171e3aac0SAndrea Arcangeli /* 271e3aac0SAndrea Arcangeli * Copyright (C) 2009 Red Hat, Inc. 371e3aac0SAndrea Arcangeli * 471e3aac0SAndrea Arcangeli * This work is licensed under the terms of the GNU GPL, version 2. See 571e3aac0SAndrea Arcangeli * the COPYING file in the top-level directory. 671e3aac0SAndrea Arcangeli */ 771e3aac0SAndrea Arcangeli 8ae3a8c1cSAndrew Morton #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 9ae3a8c1cSAndrew Morton 1071e3aac0SAndrea Arcangeli #include <linux/mm.h> 1171e3aac0SAndrea Arcangeli #include <linux/sched.h> 12f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 136a3827d7SIngo Molnar #include <linux/sched/numa_balancing.h> 1471e3aac0SAndrea Arcangeli #include <linux/highmem.h> 1571e3aac0SAndrea Arcangeli #include <linux/hugetlb.h> 1671e3aac0SAndrea Arcangeli #include <linux/mmu_notifier.h> 1771e3aac0SAndrea Arcangeli #include <linux/rmap.h> 1871e3aac0SAndrea Arcangeli #include <linux/swap.h> 1997ae1749SKirill A. Shutemov #include <linux/shrinker.h> 20ba76149fSAndrea Arcangeli #include <linux/mm_inline.h> 21e9b61f19SKirill A. Shutemov #include <linux/swapops.h> 224897c765SMatthew Wilcox #include <linux/dax.h> 23ba76149fSAndrea Arcangeli #include <linux/khugepaged.h> 24878aee7dSAndrea Arcangeli #include <linux/freezer.h> 25f25748e3SDan Williams #include <linux/pfn_t.h> 26a664b2d8SAndrea Arcangeli #include <linux/mman.h> 273565fce3SDan Williams #include <linux/memremap.h> 28325adeb5SRalf Baechle #include <linux/pagemap.h> 2949071d43SKirill A. Shutemov #include <linux/debugfs.h> 304daae3b4SMel Gorman #include <linux/migrate.h> 3143b5fbbdSSasha Levin #include <linux/hashtable.h> 326b251fc9SAndrea Arcangeli #include <linux/userfaultfd_k.h> 3333c3fc71SVladimir Davydov #include <linux/page_idle.h> 34baa355fdSKirill A. Shutemov #include <linux/shmem_fs.h> 356b31d595SMichal Hocko #include <linux/oom.h> 3697ae1749SKirill A. Shutemov 3771e3aac0SAndrea Arcangeli #include <asm/tlb.h> 3871e3aac0SAndrea Arcangeli #include <asm/pgalloc.h> 3971e3aac0SAndrea Arcangeli #include "internal.h" 4071e3aac0SAndrea Arcangeli 41ba76149fSAndrea Arcangeli /* 428bfa3f9aSJianguo Wu * By default transparent hugepage support is disabled in order that avoid 438bfa3f9aSJianguo Wu * to risk increase the memory footprint of applications without a guaranteed 448bfa3f9aSJianguo Wu * benefit. When transparent hugepage support is enabled, is for all mappings, 458bfa3f9aSJianguo Wu * and khugepaged scans all mappings. 468bfa3f9aSJianguo Wu * Defrag is invoked by khugepaged hugepage allocations and by page faults 478bfa3f9aSJianguo Wu * for all hugepage allocations. 48ba76149fSAndrea Arcangeli */ 4971e3aac0SAndrea Arcangeli unsigned long transparent_hugepage_flags __read_mostly = 5013ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 51ba76149fSAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_FLAG)| 5213ece886SAndrea Arcangeli #endif 5313ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 5413ece886SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 5513ece886SAndrea Arcangeli #endif 56444eb2a4SMel Gorman (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 5779da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 5879da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 59ba76149fSAndrea Arcangeli 609a982250SKirill A. Shutemov static struct shrinker deferred_split_shrinker; 61f000565aSAndrea Arcangeli 6297ae1749SKirill A. Shutemov static atomic_t huge_zero_refcount; 6356873f43SWang, Yalin struct page *huge_zero_page __read_mostly; 644a6c1297SKirill A. Shutemov 656fcb52a5SAaron Lu static struct page *get_huge_zero_page(void) 6697ae1749SKirill A. Shutemov { 6797ae1749SKirill A. Shutemov struct page *zero_page; 6897ae1749SKirill A. Shutemov retry: 6997ae1749SKirill A. Shutemov if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 704db0c3c2SJason Low return READ_ONCE(huge_zero_page); 7197ae1749SKirill A. Shutemov 7297ae1749SKirill A. Shutemov zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 7397ae1749SKirill A. Shutemov HPAGE_PMD_ORDER); 74d8a8e1f0SKirill A. Shutemov if (!zero_page) { 75d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 765918d10aSKirill A. Shutemov return NULL; 77d8a8e1f0SKirill A. Shutemov } 78d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC); 7997ae1749SKirill A. Shutemov preempt_disable(); 805918d10aSKirill A. Shutemov if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 8197ae1749SKirill A. Shutemov preempt_enable(); 825ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 8397ae1749SKirill A. Shutemov goto retry; 8497ae1749SKirill A. Shutemov } 8597ae1749SKirill A. Shutemov 8697ae1749SKirill A. Shutemov /* We take additional reference here. It will be put back by shrinker */ 8797ae1749SKirill A. Shutemov atomic_set(&huge_zero_refcount, 2); 8897ae1749SKirill A. Shutemov preempt_enable(); 894db0c3c2SJason Low return READ_ONCE(huge_zero_page); 9097ae1749SKirill A. Shutemov } 9197ae1749SKirill A. Shutemov 926fcb52a5SAaron Lu static void put_huge_zero_page(void) 9397ae1749SKirill A. Shutemov { 9497ae1749SKirill A. Shutemov /* 9597ae1749SKirill A. Shutemov * Counter should never go to zero here. Only shrinker can put 9697ae1749SKirill A. Shutemov * last reference. 9797ae1749SKirill A. Shutemov */ 9897ae1749SKirill A. Shutemov BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 9997ae1749SKirill A. Shutemov } 10097ae1749SKirill A. Shutemov 1016fcb52a5SAaron Lu struct page *mm_get_huge_zero_page(struct mm_struct *mm) 1026fcb52a5SAaron Lu { 1036fcb52a5SAaron Lu if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 1046fcb52a5SAaron Lu return READ_ONCE(huge_zero_page); 1056fcb52a5SAaron Lu 1066fcb52a5SAaron Lu if (!get_huge_zero_page()) 1076fcb52a5SAaron Lu return NULL; 1086fcb52a5SAaron Lu 1096fcb52a5SAaron Lu if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 1106fcb52a5SAaron Lu put_huge_zero_page(); 1116fcb52a5SAaron Lu 1126fcb52a5SAaron Lu return READ_ONCE(huge_zero_page); 1136fcb52a5SAaron Lu } 1146fcb52a5SAaron Lu 1156fcb52a5SAaron Lu void mm_put_huge_zero_page(struct mm_struct *mm) 1166fcb52a5SAaron Lu { 1176fcb52a5SAaron Lu if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 1186fcb52a5SAaron Lu put_huge_zero_page(); 1196fcb52a5SAaron Lu } 1206fcb52a5SAaron Lu 12148896466SGlauber Costa static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 12297ae1749SKirill A. Shutemov struct shrink_control *sc) 12397ae1749SKirill A. Shutemov { 12497ae1749SKirill A. Shutemov /* we can free zero page only if last reference remains */ 12597ae1749SKirill A. Shutemov return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 12648896466SGlauber Costa } 12797ae1749SKirill A. Shutemov 12848896466SGlauber Costa static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 12948896466SGlauber Costa struct shrink_control *sc) 13048896466SGlauber Costa { 13197ae1749SKirill A. Shutemov if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 1325918d10aSKirill A. Shutemov struct page *zero_page = xchg(&huge_zero_page, NULL); 1335918d10aSKirill A. Shutemov BUG_ON(zero_page == NULL); 1345ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 13548896466SGlauber Costa return HPAGE_PMD_NR; 13697ae1749SKirill A. Shutemov } 13797ae1749SKirill A. Shutemov 13897ae1749SKirill A. Shutemov return 0; 13997ae1749SKirill A. Shutemov } 14097ae1749SKirill A. Shutemov 14197ae1749SKirill A. Shutemov static struct shrinker huge_zero_page_shrinker = { 14248896466SGlauber Costa .count_objects = shrink_huge_zero_page_count, 14348896466SGlauber Costa .scan_objects = shrink_huge_zero_page_scan, 14497ae1749SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 14597ae1749SKirill A. Shutemov }; 14697ae1749SKirill A. Shutemov 14771e3aac0SAndrea Arcangeli #ifdef CONFIG_SYSFS 14871e3aac0SAndrea Arcangeli static ssize_t enabled_show(struct kobject *kobj, 14971e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 15071e3aac0SAndrea Arcangeli { 151444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 152444eb2a4SMel Gorman return sprintf(buf, "[always] madvise never\n"); 153444eb2a4SMel Gorman else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) 154444eb2a4SMel Gorman return sprintf(buf, "always [madvise] never\n"); 155444eb2a4SMel Gorman else 156444eb2a4SMel Gorman return sprintf(buf, "always madvise [never]\n"); 15771e3aac0SAndrea Arcangeli } 158444eb2a4SMel Gorman 15971e3aac0SAndrea Arcangeli static ssize_t enabled_store(struct kobject *kobj, 16071e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 16171e3aac0SAndrea Arcangeli const char *buf, size_t count) 16271e3aac0SAndrea Arcangeli { 16321440d7eSDavid Rientjes ssize_t ret = count; 164ba76149fSAndrea Arcangeli 16521440d7eSDavid Rientjes if (!memcmp("always", buf, 16621440d7eSDavid Rientjes min(sizeof("always")-1, count))) { 16721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 16821440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 16921440d7eSDavid Rientjes } else if (!memcmp("madvise", buf, 17021440d7eSDavid Rientjes min(sizeof("madvise")-1, count))) { 17121440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 17221440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 17321440d7eSDavid Rientjes } else if (!memcmp("never", buf, 17421440d7eSDavid Rientjes min(sizeof("never")-1, count))) { 17521440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 17621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 17721440d7eSDavid Rientjes } else 17821440d7eSDavid Rientjes ret = -EINVAL; 179ba76149fSAndrea Arcangeli 180ba76149fSAndrea Arcangeli if (ret > 0) { 181b46e756fSKirill A. Shutemov int err = start_stop_khugepaged(); 182ba76149fSAndrea Arcangeli if (err) 183ba76149fSAndrea Arcangeli ret = err; 184ba76149fSAndrea Arcangeli } 185ba76149fSAndrea Arcangeli return ret; 18671e3aac0SAndrea Arcangeli } 18771e3aac0SAndrea Arcangeli static struct kobj_attribute enabled_attr = 18871e3aac0SAndrea Arcangeli __ATTR(enabled, 0644, enabled_show, enabled_store); 18971e3aac0SAndrea Arcangeli 190b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_show(struct kobject *kobj, 19171e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 19271e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 19371e3aac0SAndrea Arcangeli { 194e27e6151SBen Hutchings return sprintf(buf, "%d\n", 195e27e6151SBen Hutchings !!test_bit(flag, &transparent_hugepage_flags)); 19671e3aac0SAndrea Arcangeli } 197e27e6151SBen Hutchings 198b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_store(struct kobject *kobj, 19971e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 20071e3aac0SAndrea Arcangeli const char *buf, size_t count, 20171e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 20271e3aac0SAndrea Arcangeli { 203e27e6151SBen Hutchings unsigned long value; 204e27e6151SBen Hutchings int ret; 205e27e6151SBen Hutchings 206e27e6151SBen Hutchings ret = kstrtoul(buf, 10, &value); 207e27e6151SBen Hutchings if (ret < 0) 208e27e6151SBen Hutchings return ret; 209e27e6151SBen Hutchings if (value > 1) 21071e3aac0SAndrea Arcangeli return -EINVAL; 21171e3aac0SAndrea Arcangeli 212e27e6151SBen Hutchings if (value) 213e27e6151SBen Hutchings set_bit(flag, &transparent_hugepage_flags); 214e27e6151SBen Hutchings else 215e27e6151SBen Hutchings clear_bit(flag, &transparent_hugepage_flags); 216e27e6151SBen Hutchings 21771e3aac0SAndrea Arcangeli return count; 21871e3aac0SAndrea Arcangeli } 21971e3aac0SAndrea Arcangeli 22071e3aac0SAndrea Arcangeli static ssize_t defrag_show(struct kobject *kobj, 22171e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 22271e3aac0SAndrea Arcangeli { 223444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 22421440d7eSDavid Rientjes return sprintf(buf, "[always] defer defer+madvise madvise never\n"); 225444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 22621440d7eSDavid Rientjes return sprintf(buf, "always [defer] defer+madvise madvise never\n"); 22721440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 22821440d7eSDavid Rientjes return sprintf(buf, "always defer [defer+madvise] madvise never\n"); 22921440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 23021440d7eSDavid Rientjes return sprintf(buf, "always defer defer+madvise [madvise] never\n"); 23121440d7eSDavid Rientjes return sprintf(buf, "always defer defer+madvise madvise [never]\n"); 23271e3aac0SAndrea Arcangeli } 23321440d7eSDavid Rientjes 23471e3aac0SAndrea Arcangeli static ssize_t defrag_store(struct kobject *kobj, 23571e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 23671e3aac0SAndrea Arcangeli const char *buf, size_t count) 23771e3aac0SAndrea Arcangeli { 23821440d7eSDavid Rientjes if (!memcmp("always", buf, 23921440d7eSDavid Rientjes min(sizeof("always")-1, count))) { 24021440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 24121440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 24221440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 24321440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 24421440d7eSDavid Rientjes } else if (!memcmp("defer+madvise", buf, 24521440d7eSDavid Rientjes min(sizeof("defer+madvise")-1, count))) { 24621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 24721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 24821440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 24921440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 2504fad7fb6SDavid Rientjes } else if (!memcmp("defer", buf, 2514fad7fb6SDavid Rientjes min(sizeof("defer")-1, count))) { 2524fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 2534fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 2544fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 2554fad7fb6SDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 25621440d7eSDavid Rientjes } else if (!memcmp("madvise", buf, 25721440d7eSDavid Rientjes min(sizeof("madvise")-1, count))) { 25821440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 25921440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 26021440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 26121440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 26221440d7eSDavid Rientjes } else if (!memcmp("never", buf, 26321440d7eSDavid Rientjes min(sizeof("never")-1, count))) { 26421440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 26521440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 26621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 26721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 26821440d7eSDavid Rientjes } else 26921440d7eSDavid Rientjes return -EINVAL; 27021440d7eSDavid Rientjes 27121440d7eSDavid Rientjes return count; 27271e3aac0SAndrea Arcangeli } 27371e3aac0SAndrea Arcangeli static struct kobj_attribute defrag_attr = 27471e3aac0SAndrea Arcangeli __ATTR(defrag, 0644, defrag_show, defrag_store); 27571e3aac0SAndrea Arcangeli 27679da5407SKirill A. Shutemov static ssize_t use_zero_page_show(struct kobject *kobj, 27779da5407SKirill A. Shutemov struct kobj_attribute *attr, char *buf) 27879da5407SKirill A. Shutemov { 279b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 28079da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 28179da5407SKirill A. Shutemov } 28279da5407SKirill A. Shutemov static ssize_t use_zero_page_store(struct kobject *kobj, 28379da5407SKirill A. Shutemov struct kobj_attribute *attr, const char *buf, size_t count) 28479da5407SKirill A. Shutemov { 285b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 28679da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 28779da5407SKirill A. Shutemov } 28879da5407SKirill A. Shutemov static struct kobj_attribute use_zero_page_attr = 28979da5407SKirill A. Shutemov __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 29049920d28SHugh Dickins 29149920d28SHugh Dickins static ssize_t hpage_pmd_size_show(struct kobject *kobj, 29249920d28SHugh Dickins struct kobj_attribute *attr, char *buf) 29349920d28SHugh Dickins { 29449920d28SHugh Dickins return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); 29549920d28SHugh Dickins } 29649920d28SHugh Dickins static struct kobj_attribute hpage_pmd_size_attr = 29749920d28SHugh Dickins __ATTR_RO(hpage_pmd_size); 29849920d28SHugh Dickins 29971e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 30071e3aac0SAndrea Arcangeli static ssize_t debug_cow_show(struct kobject *kobj, 30171e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 30271e3aac0SAndrea Arcangeli { 303b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 30471e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 30571e3aac0SAndrea Arcangeli } 30671e3aac0SAndrea Arcangeli static ssize_t debug_cow_store(struct kobject *kobj, 30771e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 30871e3aac0SAndrea Arcangeli const char *buf, size_t count) 30971e3aac0SAndrea Arcangeli { 310b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 31171e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 31271e3aac0SAndrea Arcangeli } 31371e3aac0SAndrea Arcangeli static struct kobj_attribute debug_cow_attr = 31471e3aac0SAndrea Arcangeli __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 31571e3aac0SAndrea Arcangeli #endif /* CONFIG_DEBUG_VM */ 31671e3aac0SAndrea Arcangeli 31771e3aac0SAndrea Arcangeli static struct attribute *hugepage_attr[] = { 31871e3aac0SAndrea Arcangeli &enabled_attr.attr, 31971e3aac0SAndrea Arcangeli &defrag_attr.attr, 32079da5407SKirill A. Shutemov &use_zero_page_attr.attr, 32149920d28SHugh Dickins &hpage_pmd_size_attr.attr, 322e496cf3dSKirill A. Shutemov #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) 3235a6e75f8SKirill A. Shutemov &shmem_enabled_attr.attr, 3245a6e75f8SKirill A. Shutemov #endif 32571e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 32671e3aac0SAndrea Arcangeli &debug_cow_attr.attr, 32771e3aac0SAndrea Arcangeli #endif 32871e3aac0SAndrea Arcangeli NULL, 32971e3aac0SAndrea Arcangeli }; 33071e3aac0SAndrea Arcangeli 3318aa95a21SArvind Yadav static const struct attribute_group hugepage_attr_group = { 33271e3aac0SAndrea Arcangeli .attrs = hugepage_attr, 333ba76149fSAndrea Arcangeli }; 334ba76149fSAndrea Arcangeli 335569e5590SShaohua Li static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 336569e5590SShaohua Li { 337569e5590SShaohua Li int err; 338569e5590SShaohua Li 339569e5590SShaohua Li *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 340569e5590SShaohua Li if (unlikely(!