171e3aac0SAndrea Arcangeli /* 271e3aac0SAndrea Arcangeli * Copyright (C) 2009 Red Hat, Inc. 371e3aac0SAndrea Arcangeli * 471e3aac0SAndrea Arcangeli * This work is licensed under the terms of the GNU GPL, version 2. See 571e3aac0SAndrea Arcangeli * the COPYING file in the top-level directory. 671e3aac0SAndrea Arcangeli */ 771e3aac0SAndrea Arcangeli 8ae3a8c1cSAndrew Morton #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 9ae3a8c1cSAndrew Morton 1071e3aac0SAndrea Arcangeli #include <linux/mm.h> 1171e3aac0SAndrea Arcangeli #include <linux/sched.h> 1271e3aac0SAndrea Arcangeli #include <linux/highmem.h> 1371e3aac0SAndrea Arcangeli #include <linux/hugetlb.h> 1471e3aac0SAndrea Arcangeli #include <linux/mmu_notifier.h> 1571e3aac0SAndrea Arcangeli #include <linux/rmap.h> 1671e3aac0SAndrea Arcangeli #include <linux/swap.h> 1797ae1749SKirill A. Shutemov #include <linux/shrinker.h> 18ba76149fSAndrea Arcangeli #include <linux/mm_inline.h> 19e9b61f19SKirill A. Shutemov #include <linux/swapops.h> 204897c765SMatthew Wilcox #include <linux/dax.h> 21ba76149fSAndrea Arcangeli #include <linux/khugepaged.h> 22878aee7dSAndrea Arcangeli #include <linux/freezer.h> 23f25748e3SDan Williams #include <linux/pfn_t.h> 24a664b2d8SAndrea Arcangeli #include <linux/mman.h> 253565fce3SDan Williams #include <linux/memremap.h> 26325adeb5SRalf Baechle #include <linux/pagemap.h> 2749071d43SKirill A. Shutemov #include <linux/debugfs.h> 284daae3b4SMel Gorman #include <linux/migrate.h> 2943b5fbbdSSasha Levin #include <linux/hashtable.h> 306b251fc9SAndrea Arcangeli #include <linux/userfaultfd_k.h> 3133c3fc71SVladimir Davydov #include <linux/page_idle.h> 32baa355fdSKirill A. Shutemov #include <linux/shmem_fs.h> 3397ae1749SKirill A. Shutemov 3471e3aac0SAndrea Arcangeli #include <asm/tlb.h> 3571e3aac0SAndrea Arcangeli #include <asm/pgalloc.h> 3671e3aac0SAndrea Arcangeli #include "internal.h" 3771e3aac0SAndrea Arcangeli 38ba76149fSAndrea Arcangeli /* 398bfa3f9aSJianguo Wu * By default transparent hugepage support is disabled in order that avoid 408bfa3f9aSJianguo Wu * to risk increase the memory footprint of applications without a guaranteed 418bfa3f9aSJianguo Wu * benefit. When transparent hugepage support is enabled, is for all mappings, 428bfa3f9aSJianguo Wu * and khugepaged scans all mappings. 438bfa3f9aSJianguo Wu * Defrag is invoked by khugepaged hugepage allocations and by page faults 448bfa3f9aSJianguo Wu * for all hugepage allocations. 45ba76149fSAndrea Arcangeli */ 4671e3aac0SAndrea Arcangeli unsigned long transparent_hugepage_flags __read_mostly = 4713ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 48ba76149fSAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_FLAG)| 4913ece886SAndrea Arcangeli #endif 5013ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 5113ece886SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 5213ece886SAndrea Arcangeli #endif 53444eb2a4SMel Gorman (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 5479da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 5579da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 56ba76149fSAndrea Arcangeli 579a982250SKirill A. Shutemov static struct shrinker deferred_split_shrinker; 58f000565aSAndrea Arcangeli 5997ae1749SKirill A. Shutemov static atomic_t huge_zero_refcount; 6056873f43SWang, Yalin struct page *huge_zero_page __read_mostly; 614a6c1297SKirill A. Shutemov 62fc437044SMatthew Wilcox struct page *get_huge_zero_page(void) 6397ae1749SKirill A. Shutemov { 6497ae1749SKirill A. Shutemov struct page *zero_page; 6597ae1749SKirill A. Shutemov retry: 6697ae1749SKirill A. Shutemov if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 674db0c3c2SJason Low return READ_ONCE(huge_zero_page); 6897ae1749SKirill A. Shutemov 6997ae1749SKirill A. Shutemov zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 7097ae1749SKirill A. Shutemov HPAGE_PMD_ORDER); 71d8a8e1f0SKirill A. Shutemov if (!zero_page) { 72d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 735918d10aSKirill A. Shutemov return NULL; 74d8a8e1f0SKirill A. Shutemov } 75d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC); 7697ae1749SKirill A. Shutemov preempt_disable(); 775918d10aSKirill A. Shutemov if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 7897ae1749SKirill A. Shutemov preempt_enable(); 795ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 8097ae1749SKirill A. Shutemov goto retry; 8197ae1749SKirill A. Shutemov } 8297ae1749SKirill A. Shutemov 8397ae1749SKirill A. Shutemov /* We take additional reference here. It will be put back by shrinker */ 8497ae1749SKirill A. Shutemov atomic_set(&huge_zero_refcount, 2); 8597ae1749SKirill A. Shutemov preempt_enable(); 864db0c3c2SJason Low return READ_ONCE(huge_zero_page); 8797ae1749SKirill A. Shutemov } 8897ae1749SKirill A. Shutemov 89aa88b68cSKirill A. Shutemov void put_huge_zero_page(void) 9097ae1749SKirill A. Shutemov { 9197ae1749SKirill A. Shutemov /* 9297ae1749SKirill A. Shutemov * Counter should never go to zero here. Only shrinker can put 9397ae1749SKirill A. Shutemov * last reference. 9497ae1749SKirill A. Shutemov */ 9597ae1749SKirill A. Shutemov BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 9697ae1749SKirill A. Shutemov } 9797ae1749SKirill A. Shutemov 9848896466SGlauber Costa static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 9997ae1749SKirill A. Shutemov struct shrink_control *sc) 10097ae1749SKirill A. Shutemov { 10197ae1749SKirill A. Shutemov /* we can free zero page only if last reference remains */ 10297ae1749SKirill A. Shutemov return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 10348896466SGlauber Costa } 10497ae1749SKirill A. Shutemov 10548896466SGlauber Costa static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 10648896466SGlauber Costa struct shrink_control *sc) 10748896466SGlauber Costa { 10897ae1749SKirill A. Shutemov if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 1095918d10aSKirill A. Shutemov struct page *zero_page = xchg(&huge_zero_page, NULL); 1105918d10aSKirill A. Shutemov BUG_ON(zero_page == NULL); 1115ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 11248896466SGlauber Costa return HPAGE_PMD_NR; 11397ae1749SKirill A. Shutemov } 11497ae1749SKirill A. Shutemov 11597ae1749SKirill A. Shutemov return 0; 11697ae1749SKirill A. Shutemov } 11797ae1749SKirill A. Shutemov 11897ae1749SKirill A. Shutemov static struct shrinker huge_zero_page_shrinker = { 11948896466SGlauber Costa .count_objects = shrink_huge_zero_page_count, 12048896466SGlauber Costa .scan_objects = shrink_huge_zero_page_scan, 12197ae1749SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 12297ae1749SKirill A. Shutemov }; 12397ae1749SKirill A. Shutemov 12471e3aac0SAndrea Arcangeli #ifdef CONFIG_SYSFS 125ba76149fSAndrea Arcangeli 126444eb2a4SMel Gorman static ssize_t triple_flag_store(struct kobject *kobj, 12771e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 12871e3aac0SAndrea Arcangeli const char *buf, size_t count, 12971e3aac0SAndrea Arcangeli enum transparent_hugepage_flag enabled, 130444eb2a4SMel Gorman enum transparent_hugepage_flag deferred, 13171e3aac0SAndrea Arcangeli enum transparent_hugepage_flag req_madv) 13271e3aac0SAndrea Arcangeli { 133444eb2a4SMel Gorman if (!memcmp("defer", buf, 134444eb2a4SMel Gorman min(sizeof("defer")-1, count))) { 135444eb2a4SMel Gorman if (enabled == deferred) 136444eb2a4SMel Gorman return -EINVAL; 137444eb2a4SMel Gorman clear_bit(enabled, &transparent_hugepage_flags); 13871e3aac0SAndrea Arcangeli clear_bit(req_madv, &transparent_hugepage_flags); 139444eb2a4SMel Gorman set_bit(deferred, &transparent_hugepage_flags); 140444eb2a4SMel Gorman } else if (!memcmp("always", buf, 141444eb2a4SMel Gorman min(sizeof("always")-1, count))) { 142444eb2a4SMel Gorman clear_bit(deferred, &transparent_hugepage_flags); 143444eb2a4SMel Gorman clear_bit(req_madv, &transparent_hugepage_flags); 144444eb2a4SMel Gorman set_bit(enabled, &transparent_hugepage_flags); 14571e3aac0SAndrea Arcangeli } else if (!memcmp("madvise", buf, 14671e3aac0SAndrea Arcangeli min(sizeof("madvise")-1, count))) { 14771e3aac0SAndrea Arcangeli clear_bit(enabled, &transparent_hugepage_flags); 148444eb2a4SMel Gorman clear_bit(deferred, &transparent_hugepage_flags); 14971e3aac0SAndrea Arcangeli set_bit(req_madv, &transparent_hugepage_flags); 15071e3aac0SAndrea Arcangeli } else if (!memcmp("never", buf, 15171e3aac0SAndrea Arcangeli min(sizeof("never")-1, count))) { 15271e3aac0SAndrea Arcangeli clear_bit(enabled, &transparent_hugepage_flags); 15371e3aac0SAndrea Arcangeli clear_bit(req_madv, &transparent_hugepage_flags); 154444eb2a4SMel Gorman clear_bit(deferred, &transparent_hugepage_flags); 15571e3aac0SAndrea Arcangeli } else 15671e3aac0SAndrea Arcangeli return -EINVAL; 15771e3aac0SAndrea Arcangeli 15871e3aac0SAndrea Arcangeli return count; 15971e3aac0SAndrea Arcangeli } 16071e3aac0SAndrea Arcangeli 16171e3aac0SAndrea Arcangeli static ssize_t enabled_show(struct kobject *kobj, 16271e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 16371e3aac0SAndrea Arcangeli { 164444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 165444eb2a4SMel Gorman return sprintf(buf, "[always] madvise never\n"); 166444eb2a4SMel Gorman else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) 167444eb2a4SMel Gorman return sprintf(buf, "always [madvise] never\n"); 168444eb2a4SMel Gorman else 169444eb2a4SMel Gorman return sprintf(buf, "always madvise [never]\n"); 17071e3aac0SAndrea Arcangeli } 171444eb2a4SMel Gorman 17271e3aac0SAndrea Arcangeli static ssize_t enabled_store(struct kobject *kobj, 17371e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 17471e3aac0SAndrea Arcangeli const char *buf, size_t count) 17571e3aac0SAndrea Arcangeli { 176ba76149fSAndrea Arcangeli ssize_t ret; 177ba76149fSAndrea Arcangeli 178444eb2a4SMel Gorman ret = triple_flag_store(kobj, attr, buf, count, 179444eb2a4SMel Gorman TRANSPARENT_HUGEPAGE_FLAG, 18071e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_FLAG, 18171e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 182ba76149fSAndrea Arcangeli 183ba76149fSAndrea Arcangeli if (ret > 0) { 184b46e756fSKirill A. Shutemov int err = start_stop_khugepaged(); 185ba76149fSAndrea Arcangeli if (err) 186ba76149fSAndrea Arcangeli ret = err; 187ba76149fSAndrea Arcangeli } 188ba76149fSAndrea Arcangeli 189ba76149fSAndrea Arcangeli return ret; 19071e3aac0SAndrea Arcangeli } 19171e3aac0SAndrea Arcangeli static struct kobj_attribute enabled_attr = 19271e3aac0SAndrea Arcangeli __ATTR(enabled, 0644, enabled_show, enabled_store); 19371e3aac0SAndrea Arcangeli 194b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_show(struct kobject *kobj, 19571e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 19671e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 19771e3aac0SAndrea Arcangeli { 198e27e6151SBen Hutchings return sprintf(buf, "%d\n", 199e27e6151SBen Hutchings !!test_bit(flag, &transparent_hugepage_flags)); 20071e3aac0SAndrea Arcangeli } 201e27e6151SBen Hutchings 202b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_store(struct kobject *kobj, 20371e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 20471e3aac0SAndrea Arcangeli const char *buf, size_t count, 20571e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 20671e3aac0SAndrea Arcangeli { 207e27e6151SBen Hutchings unsigned long value; 208e27e6151SBen Hutchings int ret; 209e27e6151SBen Hutchings 210e27e6151SBen Hutchings ret = kstrtoul(buf, 10, &value); 211e27e6151SBen Hutchings if (ret < 0) 212e27e6151SBen Hutchings return ret; 213e27e6151SBen Hutchings if (value > 1) 21471e3aac0SAndrea Arcangeli return -EINVAL; 21571e3aac0SAndrea Arcangeli 216e27e6151SBen Hutchings if (value) 217e27e6151SBen Hutchings set_bit(flag, &transparent_hugepage_flags); 218e27e6151SBen Hutchings else 219e27e6151SBen Hutchings clear_bit(flag, &transparent_hugepage_flags); 220e27e6151SBen Hutchings 22171e3aac0SAndrea Arcangeli return count; 22271e3aac0SAndrea Arcangeli } 22371e3aac0SAndrea Arcangeli 22471e3aac0SAndrea Arcangeli /* 22571e3aac0SAndrea Arcangeli * Currently defrag only disables __GFP_NOWAIT for allocation. A blind 22671e3aac0SAndrea Arcangeli * __GFP_REPEAT is too aggressive, it's never worth swapping tons of 22771e3aac0SAndrea Arcangeli * memory just to allocate one more hugepage. 