1dd450bSWill Deacon if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) 944a1dd450bSWill Deacon update_mmu_cache_pmd(vma, address, pmd); 945a1dd450bSWill Deacon 946a1dd450bSWill Deacon unlock: 947a1dd450bSWill Deacon spin_unlock(&mm->page_table_lock); 948a1dd450bSWill Deacon } 949a1dd450bSWill Deacon 95093b4796dSKirill A. Shutemov static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm, 95193b4796dSKirill A. Shutemov struct vm_area_struct *vma, unsigned long address, 95293b4796dSKirill A. Shutemov pmd_t *pmd, unsigned long haddr) 95393b4796dSKirill A. Shutemov { 95493b4796dSKirill A. Shutemov pgtable_t pgtable; 95593b4796dSKirill A. Shutemov pmd_t _pmd; 95693b4796dSKirill A. Shutemov struct page *page; 95793b4796dSKirill A. Shutemov int i, ret = 0; 95893b4796dSKirill A. Shutemov unsigned long mmun_start; /* For mmu_notifiers */ 95993b4796dSKirill A. Shutemov unsigned long mmun_end; /* For mmu_notifiers */ 96093b4796dSKirill A. Shutemov 96193b4796dSKirill A. Shutemov page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 96293b4796dSKirill A. Shutemov if (!page) { 96393b4796dSKirill A. Shutemov ret |= VM_FAULT_OOM; 96493b4796dSKirill A. Shutemov goto out; 96593b4796dSKirill A. Shutemov } 96693b4796dSKirill A. Shutemov 96793b4796dSKirill A. Shutemov if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) { 96893b4796dSKirill A. Shutemov put_page(page); 96993b4796dSKirill A. Shutemov ret |= VM_FAULT_OOM; 97093b4796dSKirill A. Shutemov goto out; 97193b4796dSKirill A. Shutemov } 97293b4796dSKirill A. Shutemov 97393b4796dSKirill A. Shutemov clear_user_highpage(page, address); 97493b4796dSKirill A. Shutemov __SetPageUptodate(page); 97593b4796dSKirill A. Shutemov 97693b4796dSKirill A. Shutemov mmun_start = haddr; 97793b4796dSKirill A. Shutemov mmun_end = haddr + HPAGE_PMD_SIZE; 97893b4796dSKirill A. Shutemov mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 97993b4796dSKirill A. Shutemov 98093b4796dSKirill A. Shutemov spin_lock(&mm->page_table_lock); 98193b4796dSKirill A. Shutemov pmdp_clear_flush(vma, haddr, pmd); 98293b4796dSKirill A. Shutemov /* leave pmd empty until pte is filled */ 98393b4796dSKirill A. Shutemov 98493b4796dSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm); 98593b4796dSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 98693b4796dSKirill A. Shutemov 98793b4796dSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 98893b4796dSKirill A. Shutemov pte_t *pte, entry; 98993b4796dSKirill A. Shutemov if (haddr == (address & PAGE_MASK)) { 99093b4796dSKirill A. Shutemov entry = mk_pte(page, vma->vm_page_prot); 99193b4796dSKirill A. Shutemov entry = maybe_mkwrite(pte_mkdirty(entry), vma); 99293b4796dSKirill A. Shutemov page_add_new_anon_rmap(page, vma, haddr); 99393b4796dSKirill A. Shutemov } else { 99493b4796dSKirill A. Shutemov entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 99593b4796dSKirill A. Shutemov entry = pte_mkspecial(entry); 99693b4796dSKirill A. Shutemov } 99793b4796dSKirill A. Shutemov pte = pte_offset_map(&_pmd, haddr); 99893b4796dSKirill A. Shutemov VM_BUG_ON(!pte_none(*pte)); 99993b4796dSKirill A. Shutemov set_pte_at(mm, haddr, pte, entry); 100093b4796dSKirill A. Shutemov pte_unmap(pte); 100193b4796dSKirill A. Shutemov } 100293b4796dSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 100393b4796dSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 100493b4796dSKirill A. Shutemov spin_unlock(&mm->page_table_lock); 100597ae1749SKirill A. Shutemov put_huge_zero_page(); 100693b4796dSKirill A. Shutemov inc_mm_counter(mm, MM_ANONPAGES); 100793b4796dSKirill A. Shutemov 100893b4796dSKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 100993b4796dSKirill A. Shutemov 101093b4796dSKirill A. Shutemov ret |= VM_FAULT_WRITE; 101193b4796dSKirill A. Shutemov out: 101293b4796dSKirill A. Shutemov return ret; 101393b4796dSKirill A. Shutemov } 101493b4796dSKirill A. Shutemov 101571e3aac0SAndrea Arcangeli static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, 101671e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 101771e3aac0SAndrea Arcangeli unsigned long address, 101871e3aac0SAndrea Arcangeli pmd_t *pmd, pmd_t orig_pmd, 101971e3aac0SAndrea Arcangeli struct page *page, 102071e3aac0SAndrea Arcangeli unsigned long haddr) 102171e3aac0SAndrea Arcangeli { 102271e3aac0SAndrea Arcangeli pgtable_t pgtable; 102371e3aac0SAndrea Arcangeli pmd_t _pmd; 102471e3aac0SAndrea Arcangeli int ret = 0, i; 102571e3aac0SAndrea Arcangeli struct page **pages; 10262ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 10272ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 102871e3aac0SAndrea Arcangeli 102971e3aac0SAndrea Arcangeli pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 103071e3aac0SAndrea Arcangeli GFP_KERNEL); 103171e3aac0SAndrea Arcangeli if (unlikely(!pages)) { 103271e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 103371e3aac0SAndrea Arcangeli goto out; 103471e3aac0SAndrea Arcangeli } 103571e3aac0SAndrea Arcangeli 103671e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 1037cc5d462fSAndi Kleen pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | 1038cc5d462fSAndi Kleen __GFP_OTHER_NODE, 103919ee151eSAndi Kleen vma, address, page_to_nid(page)); 1040b9bbfbe3SAndrea Arcangeli if (unlikely(!pages[i] || 1041b9bbfbe3SAndrea Arcangeli mem_cgroup_newpage_charge(pages[i], mm, 1042b9bbfbe3SAndrea Arcangeli GFP_KERNEL))) { 1043b9bbfbe3SAndrea Arcangeli if (pages[i]) 104471e3aac0SAndrea Arcangeli put_page(pages[i]); 1045b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_start(); 1046b9bbfbe3SAndrea Arcangeli while (--i >= 0) { 1047b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_page(pages[i]); 1048b9bbfbe3SAndrea Arcangeli put_page(pages[i]); 1049b9bbfbe3SAndrea Arcangeli } 1050b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_end(); 105171e3aac0SAndrea Arcangeli kfree(pages); 105271e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 105371e3aac0SAndrea Arcangeli goto out; 105471e3aac0SAndrea Arcangeli } 105571e3aac0SAndrea Arcangeli } 105671e3aac0SAndrea Arcangeli 105771e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 105871e3aac0SAndrea Arcangeli copy_user_highpage(pages[i], page + i, 10590089e485SHillf Danton haddr + PAGE_SIZE * i, vma); 106071e3aac0SAndrea Arcangeli __SetPageUptodate(pages[i]); 106171e3aac0SAndrea Arcangeli cond_resched(); 106271e3aac0SAndrea Arcangeli } 106371e3aac0SAndrea Arcangeli 10642ec74c3eSSagi Grimberg mmun_start = haddr; 10652ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 10662ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 10672ec74c3eSSagi Grimberg 106871e3aac0SAndrea Arcangeli spin_lock(&mm->page_table_lock); 106971e3aac0SAndrea Arcangeli if (unlikely(!pmd_same(*pmd, orig_pmd))) 107071e3aac0SAndrea Arcangeli goto out_free_pages; 107171e3aac0SAndrea Arcangeli VM_BUG_ON(!PageHead(page)); 107271e3aac0SAndrea Arcangeli 10732ec74c3eSSagi Grimberg pmdp_clear_flush(vma, haddr, pmd); 107471e3aac0SAndrea Arcangeli /* leave pmd empty until pte is filled */ 107571e3aac0SAndrea Arcangeli 1076e3ebcf64SGerald Schaefer pgtable = pgtable_trans_huge_withdraw(mm); 107771e3aac0SAndrea Arcangeli pmd_populate(mm, &_pmd, pgtable); 107871e3aac0SAndrea Arcangeli 107971e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 108071e3aac0SAndrea Arcangeli pte_t *pte, entry; 108171e3aac0SAndrea Arcangeli entry = mk_pte(pages[i], vma->vm_page_prot); 108271e3aac0SAndrea Arcangeli entry = maybe_mkwrite(pte_mkdirty(entry), vma); 108371e3aac0SAndrea Arcangeli page_add_new_anon_rmap(pages[i], vma, haddr); 108471e3aac0SAndrea Arcangeli pte = pte_offset_map(&_pmd, haddr); 108571e3aac0SAndrea Arcangeli VM_BUG_ON(!pte_none(*pte)); 108671e3aac0SAndrea Arcangeli set_pte_at(mm, haddr, pte, entry); 108771e3aac0SAndrea Arcangeli pte_unmap(pte); 108871e3aac0SAndrea Arcangeli } 108971e3aac0SAndrea Arcangeli kfree(pages); 109071e3aac0SAndrea Arcangeli 109171e3aac0SAndrea Arcangeli smp_wmb(); /* make pte visible before pmd */ 109271e3aac0SAndrea Arcangeli pmd_populate(mm, pmd, pgtable); 109371e3aac0SAndrea Arcangeli page_remove_rmap(page); 109471e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 109571e3aac0SAndrea Arcangeli 10962ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 10972ec74c3eSSagi Grimberg 109871e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 109971e3aac0SAndrea Arcangeli put_page(page); 110071e3aac0SAndrea Arcangeli 110171e3aac0SAndrea Arcangeli out: 110271e3aac0SAndrea Arcangeli return ret; 110371e3aac0SAndrea Arcangeli 110471e3aac0SAndrea Arcangeli out_free_pages: 110571e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 11062ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1107b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_start(); 1108b9bbfbe3SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 1109b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_page(pages[i]); 111071e3aac0SAndrea Arcangeli put_page(pages[i]); 1111b9bbfbe3SAndrea Arcangeli } 1112b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_end(); 111371e3aac0SAndrea Arcangeli kfree(pages); 111471e3aac0SAndrea Arcangeli goto out; 111571e3aac0SAndrea Arcangeli } 111671e3aac0SAndrea Arcangeli 111771e3aac0SAndrea Arcangeli int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 111871e3aac0SAndrea Arcangeli unsigned long address, pmd_t *pmd, pmd_t orig_pmd) 111971e3aac0SAndrea Arcangeli { 112071e3aac0SAndrea Arcangeli int ret = 0; 112193b4796dSKirill A. Shutemov struct page *page = NULL, *new_page; 112271e3aac0SAndrea Arcangeli unsigned long haddr; 11232ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 11242ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 112571e3aac0SAndrea Arcangeli 112671e3aac0SAndrea Arcangeli VM_BUG_ON(!vma->anon_vma); 112793b4796dSKirill A. Shutemov haddr = address & HPAGE_PMD_MASK; 112893b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 112993b4796dSKirill A. Shutemov goto alloc; 113071e3aac0SAndrea Arcangeli spin_lock(&mm->page_table_lock); 113171e3aac0SAndrea Arcangeli if (unlikely(!pmd_same(*pmd, orig_pmd))) 113271e3aac0SAndrea Arcangeli goto out_unlock; 113371e3aac0SAndrea Arcangeli 113471e3aac0SAndrea Arcangeli page = pmd_page(orig_pmd); 113571e3aac0SAndrea Arcangeli VM_BUG_ON(!PageCompound(page) || !PageHead(page)); 113671e3aac0SAndrea Arcangeli if (page_mapcount(page) == 1) { 113771e3aac0SAndrea Arcangeli pmd_t entry; 113871e3aac0SAndrea Arcangeli entry = pmd_mkyoung(orig_pmd); 113971e3aac0SAndrea Arcangeli entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 114071e3aac0SAndrea Arcangeli if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) 1141b113da65SDavid Miller update_mmu_cache_pmd(vma, address, pmd); 114271e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 114371e3aac0SAndrea Arcangeli goto out_unlock; 114471e3aac0SAndrea Arcangeli } 114571e3aac0SAndrea Arcangeli get_page(page); 114671e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 114793b4796dSKirill A. Shutemov alloc: 114871e3aac0SAndrea Arcangeli if (transparent_hugepage_enabled(vma) && 114971e3aac0SAndrea Arcangeli !transparent_hugepage_debug_cow()) 11500bbbc0b3SAndrea Arcangeli new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), 1151cc5d462fSAndi Kleen vma, haddr, numa_node_id(), 0); 115271e3aac0SAndrea Arcangeli else 115371e3aac0SAndrea Arcangeli new_page = NULL; 115471e3aac0SAndrea Arcangeli 115571e3aac0SAndrea Arcangeli if (unlikely(!new_page)) { 115681ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 115793b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) { 115893b4796dSKirill A. Shutemov ret = do_huge_pmd_wp_zero_page_fallback(mm, vma, 115993b4796dSKirill A. Shutemov address, pmd, haddr); 116093b4796dSKirill A. Shutemov } else { 116171e3aac0SAndrea Arcangeli ret = do_huge_pmd_wp_page_fallback(mm, vma, address, 116271e3aac0SAndrea