120c8ccb1SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 271e3aac0SAndrea Arcangeli /* 371e3aac0SAndrea Arcangeli * Copyright (C) 2009 Red Hat, Inc. 471e3aac0SAndrea Arcangeli */ 571e3aac0SAndrea Arcangeli 6ae3a8c1cSAndrew Morton #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 7ae3a8c1cSAndrew Morton 871e3aac0SAndrea Arcangeli #include <linux/mm.h> 971e3aac0SAndrea Arcangeli #include <linux/sched.h> 10fa6c0231SZi Yan #include <linux/sched/mm.h> 11f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 126a3827d7SIngo Molnar #include <linux/sched/numa_balancing.h> 1371e3aac0SAndrea Arcangeli #include <linux/highmem.h> 1471e3aac0SAndrea Arcangeli #include <linux/hugetlb.h> 1571e3aac0SAndrea Arcangeli #include <linux/mmu_notifier.h> 1671e3aac0SAndrea Arcangeli #include <linux/rmap.h> 1771e3aac0SAndrea Arcangeli #include <linux/swap.h> 1897ae1749SKirill A. Shutemov #include <linux/shrinker.h> 19ba76149fSAndrea Arcangeli #include <linux/mm_inline.h> 20e9b61f19SKirill A. Shutemov #include <linux/swapops.h> 214897c765SMatthew Wilcox #include <linux/dax.h> 22ba76149fSAndrea Arcangeli #include <linux/khugepaged.h> 23878aee7dSAndrea Arcangeli #include <linux/freezer.h> 24f25748e3SDan Williams #include <linux/pfn_t.h> 25a664b2d8SAndrea Arcangeli #include <linux/mman.h> 263565fce3SDan Williams #include <linux/memremap.h> 27325adeb5SRalf Baechle #include <linux/pagemap.h> 2849071d43SKirill A. Shutemov #include <linux/debugfs.h> 294daae3b4SMel Gorman #include <linux/migrate.h> 3043b5fbbdSSasha Levin #include <linux/hashtable.h> 316b251fc9SAndrea Arcangeli #include <linux/userfaultfd_k.h> 3233c3fc71SVladimir Davydov #include <linux/page_idle.h> 33baa355fdSKirill A. Shutemov #include <linux/shmem_fs.h> 346b31d595SMichal Hocko #include <linux/oom.h> 3598fa15f3SAnshuman Khandual #include <linux/numa.h> 36f7da677bSVlastimil Babka #include <linux/page_owner.h> 3797ae1749SKirill A. Shutemov 3871e3aac0SAndrea Arcangeli #include <asm/tlb.h> 3971e3aac0SAndrea Arcangeli #include <asm/pgalloc.h> 4071e3aac0SAndrea Arcangeli #include "internal.h" 4171e3aac0SAndrea Arcangeli 42ba76149fSAndrea Arcangeli /* 43b14d595aSMichael DeGuzis * By default, transparent hugepage support is disabled in order to avoid 44b14d595aSMichael DeGuzis * risking an increased memory footprint for applications that are not 45b14d595aSMichael DeGuzis * guaranteed to benefit from it. When transparent hugepage support is 46b14d595aSMichael DeGuzis * enabled, it is for all mappings, and khugepaged scans all mappings. 478bfa3f9aSJianguo Wu * Defrag is invoked by khugepaged hugepage allocations and by page faults 488bfa3f9aSJianguo Wu * for all hugepage allocations. 49ba76149fSAndrea Arcangeli */ 5071e3aac0SAndrea Arcangeli unsigned long transparent_hugepage_flags __read_mostly = 5113ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 52ba76149fSAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_FLAG)| 5313ece886SAndrea Arcangeli #endif 5413ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 5513ece886SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 5613ece886SAndrea Arcangeli #endif 57444eb2a4SMel Gorman (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 5879da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 5979da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 60ba76149fSAndrea Arcangeli 619a982250SKirill A. Shutemov static struct shrinker deferred_split_shrinker; 62f000565aSAndrea Arcangeli 6397ae1749SKirill A. Shutemov static atomic_t huge_zero_refcount; 6456873f43SWang, Yalin struct page *huge_zero_page __read_mostly; 654a6c1297SKirill A. Shutemov 667635d9cbSMichal Hocko bool transparent_hugepage_enabled(struct vm_area_struct *vma) 677635d9cbSMichal Hocko { 68c0630669SYang Shi /* The addr is used to check if the vma size fits */ 69c0630669SYang Shi unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE; 70c0630669SYang Shi 71c0630669SYang Shi if (!transhuge_vma_suitable(vma, addr)) 72c0630669SYang Shi return false; 737635d9cbSMichal Hocko if (vma_is_anonymous(vma)) 747635d9cbSMichal Hocko return __transparent_hugepage_enabled(vma); 75c0630669SYang Shi if (vma_is_shmem(vma)) 76c0630669SYang Shi return shmem_huge_enabled(vma); 777635d9cbSMichal Hocko 787635d9cbSMichal Hocko return false; 797635d9cbSMichal Hocko } 807635d9cbSMichal Hocko 81aaa9705bSMiaohe Lin static bool get_huge_zero_page(void) 8297ae1749SKirill A. Shutemov { 8397ae1749SKirill A. Shutemov struct page *zero_page; 8497ae1749SKirill A. Shutemov retry: 8597ae1749SKirill A. Shutemov if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 86aaa9705bSMiaohe Lin return true; 8797ae1749SKirill A. Shutemov 8897ae1749SKirill A. Shutemov zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 8997ae1749SKirill A. Shutemov HPAGE_PMD_ORDER); 90d8a8e1f0SKirill A. Shutemov if (!zero_page) { 91d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 92aaa9705bSMiaohe Lin return false; 93d8a8e1f0SKirill A. Shutemov } 94d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC); 9597ae1749SKirill A. Shutemov preempt_disable(); 965918d10aSKirill A. Shutemov if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 9797ae1749SKirill A. Shutemov preempt_enable(); 985ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 9997ae1749SKirill A. Shutemov goto retry; 10097ae1749SKirill A. Shutemov } 10197ae1749SKirill A. Shutemov 10297ae1749SKirill A. Shutemov /* We take additional reference here. It will be put back by shrinker */ 10397ae1749SKirill A. Shutemov atomic_set(&huge_zero_refcount, 2); 10497ae1749SKirill A. Shutemov preempt_enable(); 105aaa9705bSMiaohe Lin return true; 10697ae1749SKirill A. Shutemov } 10797ae1749SKirill A. Shutemov 1086fcb52a5SAaron Lu static void put_huge_zero_page(void) 10997ae1749SKirill A. Shutemov { 11097ae1749SKirill A. Shutemov /* 11197ae1749SKirill A. Shutemov * Counter should never go to zero here. Only shrinker can put 11297ae1749SKirill A. Shutemov * last reference. 11397ae1749SKirill A. Shutemov */ 11497ae1749SKirill A. Shutemov BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 11597ae1749SKirill A. Shutemov } 11697ae1749SKirill A. Shutemov 1176fcb52a5SAaron Lu struct page *mm_get_huge_zero_page(struct mm_struct *mm) 1186fcb52a5SAaron Lu { 1196fcb52a5SAaron Lu if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 1206fcb52a5SAaron Lu return READ_ONCE(huge_zero_page); 1216fcb52a5SAaron Lu 1226fcb52a5SAaron Lu if (!get_huge_zero_page()) 1236fcb52a5SAaron Lu return NULL; 1246fcb52a5SAaron Lu 1256fcb52a5SAaron Lu if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 1266fcb52a5SAaron Lu put_huge_zero_page(); 1276fcb52a5SAaron Lu 1286fcb52a5SAaron Lu return READ_ONCE(huge_zero_page); 1296fcb52a5SAaron Lu } 1306fcb52a5SAaron Lu 1316fcb52a5SAaron Lu void mm_put_huge_zero_page(struct mm_struct *mm) 1326fcb52a5SAaron Lu { 1336fcb52a5SAaron Lu if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 1346fcb52a5SAaron Lu put_huge_zero_page(); 1356fcb52a5SAaron Lu } 1366fcb52a5SAaron Lu 13748896466SGlauber Costa static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 13897ae1749SKirill A. Shutemov struct shrink_control *sc) 13997ae1749SKirill A. Shutemov { 14097ae1749SKirill A. Shutemov /* we can free zero page only if last reference remains */ 14197ae1749SKirill A. Shutemov return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 14248896466SGlauber Costa } 14397ae1749SKirill A. Shutemov 14448896466SGlauber Costa static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 14548896466SGlauber Costa struct shrink_control *sc) 14648896466SGlauber Costa { 14797ae1749SKirill A. Shutemov if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 1485918d10aSKirill A. Shutemov struct page *zero_page = xchg(&huge_zero_page, NULL); 1495918d10aSKirill A. Shutemov BUG_ON(zero_page == NULL); 1505ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 15148896466SGlauber Costa return HPAGE_PMD_NR; 15297ae1749SKirill A. Shutemov } 15397ae1749SKirill A. Shutemov 15497ae1749SKirill A. Shutemov return 0; 15597ae1749SKirill A. Shutemov } 15697ae1749SKirill A. Shutemov 15797ae1749SKirill A. Shutemov static struct shrinker huge_zero_page_shrinker = { 15848896466SGlauber Costa .count_objects = shrink_huge_zero_page_count, 15948896466SGlauber Costa .scan_objects = shrink_huge_zero_page_scan, 16097ae1749SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 16197ae1749SKirill A. Shutemov }; 16297ae1749SKirill A. Shutemov 16371e3aac0SAndrea Arcangeli #ifdef CONFIG_SYSFS 16471e3aac0SAndrea Arcangeli static ssize_t enabled_show(struct kobject *kobj, 16571e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 16671e3aac0SAndrea Arcangeli { 167bfb0ffebSJoe Perches const char *output; 168bfb0ffebSJoe Perches 169444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 170bfb0ffebSJoe Perches output = "[always] madvise never"; 171bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 172bfb0ffebSJoe Perches &transparent_hugepage_flags)) 173bfb0ffebSJoe Perches output = "always [madvise] never"; 174444eb2a4SMel Gorman else 175bfb0ffebSJoe Perches output = "always madvise [never]"; 176bfb0ffebSJoe Perches 177bfb0ffebSJoe Perches return sysfs_emit(buf, "%s\n", output); 17871e3aac0SAndrea Arcangeli } 179444eb2a4SMel Gorman 18071e3aac0SAndrea Arcangeli static ssize_t enabled_store(struct kobject *kobj, 18171e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 18271e3aac0SAndrea Arcangeli const char *buf, size_t count) 18371e3aac0SAndrea Arcangeli { 18421440d7eSDavid Rientjes ssize_t ret = count; 185ba76149fSAndrea Arcangeli 186f42f2552SDavid Rientjes if (sysfs_streq(buf, "always")) { 18721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 18821440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 189f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "madvise")) { 19021440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 19121440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 192f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "never")) { 19321440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 19421440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 19521440d7eSDavid Rientjes } else 19621440d7eSDavid Rientjes ret = -EINVAL; 197ba76149fSAndrea Arcangeli 198ba76149fSAndrea Arcangeli if (ret > 0) { 199b46e756fSKirill A. Shutemov int err = start_stop_khugepaged(); 200ba76149fSAndrea Arcangeli if (err) 201ba76149fSAndrea Arcangeli ret = err; 202ba76149fSAndrea Arcangeli } 203ba76149fSAndrea Arcangeli return ret; 20471e3aac0SAndrea Arcangeli } 20571e3aac0SAndrea Arcangeli static struct kobj_attribute enabled_attr = 20671e3aac0SAndrea Arcangeli __ATTR(enabled, 0644, enabled_show, enabled_store); 20771e3aac0SAndrea Arcangeli 208b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_show(struct kobject *kobj, 20971e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 21071e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 21171e3aac0SAndrea Arcangeli { 212bfb0ffebSJoe Perches return sysfs_emit(buf, "%d\n", 213e27e6151SBen Hutchings !!test_bit(flag, &transparent_hugepage_flags)); 21471e3aac0SAndrea Arcangeli } 215e27e6151SBen Hutchings 216b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_store(struct kobject *kobj, 21771e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 21871e3aac0SAndrea Arcangeli const char *buf, size_t count, 21971e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 22071e3aac0SAndrea Arcangeli { 221e27e6151SBen Hutchings unsigned long value; 222e27e6151SBen Hutchings int ret; 223e27e6151SBen Hutchings 224e27e6151SBen Hutchings ret = kstrtoul(buf, 10, &value); 225e27e6151SBen Hutchings if (ret < 0) 226e27e6151SBen Hutchings return ret; 227e27e6151SBen Hutchings if (value > 1) 22871e3aac0SAndrea Arcangeli return -EINVAL; 22971e3aac0SAndrea Arcangeli 230e27e6151SBen Hutchings if (value) 231e27e6151SBen Hutchings set_bit(flag, &transparent_hugepage_flags); 232e27e6151SBen Hutchings else 233e27e6151SBen Hutchings clear_bit(flag, &transparent_hugepage_flags); 234e27e6151SBen Hutchings 23571e3aac0SAndrea Arcangeli return count; 23671e3aac0SAndrea Arcangeli } 23771e3aac0SAndrea Arcangeli 23871e3aac0SAndrea Arcangeli static ssize_t defrag_show(struct kobject *kobj, 23971e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 24071e3aac0SAndrea Arcangeli { 241bfb0ffebSJoe Perches const char *output; 242bfb0ffebSJoe Perches 243bfb0ffebSJoe Perches if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, 244bfb0ffebSJoe Perches &transparent_hugepage_flags)) 245bfb0ffebSJoe Perches output = "[always] defer defer+madvise madvise never"; 246bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, 247bfb0ffebSJoe Perches &transparent_hugepage_flags)) 248bfb0ffebSJoe Perches output = "always [defer] defer+madvise madvise never"; 249bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, 250bfb0ffebSJoe Perches &transparent_hugepage_flags)) 251bfb0ffebSJoe Perches output = "always defer [defer+madvise] madvise never"; 252bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, 253bfb0ffebSJoe Perches &transparent_hugepage_flags)) 254bfb0ffebSJoe Perches output = "always defer defer+madvise [madvise] never"; 255bfb0ffebSJoe Perches else 256bfb0ffebSJoe Perches output = "always defer defer+madvise madvise [never]"; 257bfb0ffebSJoe Perches 258bfb0ffebSJoe Perches return sysfs_emit(buf, "%s\n", output); 25971e3aac0SAndrea Arcangeli } 26021440d7eSDavid Rientjes 26171e3aac0SAndrea Arcangeli static ssize_t defrag_store(struct kobject *kobj, 26271e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 26371e3aac0SAndrea Arcangeli const char *buf, size_t count) 26471e3aac0SAndrea Arcangeli { 265f42f2552SDavid Rientjes if (sysfs_streq(buf, "always")) { 26621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 26721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 26821440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 26921440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 270f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "defer+madvise")) { 27121440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 27221440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 27321440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 27421440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 275f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "defer")) { 2764fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 2774fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 2784fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 2794fad7fb6SDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 280f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "madvise")) { 28121440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 28221440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 28321440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 28421440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 285f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "never")) { 28621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 28721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 28821440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 28921440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 29021440d7eSDavid Rientjes } else 29121440d7eSDavid Rientjes return -EINVAL; 29221440d7eSDavid Rientjes 29321440d7eSDavid Rientjes return count; 29471e3aac0SAndrea Arcangeli } 29571e3aac0SAndrea Arcangeli static struct kobj_attribute defrag_attr = 29671e3aac0SAndrea Arcangeli __ATTR(defrag, 0644, defrag_show, defrag_store); 29771e3aac0SAndrea Arcangeli 29879da5407SKirill A. Shutemov static ssize_t use_zero_page_show(struct kobject *kobj, 29979da5407SKirill A. Shutemov struct kobj_attribute *attr, char *buf) 30079da5407SKirill A. Shutemov { 301b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 30279da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 30379da5407SKirill A. Shutemov } 30479da5407SKirill A. Shutemov static ssize_t use_zero_page_store(struct kobject *kobj, 30579da5407SKirill A. Shutemov struct kobj_attribute *attr, const char *buf, size_t count) 30679da5407SKirill A. Shutemov { 307b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 30879da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 30979da5407SKirill A. Shutemov } 31079da5407SKirill A. Shutemov static struct kobj_attribute use_zero_page_attr = 31179da5407SKirill A. Shutemov __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 31249920d28SHugh Dickins 31349920d28SHugh Dickins static ssize_t hpage_pmd_size_show(struct kobject *kobj, 31449920d28SHugh Dickins struct kobj_attribute *attr, char *buf) 31549920d28SHugh Dickins { 316ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE); 31749920d28SHugh Dickins } 31849920d28SHugh Dickins static struct kobj_attribute hpage_pmd_size_attr = 31949920d28SHugh Dickins __ATTR_RO(hpage_pmd_size); 32049920d28SHugh Dickins 32171e3aac0SAndrea Arcangeli static struct attribute *hugepage_attr[] = { 32271e3aac0SAndrea Arcangeli &enabled_attr.attr, 32371e3aac0SAndrea Arcangeli &defrag_attr.attr, 32479da5407SKirill A. Shutemov &use_zero_page_attr.attr, 32549920d28SHugh Dickins &hpage_pmd_size_attr.attr, 326396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 3275a6e75f8SKirill A. Shutemov &shmem_enabled_attr.attr, 3285a6e75f8SKirill A. Shutemov #endif 32971e3aac0SAndrea Arcangeli NULL, 33071e3aac0SAndrea Arcangeli }; 33171e3aac0SAndrea Arcangeli 3328aa95a21SArvind Yadav static const struct attribute_group hugepage_attr_group = { 33371e3aac0SAndrea Arcangeli .attrs = hugepage_attr, 334ba76149fSAndrea Arcangeli }; 335ba76149fSAndrea Arcangeli 336569e5590SShaohua Li static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 337569e5590SShaohua Li { 338569e5590SShaohua Li int err; 339569e5590SShaohua Li 340569e5590SShaohua Li *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 341569e5590SShaohua Li if (unlikely(!*hugepage_kobj)) { 342ae3a8c1cSAndrew Morton pr_err("failed to create transparent hugepage kobject\n"); 343569e5590SShaohua Li return -ENOMEM; 344569e5590SShaohua Li } 345569e5590SShaohua Li 346569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 347569e5590SShaohua Li if (err) { 348ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 349569e5590SShaohua Li goto delete_obj; 350569e5590SShaohua Li } 351569e5590SShaohua Li 352569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 353569e5590SShaohua Li if (err) { 354ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 355569e5590SShaohua Li goto remove_hp_group; 356569e5590SShaohua Li } 357569e5590SShaohua Li 358569e5590SShaohua Li return 0; 359569e5590SShaohua Li 360569e5590SShaohua Li remove_hp_group: 361569e5590SShaohua Li sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 362569e5590SShaohua Li delete_obj: 363569e5590SShaohua Li kobject_put(*hugepage_kobj); 364569e5590SShaohua Li return err; 365569e5590SShaohua Li } 366569e5590SShaohua Li 367569e5590SShaohua Li static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 368569e5590SShaohua Li { 369569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 370569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 371569e5590SShaohua Li kobject_put(hugepage_kobj); 372569e5590SShaohua Li } 373569e5590SShaohua Li #else 374569e5590SShaohua Li static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 375569e5590SShaohua Li { 376569e5590SShaohua Li return 0; 377569e5590SShaohua Li } 378569e5590SShaohua Li 379569e5590SShaohua Li static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 380569e5590SShaohua Li { 381569e5590SShaohua Li } 38271e3aac0SAndrea Arcangeli #endif /* CONFIG_SYSFS */ 38371e3aac0SAndrea Arcangeli 38471e3aac0SAndrea Arcangeli static int __init hugepage_init(void) 38571e3aac0SAndrea Arcangeli { 38671e3aac0SAndrea Arcangeli int err; 387569e5590SShaohua Li struct kobject *hugepage_kobj; 38871e3aac0SAndrea Arcangeli 3894b7167b9SAndrea Arcangeli if (!has_transparent_hugepage()) { 390bae84953SAneesh Kumar K.V /* 391bae84953SAneesh Kumar K.V * Hardware doesn't support hugepages, hence disable 392bae84953SAneesh Kumar K.V * DAX PMD support. 