22871e3aac0SAndrea Arcangeli */ 22971e3aac0SAndrea Arcangeli static ssize_t defrag_show(struct kobject *kobj, 23071e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 23171e3aac0SAndrea Arcangeli { 232444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 233444eb2a4SMel Gorman return sprintf(buf, "[always] defer madvise never\n"); 234444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 235444eb2a4SMel Gorman return sprintf(buf, "always [defer] madvise never\n"); 236444eb2a4SMel Gorman else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 237444eb2a4SMel Gorman return sprintf(buf, "always defer [madvise] never\n"); 238444eb2a4SMel Gorman else 239444eb2a4SMel Gorman return sprintf(buf, "always defer madvise [never]\n"); 240444eb2a4SMel Gorman 24171e3aac0SAndrea Arcangeli } 24271e3aac0SAndrea Arcangeli static ssize_t defrag_store(struct kobject *kobj, 24371e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 24471e3aac0SAndrea Arcangeli const char *buf, size_t count) 24571e3aac0SAndrea Arcangeli { 246444eb2a4SMel Gorman return triple_flag_store(kobj, attr, buf, count, 247444eb2a4SMel Gorman TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, 248444eb2a4SMel Gorman TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, 24971e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 25071e3aac0SAndrea Arcangeli } 25171e3aac0SAndrea Arcangeli static struct kobj_attribute defrag_attr = 25271e3aac0SAndrea Arcangeli __ATTR(defrag, 0644, defrag_show, defrag_store); 25371e3aac0SAndrea Arcangeli 25479da5407SKirill A. Shutemov static ssize_t use_zero_page_show(struct kobject *kobj, 25579da5407SKirill A. Shutemov struct kobj_attribute *attr, char *buf) 25679da5407SKirill A. Shutemov { 257b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 25879da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 25979da5407SKirill A. Shutemov } 26079da5407SKirill A. Shutemov static ssize_t use_zero_page_store(struct kobject *kobj, 26179da5407SKirill A. Shutemov struct kobj_attribute *attr, const char *buf, size_t count) 26279da5407SKirill A. Shutemov { 263b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 26479da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 26579da5407SKirill A. Shutemov } 26679da5407SKirill A. Shutemov static struct kobj_attribute use_zero_page_attr = 26779da5407SKirill A. Shutemov __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 26871e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 26971e3aac0SAndrea Arcangeli static ssize_t debug_cow_show(struct kobject *kobj, 27071e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 27171e3aac0SAndrea Arcangeli { 272b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 27371e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 27471e3aac0SAndrea Arcangeli } 27571e3aac0SAndrea Arcangeli static ssize_t debug_cow_store(struct kobject *kobj, 27671e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 27771e3aac0SAndrea Arcangeli const char *buf, size_t count) 27871e3aac0SAndrea Arcangeli { 279b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 28071e3aac0SAndrea Arcangeli TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 28171e3aac0SAndrea Arcangeli } 28271e3aac0SAndrea Arcangeli static struct kobj_attribute debug_cow_attr = 28371e3aac0SAndrea Arcangeli __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 28471e3aac0SAndrea Arcangeli #endif /* CONFIG_DEBUG_VM */ 28571e3aac0SAndrea Arcangeli 28671e3aac0SAndrea Arcangeli static struct attribute *hugepage_attr[] = { 28771e3aac0SAndrea Arcangeli &enabled_attr.attr, 28871e3aac0SAndrea Arcangeli &defrag_attr.attr, 28979da5407SKirill A. Shutemov &use_zero_page_attr.attr, 290e496cf3dSKirill A. Shutemov #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) 2915a6e75f8SKirill A. Shutemov &shmem_enabled_attr.attr, 2925a6e75f8SKirill A. Shutemov #endif 29371e3aac0SAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 29471e3aac0SAndrea Arcangeli &debug_cow_attr.attr, 29571e3aac0SAndrea Arcangeli #endif 29671e3aac0SAndrea Arcangeli NULL, 29771e3aac0SAndrea Arcangeli }; 29871e3aac0SAndrea Arcangeli 29971e3aac0SAndrea Arcangeli static struct attribute_group hugepage_attr_group = { 30071e3aac0SAndrea Arcangeli .attrs = hugepage_attr, 301ba76149fSAndrea Arcangeli }; 302ba76149fSAndrea Arcangeli 303569e5590SShaohua Li static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 304569e5590SShaohua Li { 305569e5590SShaohua Li int err; 306569e5590SShaohua Li 307569e5590SShaohua Li *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 308569e5590SShaohua Li if (unlikely(!*hugepage_kobj)) { 309ae3a8c1cSAndrew Morton pr_err("failed to create transparent hugepage kobject\n"); 310569e5590SShaohua Li return -ENOMEM; 311569e5590SShaohua Li } 312569e5590SShaohua Li 313569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 314569e5590SShaohua Li if (err) { 315ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 316569e5590SShaohua Li goto delete_obj; 317569e5590SShaohua Li } 318569e5590SShaohua Li 319569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 320569e5590SShaohua Li if (err) { 321ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 322569e5590SShaohua Li goto remove_hp_group; 323569e5590SShaohua Li } 324569e5590SShaohua Li 325569e5590SShaohua Li return 0; 326569e5590SShaohua Li 327569e5590SShaohua Li remove_hp_group: 328569e5590SShaohua Li sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 329569e5590SShaohua Li delete_obj: 330569e5590SShaohua Li kobject_put(*hugepage_kobj); 331569e5590SShaohua Li return err; 332569e5590SShaohua Li } 333569e5590SShaohua Li 334569e5590SShaohua Li static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 335569e5590SShaohua Li { 336569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 337569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 338569e5590SShaohua Li kobject_put(hugepage_kobj); 339569e5590SShaohua Li } 340569e5590SShaohua Li #else 341569e5590SShaohua Li static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 342569e5590SShaohua Li { 343569e5590SShaohua Li return 0; 344569e5590SShaohua Li } 345569e5590SShaohua Li 346569e5590SShaohua Li static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 347569e5590SShaohua Li { 348569e5590SShaohua Li } 34971e3aac0SAndrea Arcangeli #endif /* CONFIG_SYSFS */ 35071e3aac0SAndrea Arcangeli 35171e3aac0SAndrea Arcangeli static int __init hugepage_init(void) 35271e3aac0SAndrea Arcangeli { 35371e3aac0SAndrea Arcangeli int err; 354569e5590SShaohua Li struct kobject *hugepage_kobj; 35571e3aac0SAndrea Arcangeli 3564b7167b9SAndrea Arcangeli if (!has_transparent_hugepage()) { 3574b7167b9SAndrea Arcangeli transparent_hugepage_flags = 0; 358569e5590SShaohua Li return -EINVAL; 3594b7167b9SAndrea Arcangeli } 3604b7167b9SAndrea Arcangeli 361ff20c2e0SKirill A. Shutemov /* 362ff20c2e0SKirill A. Shutemov * hugepages can't be allocated by the buddy allocator 363ff20c2e0SKirill A. Shutemov */ 364ff20c2e0SKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 365ff20c2e0SKirill A. Shutemov /* 366ff20c2e0SKirill A. Shutemov * we use page->mapping and page->index in second tail page 367ff20c2e0SKirill A. Shutemov * as list_head: assuming THP order >= 2 368ff20c2e0SKirill A. Shutemov */ 369ff20c2e0SKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 370ff20c2e0SKirill A. Shutemov 371569e5590SShaohua Li err = hugepage_init_sysfs(&hugepage_kobj); 372569e5590SShaohua Li if (err) 37365ebb64fSKirill A. Shutemov goto err_sysfs; 374ba76149fSAndrea Arcangeli 375b46e756fSKirill A. Shutemov err = khugepaged_init(); 376ba76149fSAndrea Arcangeli if (err) 37765ebb64fSKirill A. Shutemov goto err_slab; 378ba76149fSAndrea Arcangeli 37965ebb64fSKirill A. Shutemov err = register_shrinker(&huge_zero_page_shrinker); 38065ebb64fSKirill A. Shutemov if (err) 38165ebb64fSKirill A. Shutemov goto err_hzp_shrinker; 3829a982250SKirill A. Shutemov err = register_shrinker(&deferred_split_shrinker); 3839a982250SKirill A. Shutemov if (err) 3849a982250SKirill A. Shutemov goto err_split_shrinker; 38597ae1749SKirill A. Shutemov 38697562cd2SRik van Riel /* 38797562cd2SRik van Riel * By default disable transparent hugepages on smaller systems, 38897562cd2SRik van Riel * where the extra memory used could hurt more than TLB overhead 38997562cd2SRik van Riel * is likely to save. The admin can still enable it through /sys. 39097562cd2SRik van Riel */ 39179553da2SKirill A. Shutemov if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { 39297562cd2SRik van Riel transparent_hugepage_flags = 0; 39379553da2SKirill A. Shutemov return 0; 39479553da2SKirill A. Shutemov } 39597562cd2SRik van Riel 39679553da2SKirill A. Shutemov err = start_stop_khugepaged(); 39765ebb64fSKirill A. Shutemov if (err) 39865ebb64fSKirill A. Shutemov goto err_khugepaged; 399ba76149fSAndrea Arcangeli 400569e5590SShaohua Li return 0; 40165ebb64fSKirill A. Shutemov err_khugepaged: 4029a982250SKirill A. Shutemov unregister_shrinker(&deferred_split_shrinker); 4039a982250SKirill A. Shutemov err_split_shrinker: 40465ebb64fSKirill A. Shutemov unregister_shrinker(&huge_zero_page_shrinker); 40565ebb64fSKirill A. Shutemov err_hzp_shrinker: 406b46e756fSKirill A. Shutemov khugepaged_destroy(); 40765ebb64fSKirill A. Shutemov err_slab: 408569e5590SShaohua Li hugepage_exit_sysfs(hugepage_kobj); 40965ebb64fSKirill A. Shutemov err_sysfs: 410ba76149fSAndrea Arcangeli return err; 41171e3aac0SAndrea Arcangeli } 412a64fb3cdSPaul Gortmaker subsys_initcall(hugepage_init); 41371e3aac0SAndrea Arcangeli 41471e3aac0SAndrea Arcangeli static int __init setup_transparent_hugepage(char *str) 41571e3aac0SAndrea Arcangeli { 41671e3aac0SAndrea Arcangeli int ret = 0; 41771e3aac0SAndrea Arcangeli if (!str) 41871e3aac0SAndrea Arcangeli goto out; 41971e3aac0SAndrea Arcangeli if (!strcmp(str, "always")) { 42071e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_FLAG, 42171e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 42271e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 42371e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 42471e3aac0SAndrea Arcangeli ret = 1; 42571e3aac0SAndrea Arcangeli } else if (!strcmp(str, "madvise")) { 42671e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 42771e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 42871e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 42971e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 43071e3aac0SAndrea Arcangeli ret = 1; 43171e3aac0SAndrea Arcangeli } else if (!strcmp(str, "never")) { 43271e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 43371e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 43471e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 43571e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 43671e3aac0SAndrea Arcangeli ret = 1; 43771e3aac0SAndrea Arcangeli } 43871e3aac0SAndrea Arcangeli out: 43971e3aac0SAndrea Arcangeli if (!ret) 440ae3a8c1cSAndrew Morton pr_warn("transparent_hugepage= cannot parse, ignored\n"); 44171e3aac0SAndrea