*hugepage_kobj)) { 341ae3a8c1cSAndrew Morton pr_err("failed to create transparent hugepage kobject\n"); 342569e5590SShaohua Li return -ENOMEM; 343569e5590SShaohua Li } 344569e5590SShaohua Li 345569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 346569e5590SShaohua Li if (err) { 347ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 348569e5590SShaohua Li goto delete_obj; 349569e5590SShaohua Li } 350569e5590SShaohua Li 351569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 352569e5590SShaohua Li if (err) { 353ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 354569e5590SShaohua Li goto remove_hp_group; 355569e5590SShaohua Li } 356569e5590SShaohua Li 357569e5590SShaohua Li return 0; 358569e5590SShaohua Li 359569e5590SShaohua Li remove_hp_group: 360569e5590SShaohua Li sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 361569e5590SShaohua Li delete_obj: 362569e5590SShaohua Li kobject_put(*hugepage_kobj); 363569e5590SShaohua Li return err; 364569e5590SShaohua Li } 365569e5590SShaohua Li 366569e5590SShaohua Li static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 367569e5590SShaohua Li { 368569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 369569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 370569e5590SShaohua Li kobject_put(hugepage_kobj); 371569e5590SShaohua Li } 372569e5590SShaohua Li #else 373569e5590SShaohua Li static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 374569e5590SShaohua Li { 375569e5590SShaohua Li return 0; 376569e5590SShaohua Li } 377569e5590SShaohua Li 378569e5590SShaohua Li static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 379569e5590SShaohua Li { 380569e5590SShaohua Li } 38171e3aac0SAndrea Arcangeli #endif /* CONFIG_SYSFS */ 38271e3aac0SAndrea Arcangeli 38371e3aac0SAndrea Arcangeli static int __init hugepage_init(void) 38471e3aac0SAndrea Arcangeli { 38571e3aac0SAndrea Arcangeli int err; 386569e5590SShaohua Li struct kobject *hugepage_kobj; 38771e3aac0SAndrea Arcangeli 3884b7167b9SAndrea Arcangeli if (!has_transparent_hugepage()) { 3894b7167b9SAndrea Arcangeli transparent_hugepage_flags = 0; 390569e5590SShaohua Li return -EINVAL; 3914b7167b9SAndrea Arcangeli } 3924b7167b9SAndrea Arcangeli 393ff20c2e0SKirill A. Shutemov /* 394ff20c2e0SKirill A. Shutemov * hugepages can't be allocated by the buddy allocator 395ff20c2e0SKirill A. Shutemov */ 396ff20c2e0SKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 397ff20c2e0SKirill A. Shutemov /* 398ff20c2e0SKirill A. Shutemov * we use page->mapping and page->index in second tail page 399ff20c2e0SKirill A. Shutemov * as list_head: assuming THP order >= 2 400ff20c2e0SKirill A. Shutemov */ 401ff20c2e0SKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 402ff20c2e0SKirill A. Shutemov 403569e5590SShaohua Li err = hugepage_init_sysfs(&hugepage_kobj); 404569e5590SShaohua Li if (err) 40565ebb64fSKirill A. Shutemov goto err_sysfs; 406ba76149fSAndrea Arcangeli 407b46e756fSKirill A. Shutemov err = khugepaged_init(); 408ba76149fSAndrea Arcangeli if (err) 40965ebb64fSKirill A. Shutemov goto err_slab; 410ba76149fSAndrea Arcangeli 41165ebb64fSKirill A. Shutemov err = register_shrinker(&huge_zero_page_shrinker); 41265ebb64fSKirill A. Shutemov if (err) 41365ebb64fSKirill A. Shutemov goto err_hzp_shrinker; 4149a982250SKirill A. Shutemov err = register_shrinker(&deferred_split_shrinker); 4159a982250SKirill A. Shutemov if (err) 4169a982250SKirill A. Shutemov goto err_split_shrinker; 41797ae1749SKirill A. Shutemov 41897562cd2SRik van Riel /* 41997562cd2SRik van Riel * By default disable transparent hugepages on smaller systems, 42097562cd2SRik van Riel * where the extra memory used could hurt more than TLB overhead 42197562cd2SRik van Riel * is likely to save. The admin can still enable it through /sys. 42297562cd2SRik van Riel */ 42379553da2SKirill A. Shutemov if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { 42497562cd2SRik van Riel transparent_hugepage_flags = 0; 42579553da2SKirill A. Shutemov return 0; 42679553da2SKirill A. Shutemov } 42797562cd2SRik van Riel 42879553da2SKirill A. Shutemov err = start_stop_khugepaged(); 42965ebb64fSKirill A. Shutemov if (err) 43065ebb64fSKirill A. Shutemov goto err_khugepaged; 431ba76149fSAndrea Arcangeli 432569e5590SShaohua Li return 0; 43365ebb64fSKirill A. Shutemov err_khugepaged: 4349a982250SKirill A. Shutemov unregister_shrinker(&deferred_split_shrinker); 4359a982250SKirill A. Shutemov err_split_shrinker: 43665ebb64fSKirill A. Shutemov unregister_shrinker(&huge_zero_page_shrinker); 43765ebb64fSKirill A. Shutemov err_hzp_shrinker: 438b46e756fSKirill A. Shutemov khugepaged_destroy(); 43965ebb64fSKirill A. Shutemov err_slab: 440569e5590SShaohua Li hugepage_exit_sysfs(hugepage_kobj); 44165ebb64fSKirill A. Shutemov err_sysfs: 442ba76149fSAndrea Arcangeli return err; 44371e3aac0SAndrea Arcangeli } 444a64fb3cdSPaul Gortmaker subsys_initcall(hugepage_init); 44571e3aac0SAndrea Arcangeli 44671e3aac0SAndrea Arcangeli static int __init setup_transparent_hugepage(char *str) 44771e3aac0SAndrea Arcangeli { 44871e3aac0SAndrea Arcangeli int ret = 0; 44971e3aac0SAndrea Arcangeli if (!str) 45071e3aac0SAndrea Arcangeli goto out; 45171e3aac0SAndrea Arcangeli if (!strcmp(str, "always")) { 45271e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_FLAG, 45371e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 45471e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 45571e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 45671e3aac0SAndrea Arcangeli ret = 1; 45771e3aac0SAndrea Arcangeli } else if (!strcmp(str, "madvise")) { 45871e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 45971e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 46071e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 46171e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 46271e3aac0SAndrea Arcangeli ret = 1; 46371e3aac0SAndrea Arcangeli } else if (!strcmp(str, "never")) { 46471e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 46571e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 46671e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 46771e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 46871e3aac0SAndrea Arcangeli ret = 1; 46971e3aac0SAndrea Arcangeli } 47071e3aac0SAndrea Arcangeli out: 47171e3aac0SAndrea Arcangeli if (!ret) 472ae3a8c1cSAndrew Morton pr_warn("transparent_hugepage= cannot parse, ignored\n"); 47371e3aac0SAndrea Arcangeli return ret; 47471e3aac0SAndrea Arcangeli } 47571e3aac0SAndrea Arcangeli __setup("transparent_hugepage=", setup_transparent_hugepage); 47671e3aac0SAndrea Arcangeli 477b32967ffSMel Gorman pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 47871e3aac0SAndrea Arcangeli { 47971e3aac0SAndrea Arcangeli if (likely(vma->vm_flags & VM_WRITE)) 48071e3aac0SAndrea Arcangeli pmd = pmd_mkwrite(pmd); 48171e3aac0SAndrea Arcangeli return pmd; 48271e3aac0SAndrea Arcangeli } 48371e3aac0SAndrea Arcangeli 4849a982250SKirill A. Shutemov static inline struct list_head *page_deferred_list(struct page *page) 4859a982250SKirill A. Shutemov { 4869a982250SKirill A. Shutemov /* 4879a982250SKirill A. Shutemov * ->lru in the tail pages is occupied by compound_head. 4889a982250SKirill A. Shutemov * Let's use ->mapping + ->index in the second tail page as list_head. 4899a982250SKirill A. Shutemov */ 4909a982250SKirill A. Shutemov return (struct list_head *)&page[2].mapping; 4919a982250SKirill A. Shutemov } 4929a982250SKirill A. Shutemov 4939a982250SKirill A. Shutemov void prep_transhuge_page(struct page *page) 4949a982250SKirill A. Shutemov { 4959a982250SKirill A. Shutemov /* 4969a982250SKirill A. Shutemov * we use page->mapping and page->indexlru in second tail page 4979a982250SKirill A. Shutemov * as list_head: assuming THP order >= 2 4989a982250SKirill A. Shutemov */ 4999a982250SKirill A. Shutemov 5009a982250SKirill A. Shutemov INIT_LIST_HEAD(page_deferred_list(page)); 5019a982250SKirill A. Shutemov set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 5029a982250SKirill A. Shutemov } 5039a982250SKirill A. Shutemov 50474d2fad1SToshi Kani unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len, 50574d2fad1SToshi Kani loff_t off, unsigned long flags, unsigned long size) 50674d2fad1SToshi Kani { 50774d2fad1SToshi Kani unsigned long addr; 50874d2fad1SToshi Kani loff_t off_end = off + len; 50974d2fad1SToshi Kani loff_t off_align = round_up(off, size); 51074d2fad1SToshi Kani unsigned long len_pad; 51174d2fad1SToshi Kani 51274d2fad1SToshi Kani if (off_end <= off_align || (off_end - off_align) < size) 51374d2fad1SToshi Kani return 0; 51474d2fad1SToshi Kani 51574d2fad1SToshi Kani len_pad = len + size; 51674d2fad1SToshi Kani if (len_pad < len || (off + len_pad) < off) 51774d2fad1SToshi Kani return 0; 51874d2fad1SToshi Kani 51974d2fad1SToshi Kani addr = current->mm->get_unmapped_area(filp, 0, len_pad, 52074d2fad1SToshi Kani off >> PAGE_SHIFT, flags); 52174d2fad1SToshi Kani if (IS_ERR_VALUE(addr)) 52274d2fad1SToshi Kani return 0; 52374d2fad1SToshi Kani 52474d2fad1SToshi Kani addr += (off - addr) & (size - 1); 52574d2fad1SToshi Kani return addr; 52674d2fad1SToshi Kani } 52774d2fad1SToshi Kani 52874d2fad1SToshi Kani unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 52974d2fad1SToshi Kani unsigned long len, unsigned long pgoff, unsigned long flags) 53074d2fad1SToshi Kani { 53174d2fad1SToshi Kani loff_t off = (loff_t)pgoff << PAGE_SHIFT; 53274d2fad1SToshi Kani 53374d2fad1SToshi Kani if (addr) 53474d2fad1SToshi Kani goto out; 53574d2fad1SToshi Kani if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) 53674d2fad1SToshi Kani goto out; 53774d2fad1SToshi Kani 53874d2fad1SToshi Kani addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE); 53974d2fad1SToshi Kani if (addr) 54074d2fad1SToshi Kani return addr; 54174d2fad1SToshi Kani 54274d2fad1SToshi Kani out: 54374d2fad1SToshi Kani return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 54474d2fad1SToshi Kani } 54574d2fad1SToshi Kani EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 54674d2fad1SToshi Kani 54782b0f8c3SJan Kara static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page, 548bae473a4SKirill A. Shutemov gfp_t gfp) 54971e3aac0SAndrea Arcangeli { 55082b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 55100501b53SJohannes Weiner struct mem_cgroup *memcg; 55271e3aac0SAndrea Arcangeli pgtable_t pgtable; 55382b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 5546b31d595SMichal Hocko int ret = 0; 55571e3aac0SAndrea Arcangeli 556309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page), page); 55700501b53SJohannes Weiner 558bae473a4SKirill A. Shutemov if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) { 5596b251fc9SAndrea Arcangeli put_page(page); 5606b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_FALLBACK); 5616b251fc9SAndrea Arcangeli return VM_FAULT_FALLBACK; 5626b251fc9SAndrea Arcangeli } 56371e3aac0SAndrea Arcangeli 564bae473a4SKirill A. Shutemov pgtable = pte_alloc_one(vma->vm_mm, haddr); 56500501b53SJohannes Weiner if (unlikely(!pgtable)) { 5666b31d595SMichal Hocko ret = VM_FAULT_OOM; 5676b31d595SMichal Hocko goto release; 56800501b53SJohannes Weiner } 56900501b53SJohannes Weiner 570c79b57e4SHuang Ying clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 57152f37629SMinchan Kim /* 57252f37629SMinchan Kim * The memory barrier inside __SetPageUptodate makes sure that 57352f37629SMinchan Kim * clear_huge_page writes become visible before the set_pmd_at() 57452f37629SMinchan Kim * write. 57552f37629SMinchan Kim */ 57671e3aac0SAndrea Arcangeli __SetPageUptodate(page); 57771e3aac0SAndrea Arcangeli 57882b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 57982b0f8c3SJan Kara if (unlikely(!pmd_none(*vmf->pmd))) { 5806b31d595SMichal Hocko goto unlock_release; 58171e3aac0SAndrea Arcangeli } else { 58271e3aac0SAndrea Arcangeli pmd_t entry; 5836b251fc9SAndrea Arcangeli 5846b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 5856b31d595SMichal Hocko if (ret) 5866b31d595SMichal Hocko goto unlock_release; 5876b31d595SMichal Hocko 5886b251fc9SAndrea Arcangeli /* Deliver the page fault to userland */ 5896b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 5906b251fc9SAndrea Arcangeli int ret; 5916b251fc9SAndrea Arcangeli 59282b0f8c3SJan Kara spin_unlock(vmf->ptl); 593f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(page, memcg, true); 5946b251fc9SAndrea Arcangeli put_page(page); 595bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 59682b0f8c3SJan Kara ret = handle_userfault(vmf, VM_UFFD_MISSING); 5976b251fc9SAndrea Arcangeli VM_BUG_ON(ret & VM_FAULT_FALLBACK); 5986b251fc9SAndrea Arcangeli return ret; 5996b251fc9SAndrea Arcangeli } 6006b251fc9SAndrea Arcangeli 6013122359aSKirill A. Shutemov entry = mk_huge_pmd(page, vma->vm_page_prot); 6023122359aSKirill A. Shutemov entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 603d281ee61SKirill A. Shutemov page_add_new_anon_rmap(page, vma, haddr, true); 604f627c2f5SKirill A. Shutemov mem_cgroup_commit_charge(page, memcg, false, true); 60500501b53SJohannes Weiner lru_cache_add_active_or_unevictable(page, vma); 60682b0f8c3SJan Kara pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 60782b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 608bae473a4SKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 609bae473a4SKirill A. Shutemov atomic_long_inc(&vma->vm_mm->nr_ptes); 61082b0f8c3SJan Kara spin_unlock(vmf->ptl); 6116b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_ALLOC); 61271e3aac0SAndrea Arcangeli } 61371e3aac0SAndrea Arcangeli 614aa2e878eSDavid Rientjes return 0; 6156b31d595SMichal Hocko unlock_release: 6166b31d595SMichal Hocko spin_unlock(vmf->ptl); 6176b31d595SMichal Hocko release: 6186b31d595SMichal Hocko if (pgtable) 6196b31d595SMichal Hocko pte_free(vma->vm_mm, pgtable); 6206b31d595SMichal Hocko mem_cgroup_cancel_charge(page, memcg, true); 6216b31d595SMichal Hocko put_page(page); 6226b31d595SMichal Hocko return ret; 6236b31d595SMichal Hocko 62471e3aac0SAndrea Arcangeli } 62571e3aac0SAndrea Arcangeli 626444eb2a4SMel Gorman /* 62721440d7eSDavid Rientjes * always: directly stall for all thp allocations 62821440d7eSDavid Rientjes * defer: wake kswapd and fail if not immediately available 62921440d7eSDavid Rientjes * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 63021440d7eSDavid Rientjes * fail if not immediately available 63121440d7eSDavid Rientjes * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 63221440d7eSDavid Rientjes * available 63321440d7eSDavid Rientjes * never: never stall for any thp allocation 634444eb2a4SMel Gorman */ 635444eb2a4SMel Gorman static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) 6360bbbc0b3SAndrea Arcangeli { 63721440d7eSDavid Rientjes const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); 638444eb2a4SMel Gorman 63921440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 64025160354SVlastimil Babka return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 64121440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 64221440d7eSDavid Rientjes return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 64321440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 64421440d7eSDavid Rientjes return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 64521440d7eSDavid Rientjes __GFP_KSWAPD_RECLAIM); 64621440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 64721440d7eSDavid Rientjes return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 64821440d7eSDavid Rientjes 0); 64925160354SVlastimil Babka return GFP_TRANSHUGE_LIGHT; 650444eb2a4SMel Gorman } 651444eb2a4SMel Gorman 652c4088ebdSKirill A. Shutemov /* Caller must hold page table lock. */ 653d295e341SKirill A. Shutemov static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 65497ae1749SKirill A. Shutemov struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 6555918d10aSKirill A. Shutemov struct page *zero_page) 656fc9fe822SKirill A. Shutemov { 657fc9fe822SKirill A. Shutemov pmd_t entry; 6587c414164SAndrew Morton if (!pmd_none(*pmd)) 6597c414164SAndrew Morton return false; 6605918d10aSKirill A. Shutemov entry = mk_pmd(zero_page, vma->vm_page_prot); 661fc9fe822SKirill A. Shutemov entry = pmd_mkhuge(entry); 66212c9d70bSMatthew Wilcox if (pgtable) 6636b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 664fc9fe822SKirill A. Shutemov set_pmd_at(mm, haddr, pmd, entry); 665e1f56c89SKirill A. Shutemov atomic_long_inc(&mm->nr_ptes); 6667c414164SAndrew Morton return true; 667fc9fe822SKirill A. Shutemov } 668fc9fe822SKirill A. Shutemov 66982b0f8c3SJan Kara int do_huge_pmd_anonymous_page(struct vm_fault *vmf) 67071e3aac0SAndrea Arcangeli { 67182b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 672077fcf11SAneesh Kumar K.V gfp_t gfp; 67371e3aac0SAndrea Arcangeli struct page *page; 67482b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 67571e3aac0SAndrea Arcangeli 676128ec037SKirill A. Shutemov if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 677c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 67871e3aac0SAndrea Arcangeli if (unlikely(anon_vma_prepare(vma))) 67971e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 6806d50e60cSDavid Rientjes if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 681ba76149fSAndrea Arcangeli return VM_FAULT_OOM; 68282b0f8c3SJan Kara if (!