393bae84953SAneesh Kumar K.V */ 394bae84953SAneesh Kumar K.V transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX; 395569e5590SShaohua Li return -EINVAL; 3964b7167b9SAndrea Arcangeli } 3974b7167b9SAndrea Arcangeli 398ff20c2e0SKirill A. Shutemov /* 399ff20c2e0SKirill A. Shutemov * hugepages can't be allocated by the buddy allocator 400ff20c2e0SKirill A. Shutemov */ 401ff20c2e0SKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 402ff20c2e0SKirill A. Shutemov /* 403ff20c2e0SKirill A. Shutemov * we use page->mapping and page->index in second tail page 404ff20c2e0SKirill A. Shutemov * as list_head: assuming THP order >= 2 405ff20c2e0SKirill A. Shutemov */ 406ff20c2e0SKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 407ff20c2e0SKirill A. Shutemov 408569e5590SShaohua Li err = hugepage_init_sysfs(&hugepage_kobj); 409569e5590SShaohua Li if (err) 41065ebb64fSKirill A. Shutemov goto err_sysfs; 411ba76149fSAndrea Arcangeli 412b46e756fSKirill A. Shutemov err = khugepaged_init(); 413ba76149fSAndrea Arcangeli if (err) 41465ebb64fSKirill A. Shutemov goto err_slab; 415ba76149fSAndrea Arcangeli 41665ebb64fSKirill A. Shutemov err = register_shrinker(&huge_zero_page_shrinker); 41765ebb64fSKirill A. Shutemov if (err) 41865ebb64fSKirill A. Shutemov goto err_hzp_shrinker; 4199a982250SKirill A. Shutemov err = register_shrinker(&deferred_split_shrinker); 4209a982250SKirill A. Shutemov if (err) 4219a982250SKirill A. Shutemov goto err_split_shrinker; 42297ae1749SKirill A. Shutemov 42397562cd2SRik van Riel /* 42497562cd2SRik van Riel * By default disable transparent hugepages on smaller systems, 42597562cd2SRik van Riel * where the extra memory used could hurt more than TLB overhead 42697562cd2SRik van Riel * is likely to save. The admin can still enable it through /sys. 42797562cd2SRik van Riel */ 428ca79b0c2SArun KS if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { 42997562cd2SRik van Riel transparent_hugepage_flags = 0; 43079553da2SKirill A. Shutemov return 0; 43179553da2SKirill A. Shutemov } 43297562cd2SRik van Riel 43379553da2SKirill A. Shutemov err = start_stop_khugepaged(); 43465ebb64fSKirill A. Shutemov if (err) 43565ebb64fSKirill A. Shutemov goto err_khugepaged; 436ba76149fSAndrea Arcangeli 437569e5590SShaohua Li return 0; 43865ebb64fSKirill A. Shutemov err_khugepaged: 4399a982250SKirill A. Shutemov unregister_shrinker(&deferred_split_shrinker); 4409a982250SKirill A. Shutemov err_split_shrinker: 44165ebb64fSKirill A. Shutemov unregister_shrinker(&huge_zero_page_shrinker); 44265ebb64fSKirill A. Shutemov err_hzp_shrinker: 443b46e756fSKirill A. Shutemov khugepaged_destroy(); 44465ebb64fSKirill A. Shutemov err_slab: 445569e5590SShaohua Li hugepage_exit_sysfs(hugepage_kobj); 44665ebb64fSKirill A. Shutemov err_sysfs: 447ba76149fSAndrea Arcangeli return err; 44871e3aac0SAndrea Arcangeli } 449a64fb3cdSPaul Gortmaker subsys_initcall(hugepage_init); 45071e3aac0SAndrea Arcangeli 45171e3aac0SAndrea Arcangeli static int __init setup_transparent_hugepage(char *str) 45271e3aac0SAndrea Arcangeli { 45371e3aac0SAndrea Arcangeli int ret = 0; 45471e3aac0SAndrea Arcangeli if (!str) 45571e3aac0SAndrea Arcangeli goto out; 45671e3aac0SAndrea Arcangeli if (!strcmp(str, "always")) { 45771e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_FLAG, 45871e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 45971e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 46071e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 46171e3aac0SAndrea Arcangeli ret = 1; 46271e3aac0SAndrea Arcangeli } else if (!strcmp(str, "madvise")) { 46371e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 46471e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 46571e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 46671e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 46771e3aac0SAndrea Arcangeli ret = 1; 46871e3aac0SAndrea Arcangeli } else if (!strcmp(str, "never")) { 46971e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 47071e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 47171e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 47271e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 47371e3aac0SAndrea Arcangeli ret = 1; 47471e3aac0SAndrea Arcangeli } 47571e3aac0SAndrea Arcangeli out: 47671e3aac0SAndrea Arcangeli if (!ret) 477ae3a8c1cSAndrew Morton pr_warn("transparent_hugepage= cannot parse, ignored\n"); 47871e3aac0SAndrea Arcangeli return ret; 47971e3aac0SAndrea Arcangeli } 48071e3aac0SAndrea Arcangeli __setup("transparent_hugepage=", setup_transparent_hugepage); 48171e3aac0SAndrea Arcangeli 482f55e1014SLinus Torvalds pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 48371e3aac0SAndrea Arcangeli { 484f55e1014SLinus Torvalds if (likely(vma->vm_flags & VM_WRITE)) 48571e3aac0SAndrea Arcangeli pmd = pmd_mkwrite(pmd); 48671e3aac0SAndrea Arcangeli return pmd; 48771e3aac0SAndrea Arcangeli } 48871e3aac0SAndrea Arcangeli 48987eaceb3SYang Shi #ifdef CONFIG_MEMCG 49087eaceb3SYang Shi static inline struct deferred_split *get_deferred_split_queue(struct page *page) 4919a982250SKirill A. Shutemov { 492bcfe06bfSRoman Gushchin struct mem_cgroup *memcg = page_memcg(compound_head(page)); 49387eaceb3SYang Shi struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); 49487eaceb3SYang Shi 49587eaceb3SYang Shi if (memcg) 49687eaceb3SYang Shi return &memcg->deferred_split_queue; 49787eaceb3SYang Shi else 49887eaceb3SYang Shi return &pgdat->deferred_split_queue; 4999a982250SKirill A. Shutemov } 50087eaceb3SYang Shi #else 50187eaceb3SYang Shi static inline struct deferred_split *get_deferred_split_queue(struct page *page) 50287eaceb3SYang Shi { 50387eaceb3SYang Shi struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); 50487eaceb3SYang Shi 50587eaceb3SYang Shi return &pgdat->deferred_split_queue; 50687eaceb3SYang Shi } 50787eaceb3SYang Shi #endif 5089a982250SKirill A. Shutemov 5099a982250SKirill A. Shutemov void prep_transhuge_page(struct page *page) 5109a982250SKirill A. Shutemov { 5119a982250SKirill A. Shutemov /* 5129a982250SKirill A. Shutemov * we use page->mapping and page->indexlru in second tail page 5139a982250SKirill A. Shutemov * as list_head: assuming THP order >= 2 5149a982250SKirill A. Shutemov */ 5159a982250SKirill A. Shutemov 5169a982250SKirill A. Shutemov INIT_LIST_HEAD(page_deferred_list(page)); 5179a982250SKirill A. Shutemov set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 5189a982250SKirill A. Shutemov } 5199a982250SKirill A. Shutemov 520005ba37cSSean Christopherson bool is_transparent_hugepage(struct page *page) 521005ba37cSSean Christopherson { 522005ba37cSSean Christopherson if (!PageCompound(page)) 523fa1f68ccSZou Wei return false; 524005ba37cSSean Christopherson 525005ba37cSSean Christopherson page = compound_head(page); 526005ba37cSSean Christopherson return is_huge_zero_page(page) || 527005ba37cSSean Christopherson page[1].compound_dtor == TRANSHUGE_PAGE_DTOR; 528005ba37cSSean Christopherson } 529005ba37cSSean Christopherson EXPORT_SYMBOL_GPL(is_transparent_hugepage); 530005ba37cSSean Christopherson 53197d3d0f9SKirill A. Shutemov static unsigned long __thp_get_unmapped_area(struct file *filp, 53297d3d0f9SKirill A. Shutemov unsigned long addr, unsigned long len, 53374d2fad1SToshi Kani loff_t off, unsigned long flags, unsigned long size) 53474d2fad1SToshi Kani { 53574d2fad1SToshi Kani loff_t off_end = off + len; 53674d2fad1SToshi Kani loff_t off_align = round_up(off, size); 53797d3d0f9SKirill A. Shutemov unsigned long len_pad, ret; 53874d2fad1SToshi Kani 53974d2fad1SToshi Kani if (off_end <= off_align || (off_end - off_align) < size) 54074d2fad1SToshi Kani return 0; 54174d2fad1SToshi Kani 54274d2fad1SToshi Kani len_pad = len + size; 54374d2fad1SToshi Kani if (len_pad < len || (off + len_pad) < off) 54474d2fad1SToshi Kani return 0; 54574d2fad1SToshi Kani 54697d3d0f9SKirill A. Shutemov ret = current->mm->get_unmapped_area(filp, addr, len_pad, 54774d2fad1SToshi Kani off >> PAGE_SHIFT, flags); 54897d3d0f9SKirill A. Shutemov 54997d3d0f9SKirill A. Shutemov /* 55097d3d0f9SKirill A. Shutemov * The failure might be due to length padding. The caller will retry 55197d3d0f9SKirill A. Shutemov * without the padding. 55297d3d0f9SKirill A. Shutemov */ 55397d3d0f9SKirill A. Shutemov if (IS_ERR_VALUE(ret)) 55474d2fad1SToshi Kani return 0; 55574d2fad1SToshi Kani 55697d3d0f9SKirill A. Shutemov /* 55797d3d0f9SKirill A. Shutemov * Do not try to align to THP boundary if allocation at the address 55897d3d0f9SKirill A. Shutemov * hint succeeds. 55997d3d0f9SKirill A. Shutemov */ 56097d3d0f9SKirill A. Shutemov if (ret == addr) 56174d2fad1SToshi Kani return addr; 56297d3d0f9SKirill A. Shutemov 56397d3d0f9SKirill A. Shutemov ret += (off - ret) & (size - 1); 56497d3d0f9SKirill A. Shutemov return ret; 56574d2fad1SToshi Kani } 56674d2fad1SToshi Kani 56774d2fad1SToshi Kani unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 56874d2fad1SToshi Kani unsigned long len, unsigned long pgoff, unsigned long flags) 56974d2fad1SToshi Kani { 57097d3d0f9SKirill A. Shutemov unsigned long ret; 57174d2fad1SToshi Kani loff_t off = (loff_t)pgoff << PAGE_SHIFT; 57274d2fad1SToshi Kani 57374d2fad1SToshi Kani if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) 57474d2fad1SToshi Kani goto out; 57574d2fad1SToshi Kani 57697d3d0f9SKirill A. Shutemov ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); 57797d3d0f9SKirill A. Shutemov if (ret) 57897d3d0f9SKirill A. Shutemov return ret; 57974d2fad1SToshi Kani out: 58074d2fad1SToshi Kani return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 58174d2fad1SToshi Kani } 58274d2fad1SToshi Kani EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 58374d2fad1SToshi Kani 5842b740303SSouptick Joarder static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 5852b740303SSouptick Joarder struct page *page, gfp_t gfp) 58671e3aac0SAndrea Arcangeli { 58782b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 58871e3aac0SAndrea Arcangeli pgtable_t pgtable; 58982b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 5902b740303SSouptick Joarder vm_fault_t ret = 0; 59171e3aac0SAndrea Arcangeli 592309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page), page); 59300501b53SJohannes Weiner 594d9eb1ea2SJohannes Weiner if (mem_cgroup_charge(page, vma->vm_mm, gfp)) { 5956b251fc9SAndrea Arcangeli put_page(page); 5966b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_FALLBACK); 59785b9f46eSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK_CHARGE); 5986b251fc9SAndrea Arcangeli return VM_FAULT_FALLBACK; 5996b251fc9SAndrea Arcangeli } 6009d82c694SJohannes Weiner cgroup_throttle_swaprate(page, gfp); 60171e3aac0SAndrea Arcangeli 6024cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 60300501b53SJohannes Weiner if (unlikely(!pgtable)) { 6046b31d595SMichal Hocko ret = VM_FAULT_OOM; 6056b31d595SMichal Hocko goto release; 60600501b53SJohannes Weiner } 60700501b53SJohannes Weiner 608c79b57e4SHuang Ying clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 60952f37629SMinchan Kim /* 61052f37629SMinchan Kim * The memory barrier inside __SetPageUptodate makes sure that 61152f37629SMinchan Kim * clear_huge_page writes become visible before the set_pmd_at() 61252f37629SMinchan Kim * write. 61352f37629SMinchan Kim */ 61471e3aac0SAndrea Arcangeli __SetPageUptodate(page); 61571e3aac0SAndrea Arcangeli 61682b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 61782b0f8c3SJan Kara if (unlikely(!pmd_none(*vmf->pmd))) { 6186b31d595SMichal Hocko goto unlock_release; 61971e3aac0SAndrea Arcangeli } else { 62071e3aac0SAndrea Arcangeli pmd_t entry; 6216b251fc9SAndrea Arcangeli 6226b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 6236b31d595SMichal Hocko if (ret) 6246b31d595SMichal Hocko goto unlock_release; 6256b31d595SMichal Hocko 6266b251fc9SAndrea Arcangeli /* Deliver the page fault to userland */ 6276b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 62882b0f8c3SJan Kara spin_unlock(vmf->ptl); 6296b251fc9SAndrea Arcangeli put_page(page); 630bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 6318fd5eda4SMiaohe Lin ret = handle_userfault(vmf, VM_UFFD_MISSING); 6328fd5eda4SMiaohe Lin VM_BUG_ON(ret & VM_FAULT_FALLBACK); 6338fd5eda4SMiaohe Lin return ret; 6346b251fc9SAndrea Arcangeli } 6356b251fc9SAndrea Arcangeli 6363122359aSKirill A. Shutemov entry = mk_huge_pmd(page, vma->vm_page_prot); 637f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 638d281ee61SKirill A. Shutemov page_add_new_anon_rmap(page, vma, haddr, true); 639b518154eSJoonsoo Kim lru_cache_add_inactive_or_unevictable(page, vma); 64082b0f8c3SJan Kara pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 64182b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 642fca40573SBibo Mao update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 643bae473a4SKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 644c4812909SKirill A. Shutemov mm_inc_nr_ptes(vma->vm_mm); 64582b0f8c3SJan Kara spin_unlock(vmf->ptl); 6466b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_ALLOC); 6479d82c694SJohannes Weiner count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); 64871e3aac0SAndrea Arcangeli } 64971e3aac0SAndrea Arcangeli 650aa2e878eSDavid Rientjes return 0; 6516b31d595SMichal Hocko unlock_release: 6526b31d595SMichal Hocko spin_unlock(vmf->ptl); 6536b31d595SMichal Hocko release: 6546b31d595SMichal Hocko if (pgtable) 6556b31d595SMichal Hocko pte_free(vma->vm_mm, pgtable); 6566b31d595SMichal Hocko put_page(page); 6576b31d595SMichal Hocko return ret; 6586b31d595SMichal Hocko 65971e3aac0SAndrea Arcangeli } 66071e3aac0SAndrea Arcangeli 661444eb2a4SMel Gorman /* 66221440d7eSDavid Rientjes * always: directly stall for all thp allocations 66321440d7eSDavid Rientjes * defer: wake kswapd and fail if not immediately available 66421440d7eSDavid Rientjes * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 66521440d7eSDavid Rientjes * fail if not immediately available 66621440d7eSDavid Rientjes * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 66721440d7eSDavid Rientjes * available 66821440d7eSDavid Rientjes * never: never stall for any thp allocation 669444eb2a4SMel Gorman */ 670164cc4feSRik van Riel gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma) 6710bbbc0b3SAndrea Arcangeli { 672164cc4feSRik van Riel const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE); 67389c83fb5SMichal Hocko 674ac79f78dSDavid Rientjes /* Always do synchronous compaction */ 67521440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 676a8282608SAndrea Arcangeli return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 677ac79f78dSDavid Rientjes 678ac79f78dSDavid Rientjes /* Kick kcompactd and fail quickly */ 67921440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 68019deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 681ac79f78dSDavid Rientjes 682ac79f78dSDavid Rientjes /* Synchronous compaction if madvised, otherwise kick kcompactd */ 68321440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 68419deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | 68519deb769SDavid Rientjes (vma_madvised ? __GFP_DIRECT_RECLAIM : 686ac79f78dSDavid Rientjes __GFP_KSWAPD_RECLAIM); 687ac79f78dSDavid Rientjes 688ac79f78dSDavid Rientjes /* Only do synchronous compaction if madvised */ 68921440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 69019deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | 69119deb769SDavid Rientjes (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 692ac79f78dSDavid Rientjes 69319deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT; 694444eb2a4SMel Gorman } 695444eb2a4SMel Gorman 696c4088ebdSKirill A. Shutemov /* Caller must hold page table lock. */ 6972efeb8daSMiaohe Lin static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 69897ae1749SKirill A. Shutemov struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 6995918d10aSKirill A. Shutemov struct page *zero_page) 700fc9fe822SKirill A. Shutemov { 701fc9fe822SKirill A. Shutemov pmd_t entry; 7027c414164SAndrew Morton if (!pmd_none(*pmd)) 7032efeb8daSMiaohe Lin return; 7045918d10aSKirill A. Shutemov entry = mk_pmd(zero_page, vma->vm_page_prot); 705fc9fe822SKirill A. Shutemov entry = pmd_mkhuge(entry); 70612c9d70bSMatthew Wilcox if (pgtable) 7076b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 708fc9fe822SKirill A. Shutemov set_pmd_at(mm, haddr, pmd, entry); 709c4812909SKirill A. Shutemov mm_inc_nr_ptes(mm); 710fc9fe822SKirill A. Shutemov } 711fc9fe822SKirill A. Shutemov 7122b740303SSouptick Joarder vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 71371e3aac0SAndrea Arcangeli { 71482b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 715077fcf11SAneesh Kumar K.V gfp_t gfp; 71671e3aac0SAndrea Arcangeli struct page *page; 71782b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 71871e3aac0SAndrea Arcangeli 71943675e6fSYang Shi if (!transhuge_vma_suitable(vma, haddr)) 720c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 72171e3aac0SAndrea Arcangeli if (unlikely(anon_vma_prepare(vma))) 72271e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 7236d50e60cSDavid Rientjes if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 724ba76149fSAndrea Arcangeli return VM_FAULT_OOM; 72582b0f8c3SJan Kara if (!(vmf->flags & FAULT_FLAG_WRITE) && 726bae473a4SKirill A. Shutemov !mm_forbids_zeropage(vma->vm_mm) && 72779da5407SKirill A. Shutemov transparent_hugepage_use_zero_page()) { 72880371957SKirill A. Shutemov pgtable_t pgtable; 7295918d10aSKirill A. Shutemov struct page *zero_page; 7302b740303SSouptick Joarder vm_fault_t ret; 7314cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 73280371957SKirill A. Shutemov if (unlikely(!pgtable)) 73380371957SKirill A. Shutemov return VM_FAULT_OOM; 7346fcb52a5SAaron Lu zero_page = mm_get_huge_zero_page(vma->vm_mm); 7355918d10aSKirill A. Shutemov if (unlikely(!zero_page)) { 736bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 73797ae1749SKirill A. Shutemov count_vm_event(THP_FAULT_FALLBACK); 738c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 73997ae1749SKirill A. Shutemov } 74082b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 7416b251fc9SAndrea Arcangeli ret = 0; 74282b0f8c3SJan Kara if (pmd_none(*vmf->pmd)) { 7436b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 7446b31d595SMichal Hocko if (ret) { 7456b31d595SMichal Hocko spin_unlock(vmf->ptl); 746bfe8cc1dSGerald Schaefer pte_free(vma->vm_mm, pgtable); 7476b31d595SMichal Hocko } else if (userfaultfd_missing(vma)) { 74882b0f8c3SJan Kara spin_unlock(vmf->ptl); 749bfe8cc1dSGerald Schaefer pte_free(vma->vm_mm, pgtable); 75082b0f8c3SJan Kara ret = handle_userfault(vmf, VM_UFFD_MISSING); 7516b251fc9SAndrea Arcangeli VM_BUG_ON(ret & VM_FAULT_FALLBACK); 7526b251fc9SAndrea Arcangeli } else { 753bae473a4SKirill A. Shutemov set_huge_zero_page(pgtable, vma->vm_mm, vma, 75482b0f8c3SJan Kara haddr, vmf->pmd, zero_page); 755fca40573SBibo Mao update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 75682b0f8c3SJan Kara spin_unlock(vmf->ptl); 7576b251fc9SAndrea Arcangeli } 758bfe8cc1dSGerald Schaefer } else { 75982b0f8c3SJan Kara spin_unlock(vmf->ptl); 760bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 761bfe8cc1dSGerald Schaefer } 7626b251fc9SAndrea Arcangeli return ret; 76380371957SKirill A. Shutemov } 764164cc4feSRik van Riel gfp = vma_thp_gfp_mask(vma); 76519deb769SDavid Rientjes page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 76681ab4201SAndi Kleen if (unlikely(!page)) { 76781ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 768c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 76981ab4201SAndi Kleen } 7709a982250SKirill A. Shutemov prep_transhuge_page(page); 77182b0f8c3SJan Kara return __do_huge_pmd_anonymous_page(vmf, page, gfp); 77271e3aac0SAndrea Arcangeli } 77371e3aac0SAndrea Arcangeli 774ae18d6dcSMatthew Wilcox static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 7753b6521f5SOliver O'Halloran pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 7763b6521f5SOliver O'Halloran pgtable_t pgtable) 7775cad465dSMatthew Wilcox { 7785cad465dSMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 7795cad465dSMatthew Wilcox pmd_t entry; 7805cad465dSMatthew Wilcox spinlock_t *ptl; 7815cad465dSMatthew Wilcox 7825cad465dSMatthew Wilcox ptl = pmd_lock(mm, pmd); 783c6f3c5eeSAneesh Kumar K.V if (!pmd_none(*pmd)) { 784c6f3c5eeSAneesh Kumar K.V if (write) { 785c6f3c5eeSAneesh Kumar K.V if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { 786c6f3c5eeSAneesh Kumar K.V WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); 787c6f3c5eeSAneesh Kumar K.V goto out_unlock; 788c6f3c5eeSAneesh Kumar K.V } 789c6f3c5eeSAneesh Kumar K.V entry = pmd_mkyoung(*pmd); 790c6f3c5eeSAneesh Kumar K.V entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 791c6f3c5eeSAneesh Kumar K.V if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) 792c6f3c5eeSAneesh Kumar K.V update_mmu_cache_pmd(vma, addr, pmd); 793c6f3c5eeSAneesh Kumar K.V } 794c6f3c5eeSAneesh Kumar K.V 795c6f3c5eeSAneesh Kumar K.V goto out_unlock; 796c6f3c5eeSAneesh Kumar K.V } 797c6f3c5eeSAneesh Kumar K.V 798f25748e3SDan Williams entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 799f25748e3SDan Williams if (pfn_t_devmap(pfn)) 800f25748e3SDan Williams entry = pmd_mkdevmap(entry); 8015cad465dSMatthew Wilcox if (write) { 802f55e1014SLinus Torvalds entry = pmd_mkyoung(pmd_mkdirty(entry)); 803f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(entry, vma); 8045cad465dSMatthew Wilcox } 8053b6521f5SOliver O'Halloran 8063b6521f5SOliver O'Halloran if (pgtable) { 8073b6521f5SOliver O'Halloran pgtable_trans_huge_deposit(mm, pmd, pgtable); 808c4812909SKirill A. Shutemov mm_inc_nr_ptes(mm); 809c6f3c5eeSAneesh Kumar K.V pgtable = NULL; 8103b6521f5SOliver O'Halloran } 8113b6521f5SOliver O'Halloran 8125cad465dSMatthew Wilcox set_pmd_at(mm, addr, pmd, entry); 8135cad465dSMatthew Wilcox update_mmu_cache_pmd(vma, addr, pmd); 814c6f3c5eeSAneesh Kumar K.V 815c6f3c5eeSAneesh Kumar K.V out_unlock: 8165cad465dSMatthew Wilcox spin_unlock(ptl); 817c6f3c5eeSAneesh Kumar K.V if (pgtable) 818c6f3c5eeSAneesh Kumar K.V pte_free(mm, pgtable); 8195cad465dSMatthew Wilcox } 8205cad465dSMatthew Wilcox 8219a9731b1SThomas Hellstrom (VMware) /** 8229a9731b1SThomas Hellstrom (VMware) * vmf_insert_pfn_pmd_prot - insert a pmd size pfn 8239a9731b1SThomas Hellstrom (VMware) * @vmf: Structure describing the fault 8249a9731b1SThomas Hellstrom (VMware) * @pfn: pfn to insert 8259a9731b1SThomas Hellstrom (VMware) * @pgprot: page protection to use 8269a9731b1SThomas Hellstrom (VMware) * @write: whether it's a write fault 8279a9731b1SThomas Hellstrom (VMware) * 8289a9731b1SThomas Hellstrom (VMware) * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and 8299a9731b1SThomas Hellstrom (VMware) * also consult the vmf_insert_mixed_prot() documentation when 8309a9731b1SThomas Hellstrom (VMware) * @pgprot != @vmf->vma->vm_page_prot. 8319a9731b1SThomas Hellstrom (VMware) * 8329a9731b1SThomas Hellstrom (VMware) * Return: vm_fault_t value. 