Arcangeli return ret; 44271e3aac0SAndrea Arcangeli } 44371e3aac0SAndrea Arcangeli __setup("transparent_hugepage=", setup_transparent_hugepage); 44471e3aac0SAndrea Arcangeli 445b32967ffSMel Gorman pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 44671e3aac0SAndrea Arcangeli { 44771e3aac0SAndrea Arcangeli if (likely(vma->vm_flags & VM_WRITE)) 44871e3aac0SAndrea Arcangeli pmd = pmd_mkwrite(pmd); 44971e3aac0SAndrea Arcangeli return pmd; 45071e3aac0SAndrea Arcangeli } 45171e3aac0SAndrea Arcangeli 4529a982250SKirill A. Shutemov static inline struct list_head *page_deferred_list(struct page *page) 4539a982250SKirill A. Shutemov { 4549a982250SKirill A. Shutemov /* 4559a982250SKirill A. Shutemov * ->lru in the tail pages is occupied by compound_head. 4569a982250SKirill A. Shutemov * Let's use ->mapping + ->index in the second tail page as list_head. 4579a982250SKirill A. Shutemov */ 4589a982250SKirill A. Shutemov return (struct list_head *)&page[2].mapping; 4599a982250SKirill A. Shutemov } 4609a982250SKirill A. Shutemov 4619a982250SKirill A. Shutemov void prep_transhuge_page(struct page *page) 4629a982250SKirill A. Shutemov { 4639a982250SKirill A. Shutemov /* 4649a982250SKirill A. Shutemov * we use page->mapping and page->indexlru in second tail page 4659a982250SKirill A. Shutemov * as list_head: assuming THP order >= 2 4669a982250SKirill A. Shutemov */ 4679a982250SKirill A. Shutemov 4689a982250SKirill A. Shutemov INIT_LIST_HEAD(page_deferred_list(page)); 4699a982250SKirill A. Shutemov set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 4709a982250SKirill A. Shutemov } 4719a982250SKirill A. Shutemov 472bae473a4SKirill A. Shutemov static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page, 473bae473a4SKirill A. Shutemov gfp_t gfp) 47471e3aac0SAndrea Arcangeli { 475bae473a4SKirill A. Shutemov struct vm_area_struct *vma = fe->vma; 47600501b53SJohannes Weiner struct mem_cgroup *memcg; 47771e3aac0SAndrea Arcangeli pgtable_t pgtable; 478bae473a4SKirill A. Shutemov unsigned long haddr = fe->address & HPAGE_PMD_MASK; 47971e3aac0SAndrea Arcangeli 480309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page), page); 48100501b53SJohannes Weiner 482bae473a4SKirill A. Shutemov if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) { 4836b251fc9SAndrea Arcangeli put_page(page); 4846b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_FALLBACK); 4856b251fc9SAndrea Arcangeli return VM_FAULT_FALLBACK; 4866b251fc9SAndrea Arcangeli } 48771e3aac0SAndrea Arcangeli 488bae473a4SKirill A. Shutemov pgtable = pte_alloc_one(vma->vm_mm, haddr); 48900501b53SJohannes Weiner if (unlikely(!pgtable)) { 490f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(page, memcg, true); 4916b251fc9SAndrea Arcangeli put_page(page); 49200501b53SJohannes Weiner return VM_FAULT_OOM; 49300501b53SJohannes Weiner } 49400501b53SJohannes Weiner 49571e3aac0SAndrea Arcangeli clear_huge_page(page, haddr, HPAGE_PMD_NR); 49652f37629SMinchan Kim /* 49752f37629SMinchan Kim * The memory barrier inside __SetPageUptodate makes sure that 49852f37629SMinchan Kim * clear_huge_page writes become visible before the set_pmd_at() 49952f37629SMinchan Kim * write. 50052f37629SMinchan Kim */ 50171e3aac0SAndrea Arcangeli __SetPageUptodate(page); 50271e3aac0SAndrea Arcangeli 503bae473a4SKirill A. Shutemov fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); 504bae473a4SKirill A. Shutemov if (unlikely(!pmd_none(*fe->pmd))) { 505bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 506f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(page, memcg, true); 50771e3aac0SAndrea Arcangeli put_page(page); 508bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 50971e3aac0SAndrea Arcangeli } else { 51071e3aac0SAndrea Arcangeli pmd_t entry; 5116b251fc9SAndrea Arcangeli 5126b251fc9SAndrea Arcangeli /* Deliver the page fault to userland */ 5136b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 5146b251fc9SAndrea Arcangeli int ret; 5156b251fc9SAndrea Arcangeli 516bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 517f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(page, memcg, true); 5186b251fc9SAndrea Arcangeli put_page(page); 519bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 520bae473a4SKirill A. Shutemov ret = handle_userfault(fe, VM_UFFD_MISSING); 5216b251fc9SAndrea Arcangeli VM_BUG_ON(ret & VM_FAULT_FALLBACK); 5226b251fc9SAndrea Arcangeli return ret; 5236b251fc9SAndrea Arcangeli } 5246b251fc9SAndrea Arcangeli 5253122359aSKirill A. Shutemov entry = mk_huge_pmd(page, vma->vm_page_prot); 5263122359aSKirill A. Shutemov entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 527d281ee61SKirill A. Shutemov page_add_new_anon_rmap(page, vma, haddr, true); 528f627c2f5SKirill A. Shutemov mem_cgroup_commit_charge(page, memcg, false, true); 52900501b53SJohannes Weiner lru_cache_add_active_or_unevictable(page, vma); 530bae473a4SKirill A. Shutemov pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable); 531bae473a4SKirill A. Shutemov set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry); 532bae473a4SKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 533bae473a4SKirill A. Shutemov atomic_long_inc(&vma->vm_mm->nr_ptes); 534bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 5356b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_ALLOC); 53671e3aac0SAndrea Arcangeli } 53771e3aac0SAndrea Arcangeli 538aa2e878eSDavid Rientjes return 0; 53971e3aac0SAndrea Arcangeli } 54071e3aac0SAndrea Arcangeli 541444eb2a4SMel Gorman /* 542444eb2a4SMel Gorman * If THP is set to always then directly reclaim/compact as necessary 543444eb2a4SMel Gorman * If set to defer then do no reclaim and defer to khugepaged 544444eb2a4SMel Gorman * If set to madvise and the VMA is flagged then directly reclaim/compact 545444eb2a4SMel Gorman */ 546444eb2a4SMel Gorman static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) 5470bbbc0b3SAndrea Arcangeli { 548444eb2a4SMel Gorman gfp_t reclaim_flags = 0; 549444eb2a4SMel Gorman 550444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags) && 551444eb2a4SMel Gorman (vma->vm_flags & VM_HUGEPAGE)) 552444eb2a4SMel Gorman reclaim_flags = __GFP_DIRECT_RECLAIM; 553444eb2a4SMel Gorman else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 554444eb2a4SMel Gorman reclaim_flags = __GFP_KSWAPD_RECLAIM; 555444eb2a4SMel Gorman else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 556444eb2a4SMel Gorman reclaim_flags = __GFP_DIRECT_RECLAIM; 557444eb2a4SMel Gorman 558444eb2a4SMel Gorman return GFP_TRANSHUGE | reclaim_flags; 559444eb2a4SMel Gorman } 560444eb2a4SMel Gorman 561c4088ebdSKirill A. Shutemov /* Caller must hold page table lock. */ 562d295e341SKirill A. Shutemov static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 56397ae1749SKirill A. Shutemov struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 5645918d10aSKirill A. Shutemov struct page *zero_page) 565fc9fe822SKirill A. Shutemov { 566fc9fe822SKirill A. Shutemov pmd_t entry; 5677c414164SAndrew Morton if (!pmd_none(*pmd)) 5687c414164SAndrew Morton return false; 5695918d10aSKirill A. Shutemov entry = mk_pmd(zero_page, vma->vm_page_prot); 570fc9fe822SKirill A. Shutemov entry = pmd_mkhuge(entry); 57112c9d70bSMatthew Wilcox if (pgtable) 5726b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 573fc9fe822SKirill A. Shutemov set_pmd_at(mm, haddr, pmd, entry); 574e1f56c89SKirill A. Shutemov atomic_long_inc(&mm->nr_ptes); 5757c414164SAndrew Morton return true; 576fc9fe822SKirill A. Shutemov } 577fc9fe822SKirill A. Shutemov 578bae473a4SKirill A. Shutemov int do_huge_pmd_anonymous_page(struct fault_env *fe) 57971e3aac0SAndrea Arcangeli { 580bae473a4SKirill A. Shutemov struct vm_area_struct *vma = fe->vma; 581077fcf11SAneesh Kumar K.V gfp_t gfp; 58271e3aac0SAndrea Arcangeli struct page *page; 583bae473a4SKirill A. Shutemov unsigned long haddr = fe->address & HPAGE_PMD_MASK; 58471e3aac0SAndrea Arcangeli 585128ec037SKirill A. Shutemov if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 586c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 58771e3aac0SAndrea Arcangeli if (unlikely(anon_vma_prepare(vma))) 58871e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 5896d50e60cSDavid Rientjes if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 590ba76149fSAndrea Arcangeli return VM_FAULT_OOM; 591bae473a4SKirill A. Shutemov if (!(fe->flags & FAULT_FLAG_WRITE) && 592bae473a4SKirill A. Shutemov !mm_forbids_zeropage(vma->vm_mm) && 59379da5407SKirill A. Shutemov transparent_hugepage_use_zero_page()) { 59480371957SKirill A. Shutemov pgtable_t pgtable; 5955918d10aSKirill A. Shutemov struct page *zero_page; 5963ea41e62SKirill A. Shutemov bool set; 5976b251fc9SAndrea Arcangeli int ret; 598bae473a4SKirill A. Shutemov pgtable = pte_alloc_one(vma->vm_mm, haddr); 59980371957SKirill A. Shutemov if (unlikely(!pgtable)) 60080371957SKirill A. Shutemov return VM_FAULT_OOM; 6015918d10aSKirill A. Shutemov zero_page = get_huge_zero_page(); 6025918d10aSKirill A. Shutemov if (unlikely(!zero_page)) { 603bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 60497ae1749SKirill A. Shutemov count_vm_event(THP_FAULT_FALLBACK); 605c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 60697ae1749SKirill A. Shutemov } 607bae473a4SKirill A. Shutemov fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); 6086b251fc9SAndrea Arcangeli ret = 0; 6096b251fc9SAndrea Arcangeli set = false; 610bae473a4SKirill A. Shutemov if (pmd_none(*fe->pmd)) { 6116b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 612bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 613bae473a4SKirill A. Shutemov ret = handle_userfault(fe, VM_UFFD_MISSING); 6146b251fc9SAndrea Arcangeli VM_BUG_ON(ret & VM_FAULT_FALLBACK); 6156b251fc9SAndrea Arcangeli } else { 616bae473a4SKirill A. Shutemov set_huge_zero_page(pgtable, vma->vm_mm, vma, 617bae473a4SKirill A. Shutemov haddr, fe->pmd, zero_page); 618bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 6196b251fc9SAndrea Arcangeli set = true; 6206b251fc9SAndrea Arcangeli } 6216b251fc9SAndrea Arcangeli } else 622bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 6233ea41e62SKirill A. Shutemov if (!set) { 624bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 6253ea41e62SKirill A. Shutemov put_huge_zero_page(); 6263ea41e62SKirill A. Shutemov } 6276b251fc9SAndrea Arcangeli return ret; 62880371957SKirill A. Shutemov } 629444eb2a4SMel Gorman gfp = alloc_hugepage_direct_gfpmask(vma); 630077fcf11SAneesh Kumar K.V page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 63181ab4201SAndi Kleen if (unlikely(!page)) { 63281ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 633c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 63481ab4201SAndi Kleen } 6359a982250SKirill A. Shutemov prep_transhuge_page(page); 636bae473a4SKirill A. Shutemov return __do_huge_pmd_anonymous_page(fe, page, gfp); 63771e3aac0SAndrea Arcangeli } 63871e3aac0SAndrea Arcangeli 639ae18d6dcSMatthew Wilcox static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 640f25748e3SDan Williams pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write) 6415cad465dSMatthew Wilcox { 6425cad465dSMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 6435cad465dSMatthew Wilcox pmd_t entry; 6445cad465dSMatthew Wilcox spinlock_t *ptl; 6455cad465dSMatthew Wilcox 6465cad465dSMatthew Wilcox ptl = pmd_lock(mm, pmd); 647f25748e3SDan Williams entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 648f25748e3SDan Williams if (pfn_t_devmap(pfn)) 649f25748e3SDan Williams entry = pmd_mkdevmap(entry); 6505cad465dSMatthew Wilcox if (write) { 6515cad465dSMatthew Wilcox entry = pmd_mkyoung(pmd_mkdirty(entry)); 