(vmf->flags & FAULT_FLAG_WRITE) && 683bae473a4SKirill A. Shutemov !mm_forbids_zeropage(vma->vm_mm) && 68479da5407SKirill A. Shutemov transparent_hugepage_use_zero_page()) { 68580371957SKirill A. Shutemov pgtable_t pgtable; 6865918d10aSKirill A. Shutemov struct page *zero_page; 6873ea41e62SKirill A. Shutemov bool set; 6886b251fc9SAndrea Arcangeli int ret; 689bae473a4SKirill A. Shutemov pgtable = pte_alloc_one(vma->vm_mm, haddr); 69080371957SKirill A. Shutemov if (unlikely(!pgtable)) 69180371957SKirill A. Shutemov return VM_FAULT_OOM; 6926fcb52a5SAaron Lu zero_page = mm_get_huge_zero_page(vma->vm_mm); 6935918d10aSKirill A. Shutemov if (unlikely(!zero_page)) { 694bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 69597ae1749SKirill A. Shutemov count_vm_event(THP_FAULT_FALLBACK); 696c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 69797ae1749SKirill A. Shutemov } 69882b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 6996b251fc9SAndrea Arcangeli ret = 0; 7006b251fc9SAndrea Arcangeli set = false; 70182b0f8c3SJan Kara if (pmd_none(*vmf->pmd)) { 7026b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 7036b31d595SMichal Hocko if (ret) { 7046b31d595SMichal Hocko spin_unlock(vmf->ptl); 7056b31d595SMichal Hocko } else if (userfaultfd_missing(vma)) { 70682b0f8c3SJan Kara spin_unlock(vmf->ptl); 70782b0f8c3SJan Kara ret = handle_userfault(vmf, VM_UFFD_MISSING); 7086b251fc9SAndrea Arcangeli VM_BUG_ON(ret & VM_FAULT_FALLBACK); 7096b251fc9SAndrea Arcangeli } else { 710bae473a4SKirill A. Shutemov set_huge_zero_page(pgtable, vma->vm_mm, vma, 71182b0f8c3SJan Kara haddr, vmf->pmd, zero_page); 71282b0f8c3SJan Kara spin_unlock(vmf->ptl); 7136b251fc9SAndrea Arcangeli set = true; 7146b251fc9SAndrea Arcangeli } 7156b251fc9SAndrea Arcangeli } else 71682b0f8c3SJan Kara spin_unlock(vmf->ptl); 7176fcb52a5SAaron Lu if (!set) 718bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 7196b251fc9SAndrea Arcangeli return ret; 72080371957SKirill A. Shutemov } 721444eb2a4SMel Gorman gfp = alloc_hugepage_direct_gfpmask(vma); 722077fcf11SAneesh Kumar K.V page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 72381ab4201SAndi Kleen if (unlikely(!page)) { 72481ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 725c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 72681ab4201SAndi Kleen } 7279a982250SKirill A. Shutemov prep_transhuge_page(page); 72882b0f8c3SJan Kara return __do_huge_pmd_anonymous_page(vmf, page, gfp); 72971e3aac0SAndrea Arcangeli } 73071e3aac0SAndrea Arcangeli 731ae18d6dcSMatthew Wilcox static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 7323b6521f5SOliver O'Halloran pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 7333b6521f5SOliver O'Halloran pgtable_t pgtable) 7345cad465dSMatthew Wilcox { 7355cad465dSMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 7365cad465dSMatthew Wilcox pmd_t entry; 7375cad465dSMatthew Wilcox spinlock_t *ptl; 7385cad465dSMatthew Wilcox 7395cad465dSMatthew Wilcox ptl = pmd_lock(mm, pmd); 740f25748e3SDan Williams entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 741f25748e3SDan Williams if (pfn_t_devmap(pfn)) 742f25748e3SDan Williams entry = pmd_mkdevmap(entry); 7435cad465dSMatthew Wilcox if (write) { 7445cad465dSMatthew Wilcox entry = pmd_mkyoung(pmd_mkdirty(entry)); 7455cad465dSMatthew Wilcox entry = maybe_pmd_mkwrite(entry, vma); 7465cad465dSMatthew Wilcox } 7473b6521f5SOliver O'Halloran 7483b6521f5SOliver O'Halloran if (pgtable) { 7493b6521f5SOliver O'Halloran pgtable_trans_huge_deposit(mm, pmd, pgtable); 7503b6521f5SOliver O'Halloran atomic_long_inc(&mm->nr_ptes); 7513b6521f5SOliver O'Halloran } 7523b6521f5SOliver O'Halloran 7535cad465dSMatthew Wilcox set_pmd_at(mm, addr, pmd, entry); 7545cad465dSMatthew Wilcox update_mmu_cache_pmd(vma, addr, pmd); 7555cad465dSMatthew Wilcox spin_unlock(ptl); 7565cad465dSMatthew Wilcox } 7575cad465dSMatthew Wilcox 7585cad465dSMatthew Wilcox int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 759f25748e3SDan Williams pmd_t *pmd, pfn_t pfn, bool write) 7605cad465dSMatthew Wilcox { 7615cad465dSMatthew Wilcox pgprot_t pgprot = vma->vm_page_prot; 7623b6521f5SOliver O'Halloran pgtable_t pgtable = NULL; 7635cad465dSMatthew Wilcox /* 7645cad465dSMatthew Wilcox * If we had pmd_special, we could avoid all these restrictions, 7655cad465dSMatthew Wilcox * but we need to be consistent with PTEs and architectures that 7665cad465dSMatthew Wilcox * can't support a 'special' bit. 7675cad465dSMatthew Wilcox */ 7685cad465dSMatthew Wilcox BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 7695cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 7705cad465dSMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 7715cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 772f25748e3SDan Williams BUG_ON(!pfn_t_devmap(pfn)); 7735cad465dSMatthew Wilcox 7745cad465dSMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 7755cad465dSMatthew Wilcox return VM_FAULT_SIGBUS; 776308a047cSBorislav Petkov 7773b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) { 7783b6521f5SOliver O'Halloran pgtable = pte_alloc_one(vma->vm_mm, addr); 7793b6521f5SOliver O'Halloran if (!pgtable) 7803b6521f5SOliver O'Halloran return VM_FAULT_OOM; 7813b6521f5SOliver O'Halloran } 7823b6521f5SOliver O'Halloran 783308a047cSBorislav Petkov track_pfn_insert(vma, &pgprot, pfn); 784308a047cSBorislav Petkov 7853b6521f5SOliver O'Halloran insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable); 786ae18d6dcSMatthew Wilcox return VM_FAULT_NOPAGE; 7875cad465dSMatthew Wilcox } 788dee41079SDan Williams EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 7895cad465dSMatthew Wilcox 790a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 791a00cc7d9SMatthew Wilcox static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 792a00cc7d9SMatthew Wilcox { 793a00cc7d9SMatthew Wilcox if (likely(vma->vm_flags & VM_WRITE)) 794a00cc7d9SMatthew Wilcox pud = pud_mkwrite(pud); 795a00cc7d9SMatthew Wilcox return pud; 796a00cc7d9SMatthew Wilcox } 797a00cc7d9SMatthew Wilcox 798a00cc7d9SMatthew Wilcox static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 799a00cc7d9SMatthew Wilcox pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 800a00cc7d9SMatthew Wilcox { 801a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 802a00cc7d9SMatthew Wilcox pud_t entry; 803a00cc7d9SMatthew Wilcox spinlock_t *ptl; 804a00cc7d9SMatthew Wilcox 805a00cc7d9SMatthew Wilcox ptl = pud_lock(mm, pud); 806a00cc7d9SMatthew Wilcox entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 807a00cc7d9SMatthew Wilcox if (pfn_t_devmap(pfn)) 808a00cc7d9SMatthew Wilcox entry = pud_mkdevmap(entry); 809a00cc7d9SMatthew Wilcox if (write) { 810a00cc7d9SMatthew Wilcox entry = pud_mkyoung(pud_mkdirty(entry)); 811a00cc7d9SMatthew Wilcox entry = maybe_pud_mkwrite(entry, vma); 812a00cc7d9SMatthew Wilcox } 813a00cc7d9SMatthew Wilcox set_pud_at(mm, addr, pud, entry); 814a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 815a00cc7d9SMatthew Wilcox spin_unlock(ptl); 816a00cc7d9SMatthew Wilcox } 817a00cc7d9SMatthew Wilcox 818a00cc7d9SMatthew Wilcox int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 819a00cc7d9SMatthew Wilcox pud_t *pud, pfn_t pfn, bool write) 820a00cc7d9SMatthew Wilcox { 821a00cc7d9SMatthew Wilcox pgprot_t pgprot = vma->vm_page_prot; 822a00cc7d9SMatthew Wilcox /* 823a00cc7d9SMatthew Wilcox * If we had pud_special, we could avoid all these restrictions, 824a00cc7d9SMatthew Wilcox * but we need to be consistent with PTEs and architectures that 825a00cc7d9SMatthew Wilcox * can't support a 'special' bit. 826a00cc7d9SMatthew Wilcox */ 827a00cc7d9SMatthew Wilcox BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 828a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 829a00cc7d9SMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 830a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 831a00cc7d9SMatthew Wilcox BUG_ON(!pfn_t_devmap(pfn)); 832a00cc7d9SMatthew Wilcox 833a00cc7d9SMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 834a00cc7d9SMatthew Wilcox return VM_FAULT_SIGBUS; 835a00cc7d9SMatthew Wilcox 836a00cc7d9SMatthew Wilcox track_pfn_insert(vma, &pgprot, pfn); 837a00cc7d9SMatthew Wilcox 838a00cc7d9SMatthew Wilcox insert_pfn_pud(vma, addr, pud, pfn, pgprot, write); 839a00cc7d9SMatthew Wilcox return VM_FAULT_NOPAGE; 840a00cc7d9SMatthew Wilcox } 841a00cc7d9SMatthew Wilcox EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 842a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 843a00cc7d9SMatthew Wilcox 8443565fce3SDan Williams static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 8453565fce3SDan Williams pmd_t *pmd) 8463565fce3SDan Williams { 8473565fce3SDan Williams pmd_t _pmd; 8483565fce3SDan Williams 8493565fce3SDan Williams /* 8503565fce3SDan Williams * We should set the dirty bit only for FOLL_WRITE but for now 8513565fce3SDan Williams * the dirty bit in the pmd is meaningless. And if the dirty 8523565fce3SDan Williams * bit will become meaningful and we'll only set it with 8533565fce3SDan Williams * FOLL_WRITE, an atomic set_bit will be required on the pmd to 8543565fce3SDan Williams * set the young bit, instead of the current set_pmd_at. 8553565fce3SDan Williams */ 8563565fce3SDan Williams _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 8573565fce3SDan Williams if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 8583565fce3SDan Williams pmd, _pmd, 1)) 8593565fce3SDan Williams update_mmu_cache_pmd(vma, addr, pmd); 8603565fce3SDan Williams } 8613565fce3SDan Williams 8623565fce3SDan Williams struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 8633565fce3SDan Williams pmd_t *pmd, int flags) 8643565fce3SDan Williams { 8653565fce3SDan Williams unsigned long pfn = pmd_pfn(*pmd); 8663565fce3SDan Williams struct mm_struct *mm = vma->vm_mm; 8673565fce3SDan Williams struct dev_pagemap *pgmap; 8683565fce3SDan Williams struct page *page; 8693565fce3SDan Williams 8703565fce3SDan Williams assert_spin_locked(pmd_lockptr(mm, pmd)); 8713565fce3SDan Williams 8728310d48bSKeno Fischer /* 8738310d48bSKeno Fischer * When we COW a devmap PMD entry, we split it into PTEs, so we should 8748310d48bSKeno Fischer * not be in this function with `flags & FOLL_COW` set. 8758310d48bSKeno Fischer */ 8768310d48bSKeno Fischer WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 8778310d48bSKeno Fischer 8783565fce3SDan Williams if (flags & FOLL_WRITE && !pmd_write(*pmd)) 8793565fce3SDan Williams return NULL; 8803565fce3SDan Williams 8813565fce3SDan Williams if (pmd_present(*pmd) && pmd_devmap(*pmd)) 8823565fce3SDan Williams /* pass */; 8833565fce3SDan Williams else 8843565fce3SDan Williams return NULL; 8853565fce3SDan Williams 8863565fce3SDan Williams if (flags & FOLL_TOUCH) 8873565fce3SDan Williams touch_pmd(vma, addr, pmd); 8883565fce3SDan Williams 8893565fce3SDan Williams /* 8903565fce3SDan Williams * device mapped pages can only be returned if the 8913565fce3SDan Williams * caller will manage the page reference count. 8923565fce3SDan Williams */ 8933565fce3SDan Williams if (!(flags & FOLL_GET)) 8943565fce3SDan Williams return ERR_PTR(-EEXIST); 8953565fce3SDan Williams 8963565fce3SDan Williams pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 8973565fce3SDan Williams pgmap = get_dev_pagemap(pfn, NULL); 8983565fce3SDan Williams if (!pgmap) 8993565fce3SDan Williams return ERR_PTR(-EFAULT); 9003565fce3SDan Williams page = pfn_to_page(pfn); 9013565fce3SDan Williams get_page(page); 9023565fce3SDan Williams put_dev_pagemap(pgmap); 9033565fce3SDan Williams 9043565fce3SDan Williams return page; 9053565fce3SDan Williams } 9063565fce3SDan Williams 90771e3aac0SAndrea Arcangeli int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 90871e3aac0SAndrea Arcangeli pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 90971e3aac0SAndrea Arcangeli struct vm_area_struct *vma) 91071e3aac0SAndrea Arcangeli { 911c4088ebdSKirill A. Shutemov spinlock_t *dst_ptl, *src_ptl; 91271e3aac0SAndrea Arcangeli struct page *src_page; 91371e3aac0SAndrea Arcangeli pmd_t pmd; 91412c9d70bSMatthew Wilcox pgtable_t pgtable = NULL; 915628d47ceSKirill A. Shutemov int ret = -ENOMEM; 91671e3aac0SAndrea Arcangeli 917628d47ceSKirill A. Shutemov /* Skip if can be re-fill on fault */ 918628d47ceSKirill A. Shutemov if (!vma_is_anonymous(vma)) 919628d47ceSKirill A. Shutemov return 0; 920628d47ceSKirill A. Shutemov 92171e3aac0SAndrea Arcangeli pgtable = pte_alloc_one(dst_mm, addr); 92271e3aac0SAndrea Arcangeli if (unlikely(!pgtable)) 92371e3aac0SAndrea Arcangeli goto out; 92471e3aac0SAndrea Arcangeli 925c4088ebdSKirill A. Shutemov dst_ptl = pmd_lock(dst_mm, dst_pmd); 926c4088ebdSKirill A. Shutemov src_ptl = pmd_lockptr(src_mm, src_pmd); 927c4088ebdSKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 92871e3aac0SAndrea Arcangeli 92971e3aac0SAndrea Arcangeli ret = -EAGAIN; 93071e3aac0SAndrea Arcangeli pmd = *src_pmd; 931628d47ceSKirill A. Shutemov if (unlikely(!pmd_trans_huge(pmd))) { 93271e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 93371e3aac0SAndrea Arcangeli goto out_unlock; 93471e3aac0SAndrea Arcangeli } 935fc9fe822SKirill A. Shutemov /* 936c4088ebdSKirill A. Shutemov * When page table lock is held, the huge zero pmd should not be 937fc9fe822SKirill A. Shutemov * under splitting since we don't split the page itself, only pmd to 938fc9fe822SKirill A. Shutemov * a page table. 939fc9fe822SKirill A. Shutemov */ 940fc9fe822SKirill A. Shutemov if (is_huge_zero_pmd(pmd)) { 9415918d10aSKirill A. Shutemov struct page *zero_page; 94297ae1749SKirill A. Shutemov /* 94397ae1749SKirill A. Shutemov * get_huge_zero_page() will never allocate a new page here, 94497ae1749SKirill A. Shutemov * since we already have a zero page to copy. It just takes a 94597ae1749SKirill A. Shutemov * reference. 94697ae1749SKirill A. Shutemov */ 9476fcb52a5SAaron Lu zero_page = mm_get_huge_zero_page(dst_mm); 9486b251fc9SAndrea Arcangeli set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 9495918d10aSKirill A. Shutemov zero_page); 950fc9fe822SKirill A. Shutemov ret = 0; 951fc9fe822SKirill A. Shutemov goto out_unlock; 952fc9fe822SKirill A. Shutemov } 953de466bd6SMel Gorman 95471e3aac0SAndrea Arcangeli src_page = pmd_page(pmd); 955309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 95671e3aac0SAndrea Arcangeli get_page(src_page); 95753f9263bSKirill A. Shutemov page_dup_rmap(src_page, true); 95871e3aac0SAndrea Arcangeli add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 9595c7fb56eSDan Williams atomic_long_inc(&dst_mm->nr_ptes); 9605c7fb56eSDan Williams pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 96171e3aac0SAndrea Arcangeli 96271e3aac0SAndrea Arcangeli pmdp_set_wrprotect(src_mm, addr, src_pmd); 96371e3aac0SAndrea Arcangeli pmd = pmd_mkold(pmd_wrprotect(pmd)); 96471e3aac0SAndrea Arcangeli set_pmd_at(dst_mm, addr, dst_pmd, pmd); 96571e3aac0SAndrea Arcangeli 96671e3aac0SAndrea Arcangeli ret = 0; 96771e3aac0SAndrea Arcangeli out_unlock: 968c4088ebdSKirill A. Shutemov spin_unlock(src_ptl); 969c4088ebdSKirill A. Shutemov spin_unlock(dst_ptl); 97071e3aac0SAndrea Arcangeli out: 97171e3aac0SAndrea Arcangeli return ret; 97271e3aac0SAndrea Arcangeli } 97371e3aac0SAndrea Arcangeli 974a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 975a00cc7d9SMatthew Wilcox static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 976a00cc7d9SMatthew Wilcox pud_t *pud) 977a00cc7d9SMatthew Wilcox { 978a00cc7d9SMatthew Wilcox pud_t _pud; 979a00cc7d9SMatthew Wilcox 980a00cc7d9SMatthew Wilcox /* 981a00cc7d9SMatthew Wilcox * We should set the dirty bit only for FOLL_WRITE but for now 982a00cc7d9SMatthew Wilcox * the dirty bit in the pud is meaningless. And if the dirty 983a00cc7d9SMatthew Wilcox * bit will become meaningful and we'll only set it with 984a00cc7d9SMatthew Wilcox * FOLL_WRITE, an atomic set_bit will be required on the pud to 985a00cc7d9SMatthew Wilcox * set the young bit, instead of the current set_pud_at. 986a00cc7d9SMatthew Wilcox */ 987a00cc7d9SMatthew Wilcox _pud = pud_mkyoung(pud_mkdirty(*pud)); 988a00cc7d9SMatthew Wilcox if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 989a00cc7d9SMatthew Wilcox pud, _pud, 1)) 990a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 991a00cc7d9SMatthew Wilcox } 992a00cc7d9SMatthew Wilcox 993a00cc7d9SMatthew Wilcox struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 994a00cc7d9SMatthew Wilcox pud_t *pud, int flags) 995a00cc7d9SMatthew Wilcox { 996a00cc7d9SMatthew Wilcox unsigned long pfn = pud_pfn(*pud); 997a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 998a00cc7d9SMatthew Wilcox struct dev_pagemap *pgmap; 999a00cc7d9SMatthew Wilcox struct page *page; 1000a00cc7d9SMatthew Wilcox 1001a00cc7d9SMatthew Wilcox assert_spin_locked(pud_lockptr(mm, pud)); 1002a00cc7d9SMatthew Wilcox 1003a00cc7d9SMatthew Wilcox if (flags & FOLL_WRITE && !pud_write(*pud)) 1004a00cc7d9SMatthew Wilcox return NULL; 1005a00cc7d9SMatthew Wilcox 1006a00cc7d9SMatthew Wilcox if (pud_present(*pud) && pud_devmap(*pud)) 1007a00cc7d9SMatthew Wilcox /* pass */; 1008a00cc7d9SMatthew Wilcox else 1009a00cc7d9SMatthew Wilcox return NULL; 1010a00cc7d9SMatthew Wilcox 1011a00cc7d9SMatthew Wilcox if (flags & FOLL_TOUCH) 1012a00cc7d9SMatthew Wilcox touch_pud(vma, addr, pud); 1013a00cc7d9SMatthew Wilcox 1014a00cc7d9SMatthew Wilcox /* 1015a00cc7d9SMatthew Wilcox * device mapped pages can only be returned if the 1016a00cc7d9SMatthew Wilcox * caller will manage the page reference count. 1017a00cc7d9SMatthew Wilcox */ 1018a00cc7d9SMatthew Wilcox if (!(flags & FOLL_GET)) 1019a00cc7d9SMatthew Wilcox return ERR_PTR(-EEXIST); 1020a00cc7d9SMatthew Wilcox 1021a00cc7d9SMatthew Wilcox pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1022a00cc7d9SMatthew Wilcox pgmap = get_dev_pagemap(pfn, NULL); 1023a00cc7d9SMatthew Wilcox if (!pgmap) 1024a00cc7d9SMatthew Wilcox return ERR_PTR(-EFAULT); 1025a00cc7d9SMatthew Wilcox page = pfn_to_page(pfn); 1026a00cc7d9SMatthew Wilcox get_page(page); 1027a00cc7d9SMatthew Wilcox put_dev_pagemap(pgmap); 1028a00cc7d9SMatthew Wilcox 1029a00cc7d9SMatthew Wilcox return page; 1030a00cc7d9SMatthew Wilcox } 1031a00cc7d9SMatthew Wilcox 1032a00cc7d9SMatthew Wilcox int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1033a00cc7d9SMatthew Wilcox pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1034a00cc7d9SMatthew Wilcox struct vm_area_struct *vma) 1035a00cc7d9SMatthew Wilcox { 1036a00cc7d9SMatthew Wilcox spinlock_t *dst_ptl, *src_ptl; 1037a00cc7d9SMatthew Wilcox pud_t pud; 1038a00cc7d9SMatthew Wilcox int ret; 1039a00cc7d9SMatthew Wilcox 1040a00cc7d9SMatthew Wilcox dst_ptl = pud_lock(dst_mm, dst_pud); 1041a00cc7d9SMatthew Wilcox src_ptl = pud_lockptr(src_mm, src_pud); 1042a00cc7d9SMatthew Wilcox spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1043a00cc7d9SMatthew Wilcox 1044a00cc7d9SMatthew Wilcox ret = -EAGAIN; 1045a00cc7d9SMatthew Wilcox pud = *src_pud; 1046a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1047a00cc7d9SMatthew Wilcox goto out_unlock; 1048a00cc7d9SMatthew Wilcox 1049a00cc7d9SMatthew Wilcox /* 1050a00cc7d9SMatthew Wilcox * When page table lock is held, the huge zero pud should not be 1051a00cc7d9SMatthew Wilcox * under splitting since we don't split the page itself, only pud to 1052a00cc7d9SMatthew Wilcox * a page table. 