8339a9731b1SThomas Hellstrom (VMware) */ 8349a9731b1SThomas Hellstrom (VMware) vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn, 8359a9731b1SThomas Hellstrom (VMware) pgprot_t pgprot, bool write) 8365cad465dSMatthew Wilcox { 837fce86ff5SDan Williams unsigned long addr = vmf->address & PMD_MASK; 838fce86ff5SDan Williams struct vm_area_struct *vma = vmf->vma; 8393b6521f5SOliver O'Halloran pgtable_t pgtable = NULL; 840fce86ff5SDan Williams 8415cad465dSMatthew Wilcox /* 8425cad465dSMatthew Wilcox * If we had pmd_special, we could avoid all these restrictions, 8435cad465dSMatthew Wilcox * but we need to be consistent with PTEs and architectures that 8445cad465dSMatthew Wilcox * can't support a 'special' bit. 8455cad465dSMatthew Wilcox */ 846e1fb4a08SDave Jiang BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 847e1fb4a08SDave Jiang !pfn_t_devmap(pfn)); 8485cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 8495cad465dSMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 8505cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 8515cad465dSMatthew Wilcox 8525cad465dSMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 8535cad465dSMatthew Wilcox return VM_FAULT_SIGBUS; 854308a047cSBorislav Petkov 8553b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) { 8564cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 8573b6521f5SOliver O'Halloran if (!pgtable) 8583b6521f5SOliver O'Halloran return VM_FAULT_OOM; 8593b6521f5SOliver O'Halloran } 8603b6521f5SOliver O'Halloran 861308a047cSBorislav Petkov track_pfn_insert(vma, &pgprot, pfn); 862308a047cSBorislav Petkov 863fce86ff5SDan Williams insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); 864ae18d6dcSMatthew Wilcox return VM_FAULT_NOPAGE; 8655cad465dSMatthew Wilcox } 8669a9731b1SThomas Hellstrom (VMware) EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot); 8675cad465dSMatthew Wilcox 868a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 869f55e1014SLinus Torvalds static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 870a00cc7d9SMatthew Wilcox { 871f55e1014SLinus Torvalds if (likely(vma->vm_flags & VM_WRITE)) 872a00cc7d9SMatthew Wilcox pud = pud_mkwrite(pud); 873a00cc7d9SMatthew Wilcox return pud; 874a00cc7d9SMatthew Wilcox } 875a00cc7d9SMatthew Wilcox 876a00cc7d9SMatthew Wilcox static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 877a00cc7d9SMatthew Wilcox pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 878a00cc7d9SMatthew Wilcox { 879a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 880a00cc7d9SMatthew Wilcox pud_t entry; 881a00cc7d9SMatthew Wilcox spinlock_t *ptl; 882a00cc7d9SMatthew Wilcox 883a00cc7d9SMatthew Wilcox ptl = pud_lock(mm, pud); 884c6f3c5eeSAneesh Kumar K.V if (!pud_none(*pud)) { 885c6f3c5eeSAneesh Kumar K.V if (write) { 886c6f3c5eeSAneesh Kumar K.V if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { 887c6f3c5eeSAneesh Kumar K.V WARN_ON_ONCE(!is_huge_zero_pud(*pud)); 888c6f3c5eeSAneesh Kumar K.V goto out_unlock; 889c6f3c5eeSAneesh Kumar K.V } 890c6f3c5eeSAneesh Kumar K.V entry = pud_mkyoung(*pud); 891c6f3c5eeSAneesh Kumar K.V entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); 892c6f3c5eeSAneesh Kumar K.V if (pudp_set_access_flags(vma, addr, pud, entry, 1)) 893c6f3c5eeSAneesh Kumar K.V update_mmu_cache_pud(vma, addr, pud); 894c6f3c5eeSAneesh Kumar K.V } 895c6f3c5eeSAneesh Kumar K.V goto out_unlock; 896c6f3c5eeSAneesh Kumar K.V } 897c6f3c5eeSAneesh Kumar K.V 898a00cc7d9SMatthew Wilcox entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 899a00cc7d9SMatthew Wilcox if (pfn_t_devmap(pfn)) 900a00cc7d9SMatthew Wilcox entry = pud_mkdevmap(entry); 901a00cc7d9SMatthew Wilcox if (write) { 902f55e1014SLinus Torvalds entry = pud_mkyoung(pud_mkdirty(entry)); 903f55e1014SLinus Torvalds entry = maybe_pud_mkwrite(entry, vma); 904a00cc7d9SMatthew Wilcox } 905a00cc7d9SMatthew Wilcox set_pud_at(mm, addr, pud, entry); 906a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 907c6f3c5eeSAneesh Kumar K.V 908c6f3c5eeSAneesh Kumar K.V out_unlock: 909a00cc7d9SMatthew Wilcox spin_unlock(ptl); 910a00cc7d9SMatthew Wilcox } 911a00cc7d9SMatthew Wilcox 9129a9731b1SThomas Hellstrom (VMware) /** 9139a9731b1SThomas Hellstrom (VMware) * vmf_insert_pfn_pud_prot - insert a pud size pfn 9149a9731b1SThomas Hellstrom (VMware) * @vmf: Structure describing the fault 9159a9731b1SThomas Hellstrom (VMware) * @pfn: pfn to insert 9169a9731b1SThomas Hellstrom (VMware) * @pgprot: page protection to use 9179a9731b1SThomas Hellstrom (VMware) * @write: whether it's a write fault 9189a9731b1SThomas Hellstrom (VMware) * 9199a9731b1SThomas Hellstrom (VMware) * Insert a pud size pfn. See vmf_insert_pfn() for additional info and 9209a9731b1SThomas Hellstrom (VMware) * also consult the vmf_insert_mixed_prot() documentation when 9219a9731b1SThomas Hellstrom (VMware) * @pgprot != @vmf->vma->vm_page_prot. 9229a9731b1SThomas Hellstrom (VMware) * 9239a9731b1SThomas Hellstrom (VMware) * Return: vm_fault_t value. 9249a9731b1SThomas Hellstrom (VMware) */ 9259a9731b1SThomas Hellstrom (VMware) vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn, 9269a9731b1SThomas Hellstrom (VMware) pgprot_t pgprot, bool write) 927a00cc7d9SMatthew Wilcox { 928fce86ff5SDan Williams unsigned long addr = vmf->address & PUD_MASK; 929fce86ff5SDan Williams struct vm_area_struct *vma = vmf->vma; 930fce86ff5SDan Williams 931a00cc7d9SMatthew Wilcox /* 932a00cc7d9SMatthew Wilcox * If we had pud_special, we could avoid all these restrictions, 933a00cc7d9SMatthew Wilcox * but we need to be consistent with PTEs and architectures that 934a00cc7d9SMatthew Wilcox * can't support a 'special' bit. 935a00cc7d9SMatthew Wilcox */ 93662ec0d8cSDave Jiang BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 93762ec0d8cSDave Jiang !pfn_t_devmap(pfn)); 938a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 939a00cc7d9SMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 940a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 941a00cc7d9SMatthew Wilcox 942a00cc7d9SMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 943a00cc7d9SMatthew Wilcox return VM_FAULT_SIGBUS; 944a00cc7d9SMatthew Wilcox 945a00cc7d9SMatthew Wilcox track_pfn_insert(vma, &pgprot, pfn); 946a00cc7d9SMatthew Wilcox 947fce86ff5SDan Williams insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write); 948a00cc7d9SMatthew Wilcox return VM_FAULT_NOPAGE; 949a00cc7d9SMatthew Wilcox } 9509a9731b1SThomas Hellstrom (VMware) EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot); 951a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 952a00cc7d9SMatthew Wilcox 9533565fce3SDan Williams static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 954a8f97366SKirill A. Shutemov pmd_t *pmd, int flags) 9553565fce3SDan Williams { 9563565fce3SDan Williams pmd_t _pmd; 9573565fce3SDan Williams 958a8f97366SKirill A. Shutemov _pmd = pmd_mkyoung(*pmd); 959a8f97366SKirill A. Shutemov if (flags & FOLL_WRITE) 960a8f97366SKirill A. Shutemov _pmd = pmd_mkdirty(_pmd); 9613565fce3SDan Williams if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 962a8f97366SKirill A. Shutemov pmd, _pmd, flags & FOLL_WRITE)) 9633565fce3SDan Williams update_mmu_cache_pmd(vma, addr, pmd); 9643565fce3SDan Williams } 9653565fce3SDan Williams 9663565fce3SDan Williams struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 967df06b37fSKeith Busch pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 9683565fce3SDan Williams { 9693565fce3SDan Williams unsigned long pfn = pmd_pfn(*pmd); 9703565fce3SDan Williams struct mm_struct *mm = vma->vm_mm; 9713565fce3SDan Williams struct page *page; 9723565fce3SDan Williams 9733565fce3SDan Williams assert_spin_locked(pmd_lockptr(mm, pmd)); 9743565fce3SDan Williams 9758310d48bSKeno Fischer /* 9768310d48bSKeno Fischer * When we COW a devmap PMD entry, we split it into PTEs, so we should 9778310d48bSKeno Fischer * not be in this function with `flags & FOLL_COW` set. 9788310d48bSKeno Fischer */ 9798310d48bSKeno Fischer WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 9808310d48bSKeno Fischer 9813faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 9823faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 9833faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 9843faa52c0SJohn Hubbard return NULL; 9853faa52c0SJohn Hubbard 986f6f37321SLinus Torvalds if (flags & FOLL_WRITE && !pmd_write(*pmd)) 9873565fce3SDan Williams return NULL; 9883565fce3SDan Williams 9893565fce3SDan Williams if (pmd_present(*pmd) && pmd_devmap(*pmd)) 9903565fce3SDan Williams /* pass */; 9913565fce3SDan Williams else 9923565fce3SDan Williams return NULL; 9933565fce3SDan Williams 9943565fce3SDan Williams if (flags & FOLL_TOUCH) 995a8f97366SKirill A. Shutemov touch_pmd(vma, addr, pmd, flags); 9963565fce3SDan Williams 9973565fce3SDan Williams /* 9983565fce3SDan Williams * device mapped pages can only be returned if the 9993565fce3SDan Williams * caller will manage the page reference count. 10003565fce3SDan Williams */ 10013faa52c0SJohn Hubbard if (!(flags & (FOLL_GET | FOLL_PIN))) 10023565fce3SDan Williams return ERR_PTR(-EEXIST); 10033565fce3SDan Williams 10043565fce3SDan Williams pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 1005df06b37fSKeith Busch *pgmap = get_dev_pagemap(pfn, *pgmap); 1006df06b37fSKeith Busch if (!*pgmap) 10073565fce3SDan Williams return ERR_PTR(-EFAULT); 10083565fce3SDan Williams page = pfn_to_page(pfn); 10093faa52c0SJohn Hubbard if (!try_grab_page(page, flags)) 10103faa52c0SJohn Hubbard page = ERR_PTR(-ENOMEM); 10113565fce3SDan Williams 10123565fce3SDan Williams return page; 10133565fce3SDan Williams } 10143565fce3SDan Williams 101571e3aac0SAndrea Arcangeli int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 101671e3aac0SAndrea Arcangeli pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 101771e3aac0SAndrea Arcangeli struct vm_area_struct *vma) 101871e3aac0SAndrea Arcangeli { 1019c4088ebdSKirill A. Shutemov spinlock_t *dst_ptl, *src_ptl; 102071e3aac0SAndrea Arcangeli struct page *src_page; 102171e3aac0SAndrea Arcangeli pmd_t pmd; 102212c9d70bSMatthew Wilcox pgtable_t pgtable = NULL; 1023628d47ceSKirill A. Shutemov int ret = -ENOMEM; 102471e3aac0SAndrea Arcangeli 1025628d47ceSKirill A. Shutemov /* Skip if can be re-fill on fault */ 1026628d47ceSKirill A. Shutemov if (!vma_is_anonymous(vma)) 1027628d47ceSKirill A. Shutemov return 0; 1028628d47ceSKirill A. Shutemov 10294cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(dst_mm); 103071e3aac0SAndrea Arcangeli if (unlikely(!pgtable)) 103171e3aac0SAndrea Arcangeli goto out; 103271e3aac0SAndrea Arcangeli 1033c4088ebdSKirill A. Shutemov dst_ptl = pmd_lock(dst_mm, dst_pmd); 1034c4088ebdSKirill A. Shutemov src_ptl = pmd_lockptr(src_mm, src_pmd); 1035c4088ebdSKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 103671e3aac0SAndrea Arcangeli 103771e3aac0SAndrea Arcangeli ret = -EAGAIN; 103871e3aac0SAndrea Arcangeli pmd = *src_pmd; 103984c3fc4eSZi Yan 1040b569a176SPeter Xu /* 1041b569a176SPeter Xu * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA 1042b569a176SPeter Xu * does not have the VM_UFFD_WP, which means that the uffd 1043b569a176SPeter Xu * fork event is not enabled. 1044b569a176SPeter Xu */ 1045b569a176SPeter Xu if (!(vma->vm_flags & VM_UFFD_WP)) 1046b569a176SPeter Xu pmd = pmd_clear_uffd_wp(pmd); 1047b569a176SPeter Xu 104884c3fc4eSZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 104984c3fc4eSZi Yan if (unlikely(is_swap_pmd(pmd))) { 105084c3fc4eSZi Yan swp_entry_t entry = pmd_to_swp_entry(pmd); 105184c3fc4eSZi Yan 105284c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(pmd)); 105384c3fc4eSZi Yan if (is_write_migration_entry(entry)) { 105484c3fc4eSZi Yan make_migration_entry_read(&entry); 105584c3fc4eSZi Yan pmd = swp_entry_to_pmd(entry); 1056ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*src_pmd)) 1057ab6e3d09SNaoya Horiguchi pmd = pmd_swp_mksoft_dirty(pmd); 105884c3fc4eSZi Yan set_pmd_at(src_mm, addr, src_pmd, pmd); 105984c3fc4eSZi Yan } 1060dd8a67f9SZi Yan add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1061af5b0f6aSKirill A. Shutemov mm_inc_nr_ptes(dst_mm); 1062dd8a67f9SZi Yan pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 106384c3fc4eSZi Yan set_pmd_at(dst_mm, addr, dst_pmd, pmd); 106484c3fc4eSZi Yan ret = 0; 106584c3fc4eSZi Yan goto out_unlock; 106684c3fc4eSZi Yan } 106784c3fc4eSZi Yan #endif 106884c3fc4eSZi Yan 1069628d47ceSKirill A. Shutemov if (unlikely(!pmd_trans_huge(pmd))) { 107071e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 107171e3aac0SAndrea Arcangeli goto out_unlock; 107271e3aac0SAndrea Arcangeli } 1073fc9fe822SKirill A. Shutemov /* 1074c4088ebdSKirill A. Shutemov * When page table lock is held, the huge zero pmd should not be 1075fc9fe822SKirill A. Shutemov * under splitting since we don't split the page itself, only pmd to 1076fc9fe822SKirill A. Shutemov * a page table. 1077fc9fe822SKirill A. Shutemov */ 1078fc9fe822SKirill A. Shutemov if (is_huge_zero_pmd(pmd)) { 10795918d10aSKirill A. Shutemov struct page *zero_page; 108097ae1749SKirill A. Shutemov /* 108197ae1749SKirill A. Shutemov * get_huge_zero_page() will never allocate a new page here, 108297ae1749SKirill A. Shutemov * since we already have a zero page to copy. It just takes a 108397ae1749SKirill A. Shutemov * reference. 108497ae1749SKirill A. Shutemov */ 10856fcb52a5SAaron Lu zero_page = mm_get_huge_zero_page(dst_mm); 10866b251fc9SAndrea Arcangeli set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 10875918d10aSKirill A. Shutemov zero_page); 1088fc9fe822SKirill A. Shutemov ret = 0; 1089fc9fe822SKirill A. Shutemov goto out_unlock; 1090fc9fe822SKirill A. Shutemov } 1091de466bd6SMel Gorman 109271e3aac0SAndrea Arcangeli src_page = pmd_page(pmd); 1093309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 1094d042035eSPeter Xu 1095d042035eSPeter Xu /* 1096d042035eSPeter Xu * If this page is a potentially pinned page, split and retry the fault 1097d042035eSPeter Xu * with smaller page size. Normally this should not happen because the 1098d042035eSPeter Xu * userspace should use MADV_DONTFORK upon pinned regions. This is a 1099d042035eSPeter Xu * best effort that the pinned pages won't be replaced by another 1100d042035eSPeter Xu * random page during the coming copy-on-write. 1101d042035eSPeter Xu */ 110297a7e473SPeter Xu if (unlikely(page_needs_cow_for_dma(vma, src_page))) { 1103d042035eSPeter Xu pte_free(dst_mm, pgtable); 1104d042035eSPeter Xu spin_unlock(src_ptl); 1105d042035eSPeter Xu spin_unlock(dst_ptl); 1106d042035eSPeter Xu __split_huge_pmd(vma, src_pmd, addr, false, NULL); 1107d042035eSPeter Xu return -EAGAIN; 1108d042035eSPeter Xu } 1109d042035eSPeter Xu 111071e3aac0SAndrea Arcangeli get_page(src_page); 111153f9263bSKirill A. Shutemov page_dup_rmap(src_page, true); 111271e3aac0SAndrea Arcangeli add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1113c4812909SKirill A. Shutemov mm_inc_nr_ptes(dst_mm); 11145c7fb56eSDan Williams pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 111571e3aac0SAndrea Arcangeli 111671e3aac0SAndrea Arcangeli pmdp_set_wrprotect(src_mm, addr, src_pmd); 111771e3aac0SAndrea Arcangeli pmd = pmd_mkold(pmd_wrprotect(pmd)); 111871e3aac0SAndrea Arcangeli set_pmd_at(dst_mm, addr, dst_pmd, pmd); 111971e3aac0SAndrea Arcangeli 112071e3aac0SAndrea Arcangeli ret = 0; 112171e3aac0SAndrea Arcangeli out_unlock: 1122c4088ebdSKirill A. Shutemov spin_unlock(src_ptl); 1123c4088ebdSKirill A. Shutemov spin_unlock(dst_ptl); 112471e3aac0SAndrea Arcangeli out: 112571e3aac0SAndrea Arcangeli return ret; 112671e3aac0SAndrea Arcangeli } 112771e3aac0SAndrea Arcangeli 1128a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1129a00cc7d9SMatthew Wilcox static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 1130a8f97366SKirill A. Shutemov pud_t *pud, int flags) 1131a00cc7d9SMatthew Wilcox { 1132a00cc7d9SMatthew Wilcox pud_t _pud; 1133a00cc7d9SMatthew Wilcox 1134a8f97366SKirill A. Shutemov _pud = pud_mkyoung(*pud); 1135a8f97366SKirill A. Shutemov if (flags & FOLL_WRITE) 1136a8f97366SKirill A. Shutemov _pud = pud_mkdirty(_pud); 1137a00cc7d9SMatthew Wilcox if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1138a8f97366SKirill A. Shutemov pud, _pud, flags & FOLL_WRITE)) 1139a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 1140a00cc7d9SMatthew Wilcox } 1141a00cc7d9SMatthew Wilcox 1142a00cc7d9SMatthew Wilcox struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1143df06b37fSKeith Busch pud_t *pud, int flags, struct dev_pagemap **pgmap) 1144a00cc7d9SMatthew Wilcox { 1145a00cc7d9SMatthew Wilcox unsigned long pfn = pud_pfn(*pud); 1146a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 1147a00cc7d9SMatthew Wilcox struct page *page; 1148a00cc7d9SMatthew Wilcox 1149a00cc7d9SMatthew Wilcox assert_spin_locked(pud_lockptr(mm, pud)); 1150a00cc7d9SMatthew Wilcox 1151f6f37321SLinus Torvalds if (flags & FOLL_WRITE && !pud_write(*pud)) 1152a00cc7d9SMatthew Wilcox return NULL; 1153a00cc7d9SMatthew Wilcox 11543faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 11553faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 11563faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 11573faa52c0SJohn Hubbard return NULL; 11583faa52c0SJohn Hubbard 1159a00cc7d9SMatthew Wilcox if (pud_present(*pud) && pud_devmap(*pud)) 1160a00cc7d9SMatthew Wilcox /* pass */; 1161a00cc7d9SMatthew Wilcox else 1162a00cc7d9SMatthew Wilcox return NULL; 1163a00cc7d9SMatthew Wilcox 1164a00cc7d9SMatthew Wilcox if (flags & FOLL_TOUCH) 1165a8f97366SKirill A. Shutemov touch_pud(vma, addr, pud, flags); 1166a00cc7d9SMatthew Wilcox 1167a00cc7d9SMatthew Wilcox /* 1168a00cc7d9SMatthew Wilcox * device mapped pages can only be returned if the 1169a00cc7d9SMatthew Wilcox * caller will manage the page reference count. 11703faa52c0SJohn Hubbard * 11713faa52c0SJohn Hubbard * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: 1172a00cc7d9SMatthew Wilcox */ 11733faa52c0SJohn Hubbard if (!(flags & (FOLL_GET | FOLL_PIN))) 1174a00cc7d9SMatthew Wilcox return ERR_PTR(-EEXIST); 1175a00cc7d9SMatthew Wilcox 1176a00cc7d9SMatthew Wilcox pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1177df06b37fSKeith Busch *pgmap = get_dev_pagemap(pfn, *pgmap); 1178df06b37fSKeith Busch if (!*pgmap) 1179a00cc7d9SMatthew Wilcox return ERR_PTR(-EFAULT); 1180a00cc7d9SMatthew Wilcox page = pfn_to_page(pfn); 11813faa52c0SJohn Hubbard if (!try_grab_page(page, flags)) 11823faa52c0SJohn Hubbard page = ERR_PTR(-ENOMEM); 1183a00cc7d9SMatthew Wilcox 1184a00cc7d9SMatthew Wilcox return page; 1185a00cc7d9SMatthew Wilcox } 1186a00cc7d9SMatthew Wilcox 1187a00cc7d9SMatthew Wilcox int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1188a00cc7d9SMatthew Wilcox pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1189a00cc7d9SMatthew Wilcox struct vm_area_struct *vma) 1190a00cc7d9SMatthew Wilcox { 1191a00cc7d9SMatthew Wilcox spinlock_t *dst_ptl, *src_ptl; 1192a00cc7d9SMatthew Wilcox pud_t pud; 1193a00cc7d9SMatthew Wilcox int ret; 1194a00cc7d9SMatthew Wilcox 1195a00cc7d9SMatthew Wilcox dst_ptl = pud_lock(dst_mm, dst_pud); 1196a00cc7d9SMatthew Wilcox src_ptl = pud_lockptr(src_mm, src_pud); 1197a00cc7d9SMatthew Wilcox spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1198a00cc7d9SMatthew Wilcox 1199a00cc7d9SMatthew Wilcox ret = -EAGAIN; 1200a00cc7d9SMatthew Wilcox pud = *src_pud; 1201a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1202a00cc7d9SMatthew Wilcox goto out_unlock; 1203a00cc7d9SMatthew Wilcox 1204a00cc7d9SMatthew Wilcox /* 1205a00cc7d9SMatthew Wilcox * When page table lock is held, the huge zero pud should not be 1206a00cc7d9SMatthew Wilcox * under splitting since we don't split the page itself, only pud to 1207a00cc7d9SMatthew Wilcox * a page table. 