6525cad465dSMatthew Wilcox entry = maybe_pmd_mkwrite(entry, vma); 6535cad465dSMatthew Wilcox } 6545cad465dSMatthew Wilcox set_pmd_at(mm, addr, pmd, entry); 6555cad465dSMatthew Wilcox update_mmu_cache_pmd(vma, addr, pmd); 6565cad465dSMatthew Wilcox spin_unlock(ptl); 6575cad465dSMatthew Wilcox } 6585cad465dSMatthew Wilcox 6595cad465dSMatthew Wilcox int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 660f25748e3SDan Williams pmd_t *pmd, pfn_t pfn, bool write) 6615cad465dSMatthew Wilcox { 6625cad465dSMatthew Wilcox pgprot_t pgprot = vma->vm_page_prot; 6635cad465dSMatthew Wilcox /* 6645cad465dSMatthew Wilcox * If we had pmd_special, we could avoid all these restrictions, 6655cad465dSMatthew Wilcox * but we need to be consistent with PTEs and architectures that 6665cad465dSMatthew Wilcox * can't support a 'special' bit. 6675cad465dSMatthew Wilcox */ 6685cad465dSMatthew Wilcox BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 6695cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 6705cad465dSMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 6715cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 672f25748e3SDan Williams BUG_ON(!pfn_t_devmap(pfn)); 6735cad465dSMatthew Wilcox 6745cad465dSMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 6755cad465dSMatthew Wilcox return VM_FAULT_SIGBUS; 6765cad465dSMatthew Wilcox if (track_pfn_insert(vma, &pgprot, pfn)) 6775cad465dSMatthew Wilcox return VM_FAULT_SIGBUS; 678ae18d6dcSMatthew Wilcox insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); 679ae18d6dcSMatthew Wilcox return VM_FAULT_NOPAGE; 6805cad465dSMatthew Wilcox } 681dee41079SDan Williams EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 6825cad465dSMatthew Wilcox 6833565fce3SDan Williams static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 6843565fce3SDan Williams pmd_t *pmd) 6853565fce3SDan Williams { 6863565fce3SDan Williams pmd_t _pmd; 6873565fce3SDan Williams 6883565fce3SDan Williams /* 6893565fce3SDan Williams * We should set the dirty bit only for FOLL_WRITE but for now 6903565fce3SDan Williams * the dirty bit in the pmd is meaningless. And if the dirty 6913565fce3SDan Williams * bit will become meaningful and we'll only set it with 6923565fce3SDan Williams * FOLL_WRITE, an atomic set_bit will be required on the pmd to 6933565fce3SDan Williams * set the young bit, instead of the current set_pmd_at. 6943565fce3SDan Williams */ 6953565fce3SDan Williams _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 6963565fce3SDan Williams if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 6973565fce3SDan Williams pmd, _pmd, 1)) 6983565fce3SDan Williams update_mmu_cache_pmd(vma, addr, pmd); 6993565fce3SDan Williams } 7003565fce3SDan Williams 7013565fce3SDan Williams struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 7023565fce3SDan Williams pmd_t *pmd, int flags) 7033565fce3SDan Williams { 7043565fce3SDan Williams unsigned long pfn = pmd_pfn(*pmd); 7053565fce3SDan Williams struct mm_struct *mm = vma->vm_mm; 7063565fce3SDan Williams struct dev_pagemap *pgmap; 7073565fce3SDan Williams struct page *page; 7083565fce3SDan Williams 7093565fce3SDan Williams assert_spin_locked(pmd_lockptr(mm, pmd)); 7103565fce3SDan Williams 7113565fce3SDan Williams if (flags & FOLL_WRITE && !pmd_write(*pmd)) 7123565fce3SDan Williams return NULL; 7133565fce3SDan Williams 7143565fce3SDan Williams if (pmd_present(*pmd) && pmd_devmap(*pmd)) 7153565fce3SDan Williams /* pass */; 7163565fce3SDan Williams else 7173565fce3SDan Williams return NULL; 7183565fce3SDan Williams 7193565fce3SDan Williams if (flags & FOLL_TOUCH) 7203565fce3SDan Williams touch_pmd(vma, addr, pmd); 7213565fce3SDan Williams 7223565fce3SDan Williams /* 7233565fce3SDan Williams * device mapped pages can only be returned if the 7243565fce3SDan Williams * caller will manage the page reference count. 7253565fce3SDan Williams */ 7263565fce3SDan Williams if (!(flags & FOLL_GET)) 7273565fce3SDan Williams return ERR_PTR(-EEXIST); 7283565fce3SDan Williams 7293565fce3SDan Williams pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 7303565fce3SDan Williams pgmap = get_dev_pagemap(pfn, NULL); 7313565fce3SDan Williams if (!pgmap) 7323565fce3SDan Williams return ERR_PTR(-EFAULT); 7333565fce3SDan Williams page = pfn_to_page(pfn); 7343565fce3SDan Williams get_page(page); 7353565fce3SDan Williams put_dev_pagemap(pgmap); 7363565fce3SDan Williams 7373565fce3SDan Williams return page; 7383565fce3SDan Williams } 7393565fce3SDan Williams 74071e3aac0SAndrea Arcangeli int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 74171e3aac0SAndrea Arcangeli pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 74271e3aac0SAndrea Arcangeli struct vm_area_struct *vma) 74371e3aac0SAndrea Arcangeli { 744c4088ebdSKirill A. Shutemov spinlock_t *dst_ptl, *src_ptl; 74571e3aac0SAndrea Arcangeli struct page *src_page; 74671e3aac0SAndrea Arcangeli pmd_t pmd; 74712c9d70bSMatthew Wilcox pgtable_t pgtable = NULL; 748628d47ceSKirill A. Shutemov int ret = -ENOMEM; 74971e3aac0SAndrea Arcangeli 750628d47ceSKirill A. Shutemov /* Skip if can be re-fill on fault */ 751628d47ceSKirill A. Shutemov if (!vma_is_anonymous(vma)) 752628d47ceSKirill A. Shutemov return 0; 753628d47ceSKirill A. Shutemov 75471e3aac0SAndrea Arcangeli pgtable = pte_alloc_one(dst_mm, addr); 75571e3aac0SAndrea Arcangeli if (unlikely(!pgtable)) 75671e3aac0SAndrea Arcangeli goto out; 75771e3aac0SAndrea Arcangeli 758c4088ebdSKirill A. Shutemov dst_ptl = pmd_lock(dst_mm, dst_pmd); 759c4088ebdSKirill A. Shutemov src_ptl = pmd_lockptr(src_mm, src_pmd); 760c4088ebdSKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 76171e3aac0SAndrea Arcangeli 76271e3aac0SAndrea Arcangeli ret = -EAGAIN; 76371e3aac0SAndrea Arcangeli pmd = *src_pmd; 764628d47ceSKirill A. Shutemov if (unlikely(!pmd_trans_huge(pmd))) { 76571e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 76671e3aac0SAndrea Arcangeli goto out_unlock; 76771e3aac0SAndrea Arcangeli } 768fc9fe822SKirill A. Shutemov /* 769c4088ebdSKirill A. Shutemov * When page table lock is held, the huge zero pmd should not be 770fc9fe822SKirill A. Shutemov * under splitting since we don't split the page itself, only pmd to 771fc9fe822SKirill A. Shutemov * a page table. 772fc9fe822SKirill A. Shutemov */ 773fc9fe822SKirill A. Shutemov if (is_huge_zero_pmd(pmd)) { 7745918d10aSKirill A. Shutemov struct page *zero_page; 77597ae1749SKirill A. Shutemov /* 77697ae1749SKirill A. Shutemov * get_huge_zero_page() will never allocate a new page here, 77797ae1749SKirill A. Shutemov * since we already have a zero page to copy. It just takes a 77897ae1749SKirill A. Shutemov * reference. 77997ae1749SKirill A. Shutemov */ 7805918d10aSKirill A. Shutemov zero_page = get_huge_zero_page(); 7816b251fc9SAndrea Arcangeli set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 7825918d10aSKirill A. Shutemov zero_page); 783fc9fe822SKirill A. Shutemov ret = 0; 784fc9fe822SKirill A. Shutemov goto out_unlock; 785fc9fe822SKirill A. Shutemov } 786de466bd6SMel Gorman 78771e3aac0SAndrea Arcangeli src_page = pmd_page(pmd); 788309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 78971e3aac0SAndrea Arcangeli get_page(src_page); 79053f9263bSKirill A. Shutemov page_dup_rmap(src_page, true); 79171e3aac0SAndrea Arcangeli add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 7925c7fb56eSDan Williams atomic_long_inc(&dst_mm->nr_ptes); 7935c7fb56eSDan Williams pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 79471e3aac0SAndrea Arcangeli 79571e3aac0SAndrea Arcangeli pmdp_set_wrprotect(src_mm, addr, src_pmd); 79671e3aac0SAndrea Arcangeli pmd = pmd_mkold(pmd_wrprotect(pmd)); 79771e3aac0SAndrea Arcangeli set_pmd_at(dst_mm, addr, dst_pmd, pmd); 79871e3aac0SAndrea Arcangeli 79971e3aac0SAndrea Arcangeli ret = 0; 80071e3aac0SAndrea Arcangeli out_unlock: 801c4088ebdSKirill A. Shutemov spin_unlock(src_ptl); 802c4088ebdSKirill A. Shutemov spin_unlock(dst_ptl); 80371e3aac0SAndrea Arcangeli out: 80471e3aac0SAndrea Arcangeli return ret; 80571e3aac0SAndrea Arcangeli } 80671e3aac0SAndrea Arcangeli 807bae473a4SKirill A. Shutemov void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd) 808a1dd450bSWill Deacon { 809a1dd450bSWill Deacon pmd_t entry; 810a1dd450bSWill Deacon unsigned long haddr; 811a1dd450bSWill Deacon 812bae473a4SKirill A. Shutemov fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd); 813bae473a4SKirill A. Shutemov if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) 814a1dd450bSWill Deacon goto unlock; 815a1dd450bSWill Deacon 816a1dd450bSWill Deacon entry = pmd_mkyoung(orig_pmd); 817bae473a4SKirill A. Shutemov haddr = fe->address & HPAGE_PMD_MASK; 818bae473a4SKirill A. Shutemov if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, 819bae473a4SKirill A. Shutemov fe->flags & FAULT_FLAG_WRITE)) 820bae473a4SKirill A. Shutemov update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd); 821a1dd450bSWill Deacon 822a1dd450bSWill Deacon unlock: 823bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 824a1dd450bSWill Deacon } 825a1dd450bSWill Deacon 826bae473a4SKirill A. Shutemov static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd, 827bae473a4SKirill A. Shutemov struct page *page) 82871e3aac0SAndrea Arcangeli { 829bae473a4SKirill A. Shutemov struct vm_area_struct *vma = fe->vma; 830bae473a4SKirill A. Shutemov unsigned long haddr = fe->address & HPAGE_PMD_MASK; 83100501b53SJohannes Weiner struct mem_cgroup *memcg; 83271e3aac0SAndrea Arcangeli pgtable_t pgtable; 83371e3aac0SAndrea Arcangeli pmd_t _pmd; 83471e3aac0SAndrea Arcangeli int ret = 0, i; 83571e3aac0SAndrea Arcangeli struct page **pages; 8362ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 8372ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 83871e3aac0SAndrea Arcangeli 83971e3aac0SAndrea Arcangeli pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 84071e3aac0SAndrea Arcangeli GFP_KERNEL); 84171e3aac0SAndrea Arcangeli if (unlikely(!pages)) { 84271e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 84371e3aac0SAndrea Arcangeli goto out; 84471e3aac0SAndrea Arcangeli } 84571e3aac0SAndrea Arcangeli 84671e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 847cc5d462fSAndi Kleen pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | 848bae473a4SKirill A. Shutemov __GFP_OTHER_NODE, vma, 849bae473a4SKirill A. Shutemov fe->address, page_to_nid(page)); 850b9bbfbe3SAndrea Arcangeli if (unlikely(!pages[i] || 851bae473a4SKirill A. Shutemov mem_cgroup_try_charge(pages[i], vma->vm_mm, 852bae473a4SKirill A. Shutemov GFP_KERNEL, &memcg, false))) { 853b9bbfbe3SAndrea Arcangeli if (pages[i]) 85471e3aac0SAndrea Arcangeli put_page(pages[i]); 855b9bbfbe3SAndrea Arcangeli while (--i >= 0) { 85600501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 85700501b53SJohannes Weiner set_page_private(pages[i], 0); 858f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(pages[i], memcg, 859f627c2f5SKirill A. Shutemov false); 860b9bbfbe3SAndrea Arcangeli put_page(pages[i]); 861b9bbfbe3SAndrea Arcangeli } 86271e3aac0SAndrea Arcangeli kfree(pages); 86371e3aac0SAndrea Arcangeli ret |= VM_FAULT_OOM; 86471e3aac0SAndrea Arcangeli goto out; 86571e3aac0SAndrea Arcangeli } 86600501b53SJohannes Weiner set_page_private(pages[i], (unsigned long)memcg); 86771e3aac0SAndrea Arcangeli } 86871e3aac0SAndrea Arcangeli 86971e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 87071e3aac0SAndrea Arcangeli copy_user_highpage(pages[i], page + i, 8710089e485SHillf Danton haddr + PAGE_SIZE * i, vma); 87271e3aac0SAndrea Arcangeli __SetPageUptodate(pages[i]); 87371e3aac0SAndrea Arcangeli cond_resched(); 87471e3aac0SAndrea Arcangeli } 87571e3aac0SAndrea