Arcangeli pmd, orig_pmd, page, haddr); 11631f1d06c3SDavid Rientjes if (ret & VM_FAULT_OOM) 11641f1d06c3SDavid Rientjes split_huge_page(page); 116571e3aac0SAndrea Arcangeli put_page(page); 116693b4796dSKirill A. Shutemov } 116771e3aac0SAndrea Arcangeli goto out; 116871e3aac0SAndrea Arcangeli } 116981ab4201SAndi Kleen count_vm_event(THP_FAULT_ALLOC); 117071e3aac0SAndrea Arcangeli 1171b9bbfbe3SAndrea Arcangeli if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) { 1172b9bbfbe3SAndrea Arcangeli put_page(new_page); 117393b4796dSKirill A. Shutemov if (page) { 11741f1d06c3SDavid Rientjes split_huge_page(page); 1175b9bbfbe3SAndrea Arcangeli put_page(page); 117693b4796dSKirill A. Shutemov } 1177b9bbfbe3SAndrea Arcangeli ret |= VM_FAULT_OOM; 1178b9bbfbe3SAndrea Arcangeli goto out; 1179b9bbfbe3SAndrea Arcangeli } 1180b9bbfbe3SAndrea Arcangeli 118193b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 118293b4796dSKirill A. Shutemov clear_huge_page(new_page, haddr, HPAGE_PMD_NR); 118393b4796dSKirill A. Shutemov else 118471e3aac0SAndrea Arcangeli copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 118571e3aac0SAndrea Arcangeli __SetPageUptodate(new_page); 118671e3aac0SAndrea Arcangeli 11872ec74c3eSSagi Grimberg mmun_start = haddr; 11882ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 11892ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 11902ec74c3eSSagi Grimberg 119171e3aac0SAndrea Arcangeli spin_lock(&mm->page_table_lock); 119293b4796dSKirill A. Shutemov if (page) 119371e3aac0SAndrea Arcangeli put_page(page); 1194b9bbfbe3SAndrea Arcangeli if (unlikely(!pmd_same(*pmd, orig_pmd))) { 11956f60b69dSDavid Rientjes spin_unlock(&mm->page_table_lock); 1196b9bbfbe3SAndrea Arcangeli mem_cgroup_uncharge_page(new_page); 119771e3aac0SAndrea Arcangeli put_page(new_page); 11982ec74c3eSSagi Grimberg goto out_mn; 1199b9bbfbe3SAndrea Arcangeli } else { 120071e3aac0SAndrea Arcangeli pmd_t entry; 1201b3092b3bSBob Liu entry = mk_huge_pmd(new_page, vma); 12022ec74c3eSSagi Grimberg pmdp_clear_flush(vma, haddr, pmd); 120371e3aac0SAndrea Arcangeli page_add_new_anon_rmap(new_page, vma, haddr); 120471e3aac0SAndrea Arcangeli set_pmd_at(mm, haddr, pmd, entry); 1205b113da65SDavid Miller update_mmu_cache_pmd(vma, address, pmd); 120697ae1749SKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) { 120793b4796dSKirill A. Shutemov add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 120897ae1749SKirill A. Shutemov put_huge_zero_page(); 120997ae1749SKirill A. Shutemov } else { 121093b4796dSKirill A. Shutemov VM_BUG_ON(!PageHead(page)); 121171e3aac0SAndrea Arcangeli page_remove_rmap(page); 121271e3aac0SAndrea Arcangeli put_page(page); 121393b4796dSKirill A. Shutemov } 121471e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 121571e3aac0SAndrea Arcangeli } 12162ec74c3eSSagi Grimberg spin_unlock(&mm->page_table_lock); 12172ec74c3eSSagi Grimberg out_mn: 12182ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 12192ec74c3eSSagi Grimberg out: 12202ec74c3eSSagi Grimberg return ret; 122171e3aac0SAndrea Arcangeli out_unlock: 122271e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 122371e3aac0SAndrea Arcangeli return ret; 122471e3aac0SAndrea Arcangeli } 122571e3aac0SAndrea Arcangeli 1226b676b293SDavid Rientjes struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 122771e3aac0SAndrea Arcangeli unsigned long addr, 122871e3aac0SAndrea Arcangeli pmd_t *pmd, 122971e3aac0SAndrea Arcangeli unsigned int flags) 123071e3aac0SAndrea Arcangeli { 1231b676b293SDavid Rientjes struct mm_struct *mm = vma->vm_mm; 123271e3aac0SAndrea Arcangeli struct page *page = NULL; 123371e3aac0SAndrea Arcangeli 123471e3aac0SAndrea Arcangeli assert_spin_locked(&mm->page_table_lock); 123571e3aac0SAndrea Arcangeli 123671e3aac0SAndrea Arcangeli if (flags & FOLL_WRITE && !pmd_write(*pmd)) 123771e3aac0SAndrea Arcangeli goto out; 123871e3aac0SAndrea Arcangeli 123971e3aac0SAndrea Arcangeli page = pmd_page(*pmd); 124071e3aac0SAndrea Arcangeli VM_BUG_ON(!PageHead(page)); 124171e3aac0SAndrea Arcangeli if (flags & FOLL_TOUCH) { 124271e3aac0SAndrea Arcangeli pmd_t _pmd; 124371e3aac0SAndrea Arcangeli /* 124471e3aac0SAndrea Arcangeli * We should set the dirty bit only for FOLL_WRITE but 124571e3aac0SAndrea Arcangeli * for now the dirty bit in the pmd is meaningless. 124671e3aac0SAndrea Arcangeli * And if the dirty bit will become meaningful and 124771e3aac0SAndrea Arcangeli * we'll only set it with FOLL_WRITE, an atomic 124871e3aac0SAndrea Arcangeli * set_bit will be required on the pmd to set the 124971e3aac0SAndrea Arcangeli * young bit, instead of the current set_pmd_at. 125071e3aac0SAndrea Arcangeli */ 125171e3aac0SAndrea Arcangeli _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 125271e3aac0SAndrea Arcangeli set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd); 125371e3aac0SAndrea Arcangeli } 1254b676b293SDavid Rientjes if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1255b676b293SDavid Rientjes if (page->mapping && trylock_page(page)) { 1256b676b293SDavid Rientjes lru_add_drain(); 1257b676b293SDavid Rientjes if (page->mapping) 1258b676b293SDavid Rientjes mlock_vma_page(page); 1259b676b293SDavid Rientjes unlock_page(page); 1260b676b293SDavid Rientjes } 1261b676b293SDavid Rientjes } 126271e3aac0SAndrea Arcangeli page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 126371e3aac0SAndrea Arcangeli VM_BUG_ON(!PageCompound(page)); 126471e3aac0SAndrea Arcangeli if (flags & FOLL_GET) 126570b50f94SAndrea Arcangeli get_page_foll(page); 126671e3aac0SAndrea Arcangeli 126771e3aac0SAndrea Arcangeli out: 126871e3aac0SAndrea Arcangeli return page; 126971e3aac0SAndrea Arcangeli } 127071e3aac0SAndrea Arcangeli 127171e3aac0SAndrea Arcangeli int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1272f21760b1SShaohua Li pmd_t *pmd, unsigned long addr) 127371e3aac0SAndrea Arcangeli { 127471e3aac0SAndrea Arcangeli int ret = 0; 127571e3aac0SAndrea Arcangeli 1276025c5b24SNaoya Horiguchi if (__pmd_trans_huge_lock(pmd, vma) == 1) { 127771e3aac0SAndrea Arcangeli struct page *page; 127871e3aac0SAndrea Arcangeli pgtable_t pgtable; 1279f5c8ad47SDavid Miller pmd_t orig_pmd; 1280e3ebcf64SGerald Schaefer pgtable = pgtable_trans_huge_withdraw(tlb->mm); 1281f5c8ad47SDavid Miller orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd); 1282f21760b1SShaohua Li tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1283479f0abbSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) { 1284479f0abbSKirill A. Shutemov tlb->mm->nr_ptes--; 1285479f0abbSKirill A. Shutemov spin_unlock(&tlb->mm->page_table_lock); 128697ae1749SKirill A. Shutemov put_huge_zero_page(); 1287479f0abbSKirill A. Shutemov } else { 1288479f0abbSKirill A. Shutemov page = pmd_page(orig_pmd); 128971e3aac0SAndrea Arcangeli page_remove_rmap(page); 129071e3aac0SAndrea Arcangeli VM_BUG_ON(page_mapcount(page) < 0); 129171e3aac0SAndrea Arcangeli add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 129271e3aac0SAndrea Arcangeli VM_BUG_ON(!PageHead(page)); 12931c641e84SAndrea Arcangeli tlb->mm->nr_ptes--; 129471e3aac0SAndrea Arcangeli spin_unlock(&tlb->mm->page_table_lock); 129571e3aac0SAndrea Arcangeli tlb_remove_page(tlb, page); 1296479f0abbSKirill A. Shutemov } 129771e3aac0SAndrea Arcangeli pte_free(tlb->mm, pgtable); 129871e3aac0SAndrea Arcangeli ret = 1; 129971e3aac0SAndrea Arcangeli } 130071e3aac0SAndrea Arcangeli return ret; 130171e3aac0SAndrea Arcangeli } 130271e3aac0SAndrea Arcangeli 13030ca1634dSJohannes Weiner int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 13040ca1634dSJohannes Weiner unsigned long addr, unsigned long end, 13050ca1634dSJohannes Weiner unsigned char *vec) 13060ca1634dSJohannes Weiner { 13070ca1634dSJohannes Weiner int ret = 0; 13080ca1634dSJohannes Weiner 1309025c5b24SNaoya Horiguchi if (__pmd_trans_huge_lock(pmd, vma) == 1) { 13100ca1634dSJohannes Weiner /* 13110ca1634dSJohannes Weiner * All logical pages in the range are present 13120ca1634dSJohannes Weiner * if backed by a huge page. 13130ca1634dSJohannes Weiner */ 13140ca1634dSJohannes Weiner spin_unlock(&vma->vm_mm->page_table_lock); 1315025c5b24SNaoya Horiguchi memset(vec, 1, (end - addr) >> PAGE_SHIFT); 1316025c5b24SNaoya Horiguchi ret = 1; 1317025c5b24SNaoya Horiguchi } 13180ca1634dSJohannes Weiner 13190ca1634dSJohannes Weiner return ret; 13200ca1634dSJohannes Weiner } 13210ca1634dSJohannes Weiner 132237a1c49aSAndrea Arcangeli int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, 132337a1c49aSAndrea Arcangeli unsigned long old_addr, 132437a1c49aSAndrea Arcangeli unsigned long new_addr, unsigned long old_end, 132537a1c49aSAndrea Arcangeli pmd_t *old_pmd, pmd_t *new_pmd) 132637a1c49aSAndrea Arcangeli { 132737a1c49aSAndrea Arcangeli int ret = 0; 132837a1c49aSAndrea Arcangeli pmd_t pmd; 132937a1c49aSAndrea Arcangeli 133037a1c49aSAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 133137a1c49aSAndrea Arcangeli 133237a1c49aSAndrea Arcangeli if ((old_addr & ~HPAGE_PMD_MASK) || 133337a1c49aSAndrea Arcangeli (new_addr & ~HPAGE_PMD_MASK) || 133437a1c49aSAndrea Arcangeli old_end - old_addr < HPAGE_PMD_SIZE || 133537a1c49aSAndrea Arcangeli (new_vma->vm_flags & VM_NOHUGEPAGE)) 133637a1c49aSAndrea Arcangeli goto out; 133737a1c49aSAndrea Arcangeli 133837a1c49aSAndrea Arcangeli /* 133937a1c49aSAndrea Arcangeli * The destination pmd shouldn't be established, free_pgtables() 134037a1c49aSAndrea Arcangeli * should have release it. 134137a1c49aSAndrea Arcangeli */ 134237a1c49aSAndrea Arcangeli if (WARN_ON(!pmd_none(*new_pmd))) { 134337a1c49aSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*new_pmd)); 134437a1c49aSAndrea Arcangeli goto out; 134537a1c49aSAndrea Arcangeli } 134637a1c49aSAndrea Arcangeli 1347025c5b24SNaoya Horiguchi ret = __pmd_trans_huge_lock(old_pmd, vma); 1348025c5b24SNaoya Horiguchi if (ret == 1) { 134937a1c49aSAndrea Arcangeli pmd = pmdp_get_and_clear(mm, old_addr, old_pmd); 135037a1c49aSAndrea Arcangeli VM_BUG_ON(!pmd_none(*new_pmd)); 135137a1c49aSAndrea Arcangeli set_pmd_at(mm, new_addr, new_pmd, pmd); 135237a1c49aSAndrea Arcangeli spin_unlock(&mm->page_table_lock); 135337a1c49aSAndrea Arcangeli } 135437a1c49aSAndrea Arcangeli out: 135537a1c49aSAndrea Arcangeli return ret; 135637a1c49aSAndrea Arcangeli } 135737a1c49aSAndrea Arcangeli 1358cd7548abSJohannes Weiner int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1359cd7548abSJohannes Weiner unsigned long addr, pgprot_t newprot) 1360cd7548abSJohannes Weiner { 1361cd7548abSJohannes Weiner struct mm_struct *mm = vma->vm_mm; 1362cd7548abSJohannes Weiner int ret = 0; 1363cd7548abSJohannes Weiner 1364025c5b24SNaoya Horiguchi if (__pmd_trans_huge_lock(pmd, vma) == 1) { 1365cd7548abSJohannes Weiner pmd_t entry; 1366cd7548abSJohannes Weiner entry = pmdp_get_and_clear(mm, addr, pmd); 1367cd7548abSJohannes Weiner entry = pmd_modify(entry, newprot); 1368cad7f613SKirill A. Shutemov BUG_ON(pmd_write(entry)); 1369cd7548abSJohannes Weiner set_pmd_at(mm, addr, pmd, entry); 1370cd7548abSJohannes Weiner spin_unlock(&vma->vm_mm->page_table_lock); 1371cd7548abSJohannes Weiner ret = 1; 1372cd7548abSJohannes Weiner } 1373cd7548abSJohannes Weiner 1374cd7548abSJohannes Weiner return ret; 1375cd7548abSJohannes Weiner } 1376cd7548abSJohannes Weiner 1377025c5b24SNaoya Horiguchi /* 1378025c5b24SNaoya Horiguchi * Returns 1 if a given pmd maps a stable (not under splitting) thp. 1379025c5b24SNaoya Horiguchi * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. 