1208a00cc7d9SMatthew Wilcox */ 1209a00cc7d9SMatthew Wilcox if (is_huge_zero_pud(pud)) { 1210a00cc7d9SMatthew Wilcox /* No huge zero pud yet */ 1211a00cc7d9SMatthew Wilcox } 1212a00cc7d9SMatthew Wilcox 1213d042035eSPeter Xu /* Please refer to comments in copy_huge_pmd() */ 121497a7e473SPeter Xu if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) { 1215d042035eSPeter Xu spin_unlock(src_ptl); 1216d042035eSPeter Xu spin_unlock(dst_ptl); 1217d042035eSPeter Xu __split_huge_pud(vma, src_pud, addr); 1218d042035eSPeter Xu return -EAGAIN; 1219d042035eSPeter Xu } 1220d042035eSPeter Xu 1221a00cc7d9SMatthew Wilcox pudp_set_wrprotect(src_mm, addr, src_pud); 1222a00cc7d9SMatthew Wilcox pud = pud_mkold(pud_wrprotect(pud)); 1223a00cc7d9SMatthew Wilcox set_pud_at(dst_mm, addr, dst_pud, pud); 1224a00cc7d9SMatthew Wilcox 1225a00cc7d9SMatthew Wilcox ret = 0; 1226a00cc7d9SMatthew Wilcox out_unlock: 1227a00cc7d9SMatthew Wilcox spin_unlock(src_ptl); 1228a00cc7d9SMatthew Wilcox spin_unlock(dst_ptl); 1229a00cc7d9SMatthew Wilcox return ret; 1230a00cc7d9SMatthew Wilcox } 1231a00cc7d9SMatthew Wilcox 1232a00cc7d9SMatthew Wilcox void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1233a00cc7d9SMatthew Wilcox { 1234a00cc7d9SMatthew Wilcox pud_t entry; 1235a00cc7d9SMatthew Wilcox unsigned long haddr; 1236a00cc7d9SMatthew Wilcox bool write = vmf->flags & FAULT_FLAG_WRITE; 1237a00cc7d9SMatthew Wilcox 1238a00cc7d9SMatthew Wilcox vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1239a00cc7d9SMatthew Wilcox if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1240a00cc7d9SMatthew Wilcox goto unlock; 1241a00cc7d9SMatthew Wilcox 1242a00cc7d9SMatthew Wilcox entry = pud_mkyoung(orig_pud); 1243a00cc7d9SMatthew Wilcox if (write) 1244a00cc7d9SMatthew Wilcox entry = pud_mkdirty(entry); 1245a00cc7d9SMatthew Wilcox haddr = vmf->address & HPAGE_PUD_MASK; 1246a00cc7d9SMatthew Wilcox if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1247a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1248a00cc7d9SMatthew Wilcox 1249a00cc7d9SMatthew Wilcox unlock: 1250a00cc7d9SMatthew Wilcox spin_unlock(vmf->ptl); 1251a00cc7d9SMatthew Wilcox } 1252a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1253a00cc7d9SMatthew Wilcox 125482b0f8c3SJan Kara void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1255a1dd450bSWill Deacon { 1256a1dd450bSWill Deacon pmd_t entry; 1257a1dd450bSWill Deacon unsigned long haddr; 125820f664aaSMinchan Kim bool write = vmf->flags & FAULT_FLAG_WRITE; 1259a1dd450bSWill Deacon 126082b0f8c3SJan Kara vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 126182b0f8c3SJan Kara if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1262a1dd450bSWill Deacon goto unlock; 1263a1dd450bSWill Deacon 1264a1dd450bSWill Deacon entry = pmd_mkyoung(orig_pmd); 126520f664aaSMinchan Kim if (write) 126620f664aaSMinchan Kim entry = pmd_mkdirty(entry); 126782b0f8c3SJan Kara haddr = vmf->address & HPAGE_PMD_MASK; 126820f664aaSMinchan Kim if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 126982b0f8c3SJan Kara update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1270a1dd450bSWill Deacon 1271a1dd450bSWill Deacon unlock: 127282b0f8c3SJan Kara spin_unlock(vmf->ptl); 1273a1dd450bSWill Deacon } 1274a1dd450bSWill Deacon 12752b740303SSouptick Joarder vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 127671e3aac0SAndrea Arcangeli { 127782b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 12783917c802SKirill A. Shutemov struct page *page; 127982b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 128071e3aac0SAndrea Arcangeli 128182b0f8c3SJan Kara vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 128281d1b09cSSasha Levin VM_BUG_ON_VMA(!vma->anon_vma, vma); 12833917c802SKirill A. Shutemov 128493b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 12853917c802SKirill A. Shutemov goto fallback; 12863917c802SKirill A. Shutemov 128782b0f8c3SJan Kara spin_lock(vmf->ptl); 12883917c802SKirill A. Shutemov 12893917c802SKirill A. Shutemov if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 12903917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 12913917c802SKirill A. Shutemov return 0; 12923917c802SKirill A. Shutemov } 129371e3aac0SAndrea Arcangeli 129471e3aac0SAndrea Arcangeli page = pmd_page(orig_pmd); 1295f6004e73SMiaohe Lin VM_BUG_ON_PAGE(!PageHead(page), page); 12963917c802SKirill A. Shutemov 12973917c802SKirill A. Shutemov /* Lock page for reuse_swap_page() */ 1298ba3c4ce6SHuang Ying if (!trylock_page(page)) { 1299ba3c4ce6SHuang Ying get_page(page); 1300ba3c4ce6SHuang Ying spin_unlock(vmf->ptl); 1301ba3c4ce6SHuang Ying lock_page(page); 1302ba3c4ce6SHuang Ying spin_lock(vmf->ptl); 1303ba3c4ce6SHuang Ying if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 13043917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 1305ba3c4ce6SHuang Ying unlock_page(page); 1306ba3c4ce6SHuang Ying put_page(page); 13073917c802SKirill A. Shutemov return 0; 1308ba3c4ce6SHuang Ying } 1309ba3c4ce6SHuang Ying put_page(page); 1310ba3c4ce6SHuang Ying } 13113917c802SKirill A. Shutemov 13123917c802SKirill A. Shutemov /* 13133917c802SKirill A. Shutemov * We can only reuse the page if nobody else maps the huge page or it's 13143917c802SKirill A. Shutemov * part. 13153917c802SKirill A. Shutemov */ 1316ba3c4ce6SHuang Ying if (reuse_swap_page(page, NULL)) { 131771e3aac0SAndrea Arcangeli pmd_t entry; 131871e3aac0SAndrea Arcangeli entry = pmd_mkyoung(orig_pmd); 1319f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 132082b0f8c3SJan Kara if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 132182b0f8c3SJan Kara update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1322ba3c4ce6SHuang Ying unlock_page(page); 13233917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 13243917c802SKirill A. Shutemov return VM_FAULT_WRITE; 132571e3aac0SAndrea Arcangeli } 13263917c802SKirill A. Shutemov 1327ba3c4ce6SHuang Ying unlock_page(page); 132882b0f8c3SJan Kara spin_unlock(vmf->ptl); 13293917c802SKirill A. Shutemov fallback: 13303917c802SKirill A. Shutemov __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); 13313917c802SKirill A. Shutemov return VM_FAULT_FALLBACK; 133271e3aac0SAndrea Arcangeli } 133371e3aac0SAndrea Arcangeli 13348310d48bSKeno Fischer /* 1335a308c71bSPeter Xu * FOLL_FORCE can write to even unwritable pmd's, but only 1336a308c71bSPeter Xu * after we've gone through a COW cycle and they are dirty. 13378310d48bSKeno Fischer */ 13388310d48bSKeno Fischer static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 13398310d48bSKeno Fischer { 1340a308c71bSPeter Xu return pmd_write(pmd) || 1341a308c71bSPeter Xu ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); 13428310d48bSKeno Fischer } 13438310d48bSKeno Fischer 1344b676b293SDavid Rientjes struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 134571e3aac0SAndrea Arcangeli unsigned long addr, 134671e3aac0SAndrea Arcangeli pmd_t *pmd, 134771e3aac0SAndrea Arcangeli unsigned int flags) 134871e3aac0SAndrea Arcangeli { 1349b676b293SDavid Rientjes struct mm_struct *mm = vma->vm_mm; 135071e3aac0SAndrea Arcangeli struct page *page = NULL; 135171e3aac0SAndrea Arcangeli 1352c4088ebdSKirill A. Shutemov assert_spin_locked(pmd_lockptr(mm, pmd)); 135371e3aac0SAndrea Arcangeli 13548310d48bSKeno Fischer if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 135571e3aac0SAndrea Arcangeli goto out; 135671e3aac0SAndrea Arcangeli 135785facf25SKirill A. Shutemov /* Avoid dumping huge zero page */ 135885facf25SKirill A. Shutemov if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 135985facf25SKirill A. Shutemov return ERR_PTR(-EFAULT); 136085facf25SKirill A. Shutemov 13612b4847e7SMel Gorman /* Full NUMA hinting faults to serialise migration in fault paths */ 13628a0516edSMel Gorman if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 13632b4847e7SMel Gorman goto out; 13642b4847e7SMel Gorman 136571e3aac0SAndrea Arcangeli page = pmd_page(*pmd); 1366ca120cf6SDan Williams VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 13673faa52c0SJohn Hubbard 13683faa52c0SJohn Hubbard if (!try_grab_page(page, flags)) 13693faa52c0SJohn Hubbard return ERR_PTR(-ENOMEM); 13703faa52c0SJohn Hubbard 13713565fce3SDan Williams if (flags & FOLL_TOUCH) 1372a8f97366SKirill A. Shutemov touch_pmd(vma, addr, pmd, flags); 13733faa52c0SJohn Hubbard 1374de60f5f1SEric B Munson if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1375e90309c9SKirill A. Shutemov /* 1376e90309c9SKirill A. Shutemov * We don't mlock() pte-mapped THPs. This way we can avoid 1377e90309c9SKirill A. Shutemov * leaking mlocked pages into non-VM_LOCKED VMAs. 1378e90309c9SKirill A. Shutemov * 13799a73f61bSKirill A. Shutemov * For anon THP: 13809a73f61bSKirill A. Shutemov * 1381e90309c9SKirill A. Shutemov * In most cases the pmd is the only mapping of the page as we 1382e90309c9SKirill A. Shutemov * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1383e90309c9SKirill A. Shutemov * writable private mappings in populate_vma_page_range(). 1384e90309c9SKirill A. Shutemov * 1385e90309c9SKirill A. Shutemov * The only scenario when we have the page shared here is if we 1386e90309c9SKirill A. Shutemov * mlocking read-only mapping shared over fork(). We skip 1387e90309c9SKirill A. Shutemov * mlocking such pages. 13889a73f61bSKirill A. Shutemov * 13899a73f61bSKirill A. Shutemov * For file THP: 13909a73f61bSKirill A. Shutemov * 13919a73f61bSKirill A. Shutemov * We can expect PageDoubleMap() to be stable under page lock: 13929a73f61bSKirill A. Shutemov * for file pages we set it in page_add_file_rmap(), which 13939a73f61bSKirill A. Shutemov * requires page to be locked. 1394e90309c9SKirill A. Shutemov */ 13959a73f61bSKirill A. Shutemov 13969a73f61bSKirill A. Shutemov if (PageAnon(page) && compound_mapcount(page) != 1) 13979a73f61bSKirill A. Shutemov goto skip_mlock; 13989a73f61bSKirill A. Shutemov if (PageDoubleMap(page) || !page->mapping) 13999a73f61bSKirill A. Shutemov goto skip_mlock; 14009a73f61bSKirill A. Shutemov if (!trylock_page(page)) 14019a73f61bSKirill A. Shutemov goto skip_mlock; 14029a73f61bSKirill A. Shutemov if (page->mapping && !PageDoubleMap(page)) 1403b676b293SDavid Rientjes mlock_vma_page(page); 1404b676b293SDavid Rientjes unlock_page(page); 1405b676b293SDavid Rientjes } 14069a73f61bSKirill A. Shutemov skip_mlock: 140771e3aac0SAndrea Arcangeli page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1408ca120cf6SDan Williams VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 140971e3aac0SAndrea Arcangeli 141071e3aac0SAndrea Arcangeli out: 141171e3aac0SAndrea Arcangeli return page; 141271e3aac0SAndrea Arcangeli } 141371e3aac0SAndrea Arcangeli 1414d10e63f2SMel Gorman /* NUMA hinting page fault entry point for trans huge pmds */ 14152b740303SSouptick Joarder vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1416d10e63f2SMel Gorman { 141782b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 1418b8916634SMel Gorman struct anon_vma *anon_vma = NULL; 1419b32967ffSMel Gorman struct page *page; 142082b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 142198fa15f3SAnshuman Khandual int page_nid = NUMA_NO_NODE, this_nid = numa_node_id(); 142290572890SPeter Zijlstra int target_nid, last_cpupid = -1; 14238191acbdSMel Gorman bool page_locked; 14248191acbdSMel Gorman bool migrated = false; 1425b191f9b1SMel Gorman bool was_writable; 14266688cc05SPeter Zijlstra int flags = 0; 1427d10e63f2SMel Gorman 142882b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 142982b0f8c3SJan Kara if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1430d10e63f2SMel Gorman goto out_unlock; 1431d10e63f2SMel Gorman 1432de466bd6SMel Gorman /* 1433de466bd6SMel Gorman * If there are potential migrations, wait for completion and retry 1434de466bd6SMel Gorman * without disrupting NUMA hinting information. Do not relock and 1435de466bd6SMel Gorman * check_same as the page may no longer be mapped. 1436de466bd6SMel Gorman */ 143782b0f8c3SJan Kara if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 143882b0f8c3SJan Kara page = pmd_page(*vmf->pmd); 14393c226c63SMark Rutland if (!get_page_unless_zero(page)) 14403c226c63SMark Rutland goto out_unlock; 144182b0f8c3SJan Kara spin_unlock(vmf->ptl); 144248054625SMatthew Wilcox (Oracle) put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); 1443de466bd6SMel Gorman goto out; 1444de466bd6SMel Gorman } 1445de466bd6SMel Gorman 1446d10e63f2SMel Gorman page = pmd_page(pmd); 1447a1a46184SMel Gorman BUG_ON(is_huge_zero_page(page)); 14488191acbdSMel Gorman page_nid = page_to_nid(page); 144990572890SPeter Zijlstra last_cpupid = page_cpupid_last(page); 145003c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS); 145104bb2f94SRik van Riel if (page_nid == this_nid) { 145203c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 145304bb2f94SRik van Riel flags |= TNF_FAULT_LOCAL; 145404bb2f94SRik van Riel } 14554daae3b4SMel Gorman 1456bea66fbdSMel Gorman /* See similar comment in do_numa_page for explanation */ 1457288bc549SAneesh Kumar K.V if (!pmd_savedwrite(pmd)) 14586688cc05SPeter Zijlstra flags |= TNF_NO_GROUP; 14596688cc05SPeter Zijlstra 14606688cc05SPeter Zijlstra /* 1461ff9042b1SMel Gorman * Acquire the page lock to serialise THP migrations but avoid dropping 1462ff9042b1SMel Gorman * page_table_lock if at all possible 1463ff9042b1SMel Gorman */ 1464b8916634SMel Gorman page_locked = trylock_page(page); 1465b8916634SMel Gorman target_nid = mpol_misplaced(page, vma, haddr); 1466de466bd6SMel Gorman /* Migration could have started since the pmd_trans_migrating check */ 14672b4847e7SMel Gorman if (!page_locked) { 146898fa15f3SAnshuman Khandual page_nid = NUMA_NO_NODE; 14693c226c63SMark Rutland if (!get_page_unless_zero(page)) 14703c226c63SMark Rutland goto out_unlock; 147182b0f8c3SJan Kara spin_unlock(vmf->ptl); 147248054625SMatthew Wilcox (Oracle) put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); 1473b8916634SMel Gorman goto out; 14746beb5e8bSMiaohe Lin } else if (target_nid == NUMA_NO_NODE) { 14756beb5e8bSMiaohe Lin /* There are no parallel migrations and page is in the right 14766beb5e8bSMiaohe Lin * node. Clear the numa hinting info in this pmd. 14776beb5e8bSMiaohe Lin */ 14786beb5e8bSMiaohe Lin goto clear_pmdnuma; 1479b8916634SMel Gorman } 1480b8916634SMel Gorman 14812b4847e7SMel Gorman /* 14822b4847e7SMel Gorman * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 14832b4847e7SMel Gorman * to serialises splits 14842b4847e7SMel Gorman */ 1485b8916634SMel Gorman get_page(page); 148682b0f8c3SJan Kara spin_unlock(vmf->ptl); 1487b8916634SMel Gorman anon_vma = page_lock_anon_vma_read(page); 1488b32967ffSMel Gorman 1489c69307d5SPeter Zijlstra /* Confirm the PMD did not change while page_table_lock was released */ 149082b0f8c3SJan Kara spin_lock(vmf->ptl); 149182b0f8c3SJan Kara if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1492b32967ffSMel Gorman unlock_page(page); 1493b32967ffSMel Gorman put_page(page); 149498fa15f3SAnshuman Khandual page_nid = NUMA_NO_NODE; 1495b32967ffSMel Gorman goto out_unlock; 1496b32967ffSMel Gorman } 1497ff9042b1SMel Gorman 1498c3a489caSMel Gorman /* Bail if we fail to protect against THP splits for any reason */ 1499c3a489caSMel Gorman if (unlikely(!anon_vma)) { 1500c3a489caSMel Gorman put_page(page); 150198fa15f3SAnshuman Khandual page_nid = NUMA_NO_NODE; 1502c3a489caSMel Gorman goto clear_pmdnuma; 1503c3a489caSMel Gorman } 1504c3a489caSMel Gorman 1505a54a407fSMel Gorman /* 15068b1b436dSPeter Zijlstra * Since we took the NUMA fault, we must have observed the !accessible 15078b1b436dSPeter Zijlstra * bit. Make sure all other CPUs agree with that, to avoid them 15088b1b436dSPeter Zijlstra * modifying the page we're about to migrate. 15098b1b436dSPeter Zijlstra * 15108b1b436dSPeter Zijlstra * Must be done under PTL such that we'll observe the relevant 1511ccde85baSPeter Zijlstra * inc_tlb_flush_pending(). 1512ccde85baSPeter Zijlstra * 1513ccde85baSPeter Zijlstra * We are not sure a pending tlb flush here is for a huge page 1514ccde85baSPeter Zijlstra * mapping or not. Hence use the tlb range variant 15158b1b436dSPeter Zijlstra */ 15167066f0f9SAndrea Arcangeli if (mm_tlb_flush_pending(vma->vm_mm)) { 1517ccde85baSPeter Zijlstra flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 15187066f0f9SAndrea Arcangeli /* 15197066f0f9SAndrea Arcangeli * change_huge_pmd() released the pmd lock before 15207066f0f9SAndrea Arcangeli * invalidating the secondary MMUs sharing the primary 15217066f0f9SAndrea Arcangeli * MMU pagetables (with ->invalidate_range()). The 15227066f0f9SAndrea Arcangeli * mmu_notifier_invalidate_range_end() (which 15237066f0f9SAndrea Arcangeli * internally calls ->invalidate_range()) in 15247066f0f9SAndrea Arcangeli * change_pmd_range() will run after us, so we can't 15257066f0f9SAndrea Arcangeli * rely on it here and we need an explicit invalidate. 15267066f0f9SAndrea Arcangeli */ 15277066f0f9SAndrea Arcangeli mmu_notifier_invalidate_range(vma->vm_mm, haddr, 15287066f0f9SAndrea Arcangeli haddr + HPAGE_PMD_SIZE); 15297066f0f9SAndrea Arcangeli } 15308b1b436dSPeter Zijlstra 15318b1b436dSPeter Zijlstra /* 1532a54a407fSMel Gorman * Migrate the THP to the requested node, returns with page unlocked 15338a0516edSMel Gorman * and access rights restored. 1534a54a407fSMel Gorman */ 153582b0f8c3SJan Kara spin_unlock(vmf->ptl); 15368b1b436dSPeter Zijlstra 1537bae473a4SKirill A. Shutemov migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 153882b0f8c3SJan Kara vmf->pmd, pmd, vmf->address, page, target_nid); 15396688cc05SPeter Zijlstra if (migrated) { 15406688cc05SPeter Zijlstra flags |= TNF_MIGRATED; 15418191acbdSMel Gorman page_nid = target_nid; 1542074c2381SMel Gorman } else 1543074c2381SMel Gorman flags |= TNF_MIGRATE_FAIL; 1544b32967ffSMel Gorman 15458191acbdSMel Gorman goto out; 15464daae3b4SMel Gorman clear_pmdnuma: 1547a54a407fSMel Gorman BUG_ON(!PageLocked(page)); 1548288bc549SAneesh Kumar K.V was_writable = pmd_savedwrite(pmd); 15494d942466SMel Gorman pmd = pmd_modify(pmd, vma->vm_page_prot); 1550b7b04004SMel Gorman pmd = pmd_mkyoung(pmd); 1551b191f9b1SMel Gorman if (was_writable) 1552b191f9b1SMel Gorman pmd = pmd_mkwrite(pmd); 155382b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 155482b0f8c3SJan Kara update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1555a54a407fSMel Gorman unlock_page(page); 1556d10e63f2SMel Gorman out_unlock: 155782b0f8c3SJan Kara spin_unlock(vmf->ptl); 1558b8916634SMel Gorman 1559b8916634SMel Gorman out: 1560b8916634SMel Gorman if (anon_vma) 1561b8916634SMel Gorman page_unlock_anon_vma_read(anon_vma); 1562b8916634SMel Gorman 156398fa15f3SAnshuman Khandual if (page_nid != NUMA_NO_NODE) 156482b0f8c3SJan Kara task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 15659a8b300fSAneesh Kumar K.V flags); 15668191acbdSMel Gorman 1567d10e63f2SMel Gorman return 0; 1568d10e63f2SMel Gorman } 1569d10e63f2SMel Gorman 1570319904adSHuang Ying /* 1571319904adSHuang Ying * Return true if we do MADV_FREE successfully on entire pmd page. 1572319904adSHuang Ying * Otherwise, return false. 1573319904adSHuang Ying */ 1574319904adSHuang Ying bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1575b8d3c4c3SMinchan Kim pmd_t *pmd, unsigned long addr, unsigned long next) 1576b8d3c4c3SMinchan Kim { 1577b8d3c4c3SMinchan Kim spinlock_t *ptl; 1578b8d3c4c3SMinchan Kim pmd_t orig_pmd; 1579b8d3c4c3SMinchan Kim struct page *page; 1580b8d3c4c3SMinchan Kim struct mm_struct *mm = tlb->mm; 1581319904adSHuang Ying bool ret = false; 1582b8d3c4c3SMinchan Kim 1583ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 158407e32661SAneesh Kumar K.V 1585b6ec57f4SKirill A. Shutemov ptl = pmd_trans_huge_lock(pmd, vma); 1586b6ec57f4SKirill A. Shutemov if (!ptl) 158725eedabeSLinus Torvalds goto out_unlocked; 1588b8d3c4c3SMinchan Kim 1589b8d3c4c3SMinchan Kim orig_pmd = *pmd; 1590319904adSHuang Ying if (is_huge_zero_pmd(orig_pmd)) 1591b8d3c4c3SMinchan Kim goto out; 1592b8d3c4c3SMinchan Kim 159384c3fc4eSZi Yan if (unlikely(!pmd_present(orig_pmd))) { 159484c3fc4eSZi Yan VM_BUG_ON(thp_migration_supported() && 159584c3fc4eSZi Yan !is_pmd_migration_entry(orig_pmd)); 159684c3fc4eSZi Yan goto out; 159784c3fc4eSZi Yan } 159884c3fc4eSZi Yan 1599b8d3c4c3SMinchan Kim page = pmd_page(orig_pmd); 1600b8d3c4c3SMinchan Kim /* 1601b8d3c4c3SMinchan Kim * If other processes are mapping this page, we couldn't discard 1602b8d3c4c3SMinchan Kim * the page unless they all do MADV_FREE so let's skip the page. 1603b8d3c4c3SMinchan Kim */ 1604b8d3c4c3SMinchan Kim if (page_mapcount(page) != 1) 1605b8d3c4c3SMinchan Kim goto out; 1606b8d3c4c3SMinchan Kim 1607b8d3c4c3SMinchan Kim if (!trylock_page(page)) 1608b8d3c4c3SMinchan Kim goto out; 1609b8d3c4c3SMinchan Kim 1610b8d3c4c3SMinchan Kim /* 1611b8d3c4c3SMinchan Kim * If user want to discard part-pages of THP, split it so MADV_FREE 1612b8d3c4c3SMinchan Kim * will deactivate only them. 1613b8d3c4c3SMinchan Kim */ 1614b8d3c4c3SMinchan Kim if (next - addr != HPAGE_PMD_SIZE) { 1615b8d3c4c3SMinchan Kim get_page(page); 1616b8d3c4c3SMinchan Kim spin_unlock(ptl); 16179818b8cdSHuang Ying split_huge_page(page); 1618b8d3c4c3SMinchan Kim unlock_page(page); 1619bbf29ffcSKirill A. Shutemov put_page(page); 1620b8d3c4c3SMinchan Kim goto out_unlocked; 1621b8d3c4c3SMinchan Kim } 1622b8d3c4c3SMinchan Kim 1623b8d3c4c3SMinchan Kim if (PageDirty(page)) 1624b8d3c4c3SMinchan Kim ClearPageDirty(page); 1625b8d3c4c3SMinchan Kim unlock_page(page); 1626b8d3c4c3SMinchan Kim 1627b8d3c4c3SMinchan Kim if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 162858ceeb6bSKirill A. Shutemov pmdp_invalidate(vma, addr, pmd); 1629b8d3c4c3SMinchan Kim orig_pmd = pmd_mkold(orig_pmd); 1630b8d3c4c3SMinchan Kim orig_pmd = pmd_mkclean(orig_pmd); 1631b8d3c4c3SMinchan Kim 1632b8d3c4c3SMinchan Kim set_pmd_at(mm, addr, pmd, orig_pmd); 1633b8d3c4c3SMinchan Kim tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1634b8d3c4c3SMinchan Kim } 1635802a3a92SShaohua Li 1636802a3a92SShaohua Li mark_page_lazyfree(page); 1637319904adSHuang Ying ret = true; 1638b8d3c4c3SMinchan Kim out: 1639b8d3c4c3SMinchan Kim spin_unlock(ptl); 1640b8d3c4c3SMinchan Kim out_unlocked: 1641b8d3c4c3SMinchan Kim return ret; 1642b8d3c4c3SMinchan Kim } 1643b8d3c4c3SMinchan Kim 1644953c66c2SAneesh Kumar K.V static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1645953c66c2SAneesh Kumar K.V { 1646953c66c2SAneesh Kumar K.V pgtable_t pgtable; 1647953c66c2SAneesh Kumar K.V 1648953c66c2SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1649953c66c2SAneesh Kumar K.V pte_free(mm, pgtable); 1650c4812909SKirill A. Shutemov mm_dec_nr_ptes(mm); 1651953c66c2SAneesh Kumar K.V } 1652953c66c2SAneesh Kumar K.V 165371e3aac0SAndrea Arcangeli int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1654f21760b1SShaohua Li pmd_t *pmd, unsigned long addr) 165571e3aac0SAndrea Arcangeli { 1656f5c8ad47SDavid Miller pmd_t orig_pmd; 1657da146769SKirill A. Shutemov spinlock_t *ptl; 1658da146769SKirill A. Shutemov 1659ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 166007e32661SAneesh Kumar K.V 1661b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 1662b6ec57f4SKirill A. Shutemov if (!ptl) 1663da146769SKirill A. Shutemov return 0; 1664a6bf2bb0SAneesh Kumar K.V /* 1665a6bf2bb0SAneesh Kumar K.V * For architectures like ppc64 we look at deposited pgtable 16668809aa2dSAneesh Kumar K.V * when calling pmdp_huge_get_and_clear. So do the 1667a6bf2bb0SAneesh Kumar K.V * pgtable_trans_huge_withdraw after finishing pmdp related 1668a6bf2bb0SAneesh Kumar K.V * operations. 