Arcangeli 8762ec74c3eSSagi Grimberg mmun_start = haddr; 8772ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 878bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 8792ec74c3eSSagi Grimberg 880bae473a4SKirill A. Shutemov fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); 881bae473a4SKirill A. Shutemov if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) 88271e3aac0SAndrea Arcangeli goto out_free_pages; 883309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 88471e3aac0SAndrea Arcangeli 885bae473a4SKirill A. Shutemov pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd); 88671e3aac0SAndrea Arcangeli /* leave pmd empty until pte is filled */ 88771e3aac0SAndrea Arcangeli 888bae473a4SKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd); 889bae473a4SKirill A. Shutemov pmd_populate(vma->vm_mm, &_pmd, pgtable); 89071e3aac0SAndrea Arcangeli 89171e3aac0SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 892bae473a4SKirill A. Shutemov pte_t entry; 89371e3aac0SAndrea Arcangeli entry = mk_pte(pages[i], vma->vm_page_prot); 89471e3aac0SAndrea Arcangeli entry = maybe_mkwrite(pte_mkdirty(entry), vma); 89500501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 89600501b53SJohannes Weiner set_page_private(pages[i], 0); 897bae473a4SKirill A. Shutemov page_add_new_anon_rmap(pages[i], fe->vma, haddr, false); 898f627c2f5SKirill A. Shutemov mem_cgroup_commit_charge(pages[i], memcg, false, false); 89900501b53SJohannes Weiner lru_cache_add_active_or_unevictable(pages[i], vma); 900bae473a4SKirill A. Shutemov fe->pte = pte_offset_map(&_pmd, haddr); 901bae473a4SKirill A. Shutemov VM_BUG_ON(!pte_none(*fe->pte)); 902bae473a4SKirill A. Shutemov set_pte_at(vma->vm_mm, haddr, fe->pte, entry); 903bae473a4SKirill A. Shutemov pte_unmap(fe->pte); 90471e3aac0SAndrea Arcangeli } 90571e3aac0SAndrea Arcangeli kfree(pages); 90671e3aac0SAndrea Arcangeli 90771e3aac0SAndrea Arcangeli smp_wmb(); /* make pte visible before pmd */ 908bae473a4SKirill A. Shutemov pmd_populate(vma->vm_mm, fe->pmd, pgtable); 909d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 910bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 91171e3aac0SAndrea Arcangeli 912bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 9132ec74c3eSSagi Grimberg 91471e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 91571e3aac0SAndrea Arcangeli put_page(page); 91671e3aac0SAndrea Arcangeli 91771e3aac0SAndrea Arcangeli out: 91871e3aac0SAndrea Arcangeli return ret; 91971e3aac0SAndrea Arcangeli 92071e3aac0SAndrea Arcangeli out_free_pages: 921bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 922bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 923b9bbfbe3SAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 92400501b53SJohannes Weiner memcg = (void *)page_private(pages[i]); 92500501b53SJohannes Weiner set_page_private(pages[i], 0); 926f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(pages[i], memcg, false); 92771e3aac0SAndrea Arcangeli put_page(pages[i]); 928b9bbfbe3SAndrea Arcangeli } 92971e3aac0SAndrea Arcangeli kfree(pages); 93071e3aac0SAndrea Arcangeli goto out; 93171e3aac0SAndrea Arcangeli } 93271e3aac0SAndrea Arcangeli 933bae473a4SKirill A. Shutemov int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd) 93471e3aac0SAndrea Arcangeli { 935bae473a4SKirill A. Shutemov struct vm_area_struct *vma = fe->vma; 93693b4796dSKirill A. Shutemov struct page *page = NULL, *new_page; 93700501b53SJohannes Weiner struct mem_cgroup *memcg; 938bae473a4SKirill A. Shutemov unsigned long haddr = fe->address & HPAGE_PMD_MASK; 9392ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 9402ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 9413b363692SMichal Hocko gfp_t huge_gfp; /* for allocation and charge */ 942bae473a4SKirill A. Shutemov int ret = 0; 94371e3aac0SAndrea Arcangeli 944bae473a4SKirill A. Shutemov fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd); 94581d1b09cSSasha Levin VM_BUG_ON_VMA(!vma->anon_vma, vma); 94693b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 94793b4796dSKirill A. Shutemov goto alloc; 948bae473a4SKirill A. Shutemov spin_lock(fe->ptl); 949bae473a4SKirill A. Shutemov if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) 95071e3aac0SAndrea Arcangeli goto out_unlock; 95171e3aac0SAndrea Arcangeli 95271e3aac0SAndrea Arcangeli page = pmd_page(orig_pmd); 953309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 9541f25fe20SKirill A. Shutemov /* 9551f25fe20SKirill A. Shutemov * We can only reuse the page if nobody else maps the huge page or it's 9566d0a07edSAndrea Arcangeli * part. 9571f25fe20SKirill A. Shutemov */ 9586d0a07edSAndrea Arcangeli if (page_trans_huge_mapcount(page, NULL) == 1) { 95971e3aac0SAndrea Arcangeli pmd_t entry; 96071e3aac0SAndrea Arcangeli entry = pmd_mkyoung(orig_pmd); 96171e3aac0SAndrea Arcangeli entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 962bae473a4SKirill A. Shutemov if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1)) 963bae473a4SKirill A. Shutemov update_mmu_cache_pmd(vma, fe->address, fe->pmd); 96471e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 96571e3aac0SAndrea Arcangeli goto out_unlock; 96671e3aac0SAndrea Arcangeli } 967ddc58f27SKirill A. Shutemov get_page(page); 968bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 96993b4796dSKirill A. Shutemov alloc: 97071e3aac0SAndrea Arcangeli if (transparent_hugepage_enabled(vma) && 971077fcf11SAneesh Kumar K.V !transparent_hugepage_debug_cow()) { 972444eb2a4SMel Gorman huge_gfp = alloc_hugepage_direct_gfpmask(vma); 9733b363692SMichal Hocko new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); 974077fcf11SAneesh Kumar K.V } else 97571e3aac0SAndrea Arcangeli new_page = NULL; 97671e3aac0SAndrea Arcangeli 9779a982250SKirill A. Shutemov if (likely(new_page)) { 9789a982250SKirill A. Shutemov prep_transhuge_page(new_page); 9799a982250SKirill A. Shutemov } else { 980eecc1e42SHugh Dickins if (!page) { 981bae473a4SKirill A. Shutemov split_huge_pmd(vma, fe->pmd, fe->address); 982e9b71ca9SKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 98393b4796dSKirill A. Shutemov } else { 984bae473a4SKirill A. Shutemov ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page); 9859845cbbdSKirill A. Shutemov if (ret & VM_FAULT_OOM) { 986bae473a4SKirill A. Shutemov split_huge_pmd(vma, fe->pmd, fe->address); 9879845cbbdSKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 9889845cbbdSKirill A. Shutemov } 989ddc58f27SKirill A. Shutemov put_page(page); 99093b4796dSKirill A. Shutemov } 99117766ddeSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK); 99271e3aac0SAndrea Arcangeli goto out; 99371e3aac0SAndrea Arcangeli } 99471e3aac0SAndrea Arcangeli 995bae473a4SKirill A. Shutemov if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, 996bae473a4SKirill A. Shutemov huge_gfp, &memcg, true))) { 997b9bbfbe3SAndrea Arcangeli put_page(new_page); 998bae473a4SKirill A. Shutemov split_huge_pmd(vma, fe->pmd, fe->address); 999bae473a4SKirill A. Shutemov if (page) 1000ddc58f27SKirill A. Shutemov put_page(page); 10019845cbbdSKirill A. Shutemov ret |= VM_FAULT_FALLBACK; 100217766ddeSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK); 1003b9bbfbe3SAndrea Arcangeli goto out; 1004b9bbfbe3SAndrea Arcangeli } 1005b9bbfbe3SAndrea Arcangeli 100617766ddeSDavid Rientjes count_vm_event(THP_FAULT_ALLOC); 100717766ddeSDavid Rientjes 1008eecc1e42SHugh Dickins if (!page) 100993b4796dSKirill A. Shutemov clear_huge_page(new_page, haddr, HPAGE_PMD_NR); 101093b4796dSKirill A. Shutemov else 101171e3aac0SAndrea Arcangeli copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 101271e3aac0SAndrea Arcangeli __SetPageUptodate(new_page); 101371e3aac0SAndrea Arcangeli 10142ec74c3eSSagi Grimberg mmun_start = haddr; 10152ec74c3eSSagi Grimberg mmun_end = haddr + HPAGE_PMD_SIZE; 1016bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 10172ec74c3eSSagi Grimberg 1018bae473a4SKirill A. Shutemov spin_lock(fe->ptl); 101993b4796dSKirill A. Shutemov if (page) 1020ddc58f27SKirill A. Shutemov put_page(page); 1021bae473a4SKirill A. Shutemov if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) { 1022bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 1023f627c2f5SKirill A. Shutemov mem_cgroup_cancel_charge(new_page, memcg, true); 102471e3aac0SAndrea Arcangeli put_page(new_page); 10252ec74c3eSSagi Grimberg goto out_mn; 1026b9bbfbe3SAndrea Arcangeli } else { 102771e3aac0SAndrea Arcangeli pmd_t entry; 10283122359aSKirill A. Shutemov entry = mk_huge_pmd(new_page, vma->vm_page_prot); 10293122359aSKirill A. Shutemov entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1030bae473a4SKirill A. Shutemov pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd); 1031d281ee61SKirill A. Shutemov page_add_new_anon_rmap(new_page, vma, haddr, true); 1032f627c2f5SKirill A. Shutemov mem_cgroup_commit_charge(new_page, memcg, false, true); 103300501b53SJohannes Weiner lru_cache_add_active_or_unevictable(new_page, vma); 1034bae473a4SKirill A. Shutemov set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry); 1035bae473a4SKirill A. Shutemov update_mmu_cache_pmd(vma, fe->address, fe->pmd); 1036eecc1e42SHugh Dickins if (!page) { 1037bae473a4SKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 103897ae1749SKirill A. Shutemov put_huge_zero_page(); 103997ae1749SKirill A. Shutemov } else { 1040309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1041d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 104271e3aac0SAndrea Arcangeli put_page(page); 104393b4796dSKirill A. Shutemov } 104471e3aac0SAndrea Arcangeli ret |= VM_FAULT_WRITE; 104571e3aac0SAndrea Arcangeli } 1046bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 10472ec74c3eSSagi Grimberg out_mn: 1048bae473a4SKirill A. Shutemov mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 10492ec74c3eSSagi Grimberg out: 10502ec74c3eSSagi Grimberg return ret; 105171e3aac0SAndrea Arcangeli out_unlock: 1052bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 105371e3aac0SAndrea Arcangeli return ret; 105471e3aac0SAndrea Arcangeli } 105571e3aac0SAndrea Arcangeli 1056b676b293SDavid Rientjes struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 105771e3aac0SAndrea Arcangeli unsigned long addr, 105871e3aac0SAndrea Arcangeli pmd_t *pmd, 105971e3aac0SAndrea Arcangeli unsigned int flags) 106071e3aac0SAndrea Arcangeli { 1061b676b293SDavid Rientjes struct mm_struct *mm = vma->vm_mm; 106271e3aac0SAndrea Arcangeli struct page *page = NULL; 106371e3aac0SAndrea Arcangeli 1064c4088ebdSKirill A. Shutemov assert_spin_locked(pmd_lockptr(mm, pmd)); 106571e3aac0SAndrea Arcangeli 106671e3aac0SAndrea Arcangeli if (flags & FOLL_WRITE && !pmd_write(*pmd)) 106771e3aac0SAndrea Arcangeli goto out; 106871e3aac0SAndrea Arcangeli 106985facf25SKirill A. Shutemov /* Avoid dumping huge zero page */ 107085facf25SKirill A. Shutemov if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 107185facf25SKirill A. Shutemov return ERR_PTR(-EFAULT); 107285facf25SKirill A. Shutemov 10732b4847e7SMel Gorman /* Full NUMA hinting faults to serialise migration in fault paths */ 10748a0516edSMel Gorman if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 10752b4847e7SMel Gorman goto out; 10762b4847e7SMel Gorman 107771e3aac0SAndrea Arcangeli page = pmd_page(*pmd); 1078309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 10793565fce3SDan Williams if (flags & FOLL_TOUCH) 10803565fce3SDan Williams touch_pmd(vma, addr, pmd); 1081de60f5f1SEric B Munson if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1082e90309c9SKirill A. Shutemov /* 1083e90309c9SKirill A. Shutemov * We don't mlock() pte-mapped THPs. This way we can avoid 1084e90309c9SKirill A. Shutemov * leaking mlocked pages into non-VM_LOCKED VMAs. 1085e90309c9SKirill A. Shutemov * 10869a73f61bSKirill A. Shutemov * For anon THP: 10879a73f61bSKirill A. Shutemov * 1088e90309c9SKirill A. Shutemov * In most cases the pmd is the only mapping of the page as we 1089e90309c9SKirill A. Shutemov * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1090e90309c9SKirill A. Shutemov * writable private mappings in populate_vma_page_range(). 