1380025c5b24SNaoya Horiguchi * 1381025c5b24SNaoya Horiguchi * Note that if it returns 1, this routine returns without unlocking page 1382025c5b24SNaoya Horiguchi * table locks. So callers must unlock them. 1383025c5b24SNaoya Horiguchi */ 1384025c5b24SNaoya Horiguchi int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1385025c5b24SNaoya Horiguchi { 1386025c5b24SNaoya Horiguchi spin_lock(&vma->vm_mm->page_table_lock); 1387025c5b24SNaoya Horiguchi if (likely(pmd_trans_huge(*pmd))) { 1388025c5b24SNaoya Horiguchi if (unlikely(pmd_trans_splitting(*pmd))) { 1389025c5b24SNaoya Horiguchi spin_unlock(&vma->vm_mm->page_table_lock); 1390025c5b24SNaoya Horiguchi wait_split_huge_page(vma->anon_vma, pmd); 1391025c5b24SNaoya Horiguchi return -1; 1392025c5b24SNaoya Horiguchi } else { 1393025c5b24SNaoya Horiguchi /* Thp mapped by 'pmd' is stable, so we can 1394025c5b24SNaoya Horiguchi * handle it as it is. */ 1395025c5b24SNaoya Horiguchi return 1; 1396025c5b24SNaoya Horiguchi } 1397025c5b24SNaoya Horiguchi } 1398025c5b24SNaoya Horiguchi spin_unlock(&vma->vm_mm->page_table_lock); 1399025c5b24SNaoya Horiguchi return 0; 1400025c5b24SNaoya Horiguchi } 1401025c5b24SNaoya Horiguchi 140271e3aac0SAndrea Arcangeli pmd_t *page_check_address_pmd(struct page *page, 140371e3aac0SAndrea Arcangeli struct mm_struct *mm, 140471e3aac0SAndrea Arcangeli unsigned long address, 140571e3aac0SAndrea Arcangeli enum page_check_address_pmd_flag flag) 140671e3aac0SAndrea Arcangeli { 140771e3aac0SAndrea Arcangeli pmd_t *pmd, *ret = NULL; 140871e3aac0SAndrea Arcangeli 140971e3aac0SAndrea Arcangeli if (address & ~HPAGE_PMD_MASK) 141071e3aac0SAndrea Arcangeli goto out; 141171e3aac0SAndrea Arcangeli 14126219049aSBob Liu pmd = mm_find_pmd(mm, address); 14136219049aSBob Liu if (!pmd) 141471e3aac0SAndrea Arcangeli goto out; 141571e3aac0SAndrea Arcangeli if (pmd_none(*pmd)) 141671e3aac0SAndrea Arcangeli goto out; 141771e3aac0SAndrea Arcangeli if (pmd_page(*pmd) != page) 141871e3aac0SAndrea Arcangeli goto out; 141994fcc585SAndrea Arcangeli /* 142094fcc585SAndrea Arcangeli * split_vma() may create temporary aliased mappings. There is 142194fcc585SAndrea Arcangeli * no risk as long as all huge pmd are found and have their 142294fcc585SAndrea Arcangeli * splitting bit set before __split_huge_page_refcount 142394fcc585SAndrea Arcangeli * runs. Finding the same huge pmd more than once during the 142494fcc585SAndrea Arcangeli * same rmap walk is not a problem. 142594fcc585SAndrea Arcangeli */ 142694fcc585SAndrea Arcangeli if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && 142794fcc585SAndrea Arcangeli pmd_trans_splitting(*pmd)) 142894fcc585SAndrea Arcangeli goto out; 142971e3aac0SAndrea Arcangeli if (pmd_trans_huge(*pmd)) { 143071e3aac0SAndrea Arcangeli VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && 143171e3aac0SAndrea Arcangeli !pmd_trans_splitting(*pmd)); 143271e3aac0SAndrea Arcangeli ret = pmd; 143371e3aac0SAndrea Arcangeli } 143471e3aac0SAndrea Arcangeli out: 143571e3aac0SAndrea Arcangeli return ret; 143671e3aac0SAndrea Arcangeli } 143771e3aac0SAndrea Arcangeli 143871e3aac0SAndrea Arcangeli static int __split_huge_page_splitting(struct page *page, 143971e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 144071e3aac0SAndrea Arcangeli unsigned long address) 144171e3aac0SAndrea Arcangeli { 144271e3aac0SAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 144371e3aac0SAndrea Arcangeli pmd_t *pmd; 144471e3aac0SAndrea Arcangeli int ret = 0; 14452ec74c3eSSagi Grimberg /* For mmu_notifiers */ 14462ec74c3eSSagi Grimberg const unsigned long mmun_start = address; 14472ec74c3eSSagi Grimberg const unsigned long mmun_end = address + HPAGE_PMD_SIZE; 144871e3aac0SAndrea Arcangeli 14492ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 145071e3aac0SAndrea Arcangeli spin_lock(&mm->page_table_lock); 145171e3aac0SAndrea Arcangeli pmd = page_check_address_pmd(page, mm, address, 145271e3aac0SAndrea Arcangeli PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG); 145371e3aac0SAndrea Arcangeli if (pmd) { 145471e3aac0SAndrea Arcangeli /* 145571e3aac0SAndrea Arcangeli * We can't temporarily set the pmd to null in order 145671e3aac0SAndrea Arcangeli * to split it, the pmd must remain marked huge at all 145771e3aac0SAndrea Arcangeli * times or the VM won't take the pmd_trans_huge paths 14582b575eb6SPeter Zijlstra * and it won't wait on the anon_vma->root->mutex to 145971e3aac0SAndrea Arcangeli * serialize against split_huge_page*. 146071e3aac0SAndrea Arcangeli */ 14612ec74c3eSSagi Grimberg pmdp_splitting_flush(vma, address, pmd); 146271e3aac0SAndrea Arcangeli ret = 1; 146371e3aac0SAndrea Arcangeli } 146471e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 14652ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 146671e3aac0SAndrea Arcangeli 146771e3aac0SAndrea Arcangeli return ret; 146871e3aac0SAndrea Arcangeli } 146971e3aac0SAndrea Arcangeli 147071e3aac0SAndrea Arcangeli static void __split_huge_page_refcount(struct page *page) 147171e3aac0SAndrea Arcangeli { 147271e3aac0SAndrea Arcangeli int i; 147371e3aac0SAndrea Arcangeli struct zone *zone = page_zone(page); 1474fa9add64SHugh Dickins struct lruvec *lruvec; 147570b50f94SAndrea Arcangeli int tail_count = 0; 147671e3aac0SAndrea Arcangeli 147771e3aac0SAndrea Arcangeli /* prevent PageLRU to go away from under us, and freeze lru stats */ 147871e3aac0SAndrea Arcangeli spin_lock_irq(&zone->lru_lock); 1479fa9add64SHugh Dickins lruvec = mem_cgroup_page_lruvec(page, zone); 1480fa9add64SHugh Dickins 148171e3aac0SAndrea Arcangeli compound_lock(page); 1482e94c8a9cSKAMEZAWA Hiroyuki /* complete memcg works before add pages to LRU */ 1483e94c8a9cSKAMEZAWA Hiroyuki mem_cgroup_split_huge_fixup(page); 148471e3aac0SAndrea Arcangeli 148545676885SShaohua Li for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 148671e3aac0SAndrea Arcangeli struct page *page_tail = page + i; 148771e3aac0SAndrea Arcangeli 148870b50f94SAndrea Arcangeli /* tail_page->_mapcount cannot change */ 148970b50f94SAndrea Arcangeli BUG_ON(page_mapcount(page_tail) < 0); 149070b50f94SAndrea Arcangeli tail_count += page_mapcount(page_tail); 149170b50f94SAndrea Arcangeli /* check for overflow */ 149270b50f94SAndrea Arcangeli BUG_ON(tail_count < 0); 149370b50f94SAndrea Arcangeli BUG_ON(atomic_read(&page_tail->_count) != 0); 149470b50f94SAndrea Arcangeli /* 149570b50f94SAndrea Arcangeli * tail_page->_count is zero and not changing from 149670b50f94SAndrea Arcangeli * under us. But get_page_unless_zero() may be running 149770b50f94SAndrea Arcangeli * from under us on the tail_page. If we used 149870b50f94SAndrea Arcangeli * atomic_set() below instead of atomic_add(), we 149970b50f94SAndrea Arcangeli * would then run atomic_set() concurrently with 150070b50f94SAndrea Arcangeli * get_page_unless_zero(), and atomic_set() is 150170b50f94SAndrea Arcangeli * implemented in C not using locked ops. spin_unlock 150270b50f94SAndrea Arcangeli * on x86 sometime uses locked ops because of PPro 150370b50f94SAndrea Arcangeli * errata 66, 92, so unless somebody can guarantee 150470b50f94SAndrea Arcangeli * atomic_set() here would be safe on all archs (and 150570b50f94SAndrea Arcangeli * not only on x86), it's safer to use atomic_add(). 150670b50f94SAndrea Arcangeli */ 150770b50f94SAndrea Arcangeli atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, 150870b50f94SAndrea Arcangeli &page_tail->_count); 150971e3aac0SAndrea Arcangeli 151071e3aac0SAndrea Arcangeli /* after clearing PageTail the gup refcount can be released */ 151171e3aac0SAndrea Arcangeli smp_mb(); 151271e3aac0SAndrea Arcangeli 1513a6d30dddSJin Dongming /* 1514a6d30dddSJin Dongming * retain hwpoison flag of the poisoned tail page: 1515a6d30dddSJin Dongming * fix for the unsuitable process killed on Guest Machine(KVM) 1516a6d30dddSJin Dongming * by the memory-failure. 1517a6d30dddSJin Dongming */ 1518a6d30dddSJin Dongming page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON; 151971e3aac0SAndrea Arcangeli page_tail->flags |= (page->flags & 152071e3aac0SAndrea Arcangeli ((1L << PG_referenced) | 152171e3aac0SAndrea Arcangeli (1L << PG_swapbacked) | 152271e3aac0SAndrea Arcangeli (1L << PG_mlocked) | 152371e3aac0SAndrea Arcangeli (1L << PG_uptodate))); 152471e3aac0SAndrea Arcangeli page_tail->flags |= (1L << PG_dirty); 152571e3aac0SAndrea Arcangeli 152670b50f94SAndrea Arcangeli /* clear PageTail before overwriting first_page */ 152771e3aac0SAndrea Arcangeli smp_wmb(); 152871e3aac0SAndrea Arcangeli 152971e3aac0SAndrea Arcangeli /* 153071e3aac0SAndrea Arcangeli * __split_huge_page_splitting() already set the 153171e3aac0SAndrea Arcangeli * splitting bit in all pmd that could map this 153271e3aac0SAndrea Arcangeli * hugepage, that will ensure no CPU can alter the 153371e3aac0SAndrea Arcangeli * mapcount on the head page. The mapcount is only 153471e3aac0SAndrea Arcangeli * accounted in the head page and it has to be 153571e3aac0SAndrea Arcangeli * transferred to all tail pages in the below code. So 153671e3aac0SAndrea Arcangeli * for this code to be safe, the split the mapcount 153771e3aac0SAndrea Arcangeli * can't change. But that doesn't mean userland can't 153871e3aac0SAndrea Arcangeli * keep changing and reading the page contents while 153971e3aac0SAndrea Arcangeli * we transfer the mapcount, so the pmd splitting 154071e3aac0SAndrea Arcangeli * status is achieved setting a reserved bit in the 154171e3aac0SAndrea Arcangeli * pmd, not by clearing the present bit. 154271e3aac0SAndrea Arcangeli */ 154371e3aac0SAndrea Arcangeli page_tail->_mapcount = page->_mapcount; 154471e3aac0SAndrea Arcangeli 154571e3aac0SAndrea Arcangeli BUG_ON(page_tail->mapping); 154671e3aac0SAndrea Arcangeli page_tail->mapping = page->mapping; 154771e3aac0SAndrea Arcangeli 154845676885SShaohua Li page_tail->index = page->index + i; 154971e3aac0SAndrea Arcangeli 155071e3aac0SAndrea Arcangeli BUG_ON(!PageAnon(page_tail)); 155171e3aac0SAndrea Arcangeli BUG_ON(!PageUptodate(page_tail)); 155271e3aac0SAndrea Arcangeli BUG_ON(!PageDirty(page_tail)); 155371e3aac0SAndrea Arcangeli BUG_ON(!PageSwapBacked(page_tail)); 155471e3aac0SAndrea Arcangeli 1555fa9add64SHugh Dickins lru_add_page_tail(page, page_tail, lruvec); 155671e3aac0SAndrea Arcangeli } 155770b50f94SAndrea Arcangeli atomic_sub(tail_count, &page->_count); 155870b50f94SAndrea Arcangeli BUG_ON(atomic_read(&page->_count) <= 0); 155971e3aac0SAndrea Arcangeli 1560fa9add64SHugh Dickins __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); 156179134171SAndrea Arcangeli __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR); 156279134171SAndrea Arcangeli 156371e3aac0SAndrea Arcangeli ClearPageCompound(page); 156471e3aac0SAndrea Arcangeli compound_unlock(page); 156571e3aac0SAndrea Arcangeli spin_unlock_irq(&zone->lru_lock); 156671e3aac0SAndrea Arcangeli 156771e3aac0SAndrea Arcangeli for (i = 1; i < HPAGE_PMD_NR; i++) { 156871e3aac0SAndrea Arcangeli struct page *page_tail = page + i; 156971e3aac0SAndrea Arcangeli BUG_ON(page_count(page_tail) <= 0); 157071e3aac0SAndrea Arcangeli /* 157171e3aac0SAndrea Arcangeli * Tail pages may be freed if there wasn't any mapping 157271e3aac0SAndrea Arcangeli * like if add_to_swap() is running on a lru page that 157371e3aac0SAndrea Arcangeli * had its mapping zapped. And freeing these pages 157471e3aac0SAndrea Arcangeli * requires taking the lru_lock so we do the put_page 157571e3aac0SAndrea Arcangeli * of the tail pages after the split is complete. 157671e3aac0SAndrea Arcangeli */ 157771e3aac0SAndrea Arcangeli put_page(page_tail); 157871e3aac0SAndrea Arcangeli } 157971e3aac0SAndrea Arcangeli 158071e3aac0SAndrea Arcangeli /* 158171e3aac0SAndrea Arcangeli * Only the head page (now become a regular page) is required 158271e3aac0SAndrea Arcangeli * to be pinned by the caller. 