1053a00cc7d9SMatthew Wilcox */ 1054a00cc7d9SMatthew Wilcox if (is_huge_zero_pud(pud)) { 1055a00cc7d9SMatthew Wilcox /* No huge zero pud yet */ 1056a00cc7d9SMatthew Wilcox } 1057a00cc7d9SMatthew Wilcox 1058a00cc7d9SMatthew Wilcox pudp_set_wrprotect(src_mm, addr, src_pud); 1059a00cc7d9SMatthew Wilcox pud = pud_mkold(pud_wrprotect(pud)); 1060a00cc7d9SMatthew Wilcox set_pud_at(dst_mm, addr, dst_pud, pud); 1061a00cc7d9SMatthew Wilcox 1062a00cc7d9SMatthew Wilcox ret = 0; 1063a00cc7d9SMatthew Wilcox out_unlock: 1064a00cc7d9SMatthew Wilcox spin_unlock(src_ptl); 1065a00cc7d9SMatthew Wilcox spin_unlock(dst_ptl); 1066a00cc7d9SMatthew Wilcox return ret; 1067a00cc7d9SMatthew Wilcox } 1068a00cc7d9SMatthew Wilcox 1069a00cc7d9SMatthew Wilcox void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1070a00cc7d9SMatthew Wilcox { 1071a00cc7d9SMatthew Wilcox pud_t entry; 1072a00cc7d9SMatthew Wilcox unsigned long haddr; 1073a00cc7d9SMatthew Wilcox bool write = vmf->flags & FAULT_FLAG_WRITE; 1074a00cc7d9SMatthew Wilcox 1075a00cc7d9SMatthew Wilcox vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1076a00cc7d9SMatthew Wilcox if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1077a00cc7d9SMatthew Wilcox goto unlock; 1078a00cc7d9SMatthew Wilcox 1079a00cc7d9SMatthew Wilcox entry = pud_mkyoung(orig_pud); 1080a00cc7d9SMatthew Wilcox if (write) 1081a00cc7d9SMatthew Wilcox entry = pud_mkdirty(entry); 1082a00cc7d9SMatthew Wilcox haddr = vmf->address & HPAGE_PUD_MASK; 1083a00cc7d9SMatthew Wilcox if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1084a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1085a00cc7d9SMatthew Wilcox 1086a00cc7d9SMatthew Wilcox unlock: 1087a00cc7d9SMatthew Wilcox spin_unlock(vmf->ptl); 1088a00cc7d9SMatthew Wilcox } 1089a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1090a00cc7d9SMatthew Wilcox 109182b0f8c3SJan Kara void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1092a1dd450bSWill Deacon { 1093a1dd450bSWill Deacon pmd_t entry; 1094a1dd450bSWill Deacon unsigned long haddr; 109520f664aaSMinchan Kim bool write = vmf->flags & FAULT_FLAG_WRITE; 1096a1dd450bSWill Deacon 109782b0f8c3SJan Kara vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 109882b0f8c3SJan Kara if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1099a1dd450bSWill Deacon goto unlock; 1100a1dd450bSWill Deacon 1101a1dd450bSWill Deacon entry = pmd_mkyoung(orig_pmd); 110220f664aaSMinchan Kim if (write) 110320f664aaSMinchan Kim entry = pmd_mkdirty(entry); 110482b0f8c3SJan Kara haddr = vmf->address & HPAGE_PMD_MASK; 110520f664aaSMinchan Kim if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 110682b0f8c3SJan Kara update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1107a1dd450bSWill Deacon 1108a1dd450bSWill Deacon unlock: 110982b0f8c3SJan Kara spin_unlock(vmf->ptl); 1110a1dd450bSWill Deacon } 1111a1dd450bSWill Deacon 111282b0f8c3SJan Kara static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd, 1113bae473a4SKirill A. Shutemov struct page *page) 111471e3aac0SAndrea Arcangeli { 111582b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 111682b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 111700501b53SJohannes Weiner struct mem_cgroup *memcg; 111871e3aac0SAndrea Arcangeli pgtable_t pgtable; 111971e3aac0SAndrea Arcangeli pmd_t _pmd; 112071e3aac0SAndrea Arcangeli int ret = 0, i; 112171e3aac0SAndrea Arcangeli struct page **pages; 11222ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 11232ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 112471e3aac0SAndrea Arcangeli 112571e3aac0SAndrea Arcangeli pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 112671e3aac0SAndrea Arcangeli GFP_KERNEL); 112771e3aac0SAndrea Arcangeli if (unlikely(!pages)) { 112871e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 112971e3aac0SAndrea Arcangeli goto out; 113071e3aac0SAndrea Arcangeli } 113171e3aac0SAndrea Arcangeli 113271e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 113341b6167eSMichal Hocko pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, 113482b0f8c3SJan Kara vmf->address, page_to_nid(page)); 1135b9bbfbe3SAndrea Arcangeli if (unlikely(!pages[i] || 1136bae473a4SKirill A. Shutemov mem_cgroup_try_charge(pages[i], vma->vm_mm, 1137bae473a4SKirill A. Shutemov GFP_KERNEL, &memcg, false))) { 1138b9bbfbe3SAndrea Arcangeli if (pages[i]) 113971e3aac0SAndrea Arcangeli put_page(pages[i]); 1140b9bbfbe3SAndrea Arcangeli while (--i >= 0) { 114100501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 114200501b53SJohannes Weiner set_page_private(pages[i], 0); 1143f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(pages[i], memcg, 1144f627c2f5SKirill A. Shutemov false); 1145b9bbfbe3SAndrea Arcangeli put_page(pages[i]); 1146b9bbfbe3SAndrea Arcangeli } 114771e3aac0SAndrea Arcangeli kfree(pages); 114871e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 114971e3aac0SAndrea Arcangeli goto out; 115071e3aac0SAndrea Arcangeli } 115100501b53SJohannes Weiner set_page_private(pages[i], (unsigned long)memcg); 115271e3aac0SAndrea Arcangeli } 115371e3aac0SAndrea Arcangeli 115471e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 115571e3aac0SAndrea Arcangeli copy_user_highpage(pages[i], page + i, 11560089e485SHillf Danton haddr + PAGE_SIZE * i, vma); 115771e3aac0SAndrea Arcangeli __SetPageUptodate(pages[i]); 115871e3aac0SAndrea Arcangeli cond_resched(); 115971e3aac0SAndrea Arcangeli } 116071e3aac0SAndrea Arcangeli 11612ec74c3eSSagi Grimberg mmun_start = haddr; 11622ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 1163bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 11642ec74c3eSSagi Grimberg 116582b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 116682b0f8c3SJan Kara if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 116771e3aac0SAndrea Arcangeli goto out_free_pages; 1168309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 116971e3aac0SAndrea Arcangeli 117082b0f8c3SJan Kara pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 117171e3aac0SAndrea Arcangeli /* leave pmd empty until pte is filled */ 117271e3aac0SAndrea Arcangeli 117382b0f8c3SJan Kara pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); 1174bae473a4SKirill A. Shutemov pmd_populate(vma->vm_mm, &_pmd, pgtable); 117571e3aac0SAndrea Arcangeli 117671e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1177bae473a4SKirill A. Shutemov pte_t entry; 117871e3aac0SAndrea Arcangeli entry = mk_pte(pages[i], vma->vm_page_prot); 117971e3aac0SAndrea Arcangeli entry = maybe_mkwrite(pte_mkdirty(entry), vma); 118000501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 118100501b53SJohannes Weiner set_page_private(pages[i], 0); 118282b0f8c3SJan Kara page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); 1183f627c2f5SKirill A. Shutemov mem_cgroup_commit_charge(pages[i], memcg, false, false); 118400501b53SJohannes Weiner lru_cache_add_active_or_unevictable(pages[i], vma); 118582b0f8c3SJan Kara vmf->pte = pte_offset_map(&_pmd, haddr); 118682b0f8c3SJan Kara VM_BUG_ON(!pte_none(*vmf->pte)); 118782b0f8c3SJan Kara set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); 118882b0f8c3SJan Kara pte_unmap(vmf->pte); 118971e3aac0SAndrea Arcangeli } 119071e3aac0SAndrea Arcangeli kfree(pages); 119171e3aac0SAndrea Arcangeli 119271e3aac0SAndrea Arcangeli smp_wmb(); /* make pte visible before pmd */ 119382b0f8c3SJan Kara pmd_populate(vma->vm_mm, vmf->pmd, pgtable); 1194d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 119582b0f8c3SJan Kara spin_unlock(vmf->ptl); 119671e3aac0SAndrea Arcangeli 1197bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 11982ec74c3eSSagi Grimberg 119971e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 120071e3aac0SAndrea Arcangeli put_page(page); 120171e3aac0SAndrea Arcangeli 120271e3aac0SAndrea Arcangeli out: 120371e3aac0SAndrea Arcangeli return ret; 120471e3aac0SAndrea Arcangeli 120571e3aac0SAndrea Arcangeli out_free_pages: 120682b0f8c3SJan Kara spin_unlock(vmf->ptl); 1207bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 1208b9bbfbe3SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 120900501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 121000501b53SJohannes Weiner set_page_private(pages[i], 0); 1211f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(pages[i], memcg, false); 121271e3aac0SAndrea Arcangeli put_page(pages[i]); 1213b9bbfbe3SAndrea Arcangeli } 121471e3aac0SAndrea Arcangeli kfree(pages); 121571e3aac0SAndrea Arcangeli goto out; 121671e3aac0SAndrea Arcangeli } 121771e3aac0SAndrea Arcangeli 121882b0f8c3SJan Kara int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 121971e3aac0SAndrea Arcangeli { 122082b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 122193b4796dSKirill A. Shutemov struct page *page = NULL, *new_page; 122200501b53SJohannes Weiner struct mem_cgroup *memcg; 122382b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 12242ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 12252ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 12263b363692SMichal Hocko gfp_t huge_gfp; /* for allocation and charge */ 1227bae473a4SKirill A. Shutemov int ret = 0; 122871e3aac0SAndrea Arcangeli 122982b0f8c3SJan Kara vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 123081d1b09cSSasha Levin VM_BUG_ON_VMA(!vma->anon_vma, vma); 123193b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 123293b4796dSKirill A. Shutemov goto alloc; 123382b0f8c3SJan Kara spin_lock(vmf->ptl); 123482b0f8c3SJan Kara if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 123571e3aac0SAndrea Arcangeli goto out_unlock; 123671e3aac0SAndrea Arcangeli 123771e3aac0SAndrea Arcangeli page = pmd_page(orig_pmd); 1238309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 12391f25fe20SKirill A. Shutemov /* 12401f25fe20SKirill A. Shutemov * We can only reuse the page if nobody else maps the huge page or it's 12416d0a07edSAndrea Arcangeli * part. 12421f25fe20SKirill A. Shutemov */ 1243ba3c4ce6SHuang Ying if (!trylock_page(page)) { 1244ba3c4ce6SHuang Ying get_page(page); 1245ba3c4ce6SHuang Ying spin_unlock(vmf->ptl); 1246ba3c4ce6SHuang Ying lock_page(page); 1247ba3c4ce6SHuang Ying spin_lock(vmf->ptl); 1248ba3c4ce6SHuang Ying if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1249ba3c4ce6SHuang Ying unlock_page(page); 1250ba3c4ce6SHuang Ying put_page(page); 1251ba3c4ce6SHuang Ying goto out_unlock; 1252ba3c4ce6SHuang Ying } 1253ba3c4ce6SHuang Ying put_page(page); 1254ba3c4ce6SHuang Ying } 1255ba3c4ce6SHuang Ying if (reuse_swap_page(page, NULL)) { 125671e3aac0SAndrea Arcangeli pmd_t entry; 125771e3aac0SAndrea Arcangeli entry = pmd_mkyoung(orig_pmd); 125871e3aac0SAndrea Arcangeli entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 125982b0f8c3SJan Kara if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 126082b0f8c3SJan Kara update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 126171e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 1262ba3c4ce6SHuang Ying unlock_page(page); 126371e3aac0SAndrea Arcangeli goto out_unlock; 126471e3aac0SAndrea Arcangeli } 1265ba3c4ce6SHuang Ying unlock_page(page); 1266ddc58f27SKirill A. Shutemov get_page(page); 126782b0f8c3SJan Kara spin_unlock(vmf->ptl); 126893b4796dSKirill A. Shutemov alloc: 126971e3aac0SAndrea Arcangeli if (transparent_hugepage_enabled(vma) && 1270077fcf11SAneesh Kumar K.V !transparent_hugepage_debug_cow()) { 1271444eb2a4SMel Gorman huge_gfp = alloc_hugepage_direct_gfpmask(vma); 12723b363692SMichal Hocko new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); 1273077fcf11SAneesh Kumar K.V } else 127471e3aac0SAndrea Arcangeli new_page = NULL; 127571e3aac0SAndrea Arcangeli 12769a982250SKirill A. Shutemov if (likely(new_page)) { 12779a982250SKirill A. Shutemov prep_transhuge_page(new_page); 12789a982250SKirill A. Shutemov } else { 1279eecc1e42SHugh Dickins if (!page) { 128082b0f8c3SJan Kara split_huge_pmd(vma, vmf->pmd, vmf->address); 1281e9b71ca9SKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 128293b4796dSKirill A. Shutemov } else { 128382b0f8c3SJan Kara ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); 12849845cbbdSKirill A. Shutemov if (ret & VM_FAULT_OOM) { 128582b0f8c3SJan Kara split_huge_pmd(vma, vmf->pmd, vmf->address); 12869845cbbdSKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 12879845cbbdSKirill A. Shutemov } 1288ddc58f27SKirill A. Shutemov put_page(page); 128993b4796dSKirill A. Shutemov } 129017766ddeSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK); 129171e3aac0SAndrea Arcangeli goto out; 129271e3aac0SAndrea Arcangeli } 129371e3aac0SAndrea Arcangeli 1294bae473a4SKirill A. Shutemov if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, 1295bae473a4SKirill A. Shutemov huge_gfp, &memcg, true))) { 1296b9bbfbe3SAndrea Arcangeli put_page(new_page); 129782b0f8c3SJan Kara split_huge_pmd(vma, vmf->pmd, vmf->address); 1298bae473a4SKirill A. Shutemov if (page) 1299ddc58f27SKirill A. Shutemov put_page(page); 13009845cbbdSKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 130117766ddeSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK); 1302b9bbfbe3SAndrea Arcangeli goto out; 1303b9bbfbe3SAndrea Arcangeli } 1304b9bbfbe3SAndrea Arcangeli 130517766ddeSDavid Rientjes count_vm_event(THP_FAULT_ALLOC); 130617766ddeSDavid Rientjes 1307eecc1e42SHugh Dickins if (!page) 1308c79b57e4SHuang Ying clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR); 130993b4796dSKirill A. Shutemov else 131071e3aac0SAndrea Arcangeli copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 131171e3aac0SAndrea Arcangeli __SetPageUptodate(new_page); 131271e3aac0SAndrea Arcangeli 13132ec74c3eSSagi Grimberg mmun_start = haddr; 13142ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 1315bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 13162ec74c3eSSagi Grimberg 131782b0f8c3SJan Kara spin_lock(vmf->ptl); 131893b4796dSKirill A. Shutemov if (page) 1319ddc58f27SKirill A. Shutemov put_page(page); 132082b0f8c3SJan Kara if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 132182b0f8c3SJan Kara spin_unlock(vmf->ptl); 1322f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(new_page, memcg, true); 132371e3aac0SAndrea Arcangeli put_page(new_page); 13242ec74c3eSSagi Grimberg goto out_mn; 1325b9bbfbe3SAndrea Arcangeli } else { 132671e3aac0SAndrea Arcangeli pmd_t entry; 13273122359aSKirill A. Shutemov entry = mk_huge_pmd(new_page, vma->vm_page_prot); 13283122359aSKirill A. Shutemov entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 132982b0f8c3SJan Kara pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1330d281ee61SKirill A. Shutemov page_add_new_anon_rmap(new_page, vma, haddr, true); 1331f627c2f5SKirill A. Shutemov mem_cgroup_commit_charge(new_page, memcg, false, true); 133200501b53SJohannes Weiner lru_cache_add_active_or_unevictable(new_page, vma); 133382b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 133482b0f8c3SJan Kara update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1335eecc1e42SHugh Dickins if (!page) { 1336bae473a4SKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 133797ae1749SKirill A. Shutemov } else { 1338309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1339d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 134071e3aac0SAndrea Arcangeli put_page(page); 134193b4796dSKirill A. Shutemov } 134271e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 134371e3aac0SAndrea Arcangeli } 134482b0f8c3SJan Kara spin_unlock(vmf->ptl); 13452ec74c3eSSagi Grimberg out_mn: 1346bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 13472ec74c3eSSagi Grimberg out: 13482ec74c3eSSagi Grimberg return ret; 134971e3aac0SAndrea Arcangeli out_unlock: 135082b0f8c3SJan Kara spin_unlock(vmf->ptl); 135171e3aac0SAndrea Arcangeli return ret; 135271e3aac0SAndrea Arcangeli } 135371e3aac0SAndrea Arcangeli 13548310d48bSKeno Fischer /* 13558310d48bSKeno Fischer * FOLL_FORCE can write to even unwritable pmd's, but only 13568310d48bSKeno Fischer * after we've gone through a COW cycle and they are dirty. 13578310d48bSKeno Fischer */ 13588310d48bSKeno Fischer static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 13598310d48bSKeno Fischer { 13608310d48bSKeno Fischer return pmd_write(pmd) || 13618310d48bSKeno Fischer ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); 13628310d48bSKeno Fischer } 13638310d48bSKeno Fischer 1364b676b293SDavid Rientjes struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 136571e3aac0SAndrea Arcangeli unsigned long addr, 136671e3aac0SAndrea Arcangeli pmd_t *pmd, 136771e3aac0SAndrea Arcangeli unsigned int flags) 136871e3aac0SAndrea Arcangeli { 1369b676b293SDavid Rientjes struct mm_struct *mm = vma->vm_mm; 137071e3aac0SAndrea Arcangeli struct page *page = NULL; 137171e3aac0SAndrea Arcangeli 1372c4088ebdSKirill A. Shutemov assert_spin_locked(pmd_lockptr(mm, pmd)); 137371e3aac0SAndrea Arcangeli 13748310d48bSKeno Fischer if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 137571e3aac0SAndrea Arcangeli goto out; 137671e3aac0SAndrea Arcangeli 137785facf25SKirill A. Shutemov /* Avoid dumping huge zero page */ 137885facf25SKirill A. Shutemov if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 137985facf25SKirill A. Shutemov return ERR_PTR(-EFAULT); 138085facf25SKirill A. Shutemov 13812b4847e7SMel Gorman /* Full NUMA hinting faults to serialise migration in fault paths */ 13828a0516edSMel Gorman if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 13832b4847e7SMel Gorman goto out; 13842b4847e7SMel Gorman 138571e3aac0SAndrea Arcangeli page = pmd_page(*pmd); 1386ca120cf6SDan Williams VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 13873565fce3SDan Williams if (flags & FOLL_TOUCH) 13883565fce3SDan Williams touch_pmd(vma, addr, pmd); 1389de60f5f1SEric B Munson if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1390e90309c9SKirill A. Shutemov /* 1391e90309c9SKirill A. Shutemov * We don't mlock() pte-mapped THPs. This way we can avoid 1392e90309c9SKirill A. Shutemov * leaking mlocked pages into non-VM_LOCKED VMAs. 1393e90309c9SKirill A. Shutemov * 13949a73f61bSKirill A. Shutemov * For anon THP: 13959a73f61bSKirill A. Shutemov * 1396e90309c9SKirill A. Shutemov * In most cases the pmd is the only mapping of the page as we 1397e90309c9SKirill A. Shutemov * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1398e90309c9SKirill A. Shutemov * writable private mappings in populate_vma_page_range(). 1399e90309c9SKirill A. Shutemov * 1400e90309c9SKirill A. Shutemov * The only scenario when we have the page shared here is if we 1401e90309c9SKirill A. Shutemov * mlocking read-only mapping shared over fork(). We skip 1402e90309c9SKirill A. Shutemov * mlocking such pages. 14039a73f61bSKirill A. Shutemov * 14049a73f61bSKirill A. Shutemov * For file THP: 14059a73f61bSKirill A. Shutemov * 14069a73f61bSKirill A. Shutemov * We can expect PageDoubleMap() to be stable under page lock: 14079a73f61bSKirill A. Shutemov * for file pages we set it in page_add_file_rmap(), which 14089a73f61bSKirill A. Shutemov * requires page to be locked. 1409e90309c9SKirill A. Shutemov */ 14109a73f61bSKirill A. Shutemov 14119a73f61bSKirill A. Shutemov if (PageAnon(page) && compound_mapcount(page) != 1) 14129a73f61bSKirill A. Shutemov goto skip_mlock; 14139a73f61bSKirill A. Shutemov if (PageDoubleMap(page) || !page->mapping) 14149a73f61bSKirill A. Shutemov goto skip_mlock; 14159a73f61bSKirill A. Shutemov if (!trylock_page(page)) 14169a73f61bSKirill A. Shutemov goto skip_mlock; 1417b676b293SDavid Rientjes lru_add_drain(); 14189a73f61bSKirill A. Shutemov if (page->mapping && !PageDoubleMap(page)) 1419b676b293SDavid Rientjes mlock_vma_page(page); 1420b676b293SDavid Rientjes unlock_page(page); 1421b676b293SDavid Rientjes } 14229a73f61bSKirill A. Shutemov skip_mlock: 142371e3aac0SAndrea Arcangeli page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1424ca120cf6SDan Williams VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 142571e3aac0SAndrea Arcangeli if (flags & FOLL_GET) 1426ddc58f27SKirill A. Shutemov get_page(page); 142771e3aac0SAndrea Arcangeli 142871e3aac0SAndrea Arcangeli out: 142971e3aac0SAndrea Arcangeli return page; 143071e3aac0SAndrea Arcangeli } 143171e3aac0SAndrea Arcangeli 1432d10e63f2SMel Gorman /* NUMA hinting page fault entry point for trans huge pmds */ 143382b0f8c3SJan Kara int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1434d10e63f2SMel Gorman { 143582b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 1436b8916634SMel Gorman struct anon_vma *anon_vma = NULL; 1437b32967ffSMel Gorman struct page *page; 143882b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 14398191acbdSMel Gorman int page_nid = -1, this_nid = numa_node_id(); 144090572890SPeter Zijlstra int target_nid, last_cpupid = -1; 14418191acbdSMel Gorman bool page_locked; 14428191acbdSMel Gorman bool migrated = false; 1443b191f9b1SMel Gorman bool was_writable; 14446688cc05SPeter Zijlstra int flags = 0; 1445d10e63f2SMel Gorman 144682b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 144782b0f8c3SJan Kara if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1448d10e63f2SMel Gorman goto out_unlock; 1449d10e63f2SMel Gorman 1450de466bd6SMel Gorman /* 1451de466bd6SMel Gorman * If there are potential migrations, wait for completion and retry 1452de466bd6SMel Gorman * without disrupting NUMA hinting information. Do not relock and 1453de466bd6SMel Gorman * check_same as the page may no longer be mapped. 