1091e90309c9SKirill A. Shutemov * 1092e90309c9SKirill A. Shutemov * The only scenario when we have the page shared here is if we 1093e90309c9SKirill A. Shutemov * mlocking read-only mapping shared over fork(). We skip 1094e90309c9SKirill A. Shutemov * mlocking such pages. 10959a73f61bSKirill A. Shutemov * 10969a73f61bSKirill A. Shutemov * For file THP: 10979a73f61bSKirill A. Shutemov * 10989a73f61bSKirill A. Shutemov * We can expect PageDoubleMap() to be stable under page lock: 10999a73f61bSKirill A. Shutemov * for file pages we set it in page_add_file_rmap(), which 11009a73f61bSKirill A. Shutemov * requires page to be locked. 1101e90309c9SKirill A. Shutemov */ 11029a73f61bSKirill A. Shutemov 11039a73f61bSKirill A. Shutemov if (PageAnon(page) && compound_mapcount(page) != 1) 11049a73f61bSKirill A. Shutemov goto skip_mlock; 11059a73f61bSKirill A. Shutemov if (PageDoubleMap(page) || !page->mapping) 11069a73f61bSKirill A. Shutemov goto skip_mlock; 11079a73f61bSKirill A. Shutemov if (!trylock_page(page)) 11089a73f61bSKirill A. Shutemov goto skip_mlock; 1109b676b293SDavid Rientjes lru_add_drain(); 11109a73f61bSKirill A. Shutemov if (page->mapping && !PageDoubleMap(page)) 1111b676b293SDavid Rientjes mlock_vma_page(page); 1112b676b293SDavid Rientjes unlock_page(page); 1113b676b293SDavid Rientjes } 11149a73f61bSKirill A. Shutemov skip_mlock: 111571e3aac0SAndrea Arcangeli page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1116309381feSSasha Levin VM_BUG_ON_PAGE(!PageCompound(page), page); 111771e3aac0SAndrea Arcangeli if (flags & FOLL_GET) 1118ddc58f27SKirill A. Shutemov get_page(page); 111971e3aac0SAndrea Arcangeli 112071e3aac0SAndrea Arcangeli out: 112171e3aac0SAndrea Arcangeli return page; 112271e3aac0SAndrea Arcangeli } 112371e3aac0SAndrea Arcangeli 1124d10e63f2SMel Gorman /* NUMA hinting page fault entry point for trans huge pmds */ 1125bae473a4SKirill A. Shutemov int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd) 1126d10e63f2SMel Gorman { 1127bae473a4SKirill A. Shutemov struct vm_area_struct *vma = fe->vma; 1128b8916634SMel Gorman struct anon_vma *anon_vma = NULL; 1129b32967ffSMel Gorman struct page *page; 1130bae473a4SKirill A. Shutemov unsigned long haddr = fe->address & HPAGE_PMD_MASK; 11318191acbdSMel Gorman int page_nid = -1, this_nid = numa_node_id(); 113290572890SPeter Zijlstra int target_nid, last_cpupid = -1; 11338191acbdSMel Gorman bool page_locked; 11348191acbdSMel Gorman bool migrated = false; 1135b191f9b1SMel Gorman bool was_writable; 11366688cc05SPeter Zijlstra int flags = 0; 1137d10e63f2SMel Gorman 1138c0e7cad9SMel Gorman /* A PROT_NONE fault should not end up here */ 1139c0e7cad9SMel Gorman BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); 1140c0e7cad9SMel Gorman 1141bae473a4SKirill A. Shutemov fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); 1142bae473a4SKirill A. Shutemov if (unlikely(!pmd_same(pmd, *fe->pmd))) 1143d10e63f2SMel Gorman goto out_unlock; 1144d10e63f2SMel Gorman 1145de466bd6SMel Gorman /* 1146de466bd6SMel Gorman * If there are potential migrations, wait for completion and retry 1147de466bd6SMel Gorman * without disrupting NUMA hinting information. Do not relock and 1148de466bd6SMel Gorman * check_same as the page may no longer be mapped. 1149de466bd6SMel Gorman */ 1150bae473a4SKirill A. Shutemov if (unlikely(pmd_trans_migrating(*fe->pmd))) { 1151bae473a4SKirill A. Shutemov page = pmd_page(*fe->pmd); 1152bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 11535d833062SMel Gorman wait_on_page_locked(page); 1154de466bd6SMel Gorman goto out; 1155de466bd6SMel Gorman } 1156de466bd6SMel Gorman 1157d10e63f2SMel Gorman page = pmd_page(pmd); 1158a1a46184SMel Gorman BUG_ON(is_huge_zero_page(page)); 11598191acbdSMel Gorman page_nid = page_to_nid(page); 116090572890SPeter Zijlstra last_cpupid = page_cpupid_last(page); 116103c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS); 116204bb2f94SRik van Riel if (page_nid == this_nid) { 116303c5a6e1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 116404bb2f94SRik van Riel flags |= TNF_FAULT_LOCAL; 116504bb2f94SRik van Riel } 11664daae3b4SMel Gorman 1167bea66fbdSMel Gorman /* See similar comment in do_numa_page for explanation */ 1168bea66fbdSMel Gorman if (!(vma->vm_flags & VM_WRITE)) 11696688cc05SPeter Zijlstra flags |= TNF_NO_GROUP; 11706688cc05SPeter Zijlstra 11716688cc05SPeter Zijlstra /* 1172ff9042b1SMel Gorman * Acquire the page lock to serialise THP migrations but avoid dropping 1173ff9042b1SMel Gorman * page_table_lock if at all possible 1174ff9042b1SMel Gorman */ 1175b8916634SMel Gorman page_locked = trylock_page(page); 1176b8916634SMel Gorman target_nid = mpol_misplaced(page, vma, haddr); 1177b8916634SMel Gorman if (target_nid == -1) { 1178b8916634SMel Gorman /* If the page was locked, there are no parallel migrations */ 1179a54a407fSMel Gorman if (page_locked) 1180b8916634SMel Gorman goto clear_pmdnuma; 11812b4847e7SMel Gorman } 1182cbee9f88SPeter Zijlstra 1183de466bd6SMel Gorman /* Migration could have started since the pmd_trans_migrating check */ 11842b4847e7SMel Gorman if (!page_locked) { 1185bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 1186b8916634SMel Gorman wait_on_page_locked(page); 1187a54a407fSMel Gorman page_nid = -1; 1188b8916634SMel Gorman goto out; 1189b8916634SMel Gorman } 1190b8916634SMel Gorman 11912b4847e7SMel Gorman /* 11922b4847e7SMel Gorman * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 11932b4847e7SMel Gorman * to serialises splits 11942b4847e7SMel Gorman */ 1195b8916634SMel Gorman get_page(page); 1196bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 1197b8916634SMel Gorman anon_vma = page_lock_anon_vma_read(page); 1198b32967ffSMel Gorman 1199c69307d5SPeter Zijlstra /* Confirm the PMD did not change while page_table_lock was released */ 1200bae473a4SKirill A. Shutemov spin_lock(fe->ptl); 1201bae473a4SKirill A. Shutemov if (unlikely(!pmd_same(pmd, *fe->pmd))) { 1202b32967ffSMel Gorman unlock_page(page); 1203b32967ffSMel Gorman put_page(page); 1204a54a407fSMel Gorman page_nid = -1; 1205b32967ffSMel Gorman goto out_unlock; 1206b32967ffSMel Gorman } 1207ff9042b1SMel Gorman 1208c3a489caSMel Gorman /* Bail if we fail to protect against THP splits for any reason */ 1209c3a489caSMel Gorman if (unlikely(!anon_vma)) { 1210c3a489caSMel Gorman put_page(page); 1211c3a489caSMel Gorman page_nid = -1; 1212c3a489caSMel Gorman goto clear_pmdnuma; 1213c3a489caSMel Gorman } 1214c3a489caSMel Gorman 1215a54a407fSMel Gorman /* 1216a54a407fSMel Gorman * Migrate the THP to the requested node, returns with page unlocked 12178a0516edSMel Gorman * and access rights restored. 1218a54a407fSMel Gorman */ 1219bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 1220bae473a4SKirill A. Shutemov migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 1221bae473a4SKirill A. Shutemov fe->pmd, pmd, fe->address, page, target_nid); 12226688cc05SPeter Zijlstra if (migrated) { 12236688cc05SPeter Zijlstra flags |= TNF_MIGRATED; 12248191acbdSMel Gorman page_nid = target_nid; 1225074c2381SMel Gorman } else 1226074c2381SMel Gorman flags |= TNF_MIGRATE_FAIL; 1227b32967ffSMel Gorman 12288191acbdSMel Gorman goto out; 12294daae3b4SMel Gorman clear_pmdnuma: 1230a54a407fSMel Gorman BUG_ON(!PageLocked(page)); 1231b191f9b1SMel Gorman was_writable = pmd_write(pmd); 12324d942466SMel Gorman pmd = pmd_modify(pmd, vma->vm_page_prot); 1233b7b04004SMel Gorman pmd = pmd_mkyoung(pmd); 1234b191f9b1SMel Gorman if (was_writable) 1235b191f9b1SMel Gorman pmd = pmd_mkwrite(pmd); 1236bae473a4SKirill A. Shutemov set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd); 1237bae473a4SKirill A. Shutemov update_mmu_cache_pmd(vma, fe->address, fe->pmd); 1238a54a407fSMel Gorman unlock_page(page); 1239d10e63f2SMel Gorman out_unlock: 1240bae473a4SKirill A. Shutemov spin_unlock(fe->ptl); 1241b8916634SMel Gorman 1242b8916634SMel Gorman out: 1243b8916634SMel Gorman if (anon_vma) 1244b8916634SMel Gorman page_unlock_anon_vma_read(anon_vma); 1245b8916634SMel Gorman 12468191acbdSMel Gorman if (page_nid != -1) 1247bae473a4SKirill A. Shutemov task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags); 12488191acbdSMel Gorman 1249d10e63f2SMel Gorman return 0; 1250d10e63f2SMel Gorman } 1251d10e63f2SMel Gorman 1252*319904adSHuang Ying /* 1253*319904adSHuang Ying * Return true if we do MADV_FREE successfully on entire pmd page. 1254*319904adSHuang Ying * Otherwise, return false. 1255*319904adSHuang Ying */ 1256*319904adSHuang Ying bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1257b8d3c4c3SMinchan Kim pmd_t *pmd, unsigned long addr, unsigned long next) 1258b8d3c4c3SMinchan Kim { 1259b8d3c4c3SMinchan Kim spinlock_t *ptl; 1260b8d3c4c3SMinchan Kim pmd_t orig_pmd; 1261b8d3c4c3SMinchan Kim struct page *page; 1262b8d3c4c3SMinchan Kim struct mm_struct *mm = tlb->mm; 1263*319904adSHuang Ying bool ret = false; 1264b8d3c4c3SMinchan Kim 1265b6ec57f4SKirill A. Shutemov ptl = pmd_trans_huge_lock(pmd, vma); 1266b6ec57f4SKirill A. Shutemov if (!ptl) 126725eedabeSLinus Torvalds goto out_unlocked; 1268b8d3c4c3SMinchan Kim 1269b8d3c4c3SMinchan Kim orig_pmd = *pmd; 1270*319904adSHuang Ying if (is_huge_zero_pmd(orig_pmd)) 1271b8d3c4c3SMinchan Kim goto out; 1272b8d3c4c3SMinchan Kim 1273b8d3c4c3SMinchan Kim page = pmd_page(orig_pmd); 1274b8d3c4c3SMinchan Kim /* 1275b8d3c4c3SMinchan Kim * If other processes are mapping this page, we couldn't discard 1276b8d3c4c3SMinchan Kim * the page unless they all do MADV_FREE so let's skip the page. 1277b8d3c4c3SMinchan Kim */ 1278b8d3c4c3SMinchan Kim if (page_mapcount(page) != 1) 1279b8d3c4c3SMinchan Kim goto out; 1280b8d3c4c3SMinchan Kim 1281b8d3c4c3SMinchan Kim if (!trylock_page(page)) 1282b8d3c4c3SMinchan Kim goto out; 1283b8d3c4c3SMinchan Kim 1284b8d3c4c3SMinchan Kim /* 1285b8d3c4c3SMinchan Kim * If user want to discard part-pages of THP, split it so MADV_FREE 1286b8d3c4c3SMinchan Kim * will deactivate only them. 1287b8d3c4c3SMinchan Kim */ 1288b8d3c4c3SMinchan Kim if (next - addr != HPAGE_PMD_SIZE) { 1289b8d3c4c3SMinchan Kim get_page(page); 1290b8d3c4c3SMinchan Kim spin_unlock(ptl); 12919818b8cdSHuang Ying split_huge_page(page); 1292b8d3c4c3SMinchan Kim put_page(page); 1293b8d3c4c3SMinchan Kim unlock_page(page); 1294b8d3c4c3SMinchan Kim goto out_unlocked; 1295b8d3c4c3SMinchan Kim } 1296b8d3c4c3SMinchan Kim 1297b8d3c4c3SMinchan Kim if (PageDirty(page)) 1298b8d3c4c3SMinchan Kim ClearPageDirty(page); 1299b8d3c4c3SMinchan Kim unlock_page(page); 1300b8d3c4c3SMinchan Kim 1301b8d3c4c3SMinchan Kim if (PageActive(page)) 1302b8d3c4c3SMinchan Kim deactivate_page(page); 1303b8d3c4c3SMinchan Kim 1304b8d3c4c3SMinchan Kim if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1305b8d3c4c3SMinchan Kim orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1306b8d3c4c3SMinchan Kim tlb->fullmm); 1307b8d3c4c3SMinchan Kim orig_pmd = pmd_mkold(orig_pmd); 1308b8d3c4c3SMinchan Kim orig_pmd = pmd_mkclean(orig_pmd); 1309b8d3c4c3SMinchan Kim 1310b8d3c4c3SMinchan Kim set_pmd_at(mm, addr, pmd, orig_pmd); 1311b8d3c4c3SMinchan Kim tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1312b8d3c4c3SMinchan Kim } 1313*319904adSHuang Ying ret = true; 1314b8d3c4c3SMinchan Kim out: 1315b8d3c4c3SMinchan Kim spin_unlock(ptl); 1316b8d3c4c3SMinchan Kim out_unlocked: 1317b8d3c4c3SMinchan Kim return ret; 1318b8d3c4c3SMinchan Kim } 1319b8d3c4c3SMinchan Kim 132071e3aac0SAndrea Arcangeli int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1321f21760b1SShaohua Li pmd_t *pmd, unsigned long addr) 132271e3aac0SAndrea Arcangeli { 1323f5c8ad47SDavid Miller pmd_t orig_pmd; 1324da146769SKirill A. Shutemov spinlock_t *ptl; 1325da146769SKirill A. Shutemov 1326b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 1327b6ec57f4SKirill A. Shutemov if (!ptl) 1328da146769SKirill A. Shutemov return 0; 1329a6bf2bb0SAneesh Kumar K.V /* 1330a6bf2bb0SAneesh Kumar K.V * For architectures like ppc64 we look at deposited pgtable 13318809aa2dSAneesh Kumar K.V * when calling pmdp_huge_get_and_clear. So do the 1332a6bf2bb0SAneesh Kumar K.V * pgtable_trans_huge_withdraw after finishing pmdp related 1333a6bf2bb0SAneesh Kumar K.V * operations. 