158371e3aac0SAndrea Arcangeli */ 158471e3aac0SAndrea Arcangeli BUG_ON(page_count(page) <= 0); 158571e3aac0SAndrea Arcangeli } 158671e3aac0SAndrea Arcangeli 158771e3aac0SAndrea Arcangeli static int __split_huge_page_map(struct page *page, 158871e3aac0SAndrea Arcangeli struct vm_area_struct *vma, 158971e3aac0SAndrea Arcangeli unsigned long address) 159071e3aac0SAndrea Arcangeli { 159171e3aac0SAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 159271e3aac0SAndrea Arcangeli pmd_t *pmd, _pmd; 159371e3aac0SAndrea Arcangeli int ret = 0, i; 159471e3aac0SAndrea Arcangeli pgtable_t pgtable; 159571e3aac0SAndrea Arcangeli unsigned long haddr; 159671e3aac0SAndrea Arcangeli 159771e3aac0SAndrea Arcangeli spin_lock(&mm->page_table_lock); 159871e3aac0SAndrea Arcangeli pmd = page_check_address_pmd(page, mm, address, 159971e3aac0SAndrea Arcangeli PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG); 160071e3aac0SAndrea Arcangeli if (pmd) { 1601e3ebcf64SGerald Schaefer pgtable = pgtable_trans_huge_withdraw(mm); 160271e3aac0SAndrea Arcangeli pmd_populate(mm, &_pmd, pgtable); 160371e3aac0SAndrea Arcangeli 1604e3ebcf64SGerald Schaefer haddr = address; 1605e3ebcf64SGerald Schaefer for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 160671e3aac0SAndrea Arcangeli pte_t *pte, entry; 160771e3aac0SAndrea Arcangeli BUG_ON(PageCompound(page+i)); 160871e3aac0SAndrea Arcangeli entry = mk_pte(page + i, vma->vm_page_prot); 160971e3aac0SAndrea Arcangeli entry = maybe_mkwrite(pte_mkdirty(entry), vma); 161071e3aac0SAndrea Arcangeli if (!pmd_write(*pmd)) 161171e3aac0SAndrea Arcangeli entry = pte_wrprotect(entry); 161271e3aac0SAndrea Arcangeli else 161371e3aac0SAndrea Arcangeli BUG_ON(page_mapcount(page) != 1); 161471e3aac0SAndrea Arcangeli if (!pmd_young(*pmd)) 161571e3aac0SAndrea Arcangeli entry = pte_mkold(entry); 161671e3aac0SAndrea Arcangeli pte = pte_offset_map(&_pmd, haddr); 161771e3aac0SAndrea Arcangeli BUG_ON(!pte_none(*pte)); 161871e3aac0SAndrea Arcangeli set_pte_at(mm, haddr, pte, entry); 161971e3aac0SAndrea Arcangeli pte_unmap(pte); 162071e3aac0SAndrea Arcangeli } 162171e3aac0SAndrea Arcangeli 162271e3aac0SAndrea Arcangeli smp_wmb(); /* make pte visible before pmd */ 162371e3aac0SAndrea Arcangeli /* 162471e3aac0SAndrea Arcangeli * Up to this point the pmd is present and huge and 162571e3aac0SAndrea Arcangeli * userland has the whole access to the hugepage 162671e3aac0SAndrea Arcangeli * during the split (which happens in place). If we 162771e3aac0SAndrea Arcangeli * overwrite the pmd with the not-huge version 162871e3aac0SAndrea Arcangeli * pointing to the pte here (which of course we could 162971e3aac0SAndrea Arcangeli * if all CPUs were bug free), userland could trigger 163071e3aac0SAndrea Arcangeli * a small page size TLB miss on the small sized TLB 163171e3aac0SAndrea Arcangeli * while the hugepage TLB entry is still established 163271e3aac0SAndrea Arcangeli * in the huge TLB. Some CPU doesn't like that. See 163371e3aac0SAndrea Arcangeli * http://support.amd.com/us/Processor_TechDocs/41322.pdf, 163471e3aac0SAndrea Arcangeli * Erratum 383 on page 93. Intel should be safe but is 163571e3aac0SAndrea Arcangeli * also warns that it's only safe if the permission 163671e3aac0SAndrea Arcangeli * and cache attributes of the two entries loaded in 163771e3aac0SAndrea Arcangeli * the two TLB is identical (which should be the case 163871e3aac0SAndrea Arcangeli * here). But it is generally safer to never allow 163971e3aac0SAndrea Arcangeli * small and huge TLB entries for the same virtual 164071e3aac0SAndrea Arcangeli * address to be loaded simultaneously. So instead of 164171e3aac0SAndrea Arcangeli * doing "pmd_populate(); flush_tlb_range();" we first 164271e3aac0SAndrea Arcangeli * mark the current pmd notpresent (atomically because 164371e3aac0SAndrea Arcangeli * here the pmd_trans_huge and pmd_trans_splitting 164471e3aac0SAndrea Arcangeli * must remain set at all times on the pmd until the 164571e3aac0SAndrea Arcangeli * split is complete for this pmd), then we flush the 164671e3aac0SAndrea Arcangeli * SMP TLB and finally we write the non-huge version 164771e3aac0SAndrea Arcangeli * of the pmd entry with pmd_populate. 164871e3aac0SAndrea Arcangeli */ 164946dcde73SGerald Schaefer pmdp_invalidate(vma, address, pmd); 165071e3aac0SAndrea Arcangeli pmd_populate(mm, pmd, pgtable); 165171e3aac0SAndrea Arcangeli ret = 1; 165271e3aac0SAndrea Arcangeli } 165371e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 165471e3aac0SAndrea Arcangeli 165571e3aac0SAndrea Arcangeli return ret; 165671e3aac0SAndrea Arcangeli } 165771e3aac0SAndrea Arcangeli 16582b575eb6SPeter Zijlstra /* must be called with anon_vma->root->mutex hold */ 165971e3aac0SAndrea Arcangeli static void __split_huge_page(struct page *page, 166071e3aac0SAndrea Arcangeli struct anon_vma *anon_vma) 166171e3aac0SAndrea Arcangeli { 166271e3aac0SAndrea Arcangeli int mapcount, mapcount2; 1663bf181b9fSMichel Lespinasse pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 166471e3aac0SAndrea Arcangeli struct anon_vma_chain *avc; 166571e3aac0SAndrea Arcangeli 166671e3aac0SAndrea Arcangeli BUG_ON(!PageHead(page)); 166771e3aac0SAndrea Arcangeli BUG_ON(PageTail(page)); 166871e3aac0SAndrea Arcangeli 166971e3aac0SAndrea Arcangeli mapcount = 0; 1670bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 167171e3aac0SAndrea Arcangeli struct vm_area_struct *vma = avc->vma; 167271e3aac0SAndrea Arcangeli unsigned long addr = vma_address(page, vma); 167371e3aac0SAndrea Arcangeli BUG_ON(is_vma_temporary_stack(vma)); 167471e3aac0SAndrea Arcangeli mapcount += __split_huge_page_splitting(page, vma, addr); 167571e3aac0SAndrea Arcangeli } 167605759d38SAndrea Arcangeli /* 167705759d38SAndrea Arcangeli * It is critical that new vmas are added to the tail of the 167805759d38SAndrea Arcangeli * anon_vma list. This guarantes that if copy_huge_pmd() runs 167905759d38SAndrea Arcangeli * and establishes a child pmd before 168005759d38SAndrea Arcangeli * __split_huge_page_splitting() freezes the parent pmd (so if 168105759d38SAndrea Arcangeli * we fail to prevent copy_huge_pmd() from running until the 168205759d38SAndrea Arcangeli * whole __split_huge_page() is complete), we will still see 168305759d38SAndrea Arcangeli * the newly established pmd of the child later during the 168405759d38SAndrea Arcangeli * walk, to be able to set it as pmd_trans_splitting too. 168505759d38SAndrea Arcangeli */ 168605759d38SAndrea Arcangeli if (mapcount != page_mapcount(page)) 168705759d38SAndrea Arcangeli printk(KERN_ERR "mapcount %d page_mapcount %d\n", 168805759d38SAndrea Arcangeli mapcount, page_mapcount(page)); 168971e3aac0SAndrea Arcangeli BUG_ON(mapcount != page_mapcount(page)); 169071e3aac0SAndrea Arcangeli 169171e3aac0SAndrea Arcangeli __split_huge_page_refcount(page); 169271e3aac0SAndrea Arcangeli 169371e3aac0SAndrea Arcangeli mapcount2 = 0; 1694bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 169571e3aac0SAndrea Arcangeli struct vm_area_struct *vma = avc->vma; 169671e3aac0SAndrea Arcangeli unsigned long addr = vma_address(page, vma); 169771e3aac0SAndrea Arcangeli BUG_ON(is_vma_temporary_stack(vma)); 169871e3aac0SAndrea Arcangeli mapcount2 += __split_huge_page_map(page, vma, addr); 169971e3aac0SAndrea Arcangeli } 170005759d38SAndrea Arcangeli if (mapcount != mapcount2) 170105759d38SAndrea Arcangeli printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n", 170205759d38SAndrea Arcangeli mapcount, mapcount2, page_mapcount(page)); 170371e3aac0SAndrea Arcangeli BUG_ON(mapcount != mapcount2); 170471e3aac0SAndrea Arcangeli } 170571e3aac0SAndrea Arcangeli 170671e3aac0SAndrea Arcangeli int split_huge_page(struct page *page) 170771e3aac0SAndrea Arcangeli { 170871e3aac0SAndrea Arcangeli struct anon_vma *anon_vma; 170971e3aac0SAndrea Arcangeli int ret = 1; 171071e3aac0SAndrea Arcangeli 1711c5a647d0SKirill A. Shutemov BUG_ON(is_huge_zero_pfn(page_to_pfn(page))); 171271e3aac0SAndrea Arcangeli BUG_ON(!PageAnon(page)); 171371e3aac0SAndrea Arcangeli anon_vma = page_lock_anon_vma(page); 171471e3aac0SAndrea Arcangeli if (!anon_vma) 171571e3aac0SAndrea Arcangeli goto out; 171671e3aac0SAndrea Arcangeli ret = 0; 171771e3aac0SAndrea Arcangeli if (!PageCompound(page)) 171871e3aac0SAndrea Arcangeli goto out_unlock; 171971e3aac0SAndrea Arcangeli 172071e3aac0SAndrea Arcangeli BUG_ON(!PageSwapBacked(page)); 172171e3aac0SAndrea Arcangeli __split_huge_page(page, anon_vma); 172281ab4201SAndi Kleen count_vm_event(THP_SPLIT); 172371e3aac0SAndrea Arcangeli 172471e3aac0SAndrea Arcangeli BUG_ON(PageCompound(page)); 172571e3aac0SAndrea Arcangeli out_unlock: 172671e3aac0SAndrea Arcangeli page_unlock_anon_vma(anon_vma); 172771e3aac0SAndrea Arcangeli out: 172871e3aac0SAndrea Arcangeli return ret; 172971e3aac0SAndrea Arcangeli } 173071e3aac0SAndrea Arcangeli 17314b6e1e37SKonstantin Khlebnikov #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE) 173278f11a25SAndrea Arcangeli 173360ab3244SAndrea Arcangeli int hugepage_madvise(struct vm_area_struct *vma, 173460ab3244SAndrea Arcangeli unsigned long *vm_flags, int advice) 17350af4e98bSAndrea Arcangeli { 17368e72033fSGerald Schaefer struct mm_struct *mm = vma->vm_mm; 17378e72033fSGerald Schaefer 1738a664b2d8SAndrea Arcangeli switch (advice) { 1739a664b2d8SAndrea Arcangeli case MADV_HUGEPAGE: 17400af4e98bSAndrea Arcangeli /* 17410af4e98bSAndrea Arcangeli * Be somewhat over-protective like KSM for now! 17420af4e98bSAndrea Arcangeli */ 174378f11a25SAndrea Arcangeli if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP)) 17440af4e98bSAndrea Arcangeli return -EINVAL; 17458e72033fSGerald Schaefer if (mm->def_flags & VM_NOHUGEPAGE) 17468e72033fSGerald Schaefer return -EINVAL; 1747a664b2d8SAndrea Arcangeli *vm_flags &= ~VM_NOHUGEPAGE; 17480af4e98bSAndrea Arcangeli *vm_flags |= VM_HUGEPAGE; 174960ab3244SAndrea Arcangeli /* 175060ab3244SAndrea Arcangeli * If the vma become good for khugepaged to scan, 175160ab3244SAndrea Arcangeli * register it here without waiting a page fault that 175260ab3244SAndrea Arcangeli * may not happen any time soon. 175360ab3244SAndrea Arcangeli */ 175460ab3244SAndrea Arcangeli if (unlikely(khugepaged_enter_vma_merge(vma))) 175560ab3244SAndrea Arcangeli return -ENOMEM; 1756a664b2d8SAndrea Arcangeli break; 1757a664b2d8SAndrea Arcangeli case MADV_NOHUGEPAGE: 1758a664b2d8SAndrea Arcangeli /* 1759a664b2d8SAndrea Arcangeli * Be somewhat over-protective like KSM for now! 1760a664b2d8SAndrea Arcangeli */ 176178f11a25SAndrea Arcangeli if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP)) 1762a664b2d8SAndrea Arcangeli return -EINVAL; 1763a664b2d8SAndrea Arcangeli *vm_flags &= ~VM_HUGEPAGE; 1764a664b2d8SAndrea Arcangeli *vm_flags |= VM_NOHUGEPAGE; 176560ab3244SAndrea Arcangeli /* 176660ab3244SAndrea Arcangeli * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 176760ab3244SAndrea Arcangeli * this vma even if we leave the mm registered in khugepaged if 176860ab3244SAndrea Arcangeli * it got registered before VM_NOHUGEPAGE was set. 176960ab3244SAndrea Arcangeli */ 1770a664b2d8SAndrea Arcangeli break; 1771a664b2d8SAndrea Arcangeli } 17720af4e98bSAndrea Arcangeli 17730af4e98bSAndrea Arcangeli return 0; 17740af4e98bSAndrea Arcangeli } 17750af4e98bSAndrea Arcangeli 1776ba76149fSAndrea Arcangeli static int __init khugepaged_slab_init(void) 1777ba76149fSAndrea Arcangeli { 1778ba76149fSAndrea Arcangeli mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 1779ba76149fSAndrea Arcangeli sizeof(struct mm_slot), 1780ba76149fSAndrea Arcangeli __alignof__(struct mm_slot), 0, NULL); 1781ba76149fSAndrea Arcangeli if (!mm_slot_cache) 1782ba76149fSAndrea Arcangeli return -ENOMEM; 1783ba76149fSAndrea Arcangeli 1784ba76149fSAndrea Arcangeli return 0; 1785ba76149fSAndrea Arcangeli } 1786ba76149fSAndrea Arcangeli 1787ba76149fSAndrea Arcangeli static void __init khugepaged_slab_free(void) 1788ba76149fSAndrea Arcangeli { 1789ba76149fSAndrea Arcangeli kmem_cache_destroy(mm_slot_cache); 1790ba76149fSAndrea Arcangeli mm_slot_cache = NULL; 1791ba76149fSAndrea Arcangeli } 1792ba76149fSAndrea Arcangeli 1793ba76149fSAndrea Arcangeli static inline struct mm_slot *alloc_mm_slot(void) 1794ba76149fSAndrea Arcangeli { 1795ba76149fSAndrea Arcangeli if (!mm_slot_cache) /* initialization