1454de466bd6SMel Gorman */ 145582b0f8c3SJan Kara if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 145682b0f8c3SJan Kara page = pmd_page(*vmf->pmd); 14573c226c63SMark Rutland if (!get_page_unless_zero(page)) 14583c226c63SMark Rutland goto out_unlock; 145982b0f8c3SJan Kara spin_unlock(vmf->ptl); 14605d833062SMel Gorman wait_on_page_locked(page); 14613c226c63SMark Rutland put_page(page); 1462de466bd6SMel Gorman goto out; 1463de466bd6SMel Gorman } 1464de466bd6SMel Gorman 1465d10e63f2SMel Gorman page = pmd_page(pmd); 1466a1a46184SMel Gorman BUG_ON(is_huge_zero_page(page)); 14678191acbdSMel Gorman page_nid = page_to_nid(page); 146890572890SPeter Zijlstra last_cpupid = page_cpupid_last(page); 146903c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS); 147004bb2f94SRik van Riel if (page_nid == this_nid) { 147103c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 147204bb2f94SRik van Riel flags |= TNF_FAULT_LOCAL; 147304bb2f94SRik van Riel } 14744daae3b4SMel Gorman 1475bea66fbdSMel Gorman /* See similar comment in do_numa_page for explanation */ 1476288bc549SAneesh Kumar K.V if (!pmd_savedwrite(pmd)) 14776688cc05SPeter Zijlstra flags |= TNF_NO_GROUP; 14786688cc05SPeter Zijlstra 14796688cc05SPeter Zijlstra /* 1480ff9042b1SMel Gorman * Acquire the page lock to serialise THP migrations but avoid dropping 1481ff9042b1SMel Gorman * page_table_lock if at all possible 1482ff9042b1SMel Gorman */ 1483b8916634SMel Gorman page_locked = trylock_page(page); 1484b8916634SMel Gorman target_nid = mpol_misplaced(page, vma, haddr); 1485b8916634SMel Gorman if (target_nid == -1) { 1486b8916634SMel Gorman /* If the page was locked, there are no parallel migrations */ 1487a54a407fSMel Gorman if (page_locked) 1488b8916634SMel Gorman goto clear_pmdnuma; 14892b4847e7SMel Gorman } 1490cbee9f88SPeter Zijlstra 1491de466bd6SMel Gorman /* Migration could have started since the pmd_trans_migrating check */ 14922b4847e7SMel Gorman if (!page_locked) { 14933c226c63SMark Rutland page_nid = -1; 14943c226c63SMark Rutland if (!get_page_unless_zero(page)) 14953c226c63SMark Rutland goto out_unlock; 149682b0f8c3SJan Kara spin_unlock(vmf->ptl); 1497b8916634SMel Gorman wait_on_page_locked(page); 14983c226c63SMark Rutland put_page(page); 1499b8916634SMel Gorman goto out; 1500b8916634SMel Gorman } 1501b8916634SMel Gorman 15022b4847e7SMel Gorman /* 15032b4847e7SMel Gorman * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 15042b4847e7SMel Gorman * to serialises splits 15052b4847e7SMel Gorman */ 1506b8916634SMel Gorman get_page(page); 150782b0f8c3SJan Kara spin_unlock(vmf->ptl); 1508b8916634SMel Gorman anon_vma = page_lock_anon_vma_read(page); 1509b32967ffSMel Gorman 1510c69307d5SPeter Zijlstra /* Confirm the PMD did not change while page_table_lock was released */ 151182b0f8c3SJan Kara spin_lock(vmf->ptl); 151282b0f8c3SJan Kara if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1513b32967ffSMel Gorman unlock_page(page); 1514b32967ffSMel Gorman put_page(page); 1515a54a407fSMel Gorman page_nid = -1; 1516b32967ffSMel Gorman goto out_unlock; 1517b32967ffSMel Gorman } 1518ff9042b1SMel Gorman 1519c3a489caSMel Gorman /* Bail if we fail to protect against THP splits for any reason */ 1520c3a489caSMel Gorman if (unlikely(!anon_vma)) { 1521c3a489caSMel Gorman put_page(page); 1522c3a489caSMel Gorman page_nid = -1; 1523c3a489caSMel Gorman goto clear_pmdnuma; 1524c3a489caSMel Gorman } 1525c3a489caSMel Gorman 1526a54a407fSMel Gorman /* 15278b1b436dSPeter Zijlstra * Since we took the NUMA fault, we must have observed the !accessible 15288b1b436dSPeter Zijlstra * bit. Make sure all other CPUs agree with that, to avoid them 15298b1b436dSPeter Zijlstra * modifying the page we're about to migrate. 15308b1b436dSPeter Zijlstra * 15318b1b436dSPeter Zijlstra * Must be done under PTL such that we'll observe the relevant 1532ccde85baSPeter Zijlstra * inc_tlb_flush_pending(). 1533ccde85baSPeter Zijlstra * 1534ccde85baSPeter Zijlstra * We are not sure a pending tlb flush here is for a huge page 1535ccde85baSPeter Zijlstra * mapping or not. Hence use the tlb range variant 15368b1b436dSPeter Zijlstra */ 15378b1b436dSPeter Zijlstra if (mm_tlb_flush_pending(vma->vm_mm)) 1538ccde85baSPeter Zijlstra flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 15398b1b436dSPeter Zijlstra 15408b1b436dSPeter Zijlstra /* 1541a54a407fSMel Gorman * Migrate the THP to the requested node, returns with page unlocked 15428a0516edSMel Gorman * and access rights restored. 1543a54a407fSMel Gorman */ 154482b0f8c3SJan Kara spin_unlock(vmf->ptl); 15458b1b436dSPeter Zijlstra 1546bae473a4SKirill A. Shutemov migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 154782b0f8c3SJan Kara vmf->pmd, pmd, vmf->address, page, target_nid); 15486688cc05SPeter Zijlstra if (migrated) { 15496688cc05SPeter Zijlstra flags |= TNF_MIGRATED; 15508191acbdSMel Gorman page_nid = target_nid; 1551074c2381SMel Gorman } else 1552074c2381SMel Gorman flags |= TNF_MIGRATE_FAIL; 1553b32967ffSMel Gorman 15548191acbdSMel Gorman goto out; 15554daae3b4SMel Gorman clear_pmdnuma: 1556a54a407fSMel Gorman BUG_ON(!PageLocked(page)); 1557288bc549SAneesh Kumar K.V was_writable = pmd_savedwrite(pmd); 15584d942466SMel Gorman pmd = pmd_modify(pmd, vma->vm_page_prot); 1559b7b04004SMel Gorman pmd = pmd_mkyoung(pmd); 1560b191f9b1SMel Gorman if (was_writable) 1561b191f9b1SMel Gorman pmd = pmd_mkwrite(pmd); 156282b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 156382b0f8c3SJan Kara update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1564a54a407fSMel Gorman unlock_page(page); 1565d10e63f2SMel Gorman out_unlock: 156682b0f8c3SJan Kara spin_unlock(vmf->ptl); 1567b8916634SMel Gorman 1568b8916634SMel Gorman out: 1569b8916634SMel Gorman if (anon_vma) 1570b8916634SMel Gorman page_unlock_anon_vma_read(anon_vma); 1571b8916634SMel Gorman 15728191acbdSMel Gorman if (page_nid != -1) 157382b0f8c3SJan Kara task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 15749a8b300fSAneesh Kumar K.V flags); 15758191acbdSMel Gorman 1576d10e63f2SMel Gorman return 0; 1577d10e63f2SMel Gorman } 1578d10e63f2SMel Gorman 1579319904adSHuang Ying /* 1580319904adSHuang Ying * Return true if we do MADV_FREE successfully on entire pmd page. 1581319904adSHuang Ying * Otherwise, return false. 1582319904adSHuang Ying */ 1583319904adSHuang Ying bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1584b8d3c4c3SMinchan Kim pmd_t *pmd, unsigned long addr, unsigned long next) 1585b8d3c4c3SMinchan Kim { 1586b8d3c4c3SMinchan Kim spinlock_t *ptl; 1587b8d3c4c3SMinchan Kim pmd_t orig_pmd; 1588b8d3c4c3SMinchan Kim struct page *page; 1589b8d3c4c3SMinchan Kim struct mm_struct *mm = tlb->mm; 1590319904adSHuang Ying bool ret = false; 1591b8d3c4c3SMinchan Kim 159207e32661SAneesh Kumar K.V tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 159307e32661SAneesh Kumar K.V 1594b6ec57f4SKirill A. Shutemov ptl = pmd_trans_huge_lock(pmd, vma); 1595b6ec57f4SKirill A. Shutemov if (!ptl) 159625eedabeSLinus Torvalds goto out_unlocked; 1597b8d3c4c3SMinchan Kim 1598b8d3c4c3SMinchan Kim orig_pmd = *pmd; 1599319904adSHuang Ying if (is_huge_zero_pmd(orig_pmd)) 1600b8d3c4c3SMinchan Kim goto out; 1601b8d3c4c3SMinchan Kim 1602b8d3c4c3SMinchan Kim page = pmd_page(orig_pmd); 1603b8d3c4c3SMinchan Kim /* 1604b8d3c4c3SMinchan Kim * If other processes are mapping this page, we couldn't discard 1605b8d3c4c3SMinchan Kim * the page unless they all do MADV_FREE so let's skip the page. 1606b8d3c4c3SMinchan Kim */ 1607b8d3c4c3SMinchan Kim if (page_mapcount(page) != 1) 1608b8d3c4c3SMinchan Kim goto out; 1609b8d3c4c3SMinchan Kim 1610b8d3c4c3SMinchan Kim if (!trylock_page(page)) 1611b8d3c4c3SMinchan Kim goto out; 1612b8d3c4c3SMinchan Kim 1613b8d3c4c3SMinchan Kim /* 1614b8d3c4c3SMinchan Kim * If user want to discard part-pages of THP, split it so MADV_FREE 1615b8d3c4c3SMinchan Kim * will deactivate only them. 1616b8d3c4c3SMinchan Kim */ 1617b8d3c4c3SMinchan Kim if (next - addr != HPAGE_PMD_SIZE) { 1618b8d3c4c3SMinchan Kim get_page(page); 1619b8d3c4c3SMinchan Kim spin_unlock(ptl); 16209818b8cdSHuang Ying split_huge_page(page); 1621b8d3c4c3SMinchan Kim unlock_page(page); 1622bbf29ffcSKirill A. Shutemov put_page(page); 1623b8d3c4c3SMinchan Kim goto out_unlocked; 1624b8d3c4c3SMinchan Kim } 1625b8d3c4c3SMinchan Kim 1626b8d3c4c3SMinchan Kim if (PageDirty(page)) 1627b8d3c4c3SMinchan Kim ClearPageDirty(page); 1628b8d3c4c3SMinchan Kim unlock_page(page); 1629b8d3c4c3SMinchan Kim 1630b8d3c4c3SMinchan Kim if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 163158ceeb6bSKirill A. Shutemov pmdp_invalidate(vma, addr, pmd); 1632b8d3c4c3SMinchan Kim orig_pmd = pmd_mkold(orig_pmd); 1633b8d3c4c3SMinchan Kim orig_pmd = pmd_mkclean(orig_pmd); 1634b8d3c4c3SMinchan Kim 1635b8d3c4c3SMinchan Kim set_pmd_at(mm, addr, pmd, orig_pmd); 1636b8d3c4c3SMinchan Kim tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1637b8d3c4c3SMinchan Kim } 1638802a3a92SShaohua Li 1639802a3a92SShaohua Li mark_page_lazyfree(page); 1640319904adSHuang Ying ret = true; 1641b8d3c4c3SMinchan Kim out: 1642b8d3c4c3SMinchan Kim spin_unlock(ptl); 1643b8d3c4c3SMinchan Kim out_unlocked: 1644b8d3c4c3SMinchan Kim return ret; 1645b8d3c4c3SMinchan Kim } 1646b8d3c4c3SMinchan Kim 1647953c66c2SAneesh Kumar K.V static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1648953c66c2SAneesh Kumar K.V { 1649953c66c2SAneesh Kumar K.V pgtable_t pgtable; 1650953c66c2SAneesh Kumar K.V 1651953c66c2SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1652953c66c2SAneesh Kumar K.V pte_free(mm, pgtable); 1653953c66c2SAneesh Kumar K.V atomic_long_dec(&mm->nr_ptes); 1654953c66c2SAneesh Kumar K.V } 1655953c66c2SAneesh Kumar K.V 165671e3aac0SAndrea Arcangeli int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1657f21760b1SShaohua Li pmd_t *pmd, unsigned long addr) 165871e3aac0SAndrea Arcangeli { 1659f5c8ad47SDavid Miller pmd_t orig_pmd; 1660da146769SKirill A. Shutemov spinlock_t *ptl; 1661da146769SKirill A. Shutemov 166207e32661SAneesh Kumar K.V tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 166307e32661SAneesh Kumar K.V 1664b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 1665b6ec57f4SKirill A. Shutemov if (!ptl) 1666da146769SKirill A. Shutemov return 0; 1667a6bf2bb0SAneesh Kumar K.V /* 1668a6bf2bb0SAneesh Kumar K.V * For architectures like ppc64 we look at deposited pgtable 16698809aa2dSAneesh Kumar K.V * when calling pmdp_huge_get_and_clear. So do the 1670a6bf2bb0SAneesh Kumar K.V * pgtable_trans_huge_withdraw after finishing pmdp related 1671a6bf2bb0SAneesh Kumar K.V * operations. 1672a6bf2bb0SAneesh Kumar K.V */ 16738809aa2dSAneesh Kumar K.V orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1674fcbe08d6SMartin Schwidefsky tlb->fullmm); 1675f21760b1SShaohua Li tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 16764897c765SMatthew Wilcox if (vma_is_dax(vma)) { 16773b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) 16783b6521f5SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 16794897c765SMatthew Wilcox spin_unlock(ptl); 1680da146769SKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 1681c0f2e176SAneesh Kumar K.V tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1682da146769SKirill A. Shutemov } else if (is_huge_zero_pmd(orig_pmd)) { 1683c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1684bf929152SKirill A. Shutemov spin_unlock(ptl); 1685c0f2e176SAneesh Kumar K.V tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1686479f0abbSKirill A. Shutemov } else { 1687*616b8371SZi Yan struct page *page = NULL; 1688*616b8371SZi Yan int flush_needed = 1; 1689*616b8371SZi Yan 1690*616b8371SZi Yan if (pmd_present(orig_pmd)) { 1691*616b8371SZi Yan page = pmd_page(orig_pmd); 1692d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 1693309381feSSasha Levin VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1694309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1695*616b8371SZi Yan } else if (thp_migration_supported()) { 1696*616b8371SZi Yan swp_entry_t entry; 1697*616b8371SZi Yan 1698*616b8371SZi Yan VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1699*616b8371SZi Yan entry = pmd_to_swp_entry(orig_pmd); 1700*616b8371SZi Yan page = pfn_to_page(swp_offset(entry)); 1701*616b8371SZi Yan flush_needed = 0; 1702*616b8371SZi Yan } else 1703*616b8371SZi Yan WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1704*616b8371SZi Yan 1705b5072380SKirill A. Shutemov if (PageAnon(page)) { 1706c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1707b5072380SKirill A. Shutemov add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1708b5072380SKirill A. Shutemov } else { 1709953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 1710953c66c2SAneesh Kumar K.V zap_deposited_table(tlb->mm, pmd); 1711b5072380SKirill A. Shutemov add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR); 1712b5072380SKirill A. Shutemov } 1713*616b8371SZi Yan 1714bf929152SKirill A. Shutemov spin_unlock(ptl); 1715*616b8371SZi Yan if (flush_needed) 1716e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1717479f0abbSKirill A. Shutemov } 1718da146769SKirill A. Shutemov return 1; 171971e3aac0SAndrea Arcangeli } 172071e3aac0SAndrea Arcangeli 17211dd38b6cSAneesh Kumar K.V #ifndef pmd_move_must_withdraw 17221dd38b6cSAneesh Kumar K.V static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 17231dd38b6cSAneesh Kumar K.V spinlock_t *old_pmd_ptl, 17241dd38b6cSAneesh Kumar K.V struct vm_area_struct *vma) 17251dd38b6cSAneesh Kumar K.V { 17261dd38b6cSAneesh Kumar K.V /* 17271dd38b6cSAneesh Kumar K.V * With split pmd lock we also need to move preallocated 17281dd38b6cSAneesh Kumar K.V * PTE page table if new_pmd is on different PMD page table. 17291dd38b6cSAneesh Kumar K.V * 17301dd38b6cSAneesh Kumar K.V * We also don't deposit and withdraw tables for file pages. 17311dd38b6cSAneesh Kumar K.V */ 17321dd38b6cSAneesh Kumar K.V return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 17331dd38b6cSAneesh Kumar K.V } 17341dd38b6cSAneesh Kumar K.V #endif 17351dd38b6cSAneesh Kumar K.V 1736bf8616d5SHugh Dickins bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 173737a1c49aSAndrea Arcangeli unsigned long new_addr, unsigned long old_end, 17385d190420SAaron Lu pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush) 173937a1c49aSAndrea Arcangeli { 1740bf929152SKirill A. Shutemov spinlock_t *old_ptl, *new_ptl; 174137a1c49aSAndrea Arcangeli pmd_t pmd; 174237a1c49aSAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 17435d190420SAaron Lu bool force_flush = false; 174437a1c49aSAndrea Arcangeli 174537a1c49aSAndrea Arcangeli if ((old_addr & ~HPAGE_PMD_MASK) || 174637a1c49aSAndrea Arcangeli (new_addr & ~HPAGE_PMD_MASK) || 1747bf8616d5SHugh Dickins old_end - old_addr < HPAGE_PMD_SIZE) 17484b471e88SKirill A. Shutemov return false; 174937a1c49aSAndrea Arcangeli 175037a1c49aSAndrea Arcangeli /* 175137a1c49aSAndrea Arcangeli * The destination pmd shouldn't be established, free_pgtables() 175237a1c49aSAndrea Arcangeli * should have release it. 175337a1c49aSAndrea Arcangeli */ 175437a1c49aSAndrea Arcangeli if (WARN_ON(!pmd_none(*new_pmd))) { 175537a1c49aSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*new_pmd)); 17564b471e88SKirill A. Shutemov return false; 175737a1c49aSAndrea Arcangeli } 175837a1c49aSAndrea Arcangeli 1759bf929152SKirill A. Shutemov /* 1760bf929152SKirill A. Shutemov * We don't have to worry about the ordering of src and dst 1761bf929152SKirill A. Shutemov * ptlocks because exclusive mmap_sem prevents deadlock. 1762bf929152SKirill A. Shutemov */ 1763b6ec57f4SKirill A. Shutemov old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1764b6ec57f4SKirill A. Shutemov if (old_ptl) { 1765bf929152SKirill A. Shutemov new_ptl = pmd_lockptr(mm, new_pmd); 1766bf929152SKirill A. Shutemov if (new_ptl != old_ptl) 1767bf929152SKirill A. Shutemov spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 17688809aa2dSAneesh Kumar K.V pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1769a2ce2666SAaron Lu if (pmd_present(pmd) && pmd_dirty(pmd)) 1770a2ce2666SAaron Lu force_flush = true; 177137a1c49aSAndrea Arcangeli VM_BUG_ON(!pmd_none(*new_pmd)); 17723592806cSKirill A. Shutemov 17731dd38b6cSAneesh Kumar K.V if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1774b3084f4dSAneesh Kumar K.V pgtable_t pgtable; 17753592806cSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 17763592806cSKirill A. Shutemov pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 17773592806cSKirill A. Shutemov } 1778b3084f4dSAneesh Kumar K.V set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); 1779b3084f4dSAneesh Kumar K.V if (new_ptl != old_ptl) 1780b3084f4dSAneesh Kumar K.V spin_unlock(new_ptl); 17815d190420SAaron Lu if (force_flush) 17825d190420SAaron Lu flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 17835d190420SAaron Lu else 17845d190420SAaron Lu *need_flush = true; 1785bf929152SKirill A. Shutemov spin_unlock(old_ptl); 17864b471e88SKirill A. Shutemov return true; 178737a1c49aSAndrea Arcangeli } 17884b471e88SKirill A. Shutemov return false; 178937a1c49aSAndrea Arcangeli } 179037a1c49aSAndrea Arcangeli 1791f123d74aSMel Gorman /* 1792f123d74aSMel Gorman * Returns 1793f123d74aSMel Gorman * - 0 if PMD could not be locked 1794f123d74aSMel Gorman * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1795f123d74aSMel Gorman * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1796f123d74aSMel Gorman */ 1797cd7548abSJohannes Weiner int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1798e944fd67SMel Gorman unsigned long addr, pgprot_t newprot, int prot_numa) 1799cd7548abSJohannes Weiner { 1800cd7548abSJohannes Weiner struct mm_struct *mm = vma->vm_mm; 1801bf929152SKirill A. Shutemov spinlock_t *ptl; 18020a85e51dSKirill A. Shutemov pmd_t entry; 18030a85e51dSKirill A. Shutemov bool preserve_write; 18040a85e51dSKirill A. Shutemov int ret; 1805cd7548abSJohannes Weiner 1806b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 18070a85e51dSKirill A. Shutemov if (!ptl) 18080a85e51dSKirill A. Shutemov return 0; 18090a85e51dSKirill A. Shutemov 18100a85e51dSKirill A. Shutemov preserve_write = prot_numa && pmd_write(*pmd); 1811ba68bc01SMel Gorman ret = 1; 1812e944fd67SMel Gorman 1813e944fd67SMel Gorman /* 1814e944fd67SMel Gorman * Avoid trapping faults against the zero page. The read-only 1815e944fd67SMel Gorman * data is likely to be read-cached on the local CPU and 1816e944fd67SMel Gorman * local/remote hits to the zero page are not interesting. 