1669a6bf2bb0SAneesh Kumar K.V */ 167093a98695SAneesh Kumar K.V orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, 1671fcbe08d6SMartin Schwidefsky tlb->fullmm); 1672f21760b1SShaohua Li tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 16732484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) { 16743b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) 16753b6521f5SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 16764897c765SMatthew Wilcox spin_unlock(ptl); 1677da146769SKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 1678c0f2e176SAneesh Kumar K.V tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1679da146769SKirill A. Shutemov } else if (is_huge_zero_pmd(orig_pmd)) { 1680c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1681bf929152SKirill A. Shutemov spin_unlock(ptl); 1682c0f2e176SAneesh Kumar K.V tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1683479f0abbSKirill A. Shutemov } else { 1684616b8371SZi Yan struct page *page = NULL; 1685616b8371SZi Yan int flush_needed = 1; 1686616b8371SZi Yan 1687616b8371SZi Yan if (pmd_present(orig_pmd)) { 1688616b8371SZi Yan page = pmd_page(orig_pmd); 1689d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 1690309381feSSasha Levin VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1691309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1692616b8371SZi Yan } else if (thp_migration_supported()) { 1693616b8371SZi Yan swp_entry_t entry; 1694616b8371SZi Yan 1695616b8371SZi Yan VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1696616b8371SZi Yan entry = pmd_to_swp_entry(orig_pmd); 1697a44f89dcSMiaohe Lin page = migration_entry_to_page(entry); 1698616b8371SZi Yan flush_needed = 0; 1699616b8371SZi Yan } else 1700616b8371SZi Yan WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1701616b8371SZi Yan 1702b5072380SKirill A. Shutemov if (PageAnon(page)) { 1703c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1704b5072380SKirill A. Shutemov add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1705b5072380SKirill A. Shutemov } else { 1706953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 1707953c66c2SAneesh Kumar K.V zap_deposited_table(tlb->mm, pmd); 1708fadae295SYang Shi add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1709b5072380SKirill A. Shutemov } 1710616b8371SZi Yan 1711bf929152SKirill A. Shutemov spin_unlock(ptl); 1712616b8371SZi Yan if (flush_needed) 1713e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1714479f0abbSKirill A. Shutemov } 1715da146769SKirill A. Shutemov return 1; 171671e3aac0SAndrea Arcangeli } 171771e3aac0SAndrea Arcangeli 17181dd38b6cSAneesh Kumar K.V #ifndef pmd_move_must_withdraw 17191dd38b6cSAneesh Kumar K.V static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 17201dd38b6cSAneesh Kumar K.V spinlock_t *old_pmd_ptl, 17211dd38b6cSAneesh Kumar K.V struct vm_area_struct *vma) 17221dd38b6cSAneesh Kumar K.V { 17231dd38b6cSAneesh Kumar K.V /* 17241dd38b6cSAneesh Kumar K.V * With split pmd lock we also need to move preallocated 17251dd38b6cSAneesh Kumar K.V * PTE page table if new_pmd is on different PMD page table. 17261dd38b6cSAneesh Kumar K.V * 17271dd38b6cSAneesh Kumar K.V * We also don't deposit and withdraw tables for file pages. 17281dd38b6cSAneesh Kumar K.V */ 17291dd38b6cSAneesh Kumar K.V return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 17301dd38b6cSAneesh Kumar K.V } 17311dd38b6cSAneesh Kumar K.V #endif 17321dd38b6cSAneesh Kumar K.V 1733ab6e3d09SNaoya Horiguchi static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1734ab6e3d09SNaoya Horiguchi { 1735ab6e3d09SNaoya Horiguchi #ifdef CONFIG_MEM_SOFT_DIRTY 1736ab6e3d09SNaoya Horiguchi if (unlikely(is_pmd_migration_entry(pmd))) 1737ab6e3d09SNaoya Horiguchi pmd = pmd_swp_mksoft_dirty(pmd); 1738ab6e3d09SNaoya Horiguchi else if (pmd_present(pmd)) 1739ab6e3d09SNaoya Horiguchi pmd = pmd_mksoft_dirty(pmd); 1740ab6e3d09SNaoya Horiguchi #endif 1741ab6e3d09SNaoya Horiguchi return pmd; 1742ab6e3d09SNaoya Horiguchi } 1743ab6e3d09SNaoya Horiguchi 1744bf8616d5SHugh Dickins bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1745b8aa9d9dSWei Yang unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) 174637a1c49aSAndrea Arcangeli { 1747bf929152SKirill A. Shutemov spinlock_t *old_ptl, *new_ptl; 174837a1c49aSAndrea Arcangeli pmd_t pmd; 174937a1c49aSAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 17505d190420SAaron Lu bool force_flush = false; 175137a1c49aSAndrea Arcangeli 175237a1c49aSAndrea Arcangeli /* 175337a1c49aSAndrea Arcangeli * The destination pmd shouldn't be established, free_pgtables() 175437a1c49aSAndrea Arcangeli * should have release it. 175537a1c49aSAndrea Arcangeli */ 175637a1c49aSAndrea Arcangeli if (WARN_ON(!pmd_none(*new_pmd))) { 175737a1c49aSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*new_pmd)); 17584b471e88SKirill A. Shutemov return false; 175937a1c49aSAndrea Arcangeli } 176037a1c49aSAndrea Arcangeli 1761bf929152SKirill A. Shutemov /* 1762bf929152SKirill A. Shutemov * We don't have to worry about the ordering of src and dst 1763c1e8d7c6SMichel Lespinasse * ptlocks because exclusive mmap_lock prevents deadlock. 1764bf929152SKirill A. Shutemov */ 1765b6ec57f4SKirill A. Shutemov old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1766b6ec57f4SKirill A. Shutemov if (old_ptl) { 1767bf929152SKirill A. Shutemov new_ptl = pmd_lockptr(mm, new_pmd); 1768bf929152SKirill A. Shutemov if (new_ptl != old_ptl) 1769bf929152SKirill A. Shutemov spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 17708809aa2dSAneesh Kumar K.V pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1771eb66ae03SLinus Torvalds if (pmd_present(pmd)) 1772a2ce2666SAaron Lu force_flush = true; 177337a1c49aSAndrea Arcangeli VM_BUG_ON(!pmd_none(*new_pmd)); 17743592806cSKirill A. Shutemov 17751dd38b6cSAneesh Kumar K.V if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1776b3084f4dSAneesh Kumar K.V pgtable_t pgtable; 17773592806cSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 17783592806cSKirill A. Shutemov pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 17793592806cSKirill A. Shutemov } 1780ab6e3d09SNaoya Horiguchi pmd = move_soft_dirty_pmd(pmd); 1781ab6e3d09SNaoya Horiguchi set_pmd_at(mm, new_addr, new_pmd, pmd); 17825d190420SAaron Lu if (force_flush) 17835d190420SAaron Lu flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1784eb66ae03SLinus Torvalds if (new_ptl != old_ptl) 1785eb66ae03SLinus Torvalds spin_unlock(new_ptl); 1786bf929152SKirill A. Shutemov spin_unlock(old_ptl); 17874b471e88SKirill A. Shutemov return true; 178837a1c49aSAndrea Arcangeli } 17894b471e88SKirill A. Shutemov return false; 179037a1c49aSAndrea Arcangeli } 179137a1c49aSAndrea Arcangeli 1792f123d74aSMel Gorman /* 1793f123d74aSMel Gorman * Returns 1794f123d74aSMel Gorman * - 0 if PMD could not be locked 1795f0953a1bSIngo Molnar * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary 1796f0953a1bSIngo Molnar * - HPAGE_PMD_NR if protections changed and TLB flush necessary 1797f123d74aSMel Gorman */ 1798cd7548abSJohannes Weiner int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 179958705444SPeter Xu unsigned long addr, pgprot_t newprot, unsigned long cp_flags) 1800cd7548abSJohannes Weiner { 1801cd7548abSJohannes Weiner struct mm_struct *mm = vma->vm_mm; 1802bf929152SKirill A. Shutemov spinlock_t *ptl; 18030a85e51dSKirill A. Shutemov pmd_t entry; 18040a85e51dSKirill A. Shutemov bool preserve_write; 18050a85e51dSKirill A. Shutemov int ret; 180658705444SPeter Xu bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 1807292924b2SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 1808292924b2SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 1809cd7548abSJohannes Weiner 1810b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 18110a85e51dSKirill A. Shutemov if (!ptl) 18120a85e51dSKirill A. Shutemov return 0; 18130a85e51dSKirill A. Shutemov 18140a85e51dSKirill A. Shutemov preserve_write = prot_numa && pmd_write(*pmd); 1815ba68bc01SMel Gorman ret = 1; 1816e944fd67SMel Gorman 181784c3fc4eSZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 181884c3fc4eSZi Yan if (is_swap_pmd(*pmd)) { 181984c3fc4eSZi Yan swp_entry_t entry = pmd_to_swp_entry(*pmd); 182084c3fc4eSZi Yan 182184c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 182284c3fc4eSZi Yan if (is_write_migration_entry(entry)) { 182384c3fc4eSZi Yan pmd_t newpmd; 182484c3fc4eSZi Yan /* 182584c3fc4eSZi Yan * A protection check is difficult so 182684c3fc4eSZi Yan * just be safe and disable write 182784c3fc4eSZi Yan */ 182884c3fc4eSZi Yan make_migration_entry_read(&entry); 182984c3fc4eSZi Yan newpmd = swp_entry_to_pmd(entry); 1830ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*pmd)) 1831ab6e3d09SNaoya Horiguchi newpmd = pmd_swp_mksoft_dirty(newpmd); 183284c3fc4eSZi Yan set_pmd_at(mm, addr, pmd, newpmd); 183384c3fc4eSZi Yan } 183484c3fc4eSZi Yan goto unlock; 183584c3fc4eSZi Yan } 183684c3fc4eSZi Yan #endif 183784c3fc4eSZi Yan 1838e944fd67SMel Gorman /* 1839e944fd67SMel Gorman * Avoid trapping faults against the zero page. The read-only 1840e944fd67SMel Gorman * data is likely to be read-cached on the local CPU and 1841e944fd67SMel Gorman * local/remote hits to the zero page are not interesting. 1842e944fd67SMel Gorman */ 18430a85e51dSKirill A. Shutemov if (prot_numa && is_huge_zero_pmd(*pmd)) 18440a85e51dSKirill A. Shutemov goto unlock; 1845e944fd67SMel Gorman 18460a85e51dSKirill A. Shutemov if (prot_numa && pmd_protnone(*pmd)) 18470a85e51dSKirill A. Shutemov goto unlock; 18480a85e51dSKirill A. Shutemov 1849ced10803SKirill A. Shutemov /* 18503e4e28c5SMichel Lespinasse * In case prot_numa, we are under mmap_read_lock(mm). It's critical 1851ced10803SKirill A. Shutemov * to not clear pmd intermittently to avoid race with MADV_DONTNEED 18523e4e28c5SMichel Lespinasse * which is also under mmap_read_lock(mm): 1853ced10803SKirill A. Shutemov * 1854ced10803SKirill A. Shutemov * CPU0: CPU1: 1855ced10803SKirill A. Shutemov * change_huge_pmd(prot_numa=1) 1856ced10803SKirill A. Shutemov * pmdp_huge_get_and_clear_notify() 1857ced10803SKirill A. Shutemov * madvise_dontneed() 1858ced10803SKirill A. Shutemov * zap_pmd_range() 1859ced10803SKirill A. Shutemov * pmd_trans_huge(*pmd) == 0 (without ptl) 1860ced10803SKirill A. Shutemov * // skip the pmd 1861ced10803SKirill A. Shutemov * set_pmd_at(); 1862ced10803SKirill A. Shutemov * // pmd is re-established 1863ced10803SKirill A. Shutemov * 1864ced10803SKirill A. Shutemov * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1865ced10803SKirill A. Shutemov * which may break userspace. 1866ced10803SKirill A. Shutemov * 1867ced10803SKirill A. Shutemov * pmdp_invalidate() is required to make sure we don't miss 1868ced10803SKirill A. Shutemov * dirty/young flags set by hardware. 1869ced10803SKirill A. Shutemov */ 1870a3cf988fSKirill A. Shutemov entry = pmdp_invalidate(vma, addr, pmd); 1871ced10803SKirill A. Shutemov 1872cd7548abSJohannes Weiner entry = pmd_modify(entry, newprot); 1873b191f9b1SMel Gorman if (preserve_write) 1874288bc549SAneesh Kumar K.V entry = pmd_mk_savedwrite(entry); 1875292924b2SPeter Xu if (uffd_wp) { 1876292924b2SPeter Xu entry = pmd_wrprotect(entry); 1877292924b2SPeter Xu entry = pmd_mkuffd_wp(entry); 1878292924b2SPeter Xu } else if (uffd_wp_resolve) { 1879292924b2SPeter Xu /* 1880292924b2SPeter Xu * Leave the write bit to be handled by PF interrupt 1881292924b2SPeter Xu * handler, then things like COW could be properly 1882292924b2SPeter Xu * handled. 1883292924b2SPeter Xu */ 1884292924b2SPeter Xu entry = pmd_clear_uffd_wp(entry); 1885292924b2SPeter Xu } 1886f123d74aSMel Gorman ret = HPAGE_PMD_NR; 188756eecdb9SAneesh Kumar K.V set_pmd_at(mm, addr, pmd, entry); 18880a85e51dSKirill A. Shutemov BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 18890a85e51dSKirill A. Shutemov unlock: 1890bf929152SKirill A. Shutemov spin_unlock(ptl); 1891cd7548abSJohannes Weiner return ret; 1892cd7548abSJohannes Weiner } 1893cd7548abSJohannes Weiner 1894025c5b24SNaoya Horiguchi /* 18958f19b0c0SHuang Ying * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1896025c5b24SNaoya Horiguchi * 18978f19b0c0SHuang Ying * Note that if it returns page table lock pointer, this routine returns without 18988f19b0c0SHuang Ying * unlocking page table lock. So callers must unlock it. 1899025c5b24SNaoya Horiguchi */ 1900b6ec57f4SKirill A. Shutemov spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1901025c5b24SNaoya Horiguchi { 1902b6ec57f4SKirill A. Shutemov spinlock_t *ptl; 1903b6ec57f4SKirill A. Shutemov ptl = pmd_lock(vma->vm_mm, pmd); 190484c3fc4eSZi Yan if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 190584c3fc4eSZi Yan pmd_devmap(*pmd))) 1906b6ec57f4SKirill A. Shutemov return ptl; 1907b6ec57f4SKirill A. Shutemov spin_unlock(ptl); 1908b6ec57f4SKirill A. Shutemov return NULL; 1909025c5b24SNaoya Horiguchi } 1910025c5b24SNaoya Horiguchi 1911a00cc7d9SMatthew Wilcox /* 1912a00cc7d9SMatthew Wilcox * Returns true if a given pud maps a thp, false otherwise. 1913a00cc7d9SMatthew Wilcox * 1914a00cc7d9SMatthew Wilcox * Note that if it returns true, this routine returns without unlocking page 1915a00cc7d9SMatthew Wilcox * table lock. So callers must unlock it. 1916a00cc7d9SMatthew Wilcox */ 1917a00cc7d9SMatthew Wilcox spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1918a00cc7d9SMatthew Wilcox { 1919a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1920a00cc7d9SMatthew Wilcox 1921a00cc7d9SMatthew Wilcox ptl = pud_lock(vma->vm_mm, pud); 1922a00cc7d9SMatthew Wilcox if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1923a00cc7d9SMatthew Wilcox return ptl; 1924a00cc7d9SMatthew Wilcox spin_unlock(ptl); 1925a00cc7d9SMatthew Wilcox return NULL; 1926a00cc7d9SMatthew Wilcox } 1927a00cc7d9SMatthew Wilcox 1928a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1929a00cc7d9SMatthew Wilcox int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1930a00cc7d9SMatthew Wilcox pud_t *pud, unsigned long addr) 1931a00cc7d9SMatthew Wilcox { 1932a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1933a00cc7d9SMatthew Wilcox 1934a00cc7d9SMatthew Wilcox ptl = __pud_trans_huge_lock(pud, vma); 1935a00cc7d9SMatthew Wilcox if (!ptl) 1936a00cc7d9SMatthew Wilcox return 0; 1937a00cc7d9SMatthew Wilcox /* 1938a00cc7d9SMatthew Wilcox * For architectures like ppc64 we look at deposited pgtable 1939a00cc7d9SMatthew Wilcox * when calling pudp_huge_get_and_clear. So do the 1940a00cc7d9SMatthew Wilcox * pgtable_trans_huge_withdraw after finishing pudp related 1941a00cc7d9SMatthew Wilcox * operations. 1942a00cc7d9SMatthew Wilcox */ 194370516b93SQian Cai pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm); 1944a00cc7d9SMatthew Wilcox tlb_remove_pud_tlb_entry(tlb, pud, addr); 19452484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) { 1946a00cc7d9SMatthew Wilcox spin_unlock(ptl); 1947a00cc7d9SMatthew Wilcox /* No zero page support yet */ 1948a00cc7d9SMatthew Wilcox } else { 1949a00cc7d9SMatthew Wilcox /* No support for anonymous PUD pages yet */ 1950a00cc7d9SMatthew Wilcox BUG(); 1951a00cc7d9SMatthew Wilcox } 1952a00cc7d9SMatthew Wilcox return 1; 1953a00cc7d9SMatthew Wilcox } 1954a00cc7d9SMatthew Wilcox 1955a00cc7d9SMatthew Wilcox static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 1956a00cc7d9SMatthew Wilcox unsigned long haddr) 1957a00cc7d9SMatthew Wilcox { 1958a00cc7d9SMatthew Wilcox VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 1959a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 1960a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 1961a00cc7d9SMatthew Wilcox VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 1962a00cc7d9SMatthew Wilcox 1963ce9311cfSYisheng Xie count_vm_event(THP_SPLIT_PUD); 1964a00cc7d9SMatthew Wilcox 1965a00cc7d9SMatthew Wilcox pudp_huge_clear_flush_notify(vma, haddr, pud); 1966a00cc7d9SMatthew Wilcox } 1967a00cc7d9SMatthew Wilcox 1968a00cc7d9SMatthew Wilcox void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 1969a00cc7d9SMatthew Wilcox unsigned long address) 1970a00cc7d9SMatthew Wilcox { 1971a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1972ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1973a00cc7d9SMatthew Wilcox 19747269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 19756f4f13e8SJérôme Glisse address & HPAGE_PUD_MASK, 1976ac46d4f3SJérôme Glisse (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 1977ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1978ac46d4f3SJérôme Glisse ptl = pud_lock(vma->vm_mm, pud); 1979a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 1980a00cc7d9SMatthew Wilcox goto out; 1981ac46d4f3SJérôme Glisse __split_huge_pud_locked(vma, pud, range.start); 1982a00cc7d9SMatthew Wilcox 1983a00cc7d9SMatthew Wilcox out: 1984a00cc7d9SMatthew Wilcox spin_unlock(ptl); 19854645b9feSJérôme Glisse /* 19864645b9feSJérôme Glisse * No need to double call mmu_notifier->invalidate_range() callback as 19874645b9feSJérôme Glisse * the above pudp_huge_clear_flush_notify() did already call it. 19884645b9feSJérôme Glisse */ 1989ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_only_end(&range); 1990a00cc7d9SMatthew Wilcox } 1991a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1992a00cc7d9SMatthew Wilcox 1993eef1b3baSKirill A. Shutemov static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 1994eef1b3baSKirill A. Shutemov unsigned long haddr, pmd_t *pmd) 1995eef1b3baSKirill A. Shutemov { 1996eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 1997eef1b3baSKirill A. Shutemov pgtable_t pgtable; 1998eef1b3baSKirill A. Shutemov pmd_t _pmd; 1999eef1b3baSKirill A. Shutemov int i; 2000eef1b3baSKirill A. Shutemov 20010f10851eSJérôme Glisse /* 20020f10851eSJérôme Glisse * Leave pmd empty until pte is filled note that it is fine to delay 20030f10851eSJérôme Glisse * notification until mmu_notifier_invalidate_range_end() as we are 20040f10851eSJérôme Glisse * replacing a zero pmd write protected page with a zero pte write 20050f10851eSJérôme Glisse * protected page. 20060f10851eSJérôme Glisse * 2007ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 20080f10851eSJérôme Glisse */ 20090f10851eSJérôme Glisse pmdp_huge_clear_flush(vma, haddr, pmd); 2010eef1b3baSKirill A. Shutemov 2011eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2012eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2013eef1b3baSKirill A. Shutemov 2014eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2015eef1b3baSKirill A. Shutemov pte_t *pte, entry; 2016eef1b3baSKirill A. Shutemov entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2017eef1b3baSKirill A. Shutemov entry = pte_mkspecial(entry); 2018eef1b3baSKirill A. Shutemov pte = pte_offset_map(&_pmd, haddr); 2019eef1b3baSKirill A. Shutemov VM_BUG_ON(!pte_none(*pte)); 2020eef1b3baSKirill A. Shutemov set_pte_at(mm, haddr, pte, entry); 2021eef1b3baSKirill A. Shutemov pte_unmap(pte); 2022eef1b3baSKirill A. Shutemov } 2023eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2024eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 2025eef1b3baSKirill A. Shutemov } 2026eef1b3baSKirill A. Shutemov 2027eef1b3baSKirill A. Shutemov static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2028ba988280SKirill A. Shutemov unsigned long haddr, bool freeze) 2029eef1b3baSKirill A. Shutemov { 2030eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2031eef1b3baSKirill A. Shutemov struct page *page; 2032eef1b3baSKirill A. Shutemov pgtable_t pgtable; 2033423ac9afSAneesh Kumar K.V pmd_t old_pmd, _pmd; 2034292924b2SPeter Xu bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; 20352ac015e2SKirill A. Shutemov unsigned long addr; 2036eef1b3baSKirill A. Shutemov int i; 2037eef1b3baSKirill A. Shutemov 2038eef1b3baSKirill A. Shutemov VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2039eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2040eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 204184c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 204284c3fc4eSZi Yan && !pmd_devmap(*pmd)); 2043eef1b3baSKirill A. Shutemov 2044eef1b3baSKirill A. Shutemov count_vm_event(THP_SPLIT_PMD); 2045eef1b3baSKirill A. Shutemov 2046d21b9e57SKirill A. Shutemov if (!vma_is_anonymous(vma)) { 2047*99fa8a48SHugh Dickins old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2048953c66c2SAneesh Kumar K.V /* 2049953c66c2SAneesh Kumar K.V * We are going to unmap this huge page. So 2050953c66c2SAneesh Kumar K.V * just go ahead and zap it 2051953c66c2SAneesh Kumar K.V */ 2052953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 2053953c66c2SAneesh Kumar K.V zap_deposited_table(mm, pmd); 20542484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) 2055d21b9e57SKirill A. Shutemov return; 2056*99fa8a48SHugh Dickins if (unlikely(is_pmd_migration_entry(old_pmd))) { 2057*99fa8a48SHugh Dickins swp_entry_t entry; 2058*99fa8a48SHugh Dickins 2059*99fa8a48SHugh Dickins entry = pmd_to_swp_entry(old_pmd); 2060*99fa8a48SHugh Dickins page = migration_entry_to_page(entry); 2061*99fa8a48SHugh Dickins } else { 2062*99fa8a48SHugh Dickins page = pmd_page(old_pmd); 2063*99fa8a48SHugh Dickins if (!PageDirty(page) && pmd_dirty(old_pmd)) 2064e1f1b157SHugh Dickins set_page_dirty(page); 2065*99fa8a48SHugh Dickins if (!PageReferenced(page) && pmd_young(old_pmd)) 2066d21b9e57SKirill A. Shutemov SetPageReferenced(page); 2067d21b9e57SKirill A. Shutemov page_remove_rmap(page, true); 2068d21b9e57SKirill A. Shutemov put_page(page); 2069*99fa8a48SHugh Dickins } 2070fadae295SYang Shi add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2071eef1b3baSKirill A. Shutemov return; 2072*99fa8a48SHugh Dickins } 2073*99fa8a48SHugh Dickins 2074*99fa8a48SHugh Dickins if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) { 20754645b9feSJérôme Glisse /* 20764645b9feSJérôme Glisse * FIXME: Do we want to invalidate secondary mmu by calling 20774645b9feSJérôme Glisse * mmu_notifier_invalidate_range() see comments below inside 20784645b9feSJérôme Glisse * __split_huge_pmd() ? 