failed */ 1796ba76149fSAndrea Arcangeli return NULL; 1797ba76149fSAndrea Arcangeli return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 1798ba76149fSAndrea Arcangeli } 1799ba76149fSAndrea Arcangeli 1800ba76149fSAndrea Arcangeli static inline void free_mm_slot(struct mm_slot *mm_slot) 1801ba76149fSAndrea Arcangeli { 1802ba76149fSAndrea Arcangeli kmem_cache_free(mm_slot_cache, mm_slot); 1803ba76149fSAndrea Arcangeli } 1804ba76149fSAndrea Arcangeli 1805ba76149fSAndrea Arcangeli static int __init mm_slots_hash_init(void) 1806ba76149fSAndrea Arcangeli { 1807ba76149fSAndrea Arcangeli mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head), 1808ba76149fSAndrea Arcangeli GFP_KERNEL); 1809ba76149fSAndrea Arcangeli if (!mm_slots_hash) 1810ba76149fSAndrea Arcangeli return -ENOMEM; 1811ba76149fSAndrea Arcangeli return 0; 1812ba76149fSAndrea Arcangeli } 1813ba76149fSAndrea Arcangeli 1814ba76149fSAndrea Arcangeli #if 0 1815ba76149fSAndrea Arcangeli static void __init mm_slots_hash_free(void) 1816ba76149fSAndrea Arcangeli { 1817ba76149fSAndrea Arcangeli kfree(mm_slots_hash); 1818ba76149fSAndrea Arcangeli mm_slots_hash = NULL; 1819ba76149fSAndrea Arcangeli } 1820ba76149fSAndrea Arcangeli #endif 1821ba76149fSAndrea Arcangeli 1822ba76149fSAndrea Arcangeli static struct mm_slot *get_mm_slot(struct mm_struct *mm) 1823ba76149fSAndrea Arcangeli { 1824ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 1825ba76149fSAndrea Arcangeli struct hlist_head *bucket; 1826ba76149fSAndrea Arcangeli struct hlist_node *node; 1827ba76149fSAndrea Arcangeli 1828ba76149fSAndrea Arcangeli bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) 1829ba76149fSAndrea Arcangeli % MM_SLOTS_HASH_HEADS]; 1830ba76149fSAndrea Arcangeli hlist_for_each_entry(mm_slot, node, bucket, hash) { 1831ba76149fSAndrea Arcangeli if (mm == mm_slot->mm) 1832ba76149fSAndrea Arcangeli return mm_slot; 1833ba76149fSAndrea Arcangeli } 1834ba76149fSAndrea Arcangeli return NULL; 1835ba76149fSAndrea Arcangeli } 1836ba76149fSAndrea Arcangeli 1837ba76149fSAndrea Arcangeli static void insert_to_mm_slots_hash(struct mm_struct *mm, 1838ba76149fSAndrea Arcangeli struct mm_slot *mm_slot) 1839ba76149fSAndrea Arcangeli { 1840ba76149fSAndrea Arcangeli struct hlist_head *bucket; 1841ba76149fSAndrea Arcangeli 1842ba76149fSAndrea Arcangeli bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) 1843ba76149fSAndrea Arcangeli % MM_SLOTS_HASH_HEADS]; 1844ba76149fSAndrea Arcangeli mm_slot->mm = mm; 1845ba76149fSAndrea Arcangeli hlist_add_head(&mm_slot->hash, bucket); 1846ba76149fSAndrea Arcangeli } 1847ba76149fSAndrea Arcangeli 1848ba76149fSAndrea Arcangeli static inline int khugepaged_test_exit(struct mm_struct *mm) 1849ba76149fSAndrea Arcangeli { 1850ba76149fSAndrea Arcangeli return atomic_read(&mm->mm_users) == 0; 1851ba76149fSAndrea Arcangeli } 1852ba76149fSAndrea Arcangeli 1853ba76149fSAndrea Arcangeli int __khugepaged_enter(struct mm_struct *mm) 1854ba76149fSAndrea Arcangeli { 1855ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 1856ba76149fSAndrea Arcangeli int wakeup; 1857ba76149fSAndrea Arcangeli 1858ba76149fSAndrea Arcangeli mm_slot = alloc_mm_slot(); 1859ba76149fSAndrea Arcangeli if (!mm_slot) 1860ba76149fSAndrea Arcangeli return -ENOMEM; 1861ba76149fSAndrea Arcangeli 1862ba76149fSAndrea Arcangeli /* __khugepaged_exit() must not run from under us */ 1863ba76149fSAndrea Arcangeli VM_BUG_ON(khugepaged_test_exit(mm)); 1864ba76149fSAndrea Arcangeli if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 1865ba76149fSAndrea Arcangeli free_mm_slot(mm_slot); 1866ba76149fSAndrea Arcangeli return 0; 1867ba76149fSAndrea Arcangeli } 1868ba76149fSAndrea Arcangeli 1869ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 1870ba76149fSAndrea Arcangeli insert_to_mm_slots_hash(mm, mm_slot); 1871ba76149fSAndrea Arcangeli /* 1872ba76149fSAndrea Arcangeli * Insert just behind the scanning cursor, to let the area settle 1873ba76149fSAndrea Arcangeli * down a little. 1874ba76149fSAndrea Arcangeli */ 1875ba76149fSAndrea Arcangeli wakeup = list_empty(&khugepaged_scan.mm_head); 1876ba76149fSAndrea Arcangeli list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 1877ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 1878ba76149fSAndrea Arcangeli 1879ba76149fSAndrea Arcangeli atomic_inc(&mm->mm_count); 1880ba76149fSAndrea Arcangeli if (wakeup) 1881ba76149fSAndrea Arcangeli wake_up_interruptible(&khugepaged_wait); 1882ba76149fSAndrea Arcangeli 1883ba76149fSAndrea Arcangeli return 0; 1884ba76149fSAndrea Arcangeli } 1885ba76149fSAndrea Arcangeli 1886ba76149fSAndrea Arcangeli int khugepaged_enter_vma_merge(struct vm_area_struct *vma) 1887ba76149fSAndrea Arcangeli { 1888ba76149fSAndrea Arcangeli unsigned long hstart, hend; 1889ba76149fSAndrea Arcangeli if (!vma->anon_vma) 1890ba76149fSAndrea Arcangeli /* 1891ba76149fSAndrea Arcangeli * Not yet faulted in so we will register later in the 1892ba76149fSAndrea Arcangeli * page fault if needed. 1893ba76149fSAndrea Arcangeli */ 1894ba76149fSAndrea Arcangeli return 0; 189578f11a25SAndrea Arcangeli if (vma->vm_ops) 1896ba76149fSAndrea Arcangeli /* khugepaged not yet working on file or special mappings */ 1897ba76149fSAndrea Arcangeli return 0; 1898b3b9c293SKonstantin Khlebnikov VM_BUG_ON(vma->vm_flags & VM_NO_THP); 1899ba76149fSAndrea Arcangeli hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 1900ba76149fSAndrea Arcangeli hend = vma->vm_end & HPAGE_PMD_MASK; 1901ba76149fSAndrea Arcangeli if (hstart < hend) 1902ba76149fSAndrea Arcangeli return khugepaged_enter(vma); 1903ba76149fSAndrea Arcangeli return 0; 1904ba76149fSAndrea Arcangeli } 1905ba76149fSAndrea Arcangeli 1906ba76149fSAndrea Arcangeli void __khugepaged_exit(struct mm_struct *mm) 1907ba76149fSAndrea Arcangeli { 1908ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 1909ba76149fSAndrea Arcangeli int free = 0; 1910ba76149fSAndrea Arcangeli 1911ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 1912ba76149fSAndrea Arcangeli mm_slot = get_mm_slot(mm); 1913ba76149fSAndrea Arcangeli if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 1914ba76149fSAndrea Arcangeli hlist_del(&mm_slot->hash); 1915ba76149fSAndrea Arcangeli list_del(&mm_slot->mm_node); 1916ba76149fSAndrea Arcangeli free = 1; 1917ba76149fSAndrea Arcangeli } 1918d788e80aSChris Wright spin_unlock(&khugepaged_mm_lock); 1919ba76149fSAndrea Arcangeli 1920ba76149fSAndrea Arcangeli if (free) { 1921ba76149fSAndrea Arcangeli clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1922ba76149fSAndrea Arcangeli free_mm_slot(mm_slot); 1923ba76149fSAndrea Arcangeli mmdrop(mm); 1924ba76149fSAndrea Arcangeli } else if (mm_slot) { 1925ba76149fSAndrea Arcangeli /* 1926ba76149fSAndrea Arcangeli * This is required to serialize against 1927ba76149fSAndrea Arcangeli * khugepaged_test_exit() (which is guaranteed to run 1928ba76149fSAndrea Arcangeli * under mmap sem read mode). Stop here (after we 1929ba76149fSAndrea Arcangeli * return all pagetables will be destroyed) until 1930ba76149fSAndrea Arcangeli * khugepaged has finished working on the pagetables 1931ba76149fSAndrea Arcangeli * under the mmap_sem. 1932ba76149fSAndrea Arcangeli */ 1933ba76149fSAndrea Arcangeli down_write(&mm->mmap_sem); 1934ba76149fSAndrea Arcangeli up_write(&mm->mmap_sem); 1935d788e80aSChris Wright } 1936ba76149fSAndrea Arcangeli } 1937ba76149fSAndrea Arcangeli 1938ba76149fSAndrea Arcangeli static void release_pte_page(struct page *page) 1939ba76149fSAndrea Arcangeli { 1940ba76149fSAndrea Arcangeli /* 0 stands for page_is_file_cache(page) == false */ 1941ba76149fSAndrea Arcangeli dec_zone_page_state(page, NR_ISOLATED_ANON + 0); 1942ba76149fSAndrea Arcangeli unlock_page(page); 1943ba76149fSAndrea Arcangeli putback_lru_page(page); 1944ba76149fSAndrea Arcangeli } 1945ba76149fSAndrea Arcangeli 1946ba76149fSAndrea Arcangeli static void release_pte_pages(pte_t *pte, pte_t *_pte) 1947ba76149fSAndrea Arcangeli { 1948ba76149fSAndrea Arcangeli while (--_pte >= pte) { 1949ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 1950ba76149fSAndrea Arcangeli if (!pte_none(pteval)) 1951ba76149fSAndrea Arcangeli release_pte_page(pte_page(pteval)); 1952ba76149fSAndrea Arcangeli } 1953ba76149fSAndrea Arcangeli } 1954ba76149fSAndrea Arcangeli 1955ba76149fSAndrea Arcangeli static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 1956ba76149fSAndrea Arcangeli unsigned long address, 1957ba76149fSAndrea Arcangeli pte_t *pte) 1958ba76149fSAndrea Arcangeli { 1959ba76149fSAndrea Arcangeli struct page *page; 1960ba76149fSAndrea Arcangeli pte_t *_pte; 1961344aa35cSBob Liu int referenced = 0, none = 0; 1962ba76149fSAndrea Arcangeli for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 1963ba76149fSAndrea Arcangeli _pte++, address += PAGE_SIZE) { 1964ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 1965ba76149fSAndrea Arcangeli if (pte_none(pteval)) { 1966ba76149fSAndrea Arcangeli if (++none <= khugepaged_max_ptes_none) 1967ba76149fSAndrea Arcangeli continue; 1968344aa35cSBob Liu else 1969ba76149fSAndrea Arcangeli goto out; 1970ba76149fSAndrea Arcangeli } 1971344aa35cSBob Liu if (!pte_present(pteval) || !pte_write(pteval)) 1972ba76149fSAndrea Arcangeli goto out; 1973ba76149fSAndrea Arcangeli page = vm_normal_page(vma, address, pteval); 1974344aa35cSBob Liu if (unlikely(!page)) 1975ba76149fSAndrea Arcangeli goto out; 1976344aa35cSBob Liu 1977ba76149fSAndrea Arcangeli VM_BUG_ON(PageCompound(page)); 1978ba76149fSAndrea Arcangeli BUG_ON(!PageAnon(page)); 1979ba76149fSAndrea Arcangeli VM_BUG_ON(!PageSwapBacked(page)); 1980ba76149fSAndrea Arcangeli 1981ba76149fSAndrea Arcangeli /* cannot use mapcount: can't collapse if there's a gup pin */ 1982344aa35cSBob Liu if (page_count(page) != 1) 1983ba76149fSAndrea Arcangeli goto out; 1984ba76149fSAndrea Arcangeli /* 1985ba76149fSAndrea Arcangeli * We can do it before isolate_lru_page because the 1986ba76149fSAndrea Arcangeli * page can't be freed from under us. NOTE: PG_lock 1987ba76149fSAndrea Arcangeli * is needed to serialize against split_huge_page 1988ba76149fSAndrea Arcangeli * when invoked from the VM. 1989ba76149fSAndrea Arcangeli */ 1990344aa35cSBob Liu if (!trylock_page(page)) 1991ba76149fSAndrea Arcangeli goto out; 1992ba76149fSAndrea Arcangeli /* 1993ba76149fSAndrea Arcangeli * Isolate the page to avoid collapsing an hugepage 1994ba76149fSAndrea Arcangeli * currently in use by the VM. 1995ba76149fSAndrea Arcangeli */ 1996ba76149fSAndrea Arcangeli if (isolate_lru_page(page)) { 1997ba76149fSAndrea Arcangeli unlock_page(page); 1998ba76149fSAndrea Arcangeli goto out; 1999ba76149fSAndrea Arcangeli } 2000ba76149fSAndrea Arcangeli /* 0 stands for page_is_file_cache(page) == false */ 2001ba76149fSAndrea Arcangeli inc_zone_page_state(page, NR_ISOLATED_ANON + 0); 2002ba76149fSAndrea Arcangeli VM_BUG_ON(!PageLocked(page)); 2003ba76149fSAndrea Arcangeli VM_BUG_ON(PageLRU(page)); 2004ba76149fSAndrea Arcangeli 2005ba76149fSAndrea Arcangeli /* If there is no mapped pte young don't collapse the page */ 20068ee53820SAndrea Arcangeli if (pte_young(pteval) || PageReferenced(page) || 20078ee53820SAndrea Arcangeli mmu_notifier_test_young(vma->vm_mm, address)) 2008ba76149fSAndrea Arcangeli referenced = 1; 2009ba76149fSAndrea Arcangeli } 2010344aa35cSBob Liu if (likely(referenced)) 2011344aa35cSBob Liu return 1; 2012ba76149fSAndrea Arcangeli out: 2013344aa35cSBob Liu release_pte_pages(pte, _pte); 2014344aa35cSBob Liu return 0; 2015ba76149fSAndrea Arcangeli } 2016ba76149fSAndrea Arcangeli 2017ba76149fSAndrea Arcangeli static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 2018ba76149fSAndrea Arcangeli struct vm_area_struct *vma, 2019ba76149fSAndrea Arcangeli unsigned long address, 2020ba76149fSAndrea Arcangeli spinlock_t *ptl) 2021ba76149fSAndrea Arcangeli { 2022ba76149fSAndrea Arcangeli pte_t *_pte; 2023ba76149fSAndrea Arcangeli for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { 2024ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 2025ba76149fSAndrea Arcangeli struct page *src_page; 2026ba76149fSAndrea Arcangeli 2027ba76149fSAndrea Arcangeli if (pte_none(pteval)) { 2028ba76149fSAndrea Arcangeli clear_user_highpage(page, address); 2029ba76149fSAndrea Arcangeli add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 2030ba76149fSAndrea Arcangeli } else { 2031ba76149fSAndrea Arcangeli src_page = pte_page(pteval); 2032ba76149fSAndrea Arcangeli copy_user_highpage(page, src_page, address, vma); 2033ba76149fSAndrea Arcangeli VM_BUG_ON(page_mapcount(src_page) != 1); 2034ba76149fSAndrea Arcangeli release_pte_page(src_page); 2035ba76149fSAndrea Arcangeli /* 2036ba76149fSAndrea Arcangeli * ptl mostly unnecessary, but preempt has to 2037ba76149fSAndrea Arcangeli * be disabled to update the per-cpu stats 2038ba76149fSAndrea Arcangeli * inside page_remove_rmap(). 