1817e944fd67SMel Gorman */ 18180a85e51dSKirill A. Shutemov if (prot_numa && is_huge_zero_pmd(*pmd)) 18190a85e51dSKirill A. Shutemov goto unlock; 1820e944fd67SMel Gorman 18210a85e51dSKirill A. Shutemov if (prot_numa && pmd_protnone(*pmd)) 18220a85e51dSKirill A. Shutemov goto unlock; 18230a85e51dSKirill A. Shutemov 1824ced10803SKirill A. Shutemov /* 1825ced10803SKirill A. Shutemov * In case prot_numa, we are under down_read(mmap_sem). It's critical 1826ced10803SKirill A. Shutemov * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1827ced10803SKirill A. Shutemov * which is also under down_read(mmap_sem): 1828ced10803SKirill A. Shutemov * 1829ced10803SKirill A. Shutemov * CPU0: CPU1: 1830ced10803SKirill A. Shutemov * change_huge_pmd(prot_numa=1) 1831ced10803SKirill A. Shutemov * pmdp_huge_get_and_clear_notify() 1832ced10803SKirill A. Shutemov * madvise_dontneed() 1833ced10803SKirill A. Shutemov * zap_pmd_range() 1834ced10803SKirill A. Shutemov * pmd_trans_huge(*pmd) == 0 (without ptl) 1835ced10803SKirill A. Shutemov * // skip the pmd 1836ced10803SKirill A. Shutemov * set_pmd_at(); 1837ced10803SKirill A. Shutemov * // pmd is re-established 1838ced10803SKirill A. Shutemov * 1839ced10803SKirill A. Shutemov * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1840ced10803SKirill A. Shutemov * which may break userspace. 1841ced10803SKirill A. Shutemov * 1842ced10803SKirill A. Shutemov * pmdp_invalidate() is required to make sure we don't miss 1843ced10803SKirill A. Shutemov * dirty/young flags set by hardware. 1844ced10803SKirill A. Shutemov */ 1845ced10803SKirill A. Shutemov entry = *pmd; 1846ced10803SKirill A. Shutemov pmdp_invalidate(vma, addr, pmd); 1847ced10803SKirill A. Shutemov 1848ced10803SKirill A. Shutemov /* 1849ced10803SKirill A. Shutemov * Recover dirty/young flags. It relies on pmdp_invalidate to not 1850ced10803SKirill A. Shutemov * corrupt them. 1851ced10803SKirill A. Shutemov */ 1852ced10803SKirill A. Shutemov if (pmd_dirty(*pmd)) 1853ced10803SKirill A. Shutemov entry = pmd_mkdirty(entry); 1854ced10803SKirill A. Shutemov if (pmd_young(*pmd)) 1855ced10803SKirill A. Shutemov entry = pmd_mkyoung(entry); 1856ced10803SKirill A. Shutemov 1857cd7548abSJohannes Weiner entry = pmd_modify(entry, newprot); 1858b191f9b1SMel Gorman if (preserve_write) 1859288bc549SAneesh Kumar K.V entry = pmd_mk_savedwrite(entry); 1860f123d74aSMel Gorman ret = HPAGE_PMD_NR; 186156eecdb9SAneesh Kumar K.V set_pmd_at(mm, addr, pmd, entry); 18620a85e51dSKirill A. Shutemov BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 18630a85e51dSKirill A. Shutemov unlock: 1864bf929152SKirill A. Shutemov spin_unlock(ptl); 1865cd7548abSJohannes Weiner return ret; 1866cd7548abSJohannes Weiner } 1867cd7548abSJohannes Weiner 1868025c5b24SNaoya Horiguchi /* 18698f19b0c0SHuang Ying * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1870025c5b24SNaoya Horiguchi * 18718f19b0c0SHuang Ying * Note that if it returns page table lock pointer, this routine returns without 18728f19b0c0SHuang Ying * unlocking page table lock. So callers must unlock it. 1873025c5b24SNaoya Horiguchi */ 1874b6ec57f4SKirill A. Shutemov spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1875025c5b24SNaoya Horiguchi { 1876b6ec57f4SKirill A. Shutemov spinlock_t *ptl; 1877b6ec57f4SKirill A. Shutemov ptl = pmd_lock(vma->vm_mm, pmd); 18785c7fb56eSDan Williams if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) 1879b6ec57f4SKirill A. Shutemov return ptl; 1880b6ec57f4SKirill A. Shutemov spin_unlock(ptl); 1881b6ec57f4SKirill A. Shutemov return NULL; 1882025c5b24SNaoya Horiguchi } 1883025c5b24SNaoya Horiguchi 1884a00cc7d9SMatthew Wilcox /* 1885a00cc7d9SMatthew Wilcox * Returns true if a given pud maps a thp, false otherwise. 1886a00cc7d9SMatthew Wilcox * 1887a00cc7d9SMatthew Wilcox * Note that if it returns true, this routine returns without unlocking page 1888a00cc7d9SMatthew Wilcox * table lock. So callers must unlock it. 1889a00cc7d9SMatthew Wilcox */ 1890a00cc7d9SMatthew Wilcox spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1891a00cc7d9SMatthew Wilcox { 1892a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1893a00cc7d9SMatthew Wilcox 1894a00cc7d9SMatthew Wilcox ptl = pud_lock(vma->vm_mm, pud); 1895a00cc7d9SMatthew Wilcox if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1896a00cc7d9SMatthew Wilcox return ptl; 1897a00cc7d9SMatthew Wilcox spin_unlock(ptl); 1898a00cc7d9SMatthew Wilcox return NULL; 1899a00cc7d9SMatthew Wilcox } 1900a00cc7d9SMatthew Wilcox 1901a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1902a00cc7d9SMatthew Wilcox int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1903a00cc7d9SMatthew Wilcox pud_t *pud, unsigned long addr) 1904a00cc7d9SMatthew Wilcox { 1905a00cc7d9SMatthew Wilcox pud_t orig_pud; 1906a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1907a00cc7d9SMatthew Wilcox 1908a00cc7d9SMatthew Wilcox ptl = __pud_trans_huge_lock(pud, vma); 1909a00cc7d9SMatthew Wilcox if (!ptl) 1910a00cc7d9SMatthew Wilcox return 0; 1911a00cc7d9SMatthew Wilcox /* 1912a00cc7d9SMatthew Wilcox * For architectures like ppc64 we look at deposited pgtable 1913a00cc7d9SMatthew Wilcox * when calling pudp_huge_get_and_clear. So do the 1914a00cc7d9SMatthew Wilcox * pgtable_trans_huge_withdraw after finishing pudp related 1915a00cc7d9SMatthew Wilcox * operations. 1916a00cc7d9SMatthew Wilcox */ 1917a00cc7d9SMatthew Wilcox orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, 1918a00cc7d9SMatthew Wilcox tlb->fullmm); 1919a00cc7d9SMatthew Wilcox tlb_remove_pud_tlb_entry(tlb, pud, addr); 1920a00cc7d9SMatthew Wilcox if (vma_is_dax(vma)) { 1921a00cc7d9SMatthew Wilcox spin_unlock(ptl); 1922a00cc7d9SMatthew Wilcox /* No zero page support yet */ 1923a00cc7d9SMatthew Wilcox } else { 1924a00cc7d9SMatthew Wilcox /* No support for anonymous PUD pages yet */ 1925a00cc7d9SMatthew Wilcox BUG(); 1926a00cc7d9SMatthew Wilcox } 1927a00cc7d9SMatthew Wilcox return 1; 1928a00cc7d9SMatthew Wilcox } 1929a00cc7d9SMatthew Wilcox 1930a00cc7d9SMatthew Wilcox static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 1931a00cc7d9SMatthew Wilcox unsigned long haddr) 1932a00cc7d9SMatthew Wilcox { 1933a00cc7d9SMatthew Wilcox VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 1934a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 1935a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 1936a00cc7d9SMatthew Wilcox VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 1937a00cc7d9SMatthew Wilcox 1938ce9311cfSYisheng Xie count_vm_event(THP_SPLIT_PUD); 1939a00cc7d9SMatthew Wilcox 1940a00cc7d9SMatthew Wilcox pudp_huge_clear_flush_notify(vma, haddr, pud); 1941a00cc7d9SMatthew Wilcox } 1942a00cc7d9SMatthew Wilcox 1943a00cc7d9SMatthew Wilcox void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 1944a00cc7d9SMatthew Wilcox unsigned long address) 1945a00cc7d9SMatthew Wilcox { 1946a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1947a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 1948a00cc7d9SMatthew Wilcox unsigned long haddr = address & HPAGE_PUD_MASK; 1949a00cc7d9SMatthew Wilcox 1950a00cc7d9SMatthew Wilcox mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE); 1951a00cc7d9SMatthew Wilcox ptl = pud_lock(mm, pud); 1952a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 1953a00cc7d9SMatthew Wilcox goto out; 1954a00cc7d9SMatthew Wilcox __split_huge_pud_locked(vma, pud, haddr); 1955a00cc7d9SMatthew Wilcox 1956a00cc7d9SMatthew Wilcox out: 1957a00cc7d9SMatthew Wilcox spin_unlock(ptl); 1958a00cc7d9SMatthew Wilcox mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE); 1959a00cc7d9SMatthew Wilcox } 1960a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1961a00cc7d9SMatthew Wilcox 1962eef1b3baSKirill A. Shutemov static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 1963eef1b3baSKirill A. Shutemov unsigned long haddr, pmd_t *pmd) 1964eef1b3baSKirill A. Shutemov { 1965eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 1966eef1b3baSKirill A. Shutemov pgtable_t pgtable; 1967eef1b3baSKirill A. Shutemov pmd_t _pmd; 1968eef1b3baSKirill A. Shutemov int i; 1969eef1b3baSKirill A. Shutemov 1970eef1b3baSKirill A. Shutemov /* leave pmd empty until pte is filled */ 1971eef1b3baSKirill A. Shutemov pmdp_huge_clear_flush_notify(vma, haddr, pmd); 1972eef1b3baSKirill A. Shutemov 1973eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1974eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 1975eef1b3baSKirill A. Shutemov 1976eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1977eef1b3baSKirill A. Shutemov pte_t *pte, entry; 1978eef1b3baSKirill A. Shutemov entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 1979eef1b3baSKirill A. Shutemov entry = pte_mkspecial(entry); 1980eef1b3baSKirill A. Shutemov pte = pte_offset_map(&_pmd, haddr); 1981eef1b3baSKirill A. Shutemov VM_BUG_ON(!pte_none(*pte)); 1982eef1b3baSKirill A. Shutemov set_pte_at(mm, haddr, pte, entry); 1983eef1b3baSKirill A. Shutemov pte_unmap(pte); 1984eef1b3baSKirill A. Shutemov } 1985eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 1986eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 1987eef1b3baSKirill A. Shutemov } 1988eef1b3baSKirill A. Shutemov 1989eef1b3baSKirill A. Shutemov static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 1990ba988280SKirill A. Shutemov unsigned long haddr, bool freeze) 1991eef1b3baSKirill A. Shutemov { 1992eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 1993eef1b3baSKirill A. Shutemov struct page *page; 1994eef1b3baSKirill A. Shutemov pgtable_t pgtable; 1995eef1b3baSKirill A. Shutemov pmd_t _pmd; 1996804dd150SAndrea Arcangeli bool young, write, dirty, soft_dirty; 19972ac015e2SKirill A. Shutemov unsigned long addr; 1998eef1b3baSKirill A. Shutemov int i; 1999eef1b3baSKirill A. Shutemov 2000eef1b3baSKirill A. Shutemov VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2001eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2002eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 20035c7fb56eSDan Williams VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)); 2004eef1b3baSKirill A. Shutemov 2005eef1b3baSKirill A. Shutemov count_vm_event(THP_SPLIT_PMD); 2006eef1b3baSKirill A. Shutemov 2007d21b9e57SKirill A. Shutemov if (!vma_is_anonymous(vma)) { 2008d21b9e57SKirill A. Shutemov _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2009953c66c2SAneesh Kumar K.V /* 2010953c66c2SAneesh Kumar K.V * We are going to unmap this huge page. So 2011953c66c2SAneesh Kumar K.V * just go ahead and zap it 2012953c66c2SAneesh Kumar K.V */ 2013953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 2014953c66c2SAneesh Kumar K.V zap_deposited_table(mm, pmd); 2015d21b9e57SKirill A. Shutemov if (vma_is_dax(vma)) 2016d21b9e57SKirill A. Shutemov return; 2017d21b9e57SKirill A. Shutemov page = pmd_page(_pmd); 2018d21b9e57SKirill A. Shutemov if (!PageReferenced(page) && pmd_young(_pmd)) 2019d21b9e57SKirill A. Shutemov SetPageReferenced(page); 2020d21b9e57SKirill A. Shutemov page_remove_rmap(page, true); 2021d21b9e57SKirill A. Shutemov put_page(page); 2022d21b9e57SKirill A. Shutemov add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR); 2023eef1b3baSKirill A. Shutemov return; 2024eef1b3baSKirill A. Shutemov } else if (is_huge_zero_pmd(*pmd)) { 2025eef1b3baSKirill A. Shutemov return __split_huge_zero_page_pmd(vma, haddr, pmd); 2026eef1b3baSKirill A. Shutemov } 2027eef1b3baSKirill A. Shutemov 2028eef1b3baSKirill A. Shutemov page = pmd_page(*pmd); 2029eef1b3baSKirill A. Shutemov VM_BUG_ON_PAGE(!page_count(page), page); 2030fe896d18SJoonsoo Kim page_ref_add(page, HPAGE_PMD_NR - 1); 2031eef1b3baSKirill A. Shutemov write = pmd_write(*pmd); 2032eef1b3baSKirill A. Shutemov young = pmd_young(*pmd); 2033b8d3c4c3SMinchan Kim dirty = pmd_dirty(*pmd); 2034804dd150SAndrea Arcangeli soft_dirty = pmd_soft_dirty(*pmd); 2035eef1b3baSKirill A. Shutemov 2036c777e2a8SAneesh Kumar K.V pmdp_huge_split_prepare(vma, haddr, pmd); 2037eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2038eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2039eef1b3baSKirill A. Shutemov 20402ac015e2SKirill A. Shutemov for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2041eef1b3baSKirill A. Shutemov pte_t entry, *pte; 2042eef1b3baSKirill A. Shutemov /* 2043eef1b3baSKirill A. Shutemov * Note that NUMA hinting access restrictions are not 2044eef1b3baSKirill A. Shutemov * transferred to avoid any possibility of altering 2045eef1b3baSKirill A. Shutemov * permissions across VMAs. 2046eef1b3baSKirill A. Shutemov */ 2047ba988280SKirill A. Shutemov if (freeze) { 2048ba988280SKirill A. Shutemov swp_entry_t swp_entry; 2049ba988280SKirill A. Shutemov swp_entry = make_migration_entry(page + i, write); 2050ba988280SKirill A. Shutemov entry = swp_entry_to_pte(swp_entry); 2051804dd150SAndrea Arcangeli if (soft_dirty) 2052804dd150SAndrea Arcangeli entry = pte_swp_mksoft_dirty(entry); 2053ba988280SKirill A. Shutemov } else { 20546d2329f8SAndrea Arcangeli entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2055b8d3c4c3SMinchan Kim entry = maybe_mkwrite(entry, vma); 2056eef1b3baSKirill A. Shutemov if (!write) 2057eef1b3baSKirill A. Shutemov entry = pte_wrprotect(entry); 2058eef1b3baSKirill A. Shutemov if (!young) 2059eef1b3baSKirill A. Shutemov entry = pte_mkold(entry); 2060804dd150SAndrea Arcangeli if (soft_dirty) 2061804dd150SAndrea Arcangeli entry = pte_mksoft_dirty(entry); 2062ba988280SKirill A. Shutemov } 2063b8d3c4c3SMinchan Kim if (dirty) 2064b8d3c4c3SMinchan Kim SetPageDirty(page + i); 20652ac015e2SKirill A. Shutemov pte = pte_offset_map(&_pmd, addr); 2066eef1b3baSKirill A. Shutemov BUG_ON(!pte_none(*pte)); 20672ac015e2SKirill A. Shutemov set_pte_at(mm, addr, pte, entry); 2068eef1b3baSKirill A. Shutemov atomic_inc(&page[i]._mapcount); 2069eef1b3baSKirill A. Shutemov pte_unmap(pte); 2070eef1b3baSKirill A. Shutemov } 2071eef1b3baSKirill A. Shutemov 2072eef1b3baSKirill A. Shutemov /* 2073eef1b3baSKirill A. Shutemov * Set PG_double_map before dropping compound_mapcount to avoid 2074eef1b3baSKirill A. Shutemov * false-negative page_mapped(). 2075eef1b3baSKirill A. Shutemov */ 2076eef1b3baSKirill A. Shutemov if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 2077eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 2078eef1b3baSKirill A. Shutemov atomic_inc(&page[i]._mapcount); 2079eef1b3baSKirill A. Shutemov } 2080eef1b3baSKirill A. Shutemov 2081eef1b3baSKirill A. Shutemov if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2082eef1b3baSKirill A. Shutemov /* Last compound_mapcount is gone. */ 208311fb9989SMel Gorman __dec_node_page_state(page, NR_ANON_THPS); 2084eef1b3baSKirill A. Shutemov if (TestClearPageDoubleMap(page)) { 2085eef1b3baSKirill A. Shutemov /* No need in mapcount reference anymore */ 2086eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 2087eef1b3baSKirill A. Shutemov atomic_dec(&page[i]._mapcount); 2088eef1b3baSKirill A. Shutemov } 2089eef1b3baSKirill A. Shutemov } 2090eef1b3baSKirill A. Shutemov 2091eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2092e9b61f19SKirill A. Shutemov /* 2093e9b61f19SKirill A. Shutemov * Up to this point the pmd is present and huge and userland has the 2094e9b61f19SKirill A. Shutemov * whole access to the hugepage during the split (which happens in 2095e9b61f19SKirill A. Shutemov * place). If we overwrite the pmd with the not-huge version pointing 2096e9b61f19SKirill A. Shutemov * to the pte here (which of course we could if all CPUs were bug 2097e9b61f19SKirill A. Shutemov * free), userland could trigger a small page size TLB miss on the 2098e9b61f19SKirill A. Shutemov * small sized TLB while the hugepage TLB entry is still established in 2099e9b61f19SKirill A. Shutemov * the huge TLB. Some CPU doesn't like that. 2100e9b61f19SKirill A. Shutemov * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 2101e9b61f19SKirill A. Shutemov * 383 on page 93. Intel should be safe but is also warns that it's 2102e9b61f19SKirill A. Shutemov * only safe if the permission and cache attributes of the two entries 2103e9b61f19SKirill A. Shutemov * loaded in the two TLB is identical (which should be the case here). 2104e9b61f19SKirill A. Shutemov * But it is generally safer to never allow small and huge TLB entries 2105e9b61f19SKirill A. Shutemov * for the same virtual address to be loaded simultaneously. So instead 2106e9b61f19SKirill A. Shutemov * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2107e9b61f19SKirill A. Shutemov * current pmd notpresent (atomically because here the pmd_trans_huge 2108e9b61f19SKirill A. Shutemov * and pmd_trans_splitting must remain set at all times on the pmd 2109e9b61f19SKirill A. Shutemov * until the split is complete for this pmd), then we flush the SMP TLB 2110e9b61f19SKirill A. Shutemov * and finally we write the non-huge version of the pmd entry with 2111e9b61f19SKirill A. Shutemov * pmd_populate. 