20794645b9feSJérôme Glisse * 20804645b9feSJérôme Glisse * We are going from a zero huge page write protected to zero 20814645b9feSJérôme Glisse * small page also write protected so it does not seems useful 20824645b9feSJérôme Glisse * to invalidate secondary mmu at this time. 20834645b9feSJérôme Glisse */ 2084eef1b3baSKirill A. Shutemov return __split_huge_zero_page_pmd(vma, haddr, pmd); 2085eef1b3baSKirill A. Shutemov } 2086eef1b3baSKirill A. Shutemov 2087423ac9afSAneesh Kumar K.V /* 2088423ac9afSAneesh Kumar K.V * Up to this point the pmd is present and huge and userland has the 2089423ac9afSAneesh Kumar K.V * whole access to the hugepage during the split (which happens in 2090423ac9afSAneesh Kumar K.V * place). If we overwrite the pmd with the not-huge version pointing 2091423ac9afSAneesh Kumar K.V * to the pte here (which of course we could if all CPUs were bug 2092423ac9afSAneesh Kumar K.V * free), userland could trigger a small page size TLB miss on the 2093423ac9afSAneesh Kumar K.V * small sized TLB while the hugepage TLB entry is still established in 2094423ac9afSAneesh Kumar K.V * the huge TLB. Some CPU doesn't like that. 209542742d9bSAlexander A. Klimov * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum 209642742d9bSAlexander A. Klimov * 383 on page 105. Intel should be safe but is also warns that it's 2097423ac9afSAneesh Kumar K.V * only safe if the permission and cache attributes of the two entries 2098423ac9afSAneesh Kumar K.V * loaded in the two TLB is identical (which should be the case here). 2099423ac9afSAneesh Kumar K.V * But it is generally safer to never allow small and huge TLB entries 2100423ac9afSAneesh Kumar K.V * for the same virtual address to be loaded simultaneously. So instead 2101423ac9afSAneesh Kumar K.V * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2102423ac9afSAneesh Kumar K.V * current pmd notpresent (atomically because here the pmd_trans_huge 2103423ac9afSAneesh Kumar K.V * must remain set at all times on the pmd until the split is complete 2104423ac9afSAneesh Kumar K.V * for this pmd), then we flush the SMP TLB and finally we write the 2105423ac9afSAneesh Kumar K.V * non-huge version of the pmd entry with pmd_populate. 2106423ac9afSAneesh Kumar K.V */ 2107423ac9afSAneesh Kumar K.V old_pmd = pmdp_invalidate(vma, haddr, pmd); 2108423ac9afSAneesh Kumar K.V 2109423ac9afSAneesh Kumar K.V pmd_migration = is_pmd_migration_entry(old_pmd); 21102e83ee1dSPeter Xu if (unlikely(pmd_migration)) { 211184c3fc4eSZi Yan swp_entry_t entry; 211284c3fc4eSZi Yan 2113423ac9afSAneesh Kumar K.V entry = pmd_to_swp_entry(old_pmd); 2114a44f89dcSMiaohe Lin page = migration_entry_to_page(entry); 21152e83ee1dSPeter Xu write = is_write_migration_entry(entry); 21162e83ee1dSPeter Xu young = false; 21172e83ee1dSPeter Xu soft_dirty = pmd_swp_soft_dirty(old_pmd); 2118f45ec5ffSPeter Xu uffd_wp = pmd_swp_uffd_wp(old_pmd); 21192e83ee1dSPeter Xu } else { 2120423ac9afSAneesh Kumar K.V page = pmd_page(old_pmd); 2121423ac9afSAneesh Kumar K.V if (pmd_dirty(old_pmd)) 2122423ac9afSAneesh Kumar K.V SetPageDirty(page); 2123423ac9afSAneesh Kumar K.V write = pmd_write(old_pmd); 2124423ac9afSAneesh Kumar K.V young = pmd_young(old_pmd); 2125423ac9afSAneesh Kumar K.V soft_dirty = pmd_soft_dirty(old_pmd); 2126292924b2SPeter Xu uffd_wp = pmd_uffd_wp(old_pmd); 21272e83ee1dSPeter Xu } 21282e83ee1dSPeter Xu VM_BUG_ON_PAGE(!page_count(page), page); 21292e83ee1dSPeter Xu page_ref_add(page, HPAGE_PMD_NR - 1); 2130eef1b3baSKirill A. Shutemov 2131423ac9afSAneesh Kumar K.V /* 2132423ac9afSAneesh Kumar K.V * Withdraw the table only after we mark the pmd entry invalid. 2133423ac9afSAneesh Kumar K.V * This's critical for some architectures (Power). 2134423ac9afSAneesh Kumar K.V */ 2135eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2136eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2137eef1b3baSKirill A. Shutemov 21382ac015e2SKirill A. Shutemov for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2139eef1b3baSKirill A. Shutemov pte_t entry, *pte; 2140eef1b3baSKirill A. Shutemov /* 2141eef1b3baSKirill A. Shutemov * Note that NUMA hinting access restrictions are not 2142eef1b3baSKirill A. Shutemov * transferred to avoid any possibility of altering 2143eef1b3baSKirill A. Shutemov * permissions across VMAs. 2144eef1b3baSKirill A. Shutemov */ 214584c3fc4eSZi Yan if (freeze || pmd_migration) { 2146ba988280SKirill A. Shutemov swp_entry_t swp_entry; 2147ba988280SKirill A. Shutemov swp_entry = make_migration_entry(page + i, write); 2148ba988280SKirill A. Shutemov entry = swp_entry_to_pte(swp_entry); 2149804dd150SAndrea Arcangeli if (soft_dirty) 2150804dd150SAndrea Arcangeli entry = pte_swp_mksoft_dirty(entry); 2151f45ec5ffSPeter Xu if (uffd_wp) 2152f45ec5ffSPeter Xu entry = pte_swp_mkuffd_wp(entry); 2153ba988280SKirill A. Shutemov } else { 21546d2329f8SAndrea Arcangeli entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2155b8d3c4c3SMinchan Kim entry = maybe_mkwrite(entry, vma); 2156eef1b3baSKirill A. Shutemov if (!write) 2157eef1b3baSKirill A. Shutemov entry = pte_wrprotect(entry); 2158eef1b3baSKirill A. Shutemov if (!young) 2159eef1b3baSKirill A. Shutemov entry = pte_mkold(entry); 2160804dd150SAndrea Arcangeli if (soft_dirty) 2161804dd150SAndrea Arcangeli entry = pte_mksoft_dirty(entry); 2162292924b2SPeter Xu if (uffd_wp) 2163292924b2SPeter Xu entry = pte_mkuffd_wp(entry); 2164ba988280SKirill A. Shutemov } 21652ac015e2SKirill A. Shutemov pte = pte_offset_map(&_pmd, addr); 2166eef1b3baSKirill A. Shutemov BUG_ON(!pte_none(*pte)); 21672ac015e2SKirill A. Shutemov set_pte_at(mm, addr, pte, entry); 2168ec0abae6SRalph Campbell if (!pmd_migration) 2169eef1b3baSKirill A. Shutemov atomic_inc(&page[i]._mapcount); 2170eef1b3baSKirill A. Shutemov pte_unmap(pte); 2171eef1b3baSKirill A. Shutemov } 2172eef1b3baSKirill A. Shutemov 2173ec0abae6SRalph Campbell if (!pmd_migration) { 2174eef1b3baSKirill A. Shutemov /* 2175eef1b3baSKirill A. Shutemov * Set PG_double_map before dropping compound_mapcount to avoid 2176eef1b3baSKirill A. Shutemov * false-negative page_mapped(). 2177eef1b3baSKirill A. Shutemov */ 2178ec0abae6SRalph Campbell if (compound_mapcount(page) > 1 && 2179ec0abae6SRalph Campbell !TestSetPageDoubleMap(page)) { 2180eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 2181eef1b3baSKirill A. Shutemov atomic_inc(&page[i]._mapcount); 2182eef1b3baSKirill A. Shutemov } 2183eef1b3baSKirill A. Shutemov 2184468c3982SJohannes Weiner lock_page_memcg(page); 2185eef1b3baSKirill A. Shutemov if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2186eef1b3baSKirill A. Shutemov /* Last compound_mapcount is gone. */ 218769473e5dSMuchun Song __mod_lruvec_page_state(page, NR_ANON_THPS, 218869473e5dSMuchun Song -HPAGE_PMD_NR); 2189eef1b3baSKirill A. Shutemov if (TestClearPageDoubleMap(page)) { 2190eef1b3baSKirill A. Shutemov /* No need in mapcount reference anymore */ 2191eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 2192eef1b3baSKirill A. Shutemov atomic_dec(&page[i]._mapcount); 2193eef1b3baSKirill A. Shutemov } 2194eef1b3baSKirill A. Shutemov } 2195468c3982SJohannes Weiner unlock_page_memcg(page); 2196ec0abae6SRalph Campbell } 2197eef1b3baSKirill A. Shutemov 2198eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2199eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 2200e9b61f19SKirill A. Shutemov 2201e9b61f19SKirill A. Shutemov if (freeze) { 22022ac015e2SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) { 2203e9b61f19SKirill A. Shutemov page_remove_rmap(page + i, false); 2204e9b61f19SKirill A. Shutemov put_page(page + i); 2205e9b61f19SKirill A. Shutemov } 2206e9b61f19SKirill A. Shutemov } 2207eef1b3baSKirill A. Shutemov } 2208eef1b3baSKirill A. Shutemov 2209eef1b3baSKirill A. Shutemov void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 221033f4751eSNaoya Horiguchi unsigned long address, bool freeze, struct page *page) 2211eef1b3baSKirill A. Shutemov { 2212eef1b3baSKirill A. Shutemov spinlock_t *ptl; 2213ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 22141c2f6730SHugh Dickins bool do_unlock_page = false; 2215c444eb56SAndrea Arcangeli pmd_t _pmd; 2216eef1b3baSKirill A. Shutemov 22177269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 22186f4f13e8SJérôme Glisse address & HPAGE_PMD_MASK, 2219ac46d4f3SJérôme Glisse (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 2220ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 2221ac46d4f3SJérôme Glisse ptl = pmd_lock(vma->vm_mm, pmd); 222233f4751eSNaoya Horiguchi 222333f4751eSNaoya Horiguchi /* 222433f4751eSNaoya Horiguchi * If caller asks to setup a migration entries, we need a page to check 222533f4751eSNaoya Horiguchi * pmd against. Otherwise we can end up replacing wrong page. 222633f4751eSNaoya Horiguchi */ 222733f4751eSNaoya Horiguchi VM_BUG_ON(freeze && !page); 2228c444eb56SAndrea Arcangeli if (page) { 2229c444eb56SAndrea Arcangeli VM_WARN_ON_ONCE(!PageLocked(page)); 2230c444eb56SAndrea Arcangeli if (page != pmd_page(*pmd)) 223133f4751eSNaoya Horiguchi goto out; 2232c444eb56SAndrea Arcangeli } 223333f4751eSNaoya Horiguchi 2234c444eb56SAndrea Arcangeli repeat: 22355c7fb56eSDan Williams if (pmd_trans_huge(*pmd)) { 2236c444eb56SAndrea Arcangeli if (!page) { 223733f4751eSNaoya Horiguchi page = pmd_page(*pmd); 22381c2f6730SHugh Dickins /* 22391c2f6730SHugh Dickins * An anonymous page must be locked, to ensure that a 22401c2f6730SHugh Dickins * concurrent reuse_swap_page() sees stable mapcount; 22411c2f6730SHugh Dickins * but reuse_swap_page() is not used on shmem or file, 22421c2f6730SHugh Dickins * and page lock must not be taken when zap_pmd_range() 22431c2f6730SHugh Dickins * calls __split_huge_pmd() while i_mmap_lock is held. 22441c2f6730SHugh Dickins */ 22451c2f6730SHugh Dickins if (PageAnon(page)) { 2246c444eb56SAndrea Arcangeli if (unlikely(!trylock_page(page))) { 2247c444eb56SAndrea Arcangeli get_page(page); 2248c444eb56SAndrea Arcangeli _pmd = *pmd; 2249c444eb56SAndrea Arcangeli spin_unlock(ptl); 2250c444eb56SAndrea Arcangeli lock_page(page); 2251c444eb56SAndrea Arcangeli spin_lock(ptl); 2252c444eb56SAndrea Arcangeli if (unlikely(!pmd_same(*pmd, _pmd))) { 2253c444eb56SAndrea Arcangeli unlock_page(page); 2254c444eb56SAndrea Arcangeli put_page(page); 2255c444eb56SAndrea Arcangeli page = NULL; 2256c444eb56SAndrea Arcangeli goto repeat; 2257c444eb56SAndrea Arcangeli } 2258c444eb56SAndrea Arcangeli put_page(page); 2259c444eb56SAndrea Arcangeli } 22601c2f6730SHugh Dickins do_unlock_page = true; 22611c2f6730SHugh Dickins } 2262c444eb56SAndrea Arcangeli } 2263e90309c9SKirill A. Shutemov if (PageMlocked(page)) 22645f737714SKirill A. Shutemov clear_page_mlock(page); 226584c3fc4eSZi Yan } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) 22665c7fb56eSDan Williams goto out; 2267ac46d4f3SJérôme Glisse __split_huge_pmd_locked(vma, pmd, range.start, freeze); 2268e90309c9SKirill A. Shutemov out: 2269eef1b3baSKirill A. Shutemov spin_unlock(ptl); 22701c2f6730SHugh Dickins if (do_unlock_page) 2271c444eb56SAndrea Arcangeli unlock_page(page); 22724645b9feSJérôme Glisse /* 22734645b9feSJérôme Glisse * No need to double call mmu_notifier->invalidate_range() callback. 22744645b9feSJérôme Glisse * They are 3 cases to consider inside __split_huge_pmd_locked(): 22754645b9feSJérôme Glisse * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious 22764645b9feSJérôme Glisse * 2) __split_huge_zero_page_pmd() read only zero page and any write 22774645b9feSJérôme Glisse * fault will trigger a flush_notify before pointing to a new page 22784645b9feSJérôme Glisse * (it is fine if the secondary mmu keeps pointing to the old zero 22794645b9feSJérôme Glisse * page in the meantime) 22804645b9feSJérôme Glisse * 3) Split a huge pmd into pte pointing to the same page. No need 22814645b9feSJérôme Glisse * to invalidate secondary tlb entry they are all still valid. 22824645b9feSJérôme Glisse * any further changes to individual pte will notify. So no need 22834645b9feSJérôme Glisse * to call mmu_notifier->invalidate_range() 22844645b9feSJérôme Glisse */ 2285ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_only_end(&range); 2286eef1b3baSKirill A. Shutemov } 2287eef1b3baSKirill A. Shutemov 2288fec89c10SKirill A. Shutemov void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2289fec89c10SKirill A. Shutemov bool freeze, struct page *page) 229094fcc585SAndrea Arcangeli { 2291f72e7dcdSHugh Dickins pgd_t *pgd; 2292c2febafcSKirill A. Shutemov p4d_t *p4d; 2293f72e7dcdSHugh Dickins pud_t *pud; 229494fcc585SAndrea Arcangeli pmd_t *pmd; 229594fcc585SAndrea Arcangeli 229678ddc534SKirill A. Shutemov pgd = pgd_offset(vma->vm_mm, address); 2297f72e7dcdSHugh Dickins if (!pgd_present(*pgd)) 2298f72e7dcdSHugh Dickins return; 2299f72e7dcdSHugh Dickins 2300c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, address); 2301c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 2302c2febafcSKirill A. Shutemov return; 2303c2febafcSKirill A. Shutemov 2304c2febafcSKirill A. Shutemov pud = pud_offset(p4d, address); 2305f72e7dcdSHugh Dickins if (!pud_present(*pud)) 2306f72e7dcdSHugh Dickins return; 2307f72e7dcdSHugh Dickins 2308f72e7dcdSHugh Dickins pmd = pmd_offset(pud, address); 2309fec89c10SKirill A. Shutemov 231033f4751eSNaoya Horiguchi __split_huge_pmd(vma, pmd, address, freeze, page); 231194fcc585SAndrea Arcangeli } 231294fcc585SAndrea Arcangeli 231371f9e58eSMiaohe Lin static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address) 231471f9e58eSMiaohe Lin { 231571f9e58eSMiaohe Lin /* 231671f9e58eSMiaohe Lin * If the new address isn't hpage aligned and it could previously 231771f9e58eSMiaohe Lin * contain an hugepage: check if we need to split an huge pmd. 231871f9e58eSMiaohe Lin */ 231971f9e58eSMiaohe Lin if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) && 232071f9e58eSMiaohe Lin range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE), 232171f9e58eSMiaohe Lin ALIGN(address, HPAGE_PMD_SIZE))) 232271f9e58eSMiaohe Lin split_huge_pmd_address(vma, address, false, NULL); 232371f9e58eSMiaohe Lin } 232471f9e58eSMiaohe Lin 2325e1b9996bSKirill A. Shutemov void vma_adjust_trans_huge(struct vm_area_struct *vma, 232694fcc585SAndrea Arcangeli unsigned long start, 232794fcc585SAndrea Arcangeli unsigned long end, 232894fcc585SAndrea Arcangeli long adjust_next) 232994fcc585SAndrea Arcangeli { 233071f9e58eSMiaohe Lin /* Check if we need to split start first. */ 233171f9e58eSMiaohe Lin split_huge_pmd_if_needed(vma, start); 233271f9e58eSMiaohe Lin 233371f9e58eSMiaohe Lin /* Check if we need to split end next. */ 233471f9e58eSMiaohe Lin split_huge_pmd_if_needed(vma, end); 233594fcc585SAndrea Arcangeli 233694fcc585SAndrea Arcangeli /* 233771f9e58eSMiaohe Lin * If we're also updating the vma->vm_next->vm_start, 233871f9e58eSMiaohe Lin * check if we need to split it. 233994fcc585SAndrea Arcangeli */ 234094fcc585SAndrea Arcangeli if (adjust_next > 0) { 234194fcc585SAndrea Arcangeli struct vm_area_struct *next = vma->vm_next; 234294fcc585SAndrea Arcangeli unsigned long nstart = next->vm_start; 2343f9d86a60SWei Yang nstart += adjust_next; 234471f9e58eSMiaohe Lin split_huge_pmd_if_needed(next, nstart); 234594fcc585SAndrea Arcangeli } 234694fcc585SAndrea Arcangeli } 2347e9b61f19SKirill A. Shutemov 2348906f9cdfSHugh Dickins static void unmap_page(struct page *page) 2349e9b61f19SKirill A. Shutemov { 2350013339dfSShakeel Butt enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | 2351c7ab0d2fSKirill A. Shutemov TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2352666e5a40SMinchan Kim bool unmap_success; 2353e9b61f19SKirill A. Shutemov 2354e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageHead(page), page); 2355e9b61f19SKirill A. Shutemov 2356baa355fdSKirill A. Shutemov if (PageAnon(page)) 2357b5ff8161SNaoya Horiguchi ttu_flags |= TTU_SPLIT_FREEZE; 2358baa355fdSKirill A. Shutemov 2359666e5a40SMinchan Kim unmap_success = try_to_unmap(page, ttu_flags); 2360666e5a40SMinchan Kim VM_BUG_ON_PAGE(!unmap_success, page); 2361bd56086fSKirill A. Shutemov } 2362bd56086fSKirill A. Shutemov 23638cce5475SKirill A. Shutemov static void remap_page(struct page *page, unsigned int nr) 2364e9b61f19SKirill A. Shutemov { 2365fec89c10SKirill A. Shutemov int i; 2366ace71a19SKirill A. Shutemov if (PageTransHuge(page)) { 2367ace71a19SKirill A. Shutemov remove_migration_ptes(page, page, true); 2368ace71a19SKirill A. Shutemov } else { 23698cce5475SKirill A. Shutemov for (i = 0; i < nr; i++) 2370fec89c10SKirill A. Shutemov remove_migration_ptes(page + i, page + i, true); 2371e9b61f19SKirill A. Shutemov } 2372ace71a19SKirill A. Shutemov } 2373e9b61f19SKirill A. Shutemov 237494866635SAlex Shi static void lru_add_page_tail(struct page *head, struct page *tail, 237588dcb9a3SAlex Shi struct lruvec *lruvec, struct list_head *list) 237688dcb9a3SAlex Shi { 237794866635SAlex Shi VM_BUG_ON_PAGE(!PageHead(head), head); 237894866635SAlex Shi VM_BUG_ON_PAGE(PageCompound(tail), head); 237994866635SAlex Shi VM_BUG_ON_PAGE(PageLRU(tail), head); 23806168d0daSAlex Shi lockdep_assert_held(&lruvec->lru_lock); 238188dcb9a3SAlex Shi 23826dbb5741SAlex Shi if (list) { 238388dcb9a3SAlex Shi /* page reclaim is reclaiming a huge page */ 23846dbb5741SAlex Shi VM_WARN_ON(PageLRU(head)); 238594866635SAlex Shi get_page(tail); 238694866635SAlex Shi list_add_tail(&tail->lru, list); 238788dcb9a3SAlex Shi } else { 23886dbb5741SAlex Shi /* head is still on lru (and we have it frozen) */ 23896dbb5741SAlex Shi VM_WARN_ON(!PageLRU(head)); 23906dbb5741SAlex Shi SetPageLRU(tail); 23916dbb5741SAlex Shi list_add_tail(&tail->lru, &head->lru); 239288dcb9a3SAlex Shi } 239388dcb9a3SAlex Shi } 239488dcb9a3SAlex Shi 23958df651c7SKirill A. Shutemov static void __split_huge_page_tail(struct page *head, int tail, 2396e9b61f19SKirill A. Shutemov struct lruvec *lruvec, struct list_head *list) 2397e9b61f19SKirill A. Shutemov { 2398e9b61f19SKirill A. Shutemov struct page *page_tail = head + tail; 2399e9b61f19SKirill A. Shutemov 24008df651c7SKirill A. Shutemov VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2401e9b61f19SKirill A. Shutemov 2402e9b61f19SKirill A. Shutemov /* 2403605ca5edSKonstantin Khlebnikov * Clone page flags before unfreezing refcount. 2404605ca5edSKonstantin Khlebnikov * 2405605ca5edSKonstantin Khlebnikov * After successful get_page_unless_zero() might follow flags change, 24068958b249SHaitao Shi * for example lock_page() which set PG_waiters. 2407e9b61f19SKirill A. Shutemov */ 2408e9b61f19SKirill A. Shutemov page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2409e9b61f19SKirill A. Shutemov page_tail->flags |= (head->flags & 2410e9b61f19SKirill A. Shutemov ((1L << PG_referenced) | 2411e9b61f19SKirill A. Shutemov (1L << PG_swapbacked) | 241238d8b4e6SHuang Ying (1L << PG_swapcache) | 2413e9b61f19SKirill A. Shutemov (1L << PG_mlocked) | 2414e9b61f19SKirill A. Shutemov (1L << PG_uptodate) | 2415e9b61f19SKirill A. Shutemov (1L << PG_active) | 24161899ad18SJohannes Weiner (1L << PG_workingset) | 2417e9b61f19SKirill A. Shutemov (1L << PG_locked) | 2418b8d3c4c3SMinchan Kim (1L << PG_unevictable) | 241972e6afa0SCatalin Marinas #ifdef CONFIG_64BIT 242072e6afa0SCatalin Marinas (1L << PG_arch_2) | 242172e6afa0SCatalin Marinas #endif 2422b8d3c4c3SMinchan Kim (1L << PG_dirty))); 2423e9b61f19SKirill A. Shutemov 2424173d9d9fSHugh Dickins /* ->mapping in first tail page is compound_mapcount */ 2425173d9d9fSHugh Dickins VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2426173d9d9fSHugh Dickins page_tail); 2427173d9d9fSHugh Dickins page_tail->mapping = head->mapping; 2428173d9d9fSHugh Dickins page_tail->index = head->index + tail; 2429173d9d9fSHugh Dickins 2430605ca5edSKonstantin Khlebnikov /* Page flags must be visible before we make the page non-compound. */ 2431e9b61f19SKirill A. Shutemov smp_wmb(); 2432e9b61f19SKirill A. Shutemov 2433605ca5edSKonstantin Khlebnikov /* 2434605ca5edSKonstantin Khlebnikov * Clear PageTail before unfreezing page refcount. 