1334a6bf2bb0SAneesh Kumar K.V */ 13358809aa2dSAneesh Kumar K.V orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1336fcbe08d6SMartin Schwidefsky tlb->fullmm); 1337f21760b1SShaohua Li tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 13384897c765SMatthew Wilcox if (vma_is_dax(vma)) { 13394897c765SMatthew Wilcox spin_unlock(ptl); 1340da146769SKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 1341aa88b68cSKirill A. Shutemov tlb_remove_page(tlb, pmd_page(orig_pmd)); 1342da146769SKirill A. Shutemov } else if (is_huge_zero_pmd(orig_pmd)) { 1343da146769SKirill A. Shutemov pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); 1344e1f56c89SKirill A. Shutemov atomic_long_dec(&tlb->mm->nr_ptes); 1345bf929152SKirill A. Shutemov spin_unlock(ptl); 1346aa88b68cSKirill A. Shutemov tlb_remove_page(tlb, pmd_page(orig_pmd)); 1347479f0abbSKirill A. Shutemov } else { 13484897c765SMatthew Wilcox struct page *page = pmd_page(orig_pmd); 1349d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 1350309381feSSasha Levin VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1351309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1352b5072380SKirill A. Shutemov if (PageAnon(page)) { 1353b5072380SKirill A. Shutemov pgtable_t pgtable; 1354b5072380SKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd); 1355b5072380SKirill A. Shutemov pte_free(tlb->mm, pgtable); 1356e1f56c89SKirill A. Shutemov atomic_long_dec(&tlb->mm->nr_ptes); 1357b5072380SKirill A. Shutemov add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1358b5072380SKirill A. Shutemov } else { 1359b5072380SKirill A. Shutemov add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR); 1360b5072380SKirill A. Shutemov } 1361bf929152SKirill A. Shutemov spin_unlock(ptl); 1362e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1363479f0abbSKirill A. Shutemov } 1364da146769SKirill A. Shutemov return 1; 136571e3aac0SAndrea Arcangeli } 136671e3aac0SAndrea Arcangeli 1367bf8616d5SHugh Dickins bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 136837a1c49aSAndrea Arcangeli unsigned long new_addr, unsigned long old_end, 136937a1c49aSAndrea Arcangeli pmd_t *old_pmd, pmd_t *new_pmd) 137037a1c49aSAndrea Arcangeli { 1371bf929152SKirill A. Shutemov spinlock_t *old_ptl, *new_ptl; 137237a1c49aSAndrea Arcangeli pmd_t pmd; 137337a1c49aSAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 137437a1c49aSAndrea Arcangeli 137537a1c49aSAndrea Arcangeli if ((old_addr & ~HPAGE_PMD_MASK) || 137637a1c49aSAndrea Arcangeli (new_addr & ~HPAGE_PMD_MASK) || 1377bf8616d5SHugh Dickins old_end - old_addr < HPAGE_PMD_SIZE) 13784b471e88SKirill A. Shutemov return false; 137937a1c49aSAndrea Arcangeli 138037a1c49aSAndrea Arcangeli /* 138137a1c49aSAndrea Arcangeli * The destination pmd shouldn't be established, free_pgtables() 138237a1c49aSAndrea Arcangeli * should have release it. 138337a1c49aSAndrea Arcangeli */ 138437a1c49aSAndrea Arcangeli if (WARN_ON(!pmd_none(*new_pmd))) { 138537a1c49aSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*new_pmd)); 13864b471e88SKirill A. Shutemov return false; 138737a1c49aSAndrea Arcangeli } 138837a1c49aSAndrea Arcangeli 1389bf929152SKirill A. Shutemov /* 1390bf929152SKirill A. Shutemov * We don't have to worry about the ordering of src and dst 1391bf929152SKirill A. Shutemov * ptlocks because exclusive mmap_sem prevents deadlock. 1392bf929152SKirill A. Shutemov */ 1393b6ec57f4SKirill A. Shutemov old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1394b6ec57f4SKirill A. Shutemov if (old_ptl) { 1395bf929152SKirill A. Shutemov new_ptl = pmd_lockptr(mm, new_pmd); 1396bf929152SKirill A. Shutemov if (new_ptl != old_ptl) 1397bf929152SKirill A. Shutemov spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 13988809aa2dSAneesh Kumar K.V pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 139937a1c49aSAndrea Arcangeli VM_BUG_ON(!pmd_none(*new_pmd)); 14003592806cSKirill A. Shutemov 140169a8ec2dSKirill A. Shutemov if (pmd_move_must_withdraw(new_ptl, old_ptl) && 140269a8ec2dSKirill A. Shutemov vma_is_anonymous(vma)) { 1403b3084f4dSAneesh Kumar K.V pgtable_t pgtable; 14043592806cSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 14053592806cSKirill A. Shutemov pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 14063592806cSKirill A. Shutemov } 1407b3084f4dSAneesh Kumar K.V set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); 1408b3084f4dSAneesh Kumar K.V if (new_ptl != old_ptl) 1409b3084f4dSAneesh Kumar K.V spin_unlock(new_ptl); 1410bf929152SKirill A. Shutemov spin_unlock(old_ptl); 14114b471e88SKirill A. Shutemov return true; 141237a1c49aSAndrea Arcangeli } 14134b471e88SKirill A. Shutemov return false; 141437a1c49aSAndrea Arcangeli } 141537a1c49aSAndrea Arcangeli 1416f123d74aSMel Gorman /* 1417f123d74aSMel Gorman * Returns 1418f123d74aSMel Gorman * - 0 if PMD could not be locked 1419f123d74aSMel Gorman * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1420f123d74aSMel Gorman * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1421f123d74aSMel Gorman */ 1422cd7548abSJohannes Weiner int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1423e944fd67SMel Gorman unsigned long addr, pgprot_t newprot, int prot_numa) 1424cd7548abSJohannes Weiner { 1425cd7548abSJohannes Weiner struct mm_struct *mm = vma->vm_mm; 1426bf929152SKirill A. Shutemov spinlock_t *ptl; 1427cd7548abSJohannes Weiner int ret = 0; 1428cd7548abSJohannes Weiner 1429b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 1430b6ec57f4SKirill A. Shutemov if (ptl) { 1431cd7548abSJohannes Weiner pmd_t entry; 1432b191f9b1SMel Gorman bool preserve_write = prot_numa && pmd_write(*pmd); 1433ba68bc01SMel Gorman ret = 1; 1434e944fd67SMel Gorman 1435e944fd67SMel Gorman /* 1436e944fd67SMel Gorman * Avoid trapping faults against the zero page. The read-only 1437e944fd67SMel Gorman * data is likely to be read-cached on the local CPU and 1438e944fd67SMel Gorman * local/remote hits to the zero page are not interesting. 1439e944fd67SMel Gorman */ 1440e944fd67SMel Gorman if (prot_numa && is_huge_zero_pmd(*pmd)) { 1441e944fd67SMel Gorman spin_unlock(ptl); 1442ba68bc01SMel Gorman return ret; 1443e944fd67SMel Gorman } 1444e944fd67SMel Gorman 144510c1045fSMel Gorman if (!prot_numa || !pmd_protnone(*pmd)) { 14468809aa2dSAneesh Kumar K.V entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); 1447cd7548abSJohannes Weiner entry = pmd_modify(entry, newprot); 1448b191f9b1SMel Gorman if (preserve_write) 1449b191f9b1SMel Gorman entry = pmd_mkwrite(entry); 1450f123d74aSMel Gorman ret = HPAGE_PMD_NR; 145156eecdb9SAneesh Kumar K.V set_pmd_at(mm, addr, pmd, entry); 1452b237adedSKirill A. Shutemov BUG_ON(vma_is_anonymous(vma) && !preserve_write && 1453b237adedSKirill A. Shutemov pmd_write(entry)); 145410c1045fSMel Gorman } 1455bf929152SKirill A. Shutemov spin_unlock(ptl); 1456cd7548abSJohannes Weiner } 1457cd7548abSJohannes Weiner 1458cd7548abSJohannes Weiner return ret; 1459cd7548abSJohannes Weiner } 1460cd7548abSJohannes Weiner 1461025c5b24SNaoya Horiguchi /* 14628f19b0c0SHuang Ying * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1463025c5b24SNaoya Horiguchi * 14648f19b0c0SHuang Ying * Note that if it returns page table lock pointer, this routine returns without 14658f19b0c0SHuang Ying * unlocking page table lock. So callers must unlock it. 1466025c5b24SNaoya Horiguchi */ 1467b6ec57f4SKirill A. Shutemov spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1468025c5b24SNaoya Horiguchi { 1469b6ec57f4SKirill A. Shutemov spinlock_t *ptl; 1470b6ec57f4SKirill A. Shutemov ptl = pmd_lock(vma->vm_mm, pmd); 14715c7fb56eSDan Williams if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) 1472b6ec57f4SKirill A. Shutemov return ptl; 1473b6ec57f4SKirill A. Shutemov spin_unlock(ptl); 1474b6ec57f4SKirill A. Shutemov return NULL; 1475025c5b24SNaoya Horiguchi } 1476025c5b24SNaoya Horiguchi 1477eef1b3baSKirill A. Shutemov static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 1478eef1b3baSKirill A. Shutemov unsigned long haddr, pmd_t *pmd) 1479eef1b3baSKirill A. Shutemov { 1480eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 1481eef1b3baSKirill A. Shutemov pgtable_t pgtable; 1482eef1b3baSKirill A. Shutemov pmd_t _pmd; 1483eef1b3baSKirill A. Shutemov int i; 1484eef1b3baSKirill A. Shutemov 1485eef1b3baSKirill A. Shutemov /* leave pmd empty until pte is filled */ 1486eef1b3baSKirill A. Shutemov pmdp_huge_clear_flush_notify(vma, haddr, pmd); 1487eef1b3baSKirill A. Shutemov 1488eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1489eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 1490eef1b3baSKirill A. Shutemov 1491eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1492eef1b3baSKirill A. Shutemov pte_t *pte, entry; 1493eef1b3baSKirill A. Shutemov entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 1494eef1b3baSKirill A. Shutemov entry = pte_mkspecial(entry); 1495eef1b3baSKirill A. Shutemov pte = pte_offset_map(&_pmd, haddr); 1496eef1b3baSKirill A. Shutemov VM_BUG_ON(!pte_none(*pte)); 1497eef1b3baSKirill A. Shutemov set_pte_at(mm, haddr, pte, entry); 1498eef1b3baSKirill A. Shutemov pte_unmap(pte); 1499eef1b3baSKirill A. Shutemov } 1500eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 1501eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 1502eef1b3baSKirill A. Shutemov put_huge_zero_page(); 1503eef1b3baSKirill A. Shutemov } 1504eef1b3baSKirill A. Shutemov 1505eef1b3baSKirill A. Shutemov static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 1506ba988280SKirill A. Shutemov unsigned long haddr, bool freeze) 1507eef1b3baSKirill A. Shutemov { 1508eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 1509eef1b3baSKirill A. Shutemov struct page *page; 1510eef1b3baSKirill A. Shutemov pgtable_t pgtable; 1511eef1b3baSKirill A. Shutemov pmd_t _pmd; 1512b8d3c4c3SMinchan Kim bool young, write, dirty; 15132ac015e2SKirill A. Shutemov unsigned long addr; 1514eef1b3baSKirill A. Shutemov int i; 1515eef1b3baSKirill A. Shutemov 1516eef1b3baSKirill A. Shutemov VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 1517eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 1518eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 15195c7fb56eSDan Williams VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)); 1520eef1b3baSKirill A. Shutemov 1521eef1b3baSKirill A. Shutemov count_vm_event(THP_SPLIT_PMD); 1522eef1b3baSKirill A. Shutemov 1523d21b9e57SKirill A. Shutemov if (!vma_is_anonymous(vma)) { 1524d21b9e57SKirill A. Shutemov _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 1525eef1b3baSKirill A. Shutemov if (is_huge_zero_pmd(_pmd)) 1526eef1b3baSKirill A. Shutemov put_huge_zero_page(); 1527d21b9e57SKirill A. Shutemov if (vma_is_dax(vma)) 