2112e9b61f19SKirill A. Shutemov */ 2113e9b61f19SKirill A. Shutemov pmdp_invalidate(vma, haddr, pmd); 2114eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 2115e9b61f19SKirill A. Shutemov 2116e9b61f19SKirill A. Shutemov if (freeze) { 21172ac015e2SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) { 2118e9b61f19SKirill A. Shutemov page_remove_rmap(page + i, false); 2119e9b61f19SKirill A. Shutemov put_page(page + i); 2120e9b61f19SKirill A. Shutemov } 2121e9b61f19SKirill A. Shutemov } 2122eef1b3baSKirill A. Shutemov } 2123eef1b3baSKirill A. Shutemov 2124eef1b3baSKirill A. Shutemov void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 212533f4751eSNaoya Horiguchi unsigned long address, bool freeze, struct page *page) 2126eef1b3baSKirill A. Shutemov { 2127eef1b3baSKirill A. Shutemov spinlock_t *ptl; 2128eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2129eef1b3baSKirill A. Shutemov unsigned long haddr = address & HPAGE_PMD_MASK; 2130eef1b3baSKirill A. Shutemov 2131eef1b3baSKirill A. Shutemov mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); 2132eef1b3baSKirill A. Shutemov ptl = pmd_lock(mm, pmd); 213333f4751eSNaoya Horiguchi 213433f4751eSNaoya Horiguchi /* 213533f4751eSNaoya Horiguchi * If caller asks to setup a migration entries, we need a page to check 213633f4751eSNaoya Horiguchi * pmd against. Otherwise we can end up replacing wrong page. 213733f4751eSNaoya Horiguchi */ 213833f4751eSNaoya Horiguchi VM_BUG_ON(freeze && !page); 213933f4751eSNaoya Horiguchi if (page && page != pmd_page(*pmd)) 214033f4751eSNaoya Horiguchi goto out; 214133f4751eSNaoya Horiguchi 21425c7fb56eSDan Williams if (pmd_trans_huge(*pmd)) { 214333f4751eSNaoya Horiguchi page = pmd_page(*pmd); 2144e90309c9SKirill A. Shutemov if (PageMlocked(page)) 21455f737714SKirill A. Shutemov clear_page_mlock(page); 21465c7fb56eSDan Williams } else if (!pmd_devmap(*pmd)) 21475c7fb56eSDan Williams goto out; 2148fec89c10SKirill A. Shutemov __split_huge_pmd_locked(vma, pmd, haddr, freeze); 2149e90309c9SKirill A. Shutemov out: 2150eef1b3baSKirill A. Shutemov spin_unlock(ptl); 2151eef1b3baSKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); 2152eef1b3baSKirill A. Shutemov } 2153eef1b3baSKirill A. Shutemov 2154fec89c10SKirill A. Shutemov void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2155fec89c10SKirill A. Shutemov bool freeze, struct page *page) 215694fcc585SAndrea Arcangeli { 2157f72e7dcdSHugh Dickins pgd_t *pgd; 2158c2febafcSKirill A. Shutemov p4d_t *p4d; 2159f72e7dcdSHugh Dickins pud_t *pud; 216094fcc585SAndrea Arcangeli pmd_t *pmd; 216194fcc585SAndrea Arcangeli 216278ddc534SKirill A. Shutemov pgd = pgd_offset(vma->vm_mm, address); 2163f72e7dcdSHugh Dickins if (!pgd_present(*pgd)) 2164f72e7dcdSHugh Dickins return; 2165f72e7dcdSHugh Dickins 2166c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, address); 2167c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 2168c2febafcSKirill A. Shutemov return; 2169c2febafcSKirill A. Shutemov 2170c2febafcSKirill A. Shutemov pud = pud_offset(p4d, address); 2171f72e7dcdSHugh Dickins if (!pud_present(*pud)) 2172f72e7dcdSHugh Dickins return; 2173f72e7dcdSHugh Dickins 2174f72e7dcdSHugh Dickins pmd = pmd_offset(pud, address); 2175fec89c10SKirill A. Shutemov 217633f4751eSNaoya Horiguchi __split_huge_pmd(vma, pmd, address, freeze, page); 217794fcc585SAndrea Arcangeli } 217894fcc585SAndrea Arcangeli 2179e1b9996bSKirill A. Shutemov void vma_adjust_trans_huge(struct vm_area_struct *vma, 218094fcc585SAndrea Arcangeli unsigned long start, 218194fcc585SAndrea Arcangeli unsigned long end, 218294fcc585SAndrea Arcangeli long adjust_next) 218394fcc585SAndrea Arcangeli { 218494fcc585SAndrea Arcangeli /* 218594fcc585SAndrea Arcangeli * If the new start address isn't hpage aligned and it could 218694fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 218794fcc585SAndrea Arcangeli * an huge pmd. 218894fcc585SAndrea Arcangeli */ 218994fcc585SAndrea Arcangeli if (start & ~HPAGE_PMD_MASK && 219094fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) >= vma->vm_start && 219194fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2192fec89c10SKirill A. Shutemov split_huge_pmd_address(vma, start, false, NULL); 219394fcc585SAndrea Arcangeli 219494fcc585SAndrea Arcangeli /* 219594fcc585SAndrea Arcangeli * If the new end address isn't hpage aligned and it could 219694fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 219794fcc585SAndrea Arcangeli * an huge pmd. 219894fcc585SAndrea Arcangeli */ 219994fcc585SAndrea Arcangeli if (end & ~HPAGE_PMD_MASK && 220094fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) >= vma->vm_start && 220194fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2202fec89c10SKirill A. Shutemov split_huge_pmd_address(vma, end, false, NULL); 220394fcc585SAndrea Arcangeli 220494fcc585SAndrea Arcangeli /* 220594fcc585SAndrea Arcangeli * If we're also updating the vma->vm_next->vm_start, if the new 220694fcc585SAndrea Arcangeli * vm_next->vm_start isn't page aligned and it could previously 220794fcc585SAndrea Arcangeli * contain an hugepage: check if we need to split an huge pmd. 220894fcc585SAndrea Arcangeli */ 220994fcc585SAndrea Arcangeli if (adjust_next > 0) { 221094fcc585SAndrea Arcangeli struct vm_area_struct *next = vma->vm_next; 221194fcc585SAndrea Arcangeli unsigned long nstart = next->vm_start; 221294fcc585SAndrea Arcangeli nstart += adjust_next << PAGE_SHIFT; 221394fcc585SAndrea Arcangeli if (nstart & ~HPAGE_PMD_MASK && 221494fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) >= next->vm_start && 221594fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2216fec89c10SKirill A. Shutemov split_huge_pmd_address(next, nstart, false, NULL); 221794fcc585SAndrea Arcangeli } 221894fcc585SAndrea Arcangeli } 2219e9b61f19SKirill A. Shutemov 2220fec89c10SKirill A. Shutemov static void freeze_page(struct page *page) 2221e9b61f19SKirill A. Shutemov { 2222baa355fdSKirill A. Shutemov enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 2223c7ab0d2fSKirill A. Shutemov TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2224666e5a40SMinchan Kim bool unmap_success; 2225e9b61f19SKirill A. Shutemov 2226e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageHead(page), page); 2227e9b61f19SKirill A. Shutemov 2228baa355fdSKirill A. Shutemov if (PageAnon(page)) 2229b5ff8161SNaoya Horiguchi ttu_flags |= TTU_SPLIT_FREEZE; 2230baa355fdSKirill A. Shutemov 2231666e5a40SMinchan Kim unmap_success = try_to_unmap(page, ttu_flags); 2232666e5a40SMinchan Kim VM_BUG_ON_PAGE(!unmap_success, page); 2233bd56086fSKirill A. Shutemov } 2234bd56086fSKirill A. Shutemov 2235fec89c10SKirill A. Shutemov static void unfreeze_page(struct page *page) 2236e9b61f19SKirill A. Shutemov { 2237fec89c10SKirill A. Shutemov int i; 2238ace71a19SKirill A. Shutemov if (PageTransHuge(page)) { 2239ace71a19SKirill A. Shutemov remove_migration_ptes(page, page, true); 2240ace71a19SKirill A. Shutemov } else { 2241fec89c10SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 2242fec89c10SKirill A. Shutemov remove_migration_ptes(page + i, page + i, true); 2243e9b61f19SKirill A. Shutemov } 2244ace71a19SKirill A. Shutemov } 2245e9b61f19SKirill A. Shutemov 22468df651c7SKirill A. Shutemov static void __split_huge_page_tail(struct page *head, int tail, 2247e9b61f19SKirill A. Shutemov struct lruvec *lruvec, struct list_head *list) 2248e9b61f19SKirill A. Shutemov { 2249e9b61f19SKirill A. Shutemov struct page *page_tail = head + tail; 2250e9b61f19SKirill A. Shutemov 22518df651c7SKirill A. Shutemov VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2252fe896d18SJoonsoo Kim VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail); 2253e9b61f19SKirill A. Shutemov 2254e9b61f19SKirill A. Shutemov /* 22550139aa7bSJoonsoo Kim * tail_page->_refcount is zero and not changing from under us. But 2256e9b61f19SKirill A. Shutemov * get_page_unless_zero() may be running from under us on the 2257baa355fdSKirill A. Shutemov * tail_page. If we used atomic_set() below instead of atomic_inc() or 2258baa355fdSKirill A. Shutemov * atomic_add(), we would then run atomic_set() concurrently with 2259e9b61f19SKirill A. Shutemov * get_page_unless_zero(), and atomic_set() is implemented in C not 2260e9b61f19SKirill A. Shutemov * using locked ops. spin_unlock on x86 sometime uses locked ops 2261e9b61f19SKirill A. Shutemov * because of PPro errata 66, 92, so unless somebody can guarantee 2262e9b61f19SKirill A. Shutemov * atomic_set() here would be safe on all archs (and not only on x86), 2263baa355fdSKirill A. Shutemov * it's safer to use atomic_inc()/atomic_add(). 2264e9b61f19SKirill A. Shutemov */ 226538d8b4e6SHuang Ying if (PageAnon(head) && !PageSwapCache(head)) { 2266fe896d18SJoonsoo Kim page_ref_inc(page_tail); 2267baa355fdSKirill A. Shutemov } else { 2268baa355fdSKirill A. Shutemov /* Additional pin to radix tree */ 2269baa355fdSKirill A. Shutemov page_ref_add(page_tail, 2); 2270baa355fdSKirill A. Shutemov } 2271e9b61f19SKirill A. Shutemov 2272e9b61f19SKirill A. Shutemov page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2273e9b61f19SKirill A. Shutemov page_tail->flags |= (head->flags & 2274e9b61f19SKirill A. Shutemov ((1L << PG_referenced) | 2275e9b61f19SKirill A. Shutemov (1L << PG_swapbacked) | 227638d8b4e6SHuang Ying (1L << PG_swapcache) | 2277e9b61f19SKirill A. Shutemov (1L << PG_mlocked) | 2278e9b61f19SKirill A. Shutemov (1L << PG_uptodate) | 2279e9b61f19SKirill A. Shutemov (1L << PG_active) | 2280e9b61f19SKirill A. Shutemov (1L << PG_locked) | 2281b8d3c4c3SMinchan Kim (1L << PG_unevictable) | 2282b8d3c4c3SMinchan Kim (1L << PG_dirty))); 2283e9b61f19SKirill A. Shutemov 2284e9b61f19SKirill A. Shutemov /* 2285e9b61f19SKirill A. Shutemov * After clearing PageTail the gup refcount can be released. 2286e9b61f19SKirill A. Shutemov * Page flags also must be visible before we make the page non-compound. 2287e9b61f19SKirill A. Shutemov */ 2288e9b61f19SKirill A. Shutemov smp_wmb(); 2289e9b61f19SKirill A. Shutemov 2290e9b61f19SKirill A. Shutemov clear_compound_head(page_tail); 2291e9b61f19SKirill A. Shutemov 2292e9b61f19SKirill A. Shutemov if (page_is_young(head)) 2293e9b61f19SKirill A. Shutemov set_page_young(page_tail); 2294e9b61f19SKirill A. Shutemov if (page_is_idle(head)) 2295e9b61f19SKirill A. Shutemov set_page_idle(page_tail); 2296e9b61f19SKirill A. Shutemov 2297e9b61f19SKirill A. Shutemov /* ->mapping in first tail page is compound_mapcount */ 22989a982250SKirill A. Shutemov VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2299e9b61f19SKirill A. Shutemov page_tail); 2300e9b61f19SKirill A. Shutemov page_tail->mapping = head->mapping; 2301e9b61f19SKirill A. Shutemov 2302e9b61f19SKirill A. Shutemov page_tail->index = head->index + tail; 2303e9b61f19SKirill A. Shutemov page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2304e9b61f19SKirill A. Shutemov lru_add_page_tail(head, page_tail, lruvec, list); 2305e9b61f19SKirill A. Shutemov } 2306e9b61f19SKirill A. Shutemov 2307baa355fdSKirill A. Shutemov static void __split_huge_page(struct page *page, struct list_head *list, 2308baa355fdSKirill A. Shutemov unsigned long flags) 2309e9b61f19SKirill A. Shutemov { 2310e9b61f19SKirill A. Shutemov struct page *head = compound_head(page); 2311e9b61f19SKirill A. Shutemov struct zone *zone = page_zone(head); 2312e9b61f19SKirill A. Shutemov struct lruvec *lruvec; 2313baa355fdSKirill A. Shutemov pgoff_t end = -1; 23148df651c7SKirill A. Shutemov int i; 2315e9b61f19SKirill A. Shutemov 2316599d0c95SMel Gorman lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); 2317e9b61f19SKirill A. Shutemov 2318e9b61f19SKirill A. Shutemov /* complete memcg works before add pages to LRU */ 2319e9b61f19SKirill A. Shutemov mem_cgroup_split_huge_fixup(head); 2320e9b61f19SKirill A. Shutemov 2321baa355fdSKirill A. Shutemov if (!PageAnon(page)) 2322baa355fdSKirill A. Shutemov end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE); 2323baa355fdSKirill A. Shutemov 2324baa355fdSKirill A. Shutemov for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 23258df651c7SKirill A. Shutemov __split_huge_page_tail(head, i, lruvec, list); 2326baa355fdSKirill A. Shutemov /* Some pages can be beyond i_size: drop them from page cache */ 2327baa355fdSKirill A. Shutemov if (head[i].index >= end) { 2328baa355fdSKirill A. Shutemov __ClearPageDirty(head + i); 2329baa355fdSKirill A. Shutemov __delete_from_page_cache(head + i, NULL); 2330800d8c63SKirill A. Shutemov if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2331800d8c63SKirill A. Shutemov shmem_uncharge(head->mapping->host, 1); 2332baa355fdSKirill A. Shutemov put_page(head + i); 2333baa355fdSKirill A. Shutemov } 2334baa355fdSKirill A. Shutemov } 2335e9b61f19SKirill A. Shutemov 2336e9b61f19SKirill A. Shutemov ClearPageCompound(head); 2337baa355fdSKirill A. Shutemov /* See comment in __split_huge_page_tail() */ 2338baa355fdSKirill A. Shutemov if (PageAnon(head)) { 233938d8b4e6SHuang Ying /* Additional pin to radix tree of swap cache */ 234038d8b4e6SHuang Ying if (PageSwapCache(head)) 234138d8b4e6SHuang Ying page_ref_add(head, 2); 234238d8b4e6SHuang Ying else 2343baa355fdSKirill A. Shutemov page_ref_inc(head); 2344baa355fdSKirill A. Shutemov } else { 2345baa355fdSKirill A. Shutemov /* Additional pin to radix tree */ 2346baa355fdSKirill A. Shutemov page_ref_add(head, 2); 2347baa355fdSKirill A. Shutemov spin_unlock(&head->mapping->tree_lock); 2348baa355fdSKirill A. Shutemov } 2349baa355fdSKirill A. Shutemov 2350a52633d8SMel Gorman spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2351e9b61f19SKirill A. Shutemov 2352fec89c10SKirill A. Shutemov unfreeze_page(head); 2353e9b61f19SKirill A. Shutemov 2354e9b61f19SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) { 2355e9b61f19SKirill A. Shutemov struct page *subpage = head + i; 2356e9b61f19SKirill A. Shutemov if (subpage == page) 2357e9b61f19SKirill A. Shutemov continue; 2358e9b61f19SKirill A. Shutemov unlock_page(subpage); 2359e9b61f19SKirill A. Shutemov 2360e9b61f19SKirill A. Shutemov /* 2361e9b61f19SKirill A. Shutemov * Subpages may be freed if there wasn't any mapping 2362e9b61f19SKirill A. Shutemov * like if add_to_swap() is running on a lru page that 2363e9b61f19SKirill A. Shutemov * had its mapping zapped. And freeing these pages 2364e9b61f19SKirill A. Shutemov * requires taking the lru_lock so we do the put_page 2365e9b61f19SKirill A. Shutemov * of the tail pages after the split is complete. 2366e9b61f19SKirill A. Shutemov */ 2367e9b61f19SKirill A. Shutemov put_page(subpage); 2368e9b61f19SKirill A. Shutemov } 2369e9b61f19SKirill A. Shutemov } 2370e9b61f19SKirill A. Shutemov 2371b20ce5e0SKirill A. Shutemov int total_mapcount(struct page *page) 2372b20ce5e0SKirill A. Shutemov { 2373dd78feddSKirill A. Shutemov int i, compound, ret; 2374b20ce5e0SKirill A. Shutemov 2375b20ce5e0SKirill A. Shutemov VM_BUG_ON_PAGE(PageTail(page), page); 2376b20ce5e0SKirill A. Shutemov 2377b20ce5e0SKirill A. Shutemov if (likely(!PageCompound(page))) 2378b20ce5e0SKirill A. Shutemov return atomic_read(&page->_mapcount) + 1; 2379b20ce5e0SKirill A. Shutemov 2380dd78feddSKirill A. Shutemov compound = compound_mapcount(page); 2381b20ce5e0SKirill A. Shutemov if (PageHuge(page)) 2382dd78feddSKirill A. Shutemov return compound; 2383dd78feddSKirill A. Shutemov ret = compound; 2384b20ce5e0SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 2385b20ce5e0SKirill A. Shutemov ret += atomic_read(&page[i]._mapcount) + 1; 2386dd78feddSKirill A. Shutemov /* File pages has compound_mapcount included in _mapcount */ 2387dd78feddSKirill A. Shutemov if (!PageAnon(page)) 2388dd78feddSKirill A. Shutemov return ret - compound * HPAGE_PMD_NR; 2389b20ce5e0SKirill A. Shutemov if (PageDoubleMap(page)) 2390b20ce5e0SKirill A. Shutemov ret -= HPAGE_PMD_NR; 2391b20ce5e0SKirill A. Shutemov return ret; 2392b20ce5e0SKirill A. Shutemov } 2393b20ce5e0SKirill A. Shutemov 2394e9b61f19SKirill A. Shutemov /* 23956d0a07edSAndrea Arcangeli * This calculates accurately how many mappings a transparent hugepage 23966d0a07edSAndrea Arcangeli * has (unlike page_mapcount() which isn't fully accurate). This full 23976d0a07edSAndrea Arcangeli * accuracy is primarily needed to know if copy-on-write faults can 23986d0a07edSAndrea Arcangeli * reuse the page and change the mapping to read-write instead of 23996d0a07edSAndrea Arcangeli * copying them. At the same time this returns the total_mapcount too. 24006d0a07edSAndrea Arcangeli * 24016d0a07edSAndrea Arcangeli * The function returns the highest mapcount any one of the subpages 24026d0a07edSAndrea Arcangeli * has. If the return value is one, even if different processes are 24036d0a07edSAndrea Arcangeli * mapping different subpages of the transparent hugepage, they can 24046d0a07edSAndrea Arcangeli * all reuse it, because each process is reusing a different subpage. 24056d0a07edSAndrea Arcangeli * 24066d0a07edSAndrea Arcangeli * The total_mapcount is instead counting all virtual mappings of the 24076d0a07edSAndrea Arcangeli * subpages. If the total_mapcount is equal to "one", it tells the 24086d0a07edSAndrea Arcangeli * caller all mappings belong to the same "mm" and in turn the 24096d0a07edSAndrea Arcangeli * anon_vma of the transparent hugepage can become the vma->anon_vma 24106d0a07edSAndrea Arcangeli * local one as no other process may be mapping any of the subpages. 24116d0a07edSAndrea Arcangeli * 24126d0a07edSAndrea Arcangeli * It would be more accurate to replace page_mapcount() with 24136d0a07edSAndrea Arcangeli * page_trans_huge_mapcount(), however we only use 24146d0a07edSAndrea Arcangeli * page_trans_huge_mapcount() in the copy-on-write faults where we 24156d0a07edSAndrea Arcangeli * need full accuracy to avoid breaking page pinning, because 24166d0a07edSAndrea Arcangeli * page_trans_huge_mapcount() is slower than page_mapcount(). 24176d0a07edSAndrea Arcangeli */ 24186d0a07edSAndrea Arcangeli int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 24196d0a07edSAndrea Arcangeli { 24206d0a07edSAndrea Arcangeli int i, ret, _total_mapcount, mapcount; 24216d0a07edSAndrea Arcangeli 24226d0a07edSAndrea Arcangeli /* hugetlbfs shouldn't call it */ 24236d0a07edSAndrea Arcangeli VM_BUG_ON_PAGE(PageHuge(page), page); 24246d0a07edSAndrea Arcangeli 24256d0a07edSAndrea Arcangeli if (likely(!PageTransCompound(page))) { 24266d0a07edSAndrea Arcangeli mapcount = atomic_read(&page->_mapcount) + 1; 24276d0a07edSAndrea Arcangeli if (total_mapcount) 24286d0a07edSAndrea Arcangeli *total_mapcount = mapcount; 24296d0a07edSAndrea Arcangeli return mapcount; 24306d0a07edSAndrea Arcangeli } 24316d0a07edSAndrea Arcangeli 24326d0a07edSAndrea Arcangeli page = compound_head(page); 24336d0a07edSAndrea Arcangeli 24346d0a07edSAndrea Arcangeli _total_mapcount = ret = 0; 24356d0a07edSAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 24366d0a07edSAndrea Arcangeli mapcount = atomic_read(&page[i]._mapcount) + 1; 24376d0a07edSAndrea Arcangeli ret = max(ret, mapcount); 24386d0a07edSAndrea Arcangeli _total_mapcount += mapcount; 24396d0a07edSAndrea Arcangeli } 24406d0a07edSAndrea Arcangeli if (PageDoubleMap(page)) { 24416d0a07edSAndrea Arcangeli ret -= 1; 24426d0a07edSAndrea Arcangeli _total_mapcount -= HPAGE_PMD_NR; 24436d0a07edSAndrea Arcangeli } 24446d0a07edSAndrea Arcangeli mapcount = compound_mapcount(page); 24456d0a07edSAndrea Arcangeli ret += mapcount; 24466d0a07edSAndrea Arcangeli _total_mapcount += mapcount; 24476d0a07edSAndrea Arcangeli if (total_mapcount) 24486d0a07edSAndrea Arcangeli *total_mapcount = _total_mapcount; 24496d0a07edSAndrea Arcangeli return ret; 24506d0a07edSAndrea Arcangeli } 24516d0a07edSAndrea Arcangeli 2452b8f593cdSHuang Ying /* Racy check whether the huge page can be split */ 2453b8f593cdSHuang Ying bool can_split_huge_page(struct page *page, int *pextra_pins) 2454b8f593cdSHuang Ying { 2455b8f593cdSHuang Ying int extra_pins; 2456b8f593cdSHuang Ying 2457b8f593cdSHuang Ying /* Additional pins from radix tree */ 2458b8f593cdSHuang Ying if (PageAnon(page)) 2459b8f593cdSHuang Ying extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; 2460b8f593cdSHuang Ying else 2461b8f593cdSHuang Ying extra_pins = HPAGE_PMD_NR; 2462b8f593cdSHuang Ying if (pextra_pins) 2463b8f593cdSHuang Ying *pextra_pins = extra_pins; 2464b8f593cdSHuang Ying return total_mapcount(page) == page_count(page) - extra_pins - 1; 2465b8f593cdSHuang Ying } 2466b8f593cdSHuang Ying 24676d0a07edSAndrea Arcangeli /* 2468e9b61f19SKirill A. Shutemov * This function splits huge page into normal pages. @page can point to any 2469e9b61f19SKirill A. Shutemov * subpage of huge page to split. Split doesn't change the position of @page. 2470e9b61f19SKirill A. Shutemov * 2471e9b61f19SKirill A. Shutemov * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2472e9b61f19SKirill A. Shutemov * The huge page must be locked. 