2039ba76149fSAndrea Arcangeli */ 2040ba76149fSAndrea Arcangeli spin_lock(ptl); 2041ba76149fSAndrea Arcangeli /* 2042ba76149fSAndrea Arcangeli * paravirt calls inside pte_clear here are 2043ba76149fSAndrea Arcangeli * superfluous. 2044ba76149fSAndrea Arcangeli */ 2045ba76149fSAndrea Arcangeli pte_clear(vma->vm_mm, address, _pte); 2046ba76149fSAndrea Arcangeli page_remove_rmap(src_page); 2047ba76149fSAndrea Arcangeli spin_unlock(ptl); 2048ba76149fSAndrea Arcangeli free_page_and_swap_cache(src_page); 2049ba76149fSAndrea Arcangeli } 2050ba76149fSAndrea Arcangeli 2051ba76149fSAndrea Arcangeli address += PAGE_SIZE; 2052ba76149fSAndrea Arcangeli page++; 2053ba76149fSAndrea Arcangeli } 2054ba76149fSAndrea Arcangeli } 2055ba76149fSAndrea Arcangeli 205626234f36SXiao Guangrong static void khugepaged_alloc_sleep(void) 205726234f36SXiao Guangrong { 205826234f36SXiao Guangrong wait_event_freezable_timeout(khugepaged_wait, false, 205926234f36SXiao Guangrong msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 206026234f36SXiao Guangrong } 206126234f36SXiao Guangrong 206226234f36SXiao Guangrong #ifdef CONFIG_NUMA 206326234f36SXiao Guangrong static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 206426234f36SXiao Guangrong { 206526234f36SXiao Guangrong if (IS_ERR(*hpage)) { 206626234f36SXiao Guangrong if (!*wait) 206726234f36SXiao Guangrong return false; 206826234f36SXiao Guangrong 206926234f36SXiao Guangrong *wait = false; 2070e3b4126cSXiao Guangrong *hpage = NULL; 207126234f36SXiao Guangrong khugepaged_alloc_sleep(); 207226234f36SXiao Guangrong } else if (*hpage) { 207326234f36SXiao Guangrong put_page(*hpage); 207426234f36SXiao Guangrong *hpage = NULL; 207526234f36SXiao Guangrong } 207626234f36SXiao Guangrong 207726234f36SXiao Guangrong return true; 207826234f36SXiao Guangrong } 207926234f36SXiao Guangrong 208026234f36SXiao Guangrong static struct page 208126234f36SXiao Guangrong *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 208226234f36SXiao Guangrong struct vm_area_struct *vma, unsigned long address, 208326234f36SXiao Guangrong int node) 208426234f36SXiao Guangrong { 208526234f36SXiao Guangrong VM_BUG_ON(*hpage); 208626234f36SXiao Guangrong /* 208726234f36SXiao Guangrong * Allocate the page while the vma is still valid and under 208826234f36SXiao Guangrong * the mmap_sem read mode so there is no memory allocation 208926234f36SXiao Guangrong * later when we take the mmap_sem in write mode. This is more 209026234f36SXiao Guangrong * friendly behavior (OTOH it may actually hide bugs) to 209126234f36SXiao Guangrong * filesystems in userland with daemons allocating memory in 209226234f36SXiao Guangrong * the userland I/O paths. Allocating memory with the 209326234f36SXiao Guangrong * mmap_sem in read mode is good idea also to allow greater 209426234f36SXiao Guangrong * scalability. 209526234f36SXiao Guangrong */ 209626234f36SXiao Guangrong *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address, 209726234f36SXiao Guangrong node, __GFP_OTHER_NODE); 209826234f36SXiao Guangrong 209926234f36SXiao Guangrong /* 210026234f36SXiao Guangrong * After allocating the hugepage, release the mmap_sem read lock in 210126234f36SXiao Guangrong * preparation for taking it in write mode. 210226234f36SXiao Guangrong */ 210326234f36SXiao Guangrong up_read(&mm->mmap_sem); 210426234f36SXiao Guangrong if (unlikely(!*hpage)) { 210526234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 210626234f36SXiao Guangrong *hpage = ERR_PTR(-ENOMEM); 210726234f36SXiao Guangrong return NULL; 210826234f36SXiao Guangrong } 210926234f36SXiao Guangrong 211026234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC); 211126234f36SXiao Guangrong return *hpage; 211226234f36SXiao Guangrong } 211326234f36SXiao Guangrong #else 211426234f36SXiao Guangrong static struct page *khugepaged_alloc_hugepage(bool *wait) 211526234f36SXiao Guangrong { 211626234f36SXiao Guangrong struct page *hpage; 211726234f36SXiao Guangrong 211826234f36SXiao Guangrong do { 211926234f36SXiao Guangrong hpage = alloc_hugepage(khugepaged_defrag()); 212026234f36SXiao Guangrong if (!hpage) { 212126234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 212226234f36SXiao Guangrong if (!*wait) 212326234f36SXiao Guangrong return NULL; 212426234f36SXiao Guangrong 212526234f36SXiao Guangrong *wait = false; 212626234f36SXiao Guangrong khugepaged_alloc_sleep(); 212726234f36SXiao Guangrong } else 212826234f36SXiao Guangrong count_vm_event(THP_COLLAPSE_ALLOC); 212926234f36SXiao Guangrong } while (unlikely(!hpage) && likely(khugepaged_enabled())); 213026234f36SXiao Guangrong 213126234f36SXiao Guangrong return hpage; 213226234f36SXiao Guangrong } 213326234f36SXiao Guangrong 213426234f36SXiao Guangrong static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 213526234f36SXiao Guangrong { 213626234f36SXiao Guangrong if (!*hpage) 213726234f36SXiao Guangrong *hpage = khugepaged_alloc_hugepage(wait); 213826234f36SXiao Guangrong 213926234f36SXiao Guangrong if (unlikely(!*hpage)) 214026234f36SXiao Guangrong return false; 214126234f36SXiao Guangrong 214226234f36SXiao Guangrong return true; 214326234f36SXiao Guangrong } 214426234f36SXiao Guangrong 214526234f36SXiao Guangrong static struct page 214626234f36SXiao Guangrong *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 214726234f36SXiao Guangrong struct vm_area_struct *vma, unsigned long address, 214826234f36SXiao Guangrong int node) 214926234f36SXiao Guangrong { 215026234f36SXiao Guangrong up_read(&mm->mmap_sem); 215126234f36SXiao Guangrong VM_BUG_ON(!*hpage); 215226234f36SXiao Guangrong return *hpage; 215326234f36SXiao Guangrong } 215426234f36SXiao Guangrong #endif 215526234f36SXiao Guangrong 2156fa475e51SBob Liu static bool hugepage_vma_check(struct vm_area_struct *vma) 2157fa475e51SBob Liu { 2158fa475e51SBob Liu if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || 2159fa475e51SBob Liu (vma->vm_flags & VM_NOHUGEPAGE)) 2160fa475e51SBob Liu return false; 2161fa475e51SBob Liu 2162fa475e51SBob Liu if (!vma->anon_vma || vma->vm_ops) 2163fa475e51SBob Liu return false; 2164fa475e51SBob Liu if (is_vma_temporary_stack(vma)) 2165fa475e51SBob Liu return false; 2166fa475e51SBob Liu VM_BUG_ON(vma->vm_flags & VM_NO_THP); 2167fa475e51SBob Liu return true; 2168fa475e51SBob Liu } 2169fa475e51SBob Liu 2170ba76149fSAndrea Arcangeli static void collapse_huge_page(struct mm_struct *mm, 2171ba76149fSAndrea Arcangeli unsigned long address, 2172ce83d217SAndrea Arcangeli struct page **hpage, 21735c4b4be3SAndi Kleen struct vm_area_struct *vma, 21745c4b4be3SAndi Kleen int node) 2175ba76149fSAndrea Arcangeli { 2176ba76149fSAndrea Arcangeli pmd_t *pmd, _pmd; 2177ba76149fSAndrea Arcangeli pte_t *pte; 2178ba76149fSAndrea Arcangeli pgtable_t pgtable; 2179ba76149fSAndrea Arcangeli struct page *new_page; 2180ba76149fSAndrea Arcangeli spinlock_t *ptl; 2181ba76149fSAndrea Arcangeli int isolated; 2182ba76149fSAndrea Arcangeli unsigned long hstart, hend; 21832ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 21842ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 2185ba76149fSAndrea Arcangeli 2186ba76149fSAndrea Arcangeli VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2187692e0b35SAndrea Arcangeli 218826234f36SXiao Guangrong /* release the mmap_sem read lock. */ 218926234f36SXiao Guangrong new_page = khugepaged_alloc_page(hpage, mm, vma, address, node); 219026234f36SXiao Guangrong if (!new_page) 2191ce83d217SAndrea Arcangeli return; 2192ce83d217SAndrea Arcangeli 2193420256efSXiao Guangrong if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) 2194692e0b35SAndrea Arcangeli return; 2195ba76149fSAndrea Arcangeli 2196ba76149fSAndrea Arcangeli /* 2197ba76149fSAndrea Arcangeli * Prevent all access to pagetables with the exception of 2198ba76149fSAndrea Arcangeli * gup_fast later hanlded by the ptep_clear_flush and the VM 2199ba76149fSAndrea Arcangeli * handled by the anon_vma lock + PG_lock. 2200ba76149fSAndrea Arcangeli */ 2201ba76149fSAndrea Arcangeli down_write(&mm->mmap_sem); 2202ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) 2203ba76149fSAndrea Arcangeli goto out; 2204ba76149fSAndrea Arcangeli 2205ba76149fSAndrea Arcangeli vma = find_vma(mm, address); 2206ba76149fSAndrea Arcangeli hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2207ba76149fSAndrea Arcangeli hend = vma->vm_end & HPAGE_PMD_MASK; 2208ba76149fSAndrea Arcangeli if (address < hstart || address + HPAGE_PMD_SIZE > hend) 2209ba76149fSAndrea Arcangeli goto out; 2210fa475e51SBob Liu if (!hugepage_vma_check(vma)) 2211ba76149fSAndrea Arcangeli goto out; 22126219049aSBob Liu pmd = mm_find_pmd(mm, address); 22136219049aSBob Liu if (!pmd) 2214ba76149fSAndrea Arcangeli goto out; 22156219049aSBob Liu if (pmd_trans_huge(*pmd)) 2216ba76149fSAndrea Arcangeli goto out; 2217ba76149fSAndrea Arcangeli 2218ba76149fSAndrea Arcangeli anon_vma_lock(vma->anon_vma); 2219ba76149fSAndrea Arcangeli 2220ba76149fSAndrea Arcangeli pte = pte_offset_map(pmd, address); 2221ba76149fSAndrea Arcangeli ptl = pte_lockptr(mm, pmd); 2222ba76149fSAndrea Arcangeli 22232ec74c3eSSagi Grimberg mmun_start = address; 22242ec74c3eSSagi Grimberg mmun_end = address + HPAGE_PMD_SIZE; 22252ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2226ba76149fSAndrea Arcangeli spin_lock(&mm->page_table_lock); /* probably unnecessary */ 2227ba76149fSAndrea Arcangeli /* 2228ba76149fSAndrea Arcangeli * After this gup_fast can't run anymore. This also removes 2229ba76149fSAndrea Arcangeli * any huge TLB entry from the CPU so we won't allow 2230ba76149fSAndrea Arcangeli * huge and small TLB entries for the same virtual address 2231ba76149fSAndrea Arcangeli * to avoid the risk of CPU bugs in that area. 2232ba76149fSAndrea Arcangeli */ 22332ec74c3eSSagi Grimberg _pmd = pmdp_clear_flush(vma, address, pmd); 2234ba76149fSAndrea Arcangeli spin_unlock(&mm->page_table_lock); 22352ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2236ba76149fSAndrea Arcangeli 2237ba76149fSAndrea Arcangeli spin_lock(ptl); 2238ba76149fSAndrea Arcangeli isolated = __collapse_huge_page_isolate(vma, address, pte); 2239ba76149fSAndrea Arcangeli spin_unlock(ptl); 2240ba76149fSAndrea Arcangeli 2241ba76149fSAndrea Arcangeli if (unlikely(!isolated)) { 2242453c7192SJohannes Weiner pte_unmap(pte); 2243ba76149fSAndrea Arcangeli spin_lock(&mm->page_table_lock); 2244ba76149fSAndrea Arcangeli BUG_ON(!pmd_none(*pmd)); 2245ba76149fSAndrea Arcangeli set_pmd_at(mm, address, pmd, _pmd); 2246ba76149fSAndrea Arcangeli spin_unlock(&mm->page_table_lock); 2247ba76149fSAndrea Arcangeli anon_vma_unlock(vma->anon_vma); 2248ce83d217SAndrea Arcangeli goto out; 2249ba76149fSAndrea Arcangeli } 2250ba76149fSAndrea Arcangeli 2251ba76149fSAndrea Arcangeli /* 2252ba76149fSAndrea Arcangeli * All pages are isolated and locked so anon_vma rmap 2253ba76149fSAndrea Arcangeli * can't run anymore. 