2435605ca5edSKonstantin Khlebnikov * 2436605ca5edSKonstantin Khlebnikov * After successful get_page_unless_zero() might follow put_page() 2437605ca5edSKonstantin Khlebnikov * which needs correct compound_head(). 2438605ca5edSKonstantin Khlebnikov */ 2439e9b61f19SKirill A. Shutemov clear_compound_head(page_tail); 2440e9b61f19SKirill A. Shutemov 2441605ca5edSKonstantin Khlebnikov /* Finally unfreeze refcount. Additional reference from page cache. */ 2442605ca5edSKonstantin Khlebnikov page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || 2443605ca5edSKonstantin Khlebnikov PageSwapCache(head))); 2444605ca5edSKonstantin Khlebnikov 2445e9b61f19SKirill A. Shutemov if (page_is_young(head)) 2446e9b61f19SKirill A. Shutemov set_page_young(page_tail); 2447e9b61f19SKirill A. Shutemov if (page_is_idle(head)) 2448e9b61f19SKirill A. Shutemov set_page_idle(page_tail); 2449e9b61f19SKirill A. Shutemov 2450e9b61f19SKirill A. Shutemov page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 245194723aafSMichal Hocko 245294723aafSMichal Hocko /* 245394723aafSMichal Hocko * always add to the tail because some iterators expect new 245494723aafSMichal Hocko * pages to show after the currently processed elements - e.g. 245594723aafSMichal Hocko * migrate_pages 245694723aafSMichal Hocko */ 2457e9b61f19SKirill A. Shutemov lru_add_page_tail(head, page_tail, lruvec, list); 2458e9b61f19SKirill A. Shutemov } 2459e9b61f19SKirill A. Shutemov 2460baa355fdSKirill A. Shutemov static void __split_huge_page(struct page *page, struct list_head *list, 2461b6769834SAlex Shi pgoff_t end) 2462e9b61f19SKirill A. Shutemov { 2463e9b61f19SKirill A. Shutemov struct page *head = compound_head(page); 2464e9b61f19SKirill A. Shutemov struct lruvec *lruvec; 24654101196bSMatthew Wilcox (Oracle) struct address_space *swap_cache = NULL; 24664101196bSMatthew Wilcox (Oracle) unsigned long offset = 0; 24678cce5475SKirill A. Shutemov unsigned int nr = thp_nr_pages(head); 24688df651c7SKirill A. Shutemov int i; 2469e9b61f19SKirill A. Shutemov 2470e9b61f19SKirill A. Shutemov /* complete memcg works before add pages to LRU */ 2471be6c8982SZhou Guanghui split_page_memcg(head, nr); 2472e9b61f19SKirill A. Shutemov 24734101196bSMatthew Wilcox (Oracle) if (PageAnon(head) && PageSwapCache(head)) { 24744101196bSMatthew Wilcox (Oracle) swp_entry_t entry = { .val = page_private(head) }; 24754101196bSMatthew Wilcox (Oracle) 24764101196bSMatthew Wilcox (Oracle) offset = swp_offset(entry); 24774101196bSMatthew Wilcox (Oracle) swap_cache = swap_address_space(entry); 24784101196bSMatthew Wilcox (Oracle) xa_lock(&swap_cache->i_pages); 24794101196bSMatthew Wilcox (Oracle) } 24804101196bSMatthew Wilcox (Oracle) 2481f0953a1bSIngo Molnar /* lock lru list/PageCompound, ref frozen by page_ref_freeze */ 24826168d0daSAlex Shi lruvec = lock_page_lruvec(head); 2483b6769834SAlex Shi 24848cce5475SKirill A. Shutemov for (i = nr - 1; i >= 1; i--) { 24858df651c7SKirill A. Shutemov __split_huge_page_tail(head, i, lruvec, list); 2486baa355fdSKirill A. Shutemov /* Some pages can be beyond i_size: drop them from page cache */ 2487baa355fdSKirill A. Shutemov if (head[i].index >= end) { 24882d077d4bSHugh Dickins ClearPageDirty(head + i); 2489baa355fdSKirill A. Shutemov __delete_from_page_cache(head + i, NULL); 2490800d8c63SKirill A. Shutemov if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2491800d8c63SKirill A. Shutemov shmem_uncharge(head->mapping->host, 1); 2492baa355fdSKirill A. Shutemov put_page(head + i); 24934101196bSMatthew Wilcox (Oracle) } else if (!PageAnon(page)) { 24944101196bSMatthew Wilcox (Oracle) __xa_store(&head->mapping->i_pages, head[i].index, 24954101196bSMatthew Wilcox (Oracle) head + i, 0); 24964101196bSMatthew Wilcox (Oracle) } else if (swap_cache) { 24974101196bSMatthew Wilcox (Oracle) __xa_store(&swap_cache->i_pages, offset + i, 24984101196bSMatthew Wilcox (Oracle) head + i, 0); 2499baa355fdSKirill A. Shutemov } 2500baa355fdSKirill A. Shutemov } 2501e9b61f19SKirill A. Shutemov 2502e9b61f19SKirill A. Shutemov ClearPageCompound(head); 25036168d0daSAlex Shi unlock_page_lruvec(lruvec); 2504b6769834SAlex Shi /* Caller disabled irqs, so they are still disabled here */ 2505f7da677bSVlastimil Babka 25068cce5475SKirill A. Shutemov split_page_owner(head, nr); 2507f7da677bSVlastimil Babka 2508baa355fdSKirill A. Shutemov /* See comment in __split_huge_page_tail() */ 2509baa355fdSKirill A. Shutemov if (PageAnon(head)) { 2510aa5dc07fSMatthew Wilcox /* Additional pin to swap cache */ 25114101196bSMatthew Wilcox (Oracle) if (PageSwapCache(head)) { 251238d8b4e6SHuang Ying page_ref_add(head, 2); 25134101196bSMatthew Wilcox (Oracle) xa_unlock(&swap_cache->i_pages); 25144101196bSMatthew Wilcox (Oracle) } else { 2515baa355fdSKirill A. Shutemov page_ref_inc(head); 25164101196bSMatthew Wilcox (Oracle) } 2517baa355fdSKirill A. Shutemov } else { 2518aa5dc07fSMatthew Wilcox /* Additional pin to page cache */ 2519baa355fdSKirill A. Shutemov page_ref_add(head, 2); 2520b93b0163SMatthew Wilcox xa_unlock(&head->mapping->i_pages); 2521baa355fdSKirill A. Shutemov } 2522b6769834SAlex Shi local_irq_enable(); 2523e9b61f19SKirill A. Shutemov 25248cce5475SKirill A. Shutemov remap_page(head, nr); 2525e9b61f19SKirill A. Shutemov 2526c4f9c701SHuang Ying if (PageSwapCache(head)) { 2527c4f9c701SHuang Ying swp_entry_t entry = { .val = page_private(head) }; 2528c4f9c701SHuang Ying 2529c4f9c701SHuang Ying split_swap_cluster(entry); 2530c4f9c701SHuang Ying } 2531c4f9c701SHuang Ying 25328cce5475SKirill A. Shutemov for (i = 0; i < nr; i++) { 2533e9b61f19SKirill A. Shutemov struct page *subpage = head + i; 2534e9b61f19SKirill A. Shutemov if (subpage == page) 2535e9b61f19SKirill A. Shutemov continue; 2536e9b61f19SKirill A. Shutemov unlock_page(subpage); 2537e9b61f19SKirill A. Shutemov 2538e9b61f19SKirill A. Shutemov /* 2539e9b61f19SKirill A. Shutemov * Subpages may be freed if there wasn't any mapping 2540e9b61f19SKirill A. Shutemov * like if add_to_swap() is running on a lru page that 2541e9b61f19SKirill A. Shutemov * had its mapping zapped. And freeing these pages 2542e9b61f19SKirill A. Shutemov * requires taking the lru_lock so we do the put_page 2543e9b61f19SKirill A. Shutemov * of the tail pages after the split is complete. 2544e9b61f19SKirill A. Shutemov */ 2545e9b61f19SKirill A. Shutemov put_page(subpage); 2546e9b61f19SKirill A. Shutemov } 2547e9b61f19SKirill A. Shutemov } 2548e9b61f19SKirill A. Shutemov 2549b20ce5e0SKirill A. Shutemov int total_mapcount(struct page *page) 2550b20ce5e0SKirill A. Shutemov { 255186b562b6SKirill A. Shutemov int i, compound, nr, ret; 2552b20ce5e0SKirill A. Shutemov 2553b20ce5e0SKirill A. Shutemov VM_BUG_ON_PAGE(PageTail(page), page); 2554b20ce5e0SKirill A. Shutemov 2555b20ce5e0SKirill A. Shutemov if (likely(!PageCompound(page))) 2556b20ce5e0SKirill A. Shutemov return atomic_read(&page->_mapcount) + 1; 2557b20ce5e0SKirill A. Shutemov 2558dd78feddSKirill A. Shutemov compound = compound_mapcount(page); 255986b562b6SKirill A. Shutemov nr = compound_nr(page); 2560b20ce5e0SKirill A. Shutemov if (PageHuge(page)) 2561dd78feddSKirill A. Shutemov return compound; 2562dd78feddSKirill A. Shutemov ret = compound; 256386b562b6SKirill A. Shutemov for (i = 0; i < nr; i++) 2564b20ce5e0SKirill A. Shutemov ret += atomic_read(&page[i]._mapcount) + 1; 2565dd78feddSKirill A. Shutemov /* File pages has compound_mapcount included in _mapcount */ 2566dd78feddSKirill A. Shutemov if (!PageAnon(page)) 256786b562b6SKirill A. Shutemov return ret - compound * nr; 2568b20ce5e0SKirill A. Shutemov if (PageDoubleMap(page)) 256986b562b6SKirill A. Shutemov ret -= nr; 2570b20ce5e0SKirill A. Shutemov return ret; 2571b20ce5e0SKirill A. Shutemov } 2572b20ce5e0SKirill A. Shutemov 2573e9b61f19SKirill A. Shutemov /* 25746d0a07edSAndrea Arcangeli * This calculates accurately how many mappings a transparent hugepage 25756d0a07edSAndrea Arcangeli * has (unlike page_mapcount() which isn't fully accurate). This full 25766d0a07edSAndrea Arcangeli * accuracy is primarily needed to know if copy-on-write faults can 25776d0a07edSAndrea Arcangeli * reuse the page and change the mapping to read-write instead of 25786d0a07edSAndrea Arcangeli * copying them. At the same time this returns the total_mapcount too. 25796d0a07edSAndrea Arcangeli * 25806d0a07edSAndrea Arcangeli * The function returns the highest mapcount any one of the subpages 25816d0a07edSAndrea Arcangeli * has. If the return value is one, even if different processes are 25826d0a07edSAndrea Arcangeli * mapping different subpages of the transparent hugepage, they can 25836d0a07edSAndrea Arcangeli * all reuse it, because each process is reusing a different subpage. 25846d0a07edSAndrea Arcangeli * 25856d0a07edSAndrea Arcangeli * The total_mapcount is instead counting all virtual mappings of the 25866d0a07edSAndrea Arcangeli * subpages. If the total_mapcount is equal to "one", it tells the 25876d0a07edSAndrea Arcangeli * caller all mappings belong to the same "mm" and in turn the 25886d0a07edSAndrea Arcangeli * anon_vma of the transparent hugepage can become the vma->anon_vma 25896d0a07edSAndrea Arcangeli * local one as no other process may be mapping any of the subpages. 25906d0a07edSAndrea Arcangeli * 25916d0a07edSAndrea Arcangeli * It would be more accurate to replace page_mapcount() with 25926d0a07edSAndrea Arcangeli * page_trans_huge_mapcount(), however we only use 25936d0a07edSAndrea Arcangeli * page_trans_huge_mapcount() in the copy-on-write faults where we 25946d0a07edSAndrea Arcangeli * need full accuracy to avoid breaking page pinning, because 25956d0a07edSAndrea Arcangeli * page_trans_huge_mapcount() is slower than page_mapcount(). 25966d0a07edSAndrea Arcangeli */ 25976d0a07edSAndrea Arcangeli int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 25986d0a07edSAndrea Arcangeli { 25996d0a07edSAndrea Arcangeli int i, ret, _total_mapcount, mapcount; 26006d0a07edSAndrea Arcangeli 26016d0a07edSAndrea Arcangeli /* hugetlbfs shouldn't call it */ 26026d0a07edSAndrea Arcangeli VM_BUG_ON_PAGE(PageHuge(page), page); 26036d0a07edSAndrea Arcangeli 26046d0a07edSAndrea Arcangeli if (likely(!PageTransCompound(page))) { 26056d0a07edSAndrea Arcangeli mapcount = atomic_read(&page->_mapcount) + 1; 26066d0a07edSAndrea Arcangeli if (total_mapcount) 26076d0a07edSAndrea Arcangeli *total_mapcount = mapcount; 26086d0a07edSAndrea Arcangeli return mapcount; 26096d0a07edSAndrea Arcangeli } 26106d0a07edSAndrea Arcangeli 26116d0a07edSAndrea Arcangeli page = compound_head(page); 26126d0a07edSAndrea Arcangeli 26136d0a07edSAndrea Arcangeli _total_mapcount = ret = 0; 261465dfe3c3SMatthew Wilcox (Oracle) for (i = 0; i < thp_nr_pages(page); i++) { 26156d0a07edSAndrea Arcangeli mapcount = atomic_read(&page[i]._mapcount) + 1; 26166d0a07edSAndrea Arcangeli ret = max(ret, mapcount); 26176d0a07edSAndrea Arcangeli _total_mapcount += mapcount; 26186d0a07edSAndrea Arcangeli } 26196d0a07edSAndrea Arcangeli if (PageDoubleMap(page)) { 26206d0a07edSAndrea Arcangeli ret -= 1; 262165dfe3c3SMatthew Wilcox (Oracle) _total_mapcount -= thp_nr_pages(page); 26226d0a07edSAndrea Arcangeli } 26236d0a07edSAndrea Arcangeli mapcount = compound_mapcount(page); 26246d0a07edSAndrea Arcangeli ret += mapcount; 26256d0a07edSAndrea Arcangeli _total_mapcount += mapcount; 26266d0a07edSAndrea Arcangeli if (total_mapcount) 26276d0a07edSAndrea Arcangeli *total_mapcount = _total_mapcount; 26286d0a07edSAndrea Arcangeli return ret; 26296d0a07edSAndrea Arcangeli } 26306d0a07edSAndrea Arcangeli 2631b8f593cdSHuang Ying /* Racy check whether the huge page can be split */ 2632b8f593cdSHuang Ying bool can_split_huge_page(struct page *page, int *pextra_pins) 2633b8f593cdSHuang Ying { 2634b8f593cdSHuang Ying int extra_pins; 2635b8f593cdSHuang Ying 2636aa5dc07fSMatthew Wilcox /* Additional pins from page cache */ 2637b8f593cdSHuang Ying if (PageAnon(page)) 2638e2333dadSMatthew Wilcox (Oracle) extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0; 2639b8f593cdSHuang Ying else 2640e2333dadSMatthew Wilcox (Oracle) extra_pins = thp_nr_pages(page); 2641b8f593cdSHuang Ying if (pextra_pins) 2642b8f593cdSHuang Ying *pextra_pins = extra_pins; 2643b8f593cdSHuang Ying return total_mapcount(page) == page_count(page) - extra_pins - 1; 2644b8f593cdSHuang Ying } 2645b8f593cdSHuang Ying 26466d0a07edSAndrea Arcangeli /* 2647e9b61f19SKirill A. Shutemov * This function splits huge page into normal pages. @page can point to any 2648e9b61f19SKirill A. Shutemov * subpage of huge page to split. Split doesn't change the position of @page. 2649e9b61f19SKirill A. Shutemov * 2650e9b61f19SKirill A. Shutemov * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2651e9b61f19SKirill A. Shutemov * The huge page must be locked. 2652e9b61f19SKirill A. Shutemov * 2653e9b61f19SKirill A. Shutemov * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2654e9b61f19SKirill A. Shutemov * 2655e9b61f19SKirill A. Shutemov * Both head page and tail pages will inherit mapping, flags, and so on from 2656e9b61f19SKirill A. Shutemov * the hugepage. 2657e9b61f19SKirill A. Shutemov * 2658e9b61f19SKirill A. Shutemov * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2659e9b61f19SKirill A. Shutemov * they are not mapped. 2660e9b61f19SKirill A. Shutemov * 2661e9b61f19SKirill A. Shutemov * Returns 0 if the hugepage is split successfully. 2662e9b61f19SKirill A. Shutemov * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2663e9b61f19SKirill A. Shutemov * us. 2664e9b61f19SKirill A. Shutemov */ 2665e9b61f19SKirill A. Shutemov int split_huge_page_to_list(struct page *page, struct list_head *list) 2666e9b61f19SKirill A. Shutemov { 2667e9b61f19SKirill A. Shutemov struct page *head = compound_head(page); 2668a8803e6cSWei Yang struct deferred_split *ds_queue = get_deferred_split_queue(head); 2669baa355fdSKirill A. Shutemov struct anon_vma *anon_vma = NULL; 2670baa355fdSKirill A. Shutemov struct address_space *mapping = NULL; 2671baa355fdSKirill A. Shutemov int count, mapcount, extra_pins, ret; 2672006d3ff2SHugh Dickins pgoff_t end; 2673e9b61f19SKirill A. Shutemov 2674cb829624SWei Yang VM_BUG_ON_PAGE(is_huge_zero_page(head), head); 2675a8803e6cSWei Yang VM_BUG_ON_PAGE(!PageLocked(head), head); 2676a8803e6cSWei Yang VM_BUG_ON_PAGE(!PageCompound(head), head); 2677e9b61f19SKirill A. Shutemov 2678a8803e6cSWei Yang if (PageWriteback(head)) 267959807685SHuang Ying return -EBUSY; 268059807685SHuang Ying 2681baa355fdSKirill A. Shutemov if (PageAnon(head)) { 2682e9b61f19SKirill A. Shutemov /* 2683c1e8d7c6SMichel Lespinasse * The caller does not necessarily hold an mmap_lock that would 2684baa355fdSKirill A. Shutemov * prevent the anon_vma disappearing so we first we take a 2685baa355fdSKirill A. Shutemov * reference to it and then lock the anon_vma for write. This 2686baa355fdSKirill A. Shutemov * is similar to page_lock_anon_vma_read except the write lock 2687baa355fdSKirill A. Shutemov * is taken to serialise against parallel split or collapse 2688baa355fdSKirill A. Shutemov * operations. 2689e9b61f19SKirill A. Shutemov */ 2690e9b61f19SKirill A. Shutemov anon_vma = page_get_anon_vma(head); 2691e9b61f19SKirill A. Shutemov if (!anon_vma) { 2692e9b61f19SKirill A. Shutemov ret = -EBUSY; 2693e9b61f19SKirill A. Shutemov goto out; 2694e9b61f19SKirill A. Shutemov } 2695006d3ff2SHugh Dickins end = -1; 2696baa355fdSKirill A. Shutemov mapping = NULL; 2697e9b61f19SKirill A. Shutemov anon_vma_lock_write(anon_vma); 2698baa355fdSKirill A. Shutemov } else { 2699baa355fdSKirill A. Shutemov mapping = head->mapping; 2700baa355fdSKirill A. Shutemov 2701baa355fdSKirill A. Shutemov /* Truncated ? */ 2702baa355fdSKirill A. Shutemov if (!mapping) { 2703baa355fdSKirill A. Shutemov ret = -EBUSY; 2704baa355fdSKirill A. Shutemov goto out; 2705baa355fdSKirill A. Shutemov } 2706baa355fdSKirill A. Shutemov 2707baa355fdSKirill A. Shutemov anon_vma = NULL; 2708baa355fdSKirill A. Shutemov i_mmap_lock_read(mapping); 2709006d3ff2SHugh Dickins 2710006d3ff2SHugh Dickins /* 2711006d3ff2SHugh Dickins *__split_huge_page() may need to trim off pages beyond EOF: 2712006d3ff2SHugh Dickins * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 2713006d3ff2SHugh Dickins * which cannot be nested inside the page tree lock. So note 2714006d3ff2SHugh Dickins * end now: i_size itself may be changed at any moment, but 2715006d3ff2SHugh Dickins * head page lock is good enough to serialize the trimming. 2716006d3ff2SHugh Dickins */ 2717006d3ff2SHugh Dickins end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 2718baa355fdSKirill A. Shutemov } 2719e9b61f19SKirill A. Shutemov 2720e9b61f19SKirill A. Shutemov /* 2721906f9cdfSHugh Dickins * Racy check if we can split the page, before unmap_page() will 2722e9b61f19SKirill A. Shutemov * split PMDs 2723e9b61f19SKirill A. Shutemov */ 2724b8f593cdSHuang Ying if (!can_split_huge_page(head, &extra_pins)) { 2725e9b61f19SKirill A. Shutemov ret = -EBUSY; 2726e9b61f19SKirill A. Shutemov goto out_unlock; 2727e9b61f19SKirill A. Shutemov } 2728e9b61f19SKirill A. Shutemov 2729906f9cdfSHugh Dickins unmap_page(head); 2730e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(compound_mapcount(head), head); 2731e9b61f19SKirill A. Shutemov 2732b6769834SAlex Shi /* block interrupt reentry in xa_lock and spinlock */ 2733b6769834SAlex Shi local_irq_disable(); 2734baa355fdSKirill A. Shutemov if (mapping) { 2735aa5dc07fSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, page_index(head)); 2736baa355fdSKirill A. Shutemov 2737baa355fdSKirill A. Shutemov /* 2738aa5dc07fSMatthew Wilcox * Check if the head page is present in page cache. 2739baa355fdSKirill A. Shutemov * We assume all tail are present too, if head is there. 2740baa355fdSKirill A. Shutemov */ 2741aa5dc07fSMatthew Wilcox xa_lock(&mapping->i_pages); 2742aa5dc07fSMatthew Wilcox if (xas_load(&xas) != head) 2743baa355fdSKirill A. Shutemov goto fail; 2744baa355fdSKirill A. Shutemov } 2745baa355fdSKirill A. Shutemov 27460139aa7bSJoonsoo Kim /* Prevent deferred_split_scan() touching ->_refcount */ 2747364c1eebSYang Shi spin_lock(&ds_queue->split_queue_lock); 2748e9b61f19SKirill A. Shutemov count = page_count(head); 2749e9b61f19SKirill A. Shutemov mapcount = total_mapcount(head); 2750baa355fdSKirill A. Shutemov if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 27519a982250SKirill A. Shutemov if (!list_empty(page_deferred_list(head))) { 2752364c1eebSYang Shi ds_queue->split_queue_len--; 27539a982250SKirill A. Shutemov list_del(page_deferred_list(head)); 27549a982250SKirill A. Shutemov } 2755afb97172SWei Yang spin_unlock(&ds_queue->split_queue_lock); 275606d3eff6SKirill A. Shutemov if (mapping) { 2757bf9eceadSMuchun Song int nr = thp_nr_pages(head); 2758bf9eceadSMuchun Song 2759a8803e6cSWei Yang if (PageSwapBacked(head)) 276057b2847dSMuchun Song __mod_lruvec_page_state(head, NR_SHMEM_THPS, 276157b2847dSMuchun Song -nr); 276206d3eff6SKirill A. Shutemov else 2763bf9eceadSMuchun Song __mod_lruvec_page_state(head, NR_FILE_THPS, 2764bf9eceadSMuchun Song -nr); 276506d3eff6SKirill A. Shutemov } 276606d3eff6SKirill A. Shutemov 2767b6769834SAlex Shi __split_huge_page(page, list, end); 2768e9b61f19SKirill A. Shutemov ret = 0; 2769baa355fdSKirill A. Shutemov } else { 2770baa355fdSKirill A. Shutemov if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2771e9b61f19SKirill A. Shutemov pr_alert("total_mapcount: %u, page_count(): %u\n", 2772e9b61f19SKirill A. Shutemov mapcount, count); 2773e9b61f19SKirill A. Shutemov if (PageTail(page)) 2774e9b61f19SKirill A. Shutemov dump_page(head, NULL); 2775bd56086fSKirill A. Shutemov dump_page(page, "total_mapcount(head) > 0"); 2776e9b61f19SKirill A. Shutemov BUG(); 2777baa355fdSKirill A. Shutemov } 2778364c1eebSYang Shi spin_unlock(&ds_queue->split_queue_lock); 2779baa355fdSKirill A. Shutemov fail: if (mapping) 2780b93b0163SMatthew Wilcox xa_unlock(&mapping->i_pages); 2781b6769834SAlex Shi local_irq_enable(); 27828cce5475SKirill A. Shutemov remap_page(head, thp_nr_pages(head)); 2783e9b61f19SKirill A. Shutemov ret = -EBUSY; 2784e9b61f19SKirill A. Shutemov } 2785e9b61f19SKirill A. Shutemov 2786e9b61f19SKirill A. Shutemov out_unlock: 2787baa355fdSKirill A. Shutemov if (anon_vma) { 2788e9b61f19SKirill A. Shutemov anon_vma_unlock_write(anon_vma); 2789e9b61f19SKirill A. Shutemov put_anon_vma(anon_vma); 2790baa355fdSKirill A. Shutemov } 2791baa355fdSKirill A. Shutemov if (mapping) 2792baa355fdSKirill A. Shutemov i_mmap_unlock_read(mapping); 2793e9b61f19SKirill A. Shutemov out: 2794e9b61f19SKirill A. Shutemov count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2795e9b61f19SKirill A. Shutemov return ret; 2796e9b61f19SKirill A. Shutemov } 27979a982250SKirill A. Shutemov 27989a982250SKirill A. Shutemov void free_transhuge_page(struct page *page) 27999a982250SKirill A. Shutemov { 280087eaceb3SYang Shi struct deferred_split *ds_queue = get_deferred_split_queue(page); 28019a982250SKirill A. Shutemov unsigned long flags; 28029a982250SKirill A. Shutemov 2803364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 28049a982250SKirill A. Shutemov if (!list_empty(page_deferred_list(page))) { 2805364c1eebSYang Shi ds_queue->split_queue_len--; 28069a982250SKirill A. Shutemov list_del(page_deferred_list(page)); 28079a982250SKirill A. Shutemov } 2808364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 28099a982250SKirill A. Shutemov free_compound_page(page); 28109a982250SKirill A. Shutemov } 28119a982250SKirill A. Shutemov 28129a982250SKirill A. Shutemov void deferred_split_huge_page(struct page *page) 28139a982250SKirill A. Shutemov { 281487eaceb3SYang Shi struct deferred_split *ds_queue = get_deferred_split_queue(page); 281587eaceb3SYang Shi #ifdef CONFIG_MEMCG 2816bcfe06bfSRoman Gushchin struct mem_cgroup *memcg = page_memcg(compound_head(page)); 281787eaceb3SYang Shi #endif 28189a982250SKirill A. Shutemov unsigned long flags; 28199a982250SKirill A. Shutemov 28209a982250SKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 28219a982250SKirill A. Shutemov 282287eaceb3SYang Shi /* 282387eaceb3SYang Shi * The try_to_unmap() in page reclaim path might reach here too, 282487eaceb3SYang Shi * this may cause a race condition to corrupt deferred split queue. 