2473e9b61f19SKirill A. Shutemov * 2474e9b61f19SKirill A. Shutemov * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2475e9b61f19SKirill A. Shutemov * 2476e9b61f19SKirill A. Shutemov * Both head page and tail pages will inherit mapping, flags, and so on from 2477e9b61f19SKirill A. Shutemov * the hugepage. 2478e9b61f19SKirill A. Shutemov * 2479e9b61f19SKirill A. Shutemov * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2480e9b61f19SKirill A. Shutemov * they are not mapped. 2481e9b61f19SKirill A. Shutemov * 2482e9b61f19SKirill A. Shutemov * Returns 0 if the hugepage is split successfully. 2483e9b61f19SKirill A. Shutemov * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2484e9b61f19SKirill A. Shutemov * us. 2485e9b61f19SKirill A. Shutemov */ 2486e9b61f19SKirill A. Shutemov int split_huge_page_to_list(struct page *page, struct list_head *list) 2487e9b61f19SKirill A. Shutemov { 2488e9b61f19SKirill A. Shutemov struct page *head = compound_head(page); 2489a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 2490baa355fdSKirill A. Shutemov struct anon_vma *anon_vma = NULL; 2491baa355fdSKirill A. Shutemov struct address_space *mapping = NULL; 2492baa355fdSKirill A. Shutemov int count, mapcount, extra_pins, ret; 2493d9654322SKirill A. Shutemov bool mlocked; 24940b9b6fffSKirill A. Shutemov unsigned long flags; 2495e9b61f19SKirill A. Shutemov 2496e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(is_huge_zero_page(page), page); 2497e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 2498e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageCompound(page), page); 2499e9b61f19SKirill A. Shutemov 250059807685SHuang Ying if (PageWriteback(page)) 250159807685SHuang Ying return -EBUSY; 250259807685SHuang Ying 2503baa355fdSKirill A. Shutemov if (PageAnon(head)) { 2504e9b61f19SKirill A. Shutemov /* 2505baa355fdSKirill A. Shutemov * The caller does not necessarily hold an mmap_sem that would 2506baa355fdSKirill A. Shutemov * prevent the anon_vma disappearing so we first we take a 2507baa355fdSKirill A. Shutemov * reference to it and then lock the anon_vma for write. This 2508baa355fdSKirill A. Shutemov * is similar to page_lock_anon_vma_read except the write lock 2509baa355fdSKirill A. Shutemov * is taken to serialise against parallel split or collapse 2510baa355fdSKirill A. Shutemov * operations. 2511e9b61f19SKirill A. Shutemov */ 2512e9b61f19SKirill A. Shutemov anon_vma = page_get_anon_vma(head); 2513e9b61f19SKirill A. Shutemov if (!anon_vma) { 2514e9b61f19SKirill A. Shutemov ret = -EBUSY; 2515e9b61f19SKirill A. Shutemov goto out; 2516e9b61f19SKirill A. Shutemov } 2517baa355fdSKirill A. Shutemov mapping = NULL; 2518e9b61f19SKirill A. Shutemov anon_vma_lock_write(anon_vma); 2519baa355fdSKirill A. Shutemov } else { 2520baa355fdSKirill A. Shutemov mapping = head->mapping; 2521baa355fdSKirill A. Shutemov 2522baa355fdSKirill A. Shutemov /* Truncated ? */ 2523baa355fdSKirill A. Shutemov if (!mapping) { 2524baa355fdSKirill A. Shutemov ret = -EBUSY; 2525baa355fdSKirill A. Shutemov goto out; 2526baa355fdSKirill A. Shutemov } 2527baa355fdSKirill A. Shutemov 2528baa355fdSKirill A. Shutemov anon_vma = NULL; 2529baa355fdSKirill A. Shutemov i_mmap_lock_read(mapping); 2530baa355fdSKirill A. Shutemov } 2531e9b61f19SKirill A. Shutemov 2532e9b61f19SKirill A. Shutemov /* 2533e9b61f19SKirill A. Shutemov * Racy check if we can split the page, before freeze_page() will 2534e9b61f19SKirill A. Shutemov * split PMDs 2535e9b61f19SKirill A. Shutemov */ 2536b8f593cdSHuang Ying if (!can_split_huge_page(head, &extra_pins)) { 2537e9b61f19SKirill A. Shutemov ret = -EBUSY; 2538e9b61f19SKirill A. Shutemov goto out_unlock; 2539e9b61f19SKirill A. Shutemov } 2540e9b61f19SKirill A. Shutemov 2541d9654322SKirill A. Shutemov mlocked = PageMlocked(page); 2542fec89c10SKirill A. Shutemov freeze_page(head); 2543e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(compound_mapcount(head), head); 2544e9b61f19SKirill A. Shutemov 2545d9654322SKirill A. Shutemov /* Make sure the page is not on per-CPU pagevec as it takes pin */ 2546d9654322SKirill A. Shutemov if (mlocked) 2547d9654322SKirill A. Shutemov lru_add_drain(); 2548d9654322SKirill A. Shutemov 2549baa355fdSKirill A. Shutemov /* prevent PageLRU to go away from under us, and freeze lru stats */ 2550a52633d8SMel Gorman spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); 2551baa355fdSKirill A. Shutemov 2552baa355fdSKirill A. Shutemov if (mapping) { 2553baa355fdSKirill A. Shutemov void **pslot; 2554baa355fdSKirill A. Shutemov 2555baa355fdSKirill A. Shutemov spin_lock(&mapping->tree_lock); 2556baa355fdSKirill A. Shutemov pslot = radix_tree_lookup_slot(&mapping->page_tree, 2557baa355fdSKirill A. Shutemov page_index(head)); 2558baa355fdSKirill A. Shutemov /* 2559baa355fdSKirill A. Shutemov * Check if the head page is present in radix tree. 2560baa355fdSKirill A. Shutemov * We assume all tail are present too, if head is there. 2561baa355fdSKirill A. Shutemov */ 2562baa355fdSKirill A. Shutemov if (radix_tree_deref_slot_protected(pslot, 2563baa355fdSKirill A. Shutemov &mapping->tree_lock) != head) 2564baa355fdSKirill A. Shutemov goto fail; 2565baa355fdSKirill A. Shutemov } 2566baa355fdSKirill A. Shutemov 25670139aa7bSJoonsoo Kim /* Prevent deferred_split_scan() touching ->_refcount */ 2568baa355fdSKirill A. Shutemov spin_lock(&pgdata->split_queue_lock); 2569e9b61f19SKirill A. Shutemov count = page_count(head); 2570e9b61f19SKirill A. Shutemov mapcount = total_mapcount(head); 2571baa355fdSKirill A. Shutemov if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 25729a982250SKirill A. Shutemov if (!list_empty(page_deferred_list(head))) { 2573a3d0a918SKirill A. Shutemov pgdata->split_queue_len--; 25749a982250SKirill A. Shutemov list_del(page_deferred_list(head)); 25759a982250SKirill A. Shutemov } 257665c45377SKirill A. Shutemov if (mapping) 257711fb9989SMel Gorman __dec_node_page_state(page, NR_SHMEM_THPS); 2578baa355fdSKirill A. Shutemov spin_unlock(&pgdata->split_queue_lock); 2579baa355fdSKirill A. Shutemov __split_huge_page(page, list, flags); 258059807685SHuang Ying if (PageSwapCache(head)) { 258159807685SHuang Ying swp_entry_t entry = { .val = page_private(head) }; 258259807685SHuang Ying 258359807685SHuang Ying ret = split_swap_cluster(entry); 258459807685SHuang Ying } else 2585e9b61f19SKirill A. Shutemov ret = 0; 2586baa355fdSKirill A. Shutemov } else { 2587baa355fdSKirill A. Shutemov if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2588e9b61f19SKirill A. Shutemov pr_alert("total_mapcount: %u, page_count(): %u\n", 2589e9b61f19SKirill A. Shutemov mapcount, count); 2590e9b61f19SKirill A. Shutemov if (PageTail(page)) 2591e9b61f19SKirill A. Shutemov dump_page(head, NULL); 2592bd56086fSKirill A. Shutemov dump_page(page, "total_mapcount(head) > 0"); 2593e9b61f19SKirill A. Shutemov BUG(); 2594baa355fdSKirill A. Shutemov } 2595baa355fdSKirill A. Shutemov spin_unlock(&pgdata->split_queue_lock); 2596baa355fdSKirill A. Shutemov fail: if (mapping) 2597baa355fdSKirill A. Shutemov spin_unlock(&mapping->tree_lock); 2598a52633d8SMel Gorman spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2599fec89c10SKirill A. Shutemov unfreeze_page(head); 2600e9b61f19SKirill A. Shutemov ret = -EBUSY; 2601e9b61f19SKirill A. Shutemov } 2602e9b61f19SKirill A. Shutemov 2603e9b61f19SKirill A. Shutemov out_unlock: 2604baa355fdSKirill A. Shutemov if (anon_vma) { 2605e9b61f19SKirill A. Shutemov anon_vma_unlock_write(anon_vma); 2606e9b61f19SKirill A. Shutemov put_anon_vma(anon_vma); 2607baa355fdSKirill A. Shutemov } 2608baa355fdSKirill A. Shutemov if (mapping) 2609baa355fdSKirill A. Shutemov i_mmap_unlock_read(mapping); 2610e9b61f19SKirill A. Shutemov out: 2611e9b61f19SKirill A. Shutemov count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2612e9b61f19SKirill A. Shutemov return ret; 2613e9b61f19SKirill A. Shutemov } 26149a982250SKirill A. Shutemov 26159a982250SKirill A. Shutemov void free_transhuge_page(struct page *page) 26169a982250SKirill A. Shutemov { 2617a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 26189a982250SKirill A. Shutemov unsigned long flags; 26199a982250SKirill A. Shutemov 2620a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 26219a982250SKirill A. Shutemov if (!list_empty(page_deferred_list(page))) { 2622a3d0a918SKirill A. Shutemov pgdata->split_queue_len--; 26239a982250SKirill A. Shutemov list_del(page_deferred_list(page)); 26249a982250SKirill A. Shutemov } 2625a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 26269a982250SKirill A. Shutemov free_compound_page(page); 26279a982250SKirill A. Shutemov } 26289a982250SKirill A. Shutemov 26299a982250SKirill A. Shutemov void deferred_split_huge_page(struct page *page) 26309a982250SKirill A. Shutemov { 2631a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 26329a982250SKirill A. Shutemov unsigned long flags; 26339a982250SKirill A. Shutemov 26349a982250SKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 26359a982250SKirill A. Shutemov 2636a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 26379a982250SKirill A. Shutemov if (list_empty(page_deferred_list(page))) { 2638f9719a03SKirill A. Shutemov count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2639a3d0a918SKirill A. Shutemov list_add_tail(page_deferred_list(page), &pgdata->split_queue); 2640a3d0a918SKirill A. Shutemov pgdata->split_queue_len++; 26419a982250SKirill A. Shutemov } 2642a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 26439a982250SKirill A. Shutemov } 26449a982250SKirill A. Shutemov 26459a982250SKirill A. Shutemov static unsigned long deferred_split_count(struct shrinker *shrink, 26469a982250SKirill A. Shutemov struct shrink_control *sc) 26479a982250SKirill A. Shutemov { 2648a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 2649cb8d68ecSKirill A. Shutemov return ACCESS_ONCE(pgdata->split_queue_len); 26509a982250SKirill A. Shutemov } 26519a982250SKirill A. Shutemov 26529a982250SKirill A. Shutemov static unsigned long deferred_split_scan(struct shrinker *shrink, 26539a982250SKirill A. Shutemov struct shrink_control *sc) 26549a982250SKirill A. Shutemov { 2655a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 26569a982250SKirill A. Shutemov unsigned long flags; 26579a982250SKirill A. Shutemov LIST_HEAD(list), *pos, *next; 26589a982250SKirill A. Shutemov struct page *page; 26599a982250SKirill A. Shutemov int split = 0; 26609a982250SKirill A. Shutemov 2661a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 26629a982250SKirill A. Shutemov /* Take pin on all head pages to avoid freeing them under us */ 2663ae026204SKirill A. Shutemov list_for_each_safe(pos, next, &pgdata->split_queue) { 26649a982250SKirill A. Shutemov page = list_entry((void *)pos, struct page, mapping); 26659a982250SKirill A. Shutemov page = compound_head(page); 2666e3ae1953SKirill A. Shutemov if (get_page_unless_zero(page)) { 2667e3ae1953SKirill A. Shutemov list_move(page_deferred_list(page), &list); 2668e3ae1953SKirill A. Shutemov } else { 2669e3ae1953SKirill A. Shutemov /* We lost race with put_compound_page() */ 26709a982250SKirill A. Shutemov list_del_init(page_deferred_list(page)); 2671a3d0a918SKirill A. Shutemov pgdata->split_queue_len--; 26729a982250SKirill A. Shutemov } 2673e3ae1953SKirill A. Shutemov if (!--sc->nr_to_scan) 2674e3ae1953SKirill A. Shutemov break; 26759a982250SKirill A. Shutemov } 2676a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 26779a982250SKirill A. Shutemov 26789a982250SKirill A. Shutemov list_for_each_safe(pos, next, &list) { 26799a982250SKirill A. Shutemov page = list_entry((void *)pos, struct page, mapping); 26809a982250SKirill A. Shutemov lock_page(page); 26819a982250SKirill A. Shutemov /* split_huge_page() removes page from list on success */ 26829a982250SKirill A. Shutemov if (!split_huge_page(page)) 26839a982250SKirill A. Shutemov split++; 26849a982250SKirill A. Shutemov unlock_page(page); 26859a982250SKirill A. Shutemov put_page(page); 26869a982250SKirill A. Shutemov } 26879a982250SKirill A. Shutemov 2688a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2689a3d0a918SKirill A. Shutemov list_splice_tail(&list, &pgdata->split_queue); 2690a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 26919a982250SKirill A. Shutemov 2692cb8d68ecSKirill A. Shutemov /* 2693cb8d68ecSKirill A. Shutemov * Stop shrinker if we didn't split any page, but the queue is empty. 2694cb8d68ecSKirill A. Shutemov * This can happen if pages were freed under us. 2695cb8d68ecSKirill A. Shutemov */ 2696cb8d68ecSKirill A. Shutemov if (!split && list_empty(&pgdata->split_queue)) 2697cb8d68ecSKirill A. Shutemov return SHRINK_STOP; 2698cb8d68ecSKirill A. Shutemov return split; 26999a982250SKirill A. Shutemov } 27009a982250SKirill A. Shutemov 27019a982250SKirill A. Shutemov static struct shrinker deferred_split_shrinker = { 27029a982250SKirill A. Shutemov .count_objects = deferred_split_count, 27039a982250SKirill A. Shutemov .scan_objects = deferred_split_scan, 27049a982250SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 2705a3d0a918SKirill A. Shutemov .flags = SHRINKER_NUMA_AWARE, 27069a982250SKirill A. Shutemov }; 270749071d43SKirill A. Shutemov 270849071d43SKirill A. Shutemov #ifdef CONFIG_DEBUG_FS 270949071d43SKirill A. Shutemov static int split_huge_pages_set(void *data, u64 val) 271049071d43SKirill A. Shutemov { 271149071d43SKirill A. Shutemov struct zone *zone; 271249071d43SKirill A. Shutemov struct page *page; 271349071d43SKirill A. Shutemov unsigned long pfn, max_zone_pfn; 271449071d43SKirill A. Shutemov unsigned long total = 0, split = 0; 271549071d43SKirill A. Shutemov 271649071d43SKirill A. Shutemov if (val != 1) 271749071d43SKirill A. Shutemov return -EINVAL; 271849071d43SKirill A. Shutemov 271949071d43SKirill A. Shutemov for_each_populated_zone(zone) { 272049071d43SKirill A. Shutemov max_zone_pfn = zone_end_pfn(zone); 272149071d43SKirill A. Shutemov for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 272249071d43SKirill A. Shutemov if (!pfn_valid(pfn)) 272349071d43SKirill A. Shutemov continue; 272449071d43SKirill A. Shutemov 272549071d43SKirill A. Shutemov page = pfn_to_page(pfn); 272649071d43SKirill A. Shutemov if (!get_page_unless_zero(page)) 272749071d43SKirill A. Shutemov continue; 272849071d43SKirill A. Shutemov 272949071d43SKirill A. Shutemov if (zone != page_zone(page)) 273049071d43SKirill A. Shutemov goto next; 273149071d43SKirill A. Shutemov 2732baa355fdSKirill A. Shutemov if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 273349071d43SKirill A. Shutemov goto next; 273449071d43SKirill A. Shutemov 273549071d43SKirill A. Shutemov total++; 273649071d43SKirill A. Shutemov lock_page(page); 273749071d43SKirill A. Shutemov if (!split_huge_page(page)) 273849071d43SKirill A. Shutemov split++; 273949071d43SKirill A. Shutemov unlock_page(page); 274049071d43SKirill A. Shutemov next: 274149071d43SKirill A. Shutemov put_page(page); 274249071d43SKirill A. Shutemov } 274349071d43SKirill A. Shutemov } 274449071d43SKirill A. Shutemov 2745145bdaa1SYang Shi pr_info("%lu of %lu THP split\n", split, total); 274649071d43SKirill A. Shutemov 274749071d43SKirill A. Shutemov return 0; 274849071d43SKirill A. Shutemov } 274949071d43SKirill A. Shutemov DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 275049071d43SKirill A. Shutemov "%llu\n"); 275149071d43SKirill A. Shutemov 275249071d43SKirill A. Shutemov static int __init split_huge_pages_debugfs(void) 275349071d43SKirill A. Shutemov { 275449071d43SKirill A. Shutemov void *ret; 275549071d43SKirill A. Shutemov 2756145bdaa1SYang Shi ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 275749071d43SKirill A. Shutemov &split_huge_pages_fops); 275849071d43SKirill A. Shutemov if (!ret) 275949071d43SKirill A. Shutemov pr_warn("Failed to create split_huge_pages in debugfs"); 276049071d43SKirill A. Shutemov return 0; 276149071d43SKirill A. Shutemov } 276249071d43SKirill A. Shutemov late_initcall(split_huge_pages_debugfs); 276349071d43SKirill A. Shutemov #endif 2764*616b8371SZi Yan 2765*616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2766*616b8371SZi Yan void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 2767*616b8371SZi Yan struct page *page) 2768*616b8371SZi Yan { 2769*616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 2770*616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 2771*616b8371SZi Yan unsigned long address = pvmw->address; 2772*616b8371SZi Yan pmd_t pmdval; 2773*616b8371SZi Yan swp_entry_t entry; 2774*616b8371SZi Yan 2775*616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 2776*616b8371SZi Yan return; 2777*616b8371SZi Yan 2778*616b8371SZi Yan mmu_notifier_invalidate_range_start(mm, address, 2779*616b8371SZi Yan address + HPAGE_PMD_SIZE); 2780*616b8371SZi Yan 2781*616b8371SZi Yan flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 2782*616b8371SZi Yan pmdval = *pvmw->pmd; 2783*616b8371SZi Yan pmdp_invalidate(vma, address, pvmw->pmd); 2784*616b8371SZi Yan if (pmd_dirty(pmdval)) 2785*616b8371SZi Yan set_page_dirty(page); 2786*616b8371SZi Yan entry = make_migration_entry(page, pmd_write(pmdval)); 2787*616b8371SZi Yan pmdval = swp_entry_to_pmd(entry); 2788*616b8371SZi Yan set_pmd_at(mm, address, pvmw->pmd, pmdval); 2789*616b8371SZi Yan page_remove_rmap(page, true); 2790*616b8371SZi Yan put_page(page); 2791*616b8371SZi Yan 2792*616b8371SZi Yan mmu_notifier_invalidate_range_end(mm, address, 2793*616b8371SZi Yan address + HPAGE_PMD_SIZE); 2794*616b8371SZi Yan } 2795*616b8371SZi Yan 2796*616b8371SZi Yan void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 2797*616b8371SZi Yan { 2798*616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 2799*616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 2800*616b8371SZi Yan unsigned long address = pvmw->address; 2801*616b8371SZi Yan unsigned long mmun_start = address & HPAGE_PMD_MASK; 2802*616b8371SZi Yan pmd_t pmde; 2803*616b8371SZi Yan swp_entry_t entry; 2804*616b8371SZi Yan 2805*616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 2806*616b8371SZi Yan return; 2807*616b8371SZi Yan 2808*616b8371SZi Yan entry = pmd_to_swp_entry(*pvmw->pmd); 2809*616b8371SZi Yan get_page(new); 2810*616b8371SZi Yan pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 2811*616b8371SZi Yan if (is_write_migration_entry(entry)) 2812*616b8371SZi Yan pmde = maybe_pmd_mkwrite(pmde, vma); 2813*616b8371SZi Yan 2814*616b8371SZi Yan flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 2815*616b8371SZi Yan page_add_anon_rmap(new, vma, mmun_start, true); 2816*616b8371SZi Yan set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 2817*616b8371SZi Yan if (vma->vm_flags & VM_LOCKED) 2818*616b8371SZi Yan mlock_vma_page(new); 2819*616b8371SZi Yan update_mmu_cache_pmd(vma, address, pvmw->pmd); 2820*616b8371SZi Yan } 2821*616b8371SZi Yan #endif 2822