2254ba76149fSAndrea Arcangeli */ 2255ba76149fSAndrea Arcangeli anon_vma_unlock(vma->anon_vma); 2256ba76149fSAndrea Arcangeli 2257ba76149fSAndrea Arcangeli __collapse_huge_page_copy(pte, new_page, vma, address, ptl); 2258453c7192SJohannes Weiner pte_unmap(pte); 2259ba76149fSAndrea Arcangeli __SetPageUptodate(new_page); 2260ba76149fSAndrea Arcangeli pgtable = pmd_pgtable(_pmd); 2261ba76149fSAndrea Arcangeli 2262b3092b3bSBob Liu _pmd = mk_huge_pmd(new_page, vma); 2263ba76149fSAndrea Arcangeli 2264ba76149fSAndrea Arcangeli /* 2265ba76149fSAndrea Arcangeli * spin_lock() below is not the equivalent of smp_wmb(), so 2266ba76149fSAndrea Arcangeli * this is needed to avoid the copy_huge_page writes to become 2267ba76149fSAndrea Arcangeli * visible after the set_pmd_at() write. 2268ba76149fSAndrea Arcangeli */ 2269ba76149fSAndrea Arcangeli smp_wmb(); 2270ba76149fSAndrea Arcangeli 2271ba76149fSAndrea Arcangeli spin_lock(&mm->page_table_lock); 2272ba76149fSAndrea Arcangeli BUG_ON(!pmd_none(*pmd)); 2273ba76149fSAndrea Arcangeli page_add_new_anon_rmap(new_page, vma, address); 2274ba76149fSAndrea Arcangeli set_pmd_at(mm, address, pmd, _pmd); 2275b113da65SDavid Miller update_mmu_cache_pmd(vma, address, pmd); 2276e3ebcf64SGerald Schaefer pgtable_trans_huge_deposit(mm, pgtable); 2277ba76149fSAndrea Arcangeli spin_unlock(&mm->page_table_lock); 2278ba76149fSAndrea Arcangeli 2279ba76149fSAndrea Arcangeli *hpage = NULL; 2280420256efSXiao Guangrong 2281ba76149fSAndrea Arcangeli khugepaged_pages_collapsed++; 2282ce83d217SAndrea Arcangeli out_up_write: 2283ba76149fSAndrea Arcangeli up_write(&mm->mmap_sem); 22840bbbc0b3SAndrea Arcangeli return; 22850bbbc0b3SAndrea Arcangeli 2286ce83d217SAndrea Arcangeli out: 2287678ff896SKAMEZAWA Hiroyuki mem_cgroup_uncharge_page(new_page); 2288ce83d217SAndrea Arcangeli goto out_up_write; 2289ba76149fSAndrea Arcangeli } 2290ba76149fSAndrea Arcangeli 2291ba76149fSAndrea Arcangeli static int khugepaged_scan_pmd(struct mm_struct *mm, 2292ba76149fSAndrea Arcangeli struct vm_area_struct *vma, 2293ba76149fSAndrea Arcangeli unsigned long address, 2294ba76149fSAndrea Arcangeli struct page **hpage) 2295ba76149fSAndrea Arcangeli { 2296ba76149fSAndrea Arcangeli pmd_t *pmd; 2297ba76149fSAndrea Arcangeli pte_t *pte, *_pte; 2298ba76149fSAndrea Arcangeli int ret = 0, referenced = 0, none = 0; 2299ba76149fSAndrea Arcangeli struct page *page; 2300ba76149fSAndrea Arcangeli unsigned long _address; 2301ba76149fSAndrea Arcangeli spinlock_t *ptl; 23025c4b4be3SAndi Kleen int node = -1; 2303ba76149fSAndrea Arcangeli 2304ba76149fSAndrea Arcangeli VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2305ba76149fSAndrea Arcangeli 23066219049aSBob Liu pmd = mm_find_pmd(mm, address); 23076219049aSBob Liu if (!pmd) 2308ba76149fSAndrea Arcangeli goto out; 23096219049aSBob Liu if (pmd_trans_huge(*pmd)) 2310ba76149fSAndrea Arcangeli goto out; 2311ba76149fSAndrea Arcangeli 2312ba76149fSAndrea Arcangeli pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2313ba76149fSAndrea Arcangeli for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 2314ba76149fSAndrea Arcangeli _pte++, _address += PAGE_SIZE) { 2315ba76149fSAndrea Arcangeli pte_t pteval = *_pte; 2316ba76149fSAndrea Arcangeli if (pte_none(pteval)) { 2317ba76149fSAndrea Arcangeli if (++none <= khugepaged_max_ptes_none) 2318ba76149fSAndrea Arcangeli continue; 2319ba76149fSAndrea Arcangeli else 2320ba76149fSAndrea Arcangeli goto out_unmap; 2321ba76149fSAndrea Arcangeli } 2322ba76149fSAndrea Arcangeli if (!pte_present(pteval) || !pte_write(pteval)) 2323ba76149fSAndrea Arcangeli goto out_unmap; 2324ba76149fSAndrea Arcangeli page = vm_normal_page(vma, _address, pteval); 2325ba76149fSAndrea Arcangeli if (unlikely(!page)) 2326ba76149fSAndrea Arcangeli goto out_unmap; 23275c4b4be3SAndi Kleen /* 23285c4b4be3SAndi Kleen * Chose the node of the first page. This could 23295c4b4be3SAndi Kleen * be more sophisticated and look at more pages, 23305c4b4be3SAndi Kleen * but isn't for now. 23315c4b4be3SAndi Kleen */ 23325c4b4be3SAndi Kleen if (node == -1) 23335c4b4be3SAndi Kleen node = page_to_nid(page); 2334ba76149fSAndrea Arcangeli VM_BUG_ON(PageCompound(page)); 2335ba76149fSAndrea Arcangeli if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) 2336ba76149fSAndrea Arcangeli goto out_unmap; 2337ba76149fSAndrea Arcangeli /* cannot use mapcount: can't collapse if there's a gup pin */ 2338ba76149fSAndrea Arcangeli if (page_count(page) != 1) 2339ba76149fSAndrea Arcangeli goto out_unmap; 23408ee53820SAndrea Arcangeli if (pte_young(pteval) || PageReferenced(page) || 23418ee53820SAndrea Arcangeli mmu_notifier_test_young(vma->vm_mm, address)) 2342ba76149fSAndrea Arcangeli referenced = 1; 2343ba76149fSAndrea Arcangeli } 2344ba76149fSAndrea Arcangeli if (referenced) 2345ba76149fSAndrea Arcangeli ret = 1; 2346ba76149fSAndrea Arcangeli out_unmap: 2347ba76149fSAndrea Arcangeli pte_unmap_unlock(pte, ptl); 2348ce83d217SAndrea Arcangeli if (ret) 2349ce83d217SAndrea Arcangeli /* collapse_huge_page will return with the mmap_sem released */ 23505c4b4be3SAndi Kleen collapse_huge_page(mm, address, hpage, vma, node); 2351ba76149fSAndrea Arcangeli out: 2352ba76149fSAndrea Arcangeli return ret; 2353ba76149fSAndrea Arcangeli } 2354ba76149fSAndrea Arcangeli 2355ba76149fSAndrea Arcangeli static void collect_mm_slot(struct mm_slot *mm_slot) 2356ba76149fSAndrea Arcangeli { 2357ba76149fSAndrea Arcangeli struct mm_struct *mm = mm_slot->mm; 2358ba76149fSAndrea Arcangeli 2359b9980cdcSHugh Dickins VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2360ba76149fSAndrea Arcangeli 2361ba76149fSAndrea Arcangeli if (khugepaged_test_exit(mm)) { 2362ba76149fSAndrea Arcangeli /* free mm_slot */ 2363ba76149fSAndrea Arcangeli hlist_del(&mm_slot->hash); 2364ba76149fSAndrea Arcangeli list_del(&mm_slot->mm_node); 2365ba76149fSAndrea Arcangeli 2366ba76149fSAndrea Arcangeli /* 2367ba76149fSAndrea Arcangeli * Not strictly needed because the mm exited already. 2368ba76149fSAndrea Arcangeli * 2369ba76149fSAndrea Arcangeli * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2370ba76149fSAndrea Arcangeli */ 2371ba76149fSAndrea Arcangeli 2372ba76149fSAndrea Arcangeli /* khugepaged_mm_lock actually not necessary for the below */ 2373ba76149fSAndrea Arcangeli free_mm_slot(mm_slot); 2374ba76149fSAndrea Arcangeli mmdrop(mm); 2375ba76149fSAndrea Arcangeli } 2376ba76149fSAndrea Arcangeli } 2377ba76149fSAndrea Arcangeli 2378ba76149fSAndrea Arcangeli static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 2379ba76149fSAndrea Arcangeli struct page **hpage) 23802f1da642SH Hartley Sweeten __releases(&khugepaged_mm_lock) 23812f1da642SH Hartley Sweeten __acquires(&khugepaged_mm_lock) 2382ba76149fSAndrea Arcangeli { 2383ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 2384ba76149fSAndrea Arcangeli struct mm_struct *mm; 2385ba76149fSAndrea Arcangeli struct vm_area_struct *vma; 2386ba76149fSAndrea Arcangeli int progress = 0; 2387ba76149fSAndrea Arcangeli 2388ba76149fSAndrea Arcangeli VM_BUG_ON(!pages); 2389b9980cdcSHugh Dickins VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2390ba76149fSAndrea Arcangeli 2391ba76149fSAndrea Arcangeli if (khugepaged_scan.mm_slot) 2392ba76149fSAndrea Arcangeli mm_slot = khugepaged_scan.mm_slot; 2393ba76149fSAndrea Arcangeli else { 2394ba76149fSAndrea Arcangeli mm_slot = list_entry(khugepaged_scan.mm_head.next, 2395ba76149fSAndrea Arcangeli struct mm_slot, mm_node); 2396ba76149fSAndrea Arcangeli khugepaged_scan.address = 0; 2397ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = mm_slot; 2398ba76149fSAndrea Arcangeli } 2399ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 2400ba76149fSAndrea Arcangeli 2401ba76149fSAndrea Arcangeli mm = mm_slot->mm; 2402ba76149fSAndrea Arcangeli down_read(&mm->mmap_sem); 2403ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) 2404ba76149fSAndrea Arcangeli vma = NULL; 2405ba76149fSAndrea Arcangeli else 2406ba76149fSAndrea Arcangeli vma = find_vma(mm, khugepaged_scan.address); 2407ba76149fSAndrea Arcangeli 2408ba76149fSAndrea Arcangeli progress++; 2409ba76149fSAndrea Arcangeli for (; vma; vma = vma->vm_next) { 2410ba76149fSAndrea Arcangeli unsigned long hstart, hend; 2411ba76149fSAndrea Arcangeli 2412ba76149fSAndrea Arcangeli cond_resched(); 2413ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) { 2414ba76149fSAndrea Arcangeli progress++; 2415ba76149fSAndrea Arcangeli break; 2416ba76149fSAndrea Arcangeli } 2417fa475e51SBob Liu if (!hugepage_vma_check(vma)) { 2418a7d6e4ecSAndrea Arcangeli skip: 2419ba76149fSAndrea Arcangeli progress++; 2420ba76149fSAndrea Arcangeli continue; 2421ba76149fSAndrea Arcangeli } 2422ba76149fSAndrea Arcangeli hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2423ba76149fSAndrea Arcangeli hend = vma->vm_end & HPAGE_PMD_MASK; 2424a7d6e4ecSAndrea Arcangeli if (hstart >= hend) 2425a7d6e4ecSAndrea Arcangeli goto skip; 2426a7d6e4ecSAndrea Arcangeli if (khugepaged_scan.address > hend) 2427a7d6e4ecSAndrea Arcangeli goto skip; 2428ba76149fSAndrea Arcangeli if (khugepaged_scan.address < hstart) 2429ba76149fSAndrea Arcangeli khugepaged_scan.address = hstart; 2430a7d6e4ecSAndrea Arcangeli VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2431ba76149fSAndrea Arcangeli 2432ba76149fSAndrea Arcangeli while (khugepaged_scan.address < hend) { 2433ba76149fSAndrea Arcangeli int ret; 2434ba76149fSAndrea Arcangeli cond_resched(); 2435ba76149fSAndrea Arcangeli if (unlikely(khugepaged_test_exit(mm))) 2436ba76149fSAndrea Arcangeli goto breakouterloop; 2437ba76149fSAndrea Arcangeli 2438ba76149fSAndrea Arcangeli VM_BUG_ON(khugepaged_scan.address < hstart || 2439ba76149fSAndrea Arcangeli khugepaged_scan.address + HPAGE_PMD_SIZE > 2440ba76149fSAndrea Arcangeli hend); 2441ba76149fSAndrea Arcangeli ret = khugepaged_scan_pmd(mm, vma, 2442ba76149fSAndrea Arcangeli khugepaged_scan.address, 2443ba76149fSAndrea Arcangeli hpage); 2444ba76149fSAndrea Arcangeli /* move to next address */ 2445ba76149fSAndrea Arcangeli khugepaged_scan.address += HPAGE_PMD_SIZE; 2446ba76149fSAndrea Arcangeli progress += HPAGE_PMD_NR; 2447ba76149fSAndrea Arcangeli if (ret) 2448ba76149fSAndrea Arcangeli /* we released mmap_sem so break loop */ 2449ba76149fSAndrea Arcangeli goto breakouterloop_mmap_sem; 2450ba76149fSAndrea Arcangeli if (progress >= pages) 2451ba76149fSAndrea Arcangeli goto breakouterloop; 2452ba76149fSAndrea Arcangeli } 2453ba76149fSAndrea Arcangeli } 2454ba76149fSAndrea Arcangeli breakouterloop: 2455ba76149fSAndrea Arcangeli up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ 2456ba76149fSAndrea Arcangeli breakouterloop_mmap_sem: 2457ba76149fSAndrea Arcangeli 2458ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2459a7d6e4ecSAndrea Arcangeli VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2460ba76149fSAndrea Arcangeli /* 2461ba76149fSAndrea Arcangeli * Release the current mm_slot if this mm is about to die, or 2462ba76149fSAndrea Arcangeli * if we scanned all vmas of this mm. 2463ba76149fSAndrea Arcangeli */ 2464ba76149fSAndrea Arcangeli if (khugepaged_test_exit(mm) || !vma) { 2465ba76149fSAndrea Arcangeli /* 2466ba76149fSAndrea Arcangeli * Make sure that if mm_users is reaching zero while 2467ba76149fSAndrea Arcangeli * khugepaged runs here, khugepaged_exit will find 2468ba76149fSAndrea Arcangeli * mm_slot not pointing to the exiting mm. 2469ba76149fSAndrea Arcangeli */ 2470ba76149fSAndrea Arcangeli if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2471ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = list_entry( 2472ba76149fSAndrea Arcangeli mm_slot->mm_node.next, 2473ba76149fSAndrea Arcangeli struct mm_slot, mm_node); 2474ba76149fSAndrea Arcangeli khugepaged_scan.address = 0; 2475ba76149fSAndrea Arcangeli } else { 2476ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = NULL; 2477ba76149fSAndrea Arcangeli khugepaged_full_scans++; 