282587eaceb3SYang Shi * And, if page reclaim is already handling the same page, it is 282687eaceb3SYang Shi * unnecessary to handle it again in shrinker. 282787eaceb3SYang Shi * 282887eaceb3SYang Shi * Check PageSwapCache to determine if the page is being 282987eaceb3SYang Shi * handled by page reclaim since THP swap would add the page into 283087eaceb3SYang Shi * swap cache before calling try_to_unmap(). 283187eaceb3SYang Shi */ 283287eaceb3SYang Shi if (PageSwapCache(page)) 283387eaceb3SYang Shi return; 283487eaceb3SYang Shi 2835364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 28369a982250SKirill A. Shutemov if (list_empty(page_deferred_list(page))) { 2837f9719a03SKirill A. Shutemov count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2838364c1eebSYang Shi list_add_tail(page_deferred_list(page), &ds_queue->split_queue); 2839364c1eebSYang Shi ds_queue->split_queue_len++; 284087eaceb3SYang Shi #ifdef CONFIG_MEMCG 284187eaceb3SYang Shi if (memcg) 28422bfd3637SYang Shi set_shrinker_bit(memcg, page_to_nid(page), 284387eaceb3SYang Shi deferred_split_shrinker.id); 284487eaceb3SYang Shi #endif 28459a982250SKirill A. Shutemov } 2846364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 28479a982250SKirill A. Shutemov } 28489a982250SKirill A. Shutemov 28499a982250SKirill A. Shutemov static unsigned long deferred_split_count(struct shrinker *shrink, 28509a982250SKirill A. Shutemov struct shrink_control *sc) 28519a982250SKirill A. Shutemov { 2852a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 2853364c1eebSYang Shi struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 285487eaceb3SYang Shi 285587eaceb3SYang Shi #ifdef CONFIG_MEMCG 285687eaceb3SYang Shi if (sc->memcg) 285787eaceb3SYang Shi ds_queue = &sc->memcg->deferred_split_queue; 285887eaceb3SYang Shi #endif 2859364c1eebSYang Shi return READ_ONCE(ds_queue->split_queue_len); 28609a982250SKirill A. Shutemov } 28619a982250SKirill A. Shutemov 28629a982250SKirill A. Shutemov static unsigned long deferred_split_scan(struct shrinker *shrink, 28639a982250SKirill A. Shutemov struct shrink_control *sc) 28649a982250SKirill A. Shutemov { 2865a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 2866364c1eebSYang Shi struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 28679a982250SKirill A. Shutemov unsigned long flags; 28689a982250SKirill A. Shutemov LIST_HEAD(list), *pos, *next; 28699a982250SKirill A. Shutemov struct page *page; 28709a982250SKirill A. Shutemov int split = 0; 28719a982250SKirill A. Shutemov 287287eaceb3SYang Shi #ifdef CONFIG_MEMCG 287387eaceb3SYang Shi if (sc->memcg) 287487eaceb3SYang Shi ds_queue = &sc->memcg->deferred_split_queue; 287587eaceb3SYang Shi #endif 287687eaceb3SYang Shi 2877364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 28789a982250SKirill A. Shutemov /* Take pin on all head pages to avoid freeing them under us */ 2879364c1eebSYang Shi list_for_each_safe(pos, next, &ds_queue->split_queue) { 28809a982250SKirill A. Shutemov page = list_entry((void *)pos, struct page, mapping); 28819a982250SKirill A. Shutemov page = compound_head(page); 2882e3ae1953SKirill A. Shutemov if (get_page_unless_zero(page)) { 2883e3ae1953SKirill A. Shutemov list_move(page_deferred_list(page), &list); 2884e3ae1953SKirill A. Shutemov } else { 2885e3ae1953SKirill A. Shutemov /* We lost race with put_compound_page() */ 28869a982250SKirill A. Shutemov list_del_init(page_deferred_list(page)); 2887364c1eebSYang Shi ds_queue->split_queue_len--; 28889a982250SKirill A. Shutemov } 2889e3ae1953SKirill A. Shutemov if (!--sc->nr_to_scan) 2890e3ae1953SKirill A. Shutemov break; 28919a982250SKirill A. Shutemov } 2892364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 28939a982250SKirill A. Shutemov 28949a982250SKirill A. Shutemov list_for_each_safe(pos, next, &list) { 28959a982250SKirill A. Shutemov page = list_entry((void *)pos, struct page, mapping); 2896fa41b900SKirill A. Shutemov if (!trylock_page(page)) 2897fa41b900SKirill A. Shutemov goto next; 28989a982250SKirill A. Shutemov /* split_huge_page() removes page from list on success */ 28999a982250SKirill A. Shutemov if (!split_huge_page(page)) 29009a982250SKirill A. Shutemov split++; 29019a982250SKirill A. Shutemov unlock_page(page); 2902fa41b900SKirill A. Shutemov next: 29039a982250SKirill A. Shutemov put_page(page); 29049a982250SKirill A. Shutemov } 29059a982250SKirill A. Shutemov 2906364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2907364c1eebSYang Shi list_splice_tail(&list, &ds_queue->split_queue); 2908364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 29099a982250SKirill A. Shutemov 2910cb8d68ecSKirill A. Shutemov /* 2911cb8d68ecSKirill A. Shutemov * Stop shrinker if we didn't split any page, but the queue is empty. 2912cb8d68ecSKirill A. Shutemov * This can happen if pages were freed under us. 2913cb8d68ecSKirill A. Shutemov */ 2914364c1eebSYang Shi if (!split && list_empty(&ds_queue->split_queue)) 2915cb8d68ecSKirill A. Shutemov return SHRINK_STOP; 2916cb8d68ecSKirill A. Shutemov return split; 29179a982250SKirill A. Shutemov } 29189a982250SKirill A. Shutemov 29199a982250SKirill A. Shutemov static struct shrinker deferred_split_shrinker = { 29209a982250SKirill A. Shutemov .count_objects = deferred_split_count, 29219a982250SKirill A. Shutemov .scan_objects = deferred_split_scan, 29229a982250SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 292387eaceb3SYang Shi .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE | 292487eaceb3SYang Shi SHRINKER_NONSLAB, 29259a982250SKirill A. Shutemov }; 292649071d43SKirill A. Shutemov 292749071d43SKirill A. Shutemov #ifdef CONFIG_DEBUG_FS 2928fa6c0231SZi Yan static void split_huge_pages_all(void) 292949071d43SKirill A. Shutemov { 293049071d43SKirill A. Shutemov struct zone *zone; 293149071d43SKirill A. Shutemov struct page *page; 293249071d43SKirill A. Shutemov unsigned long pfn, max_zone_pfn; 293349071d43SKirill A. Shutemov unsigned long total = 0, split = 0; 293449071d43SKirill A. Shutemov 2935fa6c0231SZi Yan pr_debug("Split all THPs\n"); 293649071d43SKirill A. Shutemov for_each_populated_zone(zone) { 293749071d43SKirill A. Shutemov max_zone_pfn = zone_end_pfn(zone); 293849071d43SKirill A. Shutemov for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 293949071d43SKirill A. Shutemov if (!pfn_valid(pfn)) 294049071d43SKirill A. Shutemov continue; 294149071d43SKirill A. Shutemov 294249071d43SKirill A. Shutemov page = pfn_to_page(pfn); 294349071d43SKirill A. Shutemov if (!get_page_unless_zero(page)) 294449071d43SKirill A. Shutemov continue; 294549071d43SKirill A. Shutemov 294649071d43SKirill A. Shutemov if (zone != page_zone(page)) 294749071d43SKirill A. Shutemov goto next; 294849071d43SKirill A. Shutemov 2949baa355fdSKirill A. Shutemov if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 295049071d43SKirill A. Shutemov goto next; 295149071d43SKirill A. Shutemov 295249071d43SKirill A. Shutemov total++; 295349071d43SKirill A. Shutemov lock_page(page); 295449071d43SKirill A. Shutemov if (!split_huge_page(page)) 295549071d43SKirill A. Shutemov split++; 295649071d43SKirill A. Shutemov unlock_page(page); 295749071d43SKirill A. Shutemov next: 295849071d43SKirill A. Shutemov put_page(page); 2959fa6c0231SZi Yan cond_resched(); 296049071d43SKirill A. Shutemov } 296149071d43SKirill A. Shutemov } 296249071d43SKirill A. Shutemov 2963fa6c0231SZi Yan pr_debug("%lu of %lu THP split\n", split, total); 296449071d43SKirill A. Shutemov } 2965fa6c0231SZi Yan 2966fa6c0231SZi Yan static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma) 2967fa6c0231SZi Yan { 2968fa6c0231SZi Yan return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) || 2969fa6c0231SZi Yan is_vm_hugetlb_page(vma); 2970fa6c0231SZi Yan } 2971fa6c0231SZi Yan 2972fa6c0231SZi Yan static int split_huge_pages_pid(int pid, unsigned long vaddr_start, 2973fa6c0231SZi Yan unsigned long vaddr_end) 2974fa6c0231SZi Yan { 2975fa6c0231SZi Yan int ret = 0; 2976fa6c0231SZi Yan struct task_struct *task; 2977fa6c0231SZi Yan struct mm_struct *mm; 2978fa6c0231SZi Yan unsigned long total = 0, split = 0; 2979fa6c0231SZi Yan unsigned long addr; 2980fa6c0231SZi Yan 2981fa6c0231SZi Yan vaddr_start &= PAGE_MASK; 2982fa6c0231SZi Yan vaddr_end &= PAGE_MASK; 2983fa6c0231SZi Yan 2984fa6c0231SZi Yan /* Find the task_struct from pid */ 2985fa6c0231SZi Yan rcu_read_lock(); 2986fa6c0231SZi Yan task = find_task_by_vpid(pid); 2987fa6c0231SZi Yan if (!task) { 2988fa6c0231SZi Yan rcu_read_unlock(); 2989fa6c0231SZi Yan ret = -ESRCH; 2990fa6c0231SZi Yan goto out; 2991fa6c0231SZi Yan } 2992fa6c0231SZi Yan get_task_struct(task); 2993fa6c0231SZi Yan rcu_read_unlock(); 2994fa6c0231SZi Yan 2995fa6c0231SZi Yan /* Find the mm_struct */ 2996fa6c0231SZi Yan mm = get_task_mm(task); 2997fa6c0231SZi Yan put_task_struct(task); 2998fa6c0231SZi Yan 2999fa6c0231SZi Yan if (!mm) { 3000fa6c0231SZi Yan ret = -EINVAL; 3001fa6c0231SZi Yan goto out; 3002fa6c0231SZi Yan } 3003fa6c0231SZi Yan 3004fa6c0231SZi Yan pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n", 3005fa6c0231SZi Yan pid, vaddr_start, vaddr_end); 3006fa6c0231SZi Yan 3007fa6c0231SZi Yan mmap_read_lock(mm); 3008fa6c0231SZi Yan /* 3009fa6c0231SZi Yan * always increase addr by PAGE_SIZE, since we could have a PTE page 3010fa6c0231SZi Yan * table filled with PTE-mapped THPs, each of which is distinct. 3011fa6c0231SZi Yan */ 3012fa6c0231SZi Yan for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) { 3013fa6c0231SZi Yan struct vm_area_struct *vma = find_vma(mm, addr); 3014fa6c0231SZi Yan unsigned int follflags; 3015fa6c0231SZi Yan struct page *page; 3016fa6c0231SZi Yan 3017fa6c0231SZi Yan if (!vma || addr < vma->vm_start) 3018fa6c0231SZi Yan break; 3019fa6c0231SZi Yan 3020fa6c0231SZi Yan /* skip special VMA and hugetlb VMA */ 3021fa6c0231SZi Yan if (vma_not_suitable_for_thp_split(vma)) { 3022fa6c0231SZi Yan addr = vma->vm_end; 3023fa6c0231SZi Yan continue; 3024fa6c0231SZi Yan } 3025fa6c0231SZi Yan 3026fa6c0231SZi Yan /* FOLL_DUMP to ignore special (like zero) pages */ 3027fa6c0231SZi Yan follflags = FOLL_GET | FOLL_DUMP; 3028fa6c0231SZi Yan page = follow_page(vma, addr, follflags); 3029fa6c0231SZi Yan 3030fa6c0231SZi Yan if (IS_ERR(page)) 3031fa6c0231SZi Yan continue; 3032fa6c0231SZi Yan if (!page) 3033fa6c0231SZi Yan continue; 3034fa6c0231SZi Yan 3035fa6c0231SZi Yan if (!is_transparent_hugepage(page)) 3036fa6c0231SZi Yan goto next; 3037fa6c0231SZi Yan 3038fa6c0231SZi Yan total++; 3039fa6c0231SZi Yan if (!can_split_huge_page(compound_head(page), NULL)) 3040fa6c0231SZi Yan goto next; 3041fa6c0231SZi Yan 3042fa6c0231SZi Yan if (!trylock_page(page)) 3043fa6c0231SZi Yan goto next; 3044fa6c0231SZi Yan 3045fa6c0231SZi Yan if (!split_huge_page(page)) 3046fa6c0231SZi Yan split++; 3047fa6c0231SZi Yan 3048fa6c0231SZi Yan unlock_page(page); 3049fa6c0231SZi Yan next: 3050fa6c0231SZi Yan put_page(page); 3051fa6c0231SZi Yan cond_resched(); 3052fa6c0231SZi Yan } 3053fa6c0231SZi Yan mmap_read_unlock(mm); 3054fa6c0231SZi Yan mmput(mm); 3055fa6c0231SZi Yan 3056fa6c0231SZi Yan pr_debug("%lu of %lu THP split\n", split, total); 3057fa6c0231SZi Yan 3058fa6c0231SZi Yan out: 3059fa6c0231SZi Yan return ret; 3060fa6c0231SZi Yan } 3061fa6c0231SZi Yan 3062fbe37501SZi Yan static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start, 3063fbe37501SZi Yan pgoff_t off_end) 3064fbe37501SZi Yan { 3065fbe37501SZi Yan struct filename *file; 3066fbe37501SZi Yan struct file *candidate; 3067fbe37501SZi Yan struct address_space *mapping; 3068fbe37501SZi Yan int ret = -EINVAL; 3069fbe37501SZi Yan pgoff_t index; 3070fbe37501SZi Yan int nr_pages = 1; 3071fbe37501SZi Yan unsigned long total = 0, split = 0; 3072fbe37501SZi Yan 3073fbe37501SZi Yan file = getname_kernel(file_path); 3074fbe37501SZi Yan if (IS_ERR(file)) 3075fbe37501SZi Yan return ret; 3076fbe37501SZi Yan 3077fbe37501SZi Yan candidate = file_open_name(file, O_RDONLY, 0); 3078fbe37501SZi Yan if (IS_ERR(candidate)) 3079fbe37501SZi Yan goto out; 3080fbe37501SZi Yan 3081fbe37501SZi Yan pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n", 3082fbe37501SZi Yan file_path, off_start, off_end); 3083fbe37501SZi Yan 3084fbe37501SZi Yan mapping = candidate->f_mapping; 3085fbe37501SZi Yan 3086fbe37501SZi Yan for (index = off_start; index < off_end; index += nr_pages) { 3087fbe37501SZi Yan struct page *fpage = pagecache_get_page(mapping, index, 3088fbe37501SZi Yan FGP_ENTRY | FGP_HEAD, 0); 3089fbe37501SZi Yan 3090fbe37501SZi Yan nr_pages = 1; 3091fbe37501SZi Yan if (xa_is_value(fpage) || !fpage) 3092fbe37501SZi Yan continue; 3093fbe37501SZi Yan 3094fbe37501SZi Yan if (!is_transparent_hugepage(fpage)) 3095fbe37501SZi Yan goto next; 3096fbe37501SZi Yan 3097fbe37501SZi Yan total++; 3098fbe37501SZi Yan nr_pages = thp_nr_pages(fpage); 3099fbe37501SZi Yan 3100fbe37501SZi Yan if (!trylock_page(fpage)) 3101fbe37501SZi Yan goto next; 3102fbe37501SZi Yan 3103fbe37501SZi Yan if (!split_huge_page(fpage)) 3104fbe37501SZi Yan split++; 3105fbe37501SZi Yan 3106fbe37501SZi Yan unlock_page(fpage); 3107fbe37501SZi Yan next: 3108fbe37501SZi Yan put_page(fpage); 3109fbe37501SZi Yan cond_resched(); 3110fbe37501SZi Yan } 3111fbe37501SZi Yan 3112fbe37501SZi Yan filp_close(candidate, NULL); 3113fbe37501SZi Yan ret = 0; 3114fbe37501SZi Yan 3115fbe37501SZi Yan pr_debug("%lu of %lu file-backed THP split\n", split, total); 3116fbe37501SZi Yan out: 3117fbe37501SZi Yan putname(file); 3118fbe37501SZi Yan return ret; 3119fbe37501SZi Yan } 3120fbe37501SZi Yan 3121fa6c0231SZi Yan #define MAX_INPUT_BUF_SZ 255 3122fa6c0231SZi Yan 3123fa6c0231SZi Yan static ssize_t split_huge_pages_write(struct file *file, const char __user *buf, 3124fa6c0231SZi Yan size_t count, loff_t *ppops) 3125fa6c0231SZi Yan { 3126fa6c0231SZi Yan static DEFINE_MUTEX(split_debug_mutex); 3127fa6c0231SZi Yan ssize_t ret; 3128fbe37501SZi Yan /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */ 3129fbe37501SZi Yan char input_buf[MAX_INPUT_BUF_SZ]; 3130fa6c0231SZi Yan int pid; 3131fa6c0231SZi Yan unsigned long vaddr_start, vaddr_end; 3132fa6c0231SZi Yan 3133fa6c0231SZi Yan ret = mutex_lock_interruptible(&split_debug_mutex); 3134fa6c0231SZi Yan if (ret) 3135fa6c0231SZi Yan return ret; 3136fa6c0231SZi Yan 3137fa6c0231SZi Yan ret = -EFAULT; 3138fa6c0231SZi Yan 3139fa6c0231SZi Yan memset(input_buf, 0, MAX_INPUT_BUF_SZ); 3140fa6c0231SZi Yan if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ))) 3141fa6c0231SZi Yan goto out; 3142fa6c0231SZi Yan 3143fa6c0231SZi Yan input_buf[MAX_INPUT_BUF_SZ - 1] = '\0'; 3144fbe37501SZi Yan 3145fbe37501SZi Yan if (input_buf[0] == '/') { 3146fbe37501SZi Yan char *tok; 3147fbe37501SZi Yan char *buf = input_buf; 3148fbe37501SZi Yan char file_path[MAX_INPUT_BUF_SZ]; 3149fbe37501SZi Yan pgoff_t off_start = 0, off_end = 0; 3150fbe37501SZi Yan size_t input_len = strlen(input_buf); 3151fbe37501SZi Yan 3152fbe37501SZi Yan tok = strsep(&buf, ","); 3153fbe37501SZi Yan if (tok) { 3154fbe37501SZi Yan strncpy(file_path, tok, MAX_INPUT_BUF_SZ); 3155fbe37501SZi Yan } else { 3156fbe37501SZi Yan ret = -EINVAL; 3157fbe37501SZi Yan goto out; 3158fbe37501SZi Yan } 3159fbe37501SZi Yan 3160fbe37501SZi Yan ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end); 3161fbe37501SZi Yan if (ret != 2) { 3162fbe37501SZi Yan ret = -EINVAL; 3163fbe37501SZi Yan goto out; 3164fbe37501SZi Yan } 3165fbe37501SZi Yan ret = split_huge_pages_in_file(file_path, off_start, off_end); 3166fbe37501SZi Yan if (!ret) 3167fbe37501SZi Yan ret = input_len; 3168fbe37501SZi Yan 3169fbe37501SZi Yan goto out; 3170fbe37501SZi Yan } 3171fbe37501SZi Yan 3172fa6c0231SZi Yan ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end); 3173fa6c0231SZi Yan if (ret == 1 && pid == 1) { 3174fa6c0231SZi Yan split_huge_pages_all(); 3175fa6c0231SZi Yan ret = strlen(input_buf); 3176fa6c0231SZi Yan goto out; 3177fa6c0231SZi Yan } else if (ret != 3) { 3178fa6c0231SZi Yan ret = -EINVAL; 3179fa6c0231SZi Yan goto out; 3180fa6c0231SZi Yan } 3181fa6c0231SZi Yan 3182fa6c0231SZi Yan ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end); 3183fa6c0231SZi Yan if (!ret) 3184fa6c0231SZi Yan ret = strlen(input_buf); 3185fa6c0231SZi Yan out: 3186fa6c0231SZi Yan mutex_unlock(&split_debug_mutex); 3187fa6c0231SZi Yan return ret; 3188fa6c0231SZi Yan 3189fa6c0231SZi Yan } 3190fa6c0231SZi Yan 3191fa6c0231SZi Yan static const struct file_operations split_huge_pages_fops = { 3192fa6c0231SZi Yan .owner = THIS_MODULE, 3193fa6c0231SZi Yan .write = split_huge_pages_write, 3194fa6c0231SZi Yan .llseek = no_llseek, 3195fa6c0231SZi Yan }; 319649071d43SKirill A. Shutemov 319749071d43SKirill A. Shutemov static int __init split_huge_pages_debugfs(void) 319849071d43SKirill A. Shutemov { 3199d9f7979cSGreg Kroah-Hartman debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 320049071d43SKirill A. Shutemov &split_huge_pages_fops); 320149071d43SKirill A. Shutemov return 0; 320249071d43SKirill A. Shutemov } 320349071d43SKirill A. Shutemov late_initcall(split_huge_pages_debugfs); 320449071d43SKirill A. Shutemov #endif 3205616b8371SZi Yan 3206616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 3207616b8371SZi Yan void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 3208616b8371SZi Yan struct page *page) 3209616b8371SZi Yan { 3210616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 3211616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 3212616b8371SZi Yan unsigned long address = pvmw->address; 3213616b8371SZi Yan pmd_t pmdval; 3214616b8371SZi Yan swp_entry_t entry; 3215ab6e3d09SNaoya Horiguchi pmd_t pmdswp; 3216616b8371SZi Yan 3217616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 3218616b8371SZi Yan return; 3219616b8371SZi Yan 3220616b8371SZi Yan flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 32218a8683adSHuang Ying pmdval = pmdp_invalidate(vma, address, pvmw->pmd); 3222616b8371SZi Yan if (pmd_dirty(pmdval)) 3223616b8371SZi Yan set_page_dirty(page); 3224616b8371SZi Yan entry = make_migration_entry(page, pmd_write(pmdval)); 3225ab6e3d09SNaoya Horiguchi pmdswp = swp_entry_to_pmd(entry); 3226ab6e3d09SNaoya Horiguchi if (pmd_soft_dirty(pmdval)) 3227ab6e3d09SNaoya Horiguchi pmdswp = pmd_swp_mksoft_dirty(pmdswp); 3228ab6e3d09SNaoya Horiguchi set_pmd_at(mm, address, pvmw->pmd, pmdswp); 3229616b8371SZi Yan page_remove_rmap(page, true); 3230616b8371SZi Yan put_page(page); 3231616b8371SZi Yan } 3232616b8371SZi Yan 3233616b8371SZi Yan void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 3234616b8371SZi Yan { 3235616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 3236616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 3237616b8371SZi Yan unsigned long address = pvmw->address; 3238616b8371SZi Yan unsigned long mmun_start = address & HPAGE_PMD_MASK; 3239616b8371SZi Yan pmd_t pmde; 3240616b8371SZi Yan swp_entry_t entry; 3241616b8371SZi Yan 3242616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 3243616b8371SZi Yan return; 3244616b8371SZi Yan 3245616b8371SZi Yan entry = pmd_to_swp_entry(*pvmw->pmd); 3246616b8371SZi Yan get_page(new); 3247616b8371SZi Yan pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 3248ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*pvmw->pmd)) 3249ab6e3d09SNaoya Horiguchi pmde = pmd_mksoft_dirty(pmde); 3250616b8371SZi Yan if (is_write_migration_entry(entry)) 3251f55e1014SLinus Torvalds pmde = maybe_pmd_mkwrite(pmde, vma); 3252616b8371SZi Yan 3253616b8371SZi Yan flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 3254e71769aeSNaoya Horiguchi if (PageAnon(new)) 3255616b8371SZi Yan page_add_anon_rmap(new, vma, mmun_start, true); 3256e71769aeSNaoya Horiguchi else 3257e71769aeSNaoya Horiguchi page_add_file_rmap(new, true); 3258616b8371SZi Yan set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 3259e125fe40SKirill A. Shutemov if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) 3260616b8371SZi Yan mlock_vma_page(new); 3261616b8371SZi Yan update_mmu_cache_pmd(vma, address, pvmw->pmd); 3262616b8371SZi Yan } 3263616b8371SZi Yan #endif 3264