2478ba76149fSAndrea Arcangeli } 2479ba76149fSAndrea Arcangeli 2480ba76149fSAndrea Arcangeli collect_mm_slot(mm_slot); 2481ba76149fSAndrea Arcangeli } 2482ba76149fSAndrea Arcangeli 2483ba76149fSAndrea Arcangeli return progress; 2484ba76149fSAndrea Arcangeli } 2485ba76149fSAndrea Arcangeli 2486ba76149fSAndrea Arcangeli static int khugepaged_has_work(void) 2487ba76149fSAndrea Arcangeli { 2488ba76149fSAndrea Arcangeli return !list_empty(&khugepaged_scan.mm_head) && 2489ba76149fSAndrea Arcangeli khugepaged_enabled(); 2490ba76149fSAndrea Arcangeli } 2491ba76149fSAndrea Arcangeli 2492ba76149fSAndrea Arcangeli static int khugepaged_wait_event(void) 2493ba76149fSAndrea Arcangeli { 2494ba76149fSAndrea Arcangeli return !list_empty(&khugepaged_scan.mm_head) || 24952017c0bfSXiao Guangrong kthread_should_stop(); 2496ba76149fSAndrea Arcangeli } 2497ba76149fSAndrea Arcangeli 2498d516904bSXiao Guangrong static void khugepaged_do_scan(void) 2499d516904bSXiao Guangrong { 2500d516904bSXiao Guangrong struct page *hpage = NULL; 2501ba76149fSAndrea Arcangeli unsigned int progress = 0, pass_through_head = 0; 2502ba76149fSAndrea Arcangeli unsigned int pages = khugepaged_pages_to_scan; 2503d516904bSXiao Guangrong bool wait = true; 2504ba76149fSAndrea Arcangeli 2505ba76149fSAndrea Arcangeli barrier(); /* write khugepaged_pages_to_scan to local stack */ 2506ba76149fSAndrea Arcangeli 2507ba76149fSAndrea Arcangeli while (progress < pages) { 250826234f36SXiao Guangrong if (!khugepaged_prealloc_page(&hpage, &wait)) 250926234f36SXiao Guangrong break; 2510d516904bSXiao Guangrong 2511420256efSXiao Guangrong cond_resched(); 2512ba76149fSAndrea Arcangeli 2513878aee7dSAndrea Arcangeli if (unlikely(kthread_should_stop() || freezing(current))) 2514878aee7dSAndrea Arcangeli break; 2515878aee7dSAndrea Arcangeli 2516ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2517ba76149fSAndrea Arcangeli if (!khugepaged_scan.mm_slot) 2518ba76149fSAndrea Arcangeli pass_through_head++; 2519ba76149fSAndrea Arcangeli if (khugepaged_has_work() && 2520ba76149fSAndrea Arcangeli pass_through_head < 2) 2521ba76149fSAndrea Arcangeli progress += khugepaged_scan_mm_slot(pages - progress, 2522d516904bSXiao Guangrong &hpage); 2523ba76149fSAndrea Arcangeli else 2524ba76149fSAndrea Arcangeli progress = pages; 2525ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 2526ba76149fSAndrea Arcangeli } 2527ba76149fSAndrea Arcangeli 2528d516904bSXiao Guangrong if (!IS_ERR_OR_NULL(hpage)) 2529d516904bSXiao Guangrong put_page(hpage); 2530ba76149fSAndrea Arcangeli } 25310bbbc0b3SAndrea Arcangeli 25322017c0bfSXiao Guangrong static void khugepaged_wait_work(void) 25332017c0bfSXiao Guangrong { 25342017c0bfSXiao Guangrong try_to_freeze(); 25352017c0bfSXiao Guangrong 25362017c0bfSXiao Guangrong if (khugepaged_has_work()) { 25372017c0bfSXiao Guangrong if (!khugepaged_scan_sleep_millisecs) 25382017c0bfSXiao Guangrong return; 25392017c0bfSXiao Guangrong 25402017c0bfSXiao Guangrong wait_event_freezable_timeout(khugepaged_wait, 25412017c0bfSXiao Guangrong kthread_should_stop(), 25422017c0bfSXiao Guangrong msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); 25432017c0bfSXiao Guangrong return; 25442017c0bfSXiao Guangrong } 25452017c0bfSXiao Guangrong 25462017c0bfSXiao Guangrong if (khugepaged_enabled()) 25472017c0bfSXiao Guangrong wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 25482017c0bfSXiao Guangrong } 25492017c0bfSXiao Guangrong 2550ba76149fSAndrea Arcangeli static int khugepaged(void *none) 2551ba76149fSAndrea Arcangeli { 2552ba76149fSAndrea Arcangeli struct mm_slot *mm_slot; 2553ba76149fSAndrea Arcangeli 2554878aee7dSAndrea Arcangeli set_freezable(); 2555ba76149fSAndrea Arcangeli set_user_nice(current, 19); 2556ba76149fSAndrea Arcangeli 2557b7231789SXiao Guangrong while (!kthread_should_stop()) { 2558b7231789SXiao Guangrong khugepaged_do_scan(); 2559b7231789SXiao Guangrong khugepaged_wait_work(); 2560b7231789SXiao Guangrong } 2561ba76149fSAndrea Arcangeli 2562ba76149fSAndrea Arcangeli spin_lock(&khugepaged_mm_lock); 2563ba76149fSAndrea Arcangeli mm_slot = khugepaged_scan.mm_slot; 2564ba76149fSAndrea Arcangeli khugepaged_scan.mm_slot = NULL; 2565ba76149fSAndrea Arcangeli if (mm_slot) 2566ba76149fSAndrea Arcangeli collect_mm_slot(mm_slot); 2567ba76149fSAndrea Arcangeli spin_unlock(&khugepaged_mm_lock); 2568ba76149fSAndrea Arcangeli return 0; 2569ba76149fSAndrea Arcangeli } 2570ba76149fSAndrea Arcangeli 2571c5a647d0SKirill A. Shutemov static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2572c5a647d0SKirill A. Shutemov unsigned long haddr, pmd_t *pmd) 2573c5a647d0SKirill A. Shutemov { 2574c5a647d0SKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2575c5a647d0SKirill A. Shutemov pgtable_t pgtable; 2576c5a647d0SKirill A. Shutemov pmd_t _pmd; 2577c5a647d0SKirill A. Shutemov int i; 2578c5a647d0SKirill A. Shutemov 2579c5a647d0SKirill A. Shutemov pmdp_clear_flush(vma, haddr, pmd); 2580c5a647d0SKirill A. Shutemov /* leave pmd empty until pte is filled */ 2581c5a647d0SKirill A. Shutemov 2582c5a647d0SKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm); 2583c5a647d0SKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2584c5a647d0SKirill A. Shutemov 2585c5a647d0SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2586c5a647d0SKirill A. Shutemov pte_t *pte, entry; 2587c5a647d0SKirill A. Shutemov entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2588c5a647d0SKirill A. Shutemov entry = pte_mkspecial(entry); 2589c5a647d0SKirill A. Shutemov pte = pte_offset_map(&_pmd, haddr); 2590c5a647d0SKirill A. Shutemov VM_BUG_ON(!pte_none(*pte)); 2591c5a647d0SKirill A. Shutemov set_pte_at(mm, haddr, pte, entry); 2592c5a647d0SKirill A. Shutemov pte_unmap(pte); 2593c5a647d0SKirill A. Shutemov } 2594c5a647d0SKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2595c5a647d0SKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 259697ae1749SKirill A. Shutemov put_huge_zero_page(); 2597c5a647d0SKirill A. Shutemov } 2598c5a647d0SKirill A. Shutemov 2599e180377fSKirill A. Shutemov void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, 2600e180377fSKirill A. Shutemov pmd_t *pmd) 260171e3aac0SAndrea Arcangeli { 260271e3aac0SAndrea Arcangeli struct page *page; 2603e180377fSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2604c5a647d0SKirill A. Shutemov unsigned long haddr = address & HPAGE_PMD_MASK; 2605c5a647d0SKirill A. Shutemov unsigned long mmun_start; /* For mmu_notifiers */ 2606c5a647d0SKirill A. Shutemov unsigned long mmun_end; /* For mmu_notifiers */ 2607e180377fSKirill A. Shutemov 2608e180377fSKirill A. Shutemov BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); 260971e3aac0SAndrea Arcangeli 2610c5a647d0SKirill A. Shutemov mmun_start = haddr; 2611c5a647d0SKirill A. Shutemov mmun_end = haddr + HPAGE_PMD_SIZE; 2612c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 261371e3aac0SAndrea Arcangeli spin_lock(&mm->page_table_lock); 261471e3aac0SAndrea Arcangeli if (unlikely(!pmd_trans_huge(*pmd))) { 261571e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 2616c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2617c5a647d0SKirill A. Shutemov return; 2618c5a647d0SKirill A. Shutemov } 2619c5a647d0SKirill A. Shutemov if (is_huge_zero_pmd(*pmd)) { 2620c5a647d0SKirill A. Shutemov __split_huge_zero_page_pmd(vma, haddr, pmd); 2621c5a647d0SKirill A. Shutemov spin_unlock(&mm->page_table_lock); 2622c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 262371e3aac0SAndrea Arcangeli return; 262471e3aac0SAndrea Arcangeli } 262571e3aac0SAndrea Arcangeli page = pmd_page(*pmd); 262671e3aac0SAndrea Arcangeli VM_BUG_ON(!page_count(page)); 262771e3aac0SAndrea Arcangeli get_page(page); 262871e3aac0SAndrea Arcangeli spin_unlock(&mm->page_table_lock); 2629c5a647d0SKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 263071e3aac0SAndrea Arcangeli 263171e3aac0SAndrea Arcangeli split_huge_page(page); 263271e3aac0SAndrea Arcangeli 263371e3aac0SAndrea Arcangeli put_page(page); 263471e3aac0SAndrea Arcangeli BUG_ON(pmd_trans_huge(*pmd)); 263571e3aac0SAndrea Arcangeli } 263694fcc585SAndrea Arcangeli 2637e180377fSKirill A. Shutemov void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address, 2638e180377fSKirill A. Shutemov pmd_t *pmd) 2639e180377fSKirill A. Shutemov { 2640e180377fSKirill A. Shutemov struct vm_area_struct *vma; 2641e180377fSKirill A. Shutemov 2642e180377fSKirill A. Shutemov vma = find_vma(mm, address); 2643e180377fSKirill A. Shutemov BUG_ON(vma == NULL); 2644e180377fSKirill A. Shutemov split_huge_page_pmd(vma, address, pmd); 2645e180377fSKirill A. Shutemov } 2646e180377fSKirill A. Shutemov 264794fcc585SAndrea Arcangeli static void split_huge_page_address(struct mm_struct *mm, 264894fcc585SAndrea Arcangeli unsigned long address) 264994fcc585SAndrea Arcangeli { 265094fcc585SAndrea Arcangeli pmd_t *pmd; 265194fcc585SAndrea Arcangeli 265294fcc585SAndrea Arcangeli VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); 265394fcc585SAndrea Arcangeli 26546219049aSBob Liu pmd = mm_find_pmd(mm, address); 26556219049aSBob Liu if (!pmd) 265694fcc585SAndrea Arcangeli return; 265794fcc585SAndrea Arcangeli /* 265894fcc585SAndrea Arcangeli * Caller holds the mmap_sem write mode, so a huge pmd cannot 265994fcc585SAndrea Arcangeli * materialize from under us. 266094fcc585SAndrea Arcangeli */ 2661e180377fSKirill A. Shutemov split_huge_page_pmd_mm(mm, address, pmd); 266294fcc585SAndrea Arcangeli } 266394fcc585SAndrea Arcangeli 266494fcc585SAndrea Arcangeli void __vma_adjust_trans_huge(struct vm_area_struct *vma, 266594fcc585SAndrea Arcangeli unsigned long start, 266694fcc585SAndrea Arcangeli unsigned long end, 266794fcc585SAndrea Arcangeli long adjust_next) 266894fcc585SAndrea Arcangeli { 266994fcc585SAndrea Arcangeli /* 267094fcc585SAndrea Arcangeli * If the new start address isn't hpage aligned and it could 267194fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 267294fcc585SAndrea Arcangeli * an huge pmd. 267394fcc585SAndrea Arcangeli */ 267494fcc585SAndrea Arcangeli if (start & ~HPAGE_PMD_MASK && 267594fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) >= vma->vm_start && 267694fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 267794fcc585SAndrea Arcangeli split_huge_page_address(vma->vm_mm, start); 267894fcc585SAndrea Arcangeli 267994fcc585SAndrea Arcangeli /* 268094fcc585SAndrea Arcangeli * If the new end address isn't hpage aligned and it could 268194fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 268294fcc585SAndrea Arcangeli * an huge pmd. 268394fcc585SAndrea Arcangeli */ 268494fcc585SAndrea Arcangeli if (end & ~HPAGE_PMD_MASK && 268594fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) >= vma->vm_start && 268694fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 268794fcc585SAndrea Arcangeli split_huge_page_address(vma->vm_mm, end); 268894fcc585SAndrea Arcangeli 268994fcc585SAndrea Arcangeli /* 269094fcc585SAndrea Arcangeli * If we're also updating the vma->vm_next->vm_start, if the new 269194fcc585SAndrea Arcangeli * vm_next->vm_start isn't page aligned and it could previously 269294fcc585SAndrea Arcangeli * contain an hugepage: check if we need to split an huge pmd. 269394fcc585SAndrea Arcangeli */ 269494fcc585SAndrea Arcangeli if (adjust_next > 0) { 269594fcc585SAndrea Arcangeli struct vm_area_struct *next = vma->vm_next; 269694fcc585SAndrea Arcangeli unsigned long nstart = next->vm_start; 269794fcc585SAndrea Arcangeli nstart += adjust_next << PAGE_SHIFT; 269894fcc585SAndrea Arcangeli if (nstart & ~HPAGE_PMD_MASK && 269994fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) >= next->vm_start && 270094fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 270194fcc585SAndrea Arcangeli split_huge_page_address(next->vm_mm, nstart); 270294fcc585SAndrea Arcangeli } 270394fcc585SAndrea Arcangeli } 2704