1528d21b9e57SKirill A. Shutemov return; 1529d21b9e57SKirill A. Shutemov page = pmd_page(_pmd); 1530d21b9e57SKirill A. Shutemov if (!PageReferenced(page) && pmd_young(_pmd)) 1531d21b9e57SKirill A. Shutemov SetPageReferenced(page); 1532d21b9e57SKirill A. Shutemov page_remove_rmap(page, true); 1533d21b9e57SKirill A. Shutemov put_page(page); 1534d21b9e57SKirill A. Shutemov add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR); 1535eef1b3baSKirill A. Shutemov return; 1536eef1b3baSKirill A. Shutemov } else if (is_huge_zero_pmd(*pmd)) { 1537eef1b3baSKirill A. Shutemov return __split_huge_zero_page_pmd(vma, haddr, pmd); 1538eef1b3baSKirill A. Shutemov } 1539eef1b3baSKirill A. Shutemov 1540eef1b3baSKirill A. Shutemov page = pmd_page(*pmd); 1541eef1b3baSKirill A. Shutemov VM_BUG_ON_PAGE(!page_count(page), page); 1542fe896d18SJoonsoo Kim page_ref_add(page, HPAGE_PMD_NR - 1); 1543eef1b3baSKirill A. Shutemov write = pmd_write(*pmd); 1544eef1b3baSKirill A. Shutemov young = pmd_young(*pmd); 1545b8d3c4c3SMinchan Kim dirty = pmd_dirty(*pmd); 1546eef1b3baSKirill A. Shutemov 1547c777e2a8SAneesh Kumar K.V pmdp_huge_split_prepare(vma, haddr, pmd); 1548eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1549eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 1550eef1b3baSKirill A. Shutemov 15512ac015e2SKirill A. Shutemov for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 1552eef1b3baSKirill A. Shutemov pte_t entry, *pte; 1553eef1b3baSKirill A. Shutemov /* 1554eef1b3baSKirill A. Shutemov * Note that NUMA hinting access restrictions are not 1555eef1b3baSKirill A. Shutemov * transferred to avoid any possibility of altering 1556eef1b3baSKirill A. Shutemov * permissions across VMAs. 1557eef1b3baSKirill A. Shutemov */ 1558ba988280SKirill A. Shutemov if (freeze) { 1559ba988280SKirill A. Shutemov swp_entry_t swp_entry; 1560ba988280SKirill A. Shutemov swp_entry = make_migration_entry(page + i, write); 1561ba988280SKirill A. Shutemov entry = swp_entry_to_pte(swp_entry); 1562ba988280SKirill A. Shutemov } else { 1563eef1b3baSKirill A. Shutemov entry = mk_pte(page + i, vma->vm_page_prot); 1564b8d3c4c3SMinchan Kim entry = maybe_mkwrite(entry, vma); 1565eef1b3baSKirill A. Shutemov if (!write) 1566eef1b3baSKirill A. Shutemov entry = pte_wrprotect(entry); 1567eef1b3baSKirill A. Shutemov if (!young) 1568eef1b3baSKirill A. Shutemov entry = pte_mkold(entry); 1569ba988280SKirill A. Shutemov } 1570b8d3c4c3SMinchan Kim if (dirty) 1571b8d3c4c3SMinchan Kim SetPageDirty(page + i); 15722ac015e2SKirill A. Shutemov pte = pte_offset_map(&_pmd, addr); 1573eef1b3baSKirill A. Shutemov BUG_ON(!pte_none(*pte)); 15742ac015e2SKirill A. Shutemov set_pte_at(mm, addr, pte, entry); 1575eef1b3baSKirill A. Shutemov atomic_inc(&page[i]._mapcount); 1576eef1b3baSKirill A. Shutemov pte_unmap(pte); 1577eef1b3baSKirill A. Shutemov } 1578eef1b3baSKirill A. Shutemov 1579eef1b3baSKirill A. Shutemov /* 1580eef1b3baSKirill A. Shutemov * Set PG_double_map before dropping compound_mapcount to avoid 1581eef1b3baSKirill A. Shutemov * false-negative page_mapped(). 1582eef1b3baSKirill A. Shutemov */ 1583eef1b3baSKirill A. Shutemov if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 1584eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 1585eef1b3baSKirill A. Shutemov atomic_inc(&page[i]._mapcount); 1586eef1b3baSKirill A. Shutemov } 1587eef1b3baSKirill A. Shutemov 1588eef1b3baSKirill A. Shutemov if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 1589eef1b3baSKirill A. Shutemov /* Last compound_mapcount is gone. */ 159011fb9989SMel Gorman __dec_node_page_state(page, NR_ANON_THPS); 1591eef1b3baSKirill A. Shutemov if (TestClearPageDoubleMap(page)) { 1592eef1b3baSKirill A. Shutemov /* No need in mapcount reference anymore */ 1593eef1b3baSKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 1594eef1b3baSKirill A. Shutemov atomic_dec(&page[i]._mapcount); 1595eef1b3baSKirill A. Shutemov } 1596eef1b3baSKirill A. Shutemov } 1597eef1b3baSKirill A. Shutemov 1598eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 1599e9b61f19SKirill A. Shutemov /* 1600e9b61f19SKirill A. Shutemov * Up to this point the pmd is present and huge and userland has the 1601e9b61f19SKirill A. Shutemov * whole access to the hugepage during the split (which happens in 1602e9b61f19SKirill A. Shutemov * place). If we overwrite the pmd with the not-huge version pointing 1603e9b61f19SKirill A. Shutemov * to the pte here (which of course we could if all CPUs were bug 1604e9b61f19SKirill A. Shutemov * free), userland could trigger a small page size TLB miss on the 1605e9b61f19SKirill A. Shutemov * small sized TLB while the hugepage TLB entry is still established in 1606e9b61f19SKirill A. Shutemov * the huge TLB. Some CPU doesn't like that. 1607e9b61f19SKirill A. Shutemov * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 1608e9b61f19SKirill A. Shutemov * 383 on page 93. Intel should be safe but is also warns that it's 1609e9b61f19SKirill A. Shutemov * only safe if the permission and cache attributes of the two entries 1610e9b61f19SKirill A. Shutemov * loaded in the two TLB is identical (which should be the case here). 1611e9b61f19SKirill A. Shutemov * But it is generally safer to never allow small and huge TLB entries 1612e9b61f19SKirill A. Shutemov * for the same virtual address to be loaded simultaneously. So instead 1613e9b61f19SKirill A. Shutemov * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 1614e9b61f19SKirill A. Shutemov * current pmd notpresent (atomically because here the pmd_trans_huge 1615e9b61f19SKirill A. Shutemov * and pmd_trans_splitting must remain set at all times on the pmd 1616e9b61f19SKirill A. Shutemov * until the split is complete for this pmd), then we flush the SMP TLB 1617e9b61f19SKirill A. Shutemov * and finally we write the non-huge version of the pmd entry with 1618e9b61f19SKirill A. Shutemov * pmd_populate. 1619e9b61f19SKirill A. Shutemov */ 1620e9b61f19SKirill A. Shutemov pmdp_invalidate(vma, haddr, pmd); 1621eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 1622e9b61f19SKirill A. Shutemov 1623e9b61f19SKirill A. Shutemov if (freeze) { 16242ac015e2SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) { 1625e9b61f19SKirill A. Shutemov page_remove_rmap(page + i, false); 1626e9b61f19SKirill A. Shutemov put_page(page + i); 1627e9b61f19SKirill A. Shutemov } 1628e9b61f19SKirill A. Shutemov } 1629eef1b3baSKirill A. Shutemov } 1630eef1b3baSKirill A. Shutemov 1631eef1b3baSKirill A. Shutemov void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 163233f4751eSNaoya Horiguchi unsigned long address, bool freeze, struct page *page) 1633eef1b3baSKirill A. Shutemov { 1634eef1b3baSKirill A. Shutemov spinlock_t *ptl; 1635eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 1636eef1b3baSKirill A. Shutemov unsigned long haddr = address & HPAGE_PMD_MASK; 1637eef1b3baSKirill A. Shutemov 1638eef1b3baSKirill A. Shutemov mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); 1639eef1b3baSKirill A. Shutemov ptl = pmd_lock(mm, pmd); 164033f4751eSNaoya Horiguchi 164133f4751eSNaoya Horiguchi /* 164233f4751eSNaoya Horiguchi * If caller asks to setup a migration entries, we need a page to check 164333f4751eSNaoya Horiguchi * pmd against. Otherwise we can end up replacing wrong page. 164433f4751eSNaoya Horiguchi */ 164533f4751eSNaoya Horiguchi VM_BUG_ON(freeze && !page); 164633f4751eSNaoya Horiguchi if (page && page != pmd_page(*pmd)) 164733f4751eSNaoya Horiguchi goto out; 164833f4751eSNaoya Horiguchi 16495c7fb56eSDan Williams if (pmd_trans_huge(*pmd)) { 165033f4751eSNaoya Horiguchi page = pmd_page(*pmd); 1651e90309c9SKirill A. Shutemov if (PageMlocked(page)) 16525f737714SKirill A. Shutemov clear_page_mlock(page); 16535c7fb56eSDan Williams } else if (!pmd_devmap(*pmd)) 16545c7fb56eSDan Williams goto out; 1655fec89c10SKirill A. Shutemov __split_huge_pmd_locked(vma, pmd, haddr, freeze); 1656e90309c9SKirill A. Shutemov out: 1657eef1b3baSKirill A. Shutemov spin_unlock(ptl); 1658eef1b3baSKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); 1659eef1b3baSKirill A. Shutemov } 1660eef1b3baSKirill A. Shutemov 1661fec89c10SKirill A. Shutemov void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 1662fec89c10SKirill A. Shutemov bool freeze, struct page *page) 166394fcc585SAndrea Arcangeli { 1664f72e7dcdSHugh Dickins pgd_t *pgd; 1665f72e7dcdSHugh Dickins pud_t *pud; 166694fcc585SAndrea Arcangeli pmd_t *pmd; 166794fcc585SAndrea Arcangeli 166878ddc534SKirill A. Shutemov pgd = pgd_offset(vma->vm_mm, address); 1669f72e7dcdSHugh Dickins if (!pgd_present(*pgd)) 1670f72e7dcdSHugh Dickins return; 1671f72e7dcdSHugh Dickins 1672f72e7dcdSHugh Dickins pud = pud_offset(pgd, address); 1673f72e7dcdSHugh Dickins if (!pud_present(*pud)) 1674f72e7dcdSHugh Dickins return; 1675f72e7dcdSHugh Dickins 1676f72e7dcdSHugh Dickins pmd = pmd_offset(pud, address); 1677fec89c10SKirill A. Shutemov 167833f4751eSNaoya Horiguchi __split_huge_pmd(vma, pmd, address, freeze, page); 167994fcc585SAndrea Arcangeli } 168094fcc585SAndrea Arcangeli 1681e1b9996bSKirill A. Shutemov void vma_adjust_trans_huge(struct vm_area_struct *vma, 168294fcc585SAndrea Arcangeli unsigned long start, 168394fcc585SAndrea Arcangeli unsigned long end, 168494fcc585SAndrea Arcangeli long adjust_next) 168594fcc585SAndrea Arcangeli { 168694fcc585SAndrea Arcangeli /* 168794fcc585SAndrea Arcangeli * If the new start address isn't hpage aligned and it could 168894fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 168994fcc585SAndrea Arcangeli * an huge pmd. 169094fcc585SAndrea Arcangeli */ 169194fcc585SAndrea Arcangeli if (start & ~HPAGE_PMD_MASK && 169294fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) >= vma->vm_start && 169394fcc585SAndrea Arcangeli (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 1694fec89c10SKirill A. Shutemov split_huge_pmd_address(vma, start, false, NULL); 169594fcc585SAndrea Arcangeli 169694fcc585SAndrea Arcangeli /* 169794fcc585SAndrea Arcangeli * If the new end address isn't hpage aligned and it could 169894fcc585SAndrea Arcangeli * previously contain an hugepage: check if we need to split 169994fcc585SAndrea Arcangeli * an huge pmd. 170094fcc585SAndrea Arcangeli */ 170194fcc585SAndrea Arcangeli if (end & ~HPAGE_PMD_MASK && 170294fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) >= vma->vm_start && 170394fcc585SAndrea Arcangeli (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 1704fec89c10SKirill A. Shutemov split_huge_pmd_address(vma, end, false, NULL); 170594fcc585SAndrea Arcangeli 170694fcc585SAndrea Arcangeli /* 170794fcc585SAndrea Arcangeli * If we're also updating the vma->vm_next->vm_start, if the new 170894fcc585SAndrea Arcangeli * vm_next->vm_start isn't page aligned and it could previously 170994fcc585SAndrea Arcangeli * contain an hugepage: check if we need to split an huge pmd. 171094fcc585SAndrea Arcangeli */ 171194fcc585SAndrea Arcangeli if (adjust_next > 0) { 171294fcc585SAndrea Arcangeli struct vm_area_struct *next = vma->vm_next; 171394fcc585SAndrea Arcangeli unsigned long nstart = next->vm_start; 171494fcc585SAndrea Arcangeli nstart += adjust_next << PAGE_SHIFT; 171594fcc585SAndrea Arcangeli if (nstart & ~HPAGE_PMD_MASK && 171694fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) >= next->vm_start && 171794fcc585SAndrea Arcangeli (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 1718fec89c10SKirill A. Shutemov split_huge_pmd_address(next, nstart, false, NULL); 171994fcc585SAndrea Arcangeli } 172094fcc585SAndrea Arcangeli } 1721e9b61f19SKirill A. Shutemov 1722fec89c10SKirill A. Shutemov static void freeze_page(struct page *page) 1723e9b61f19SKirill A. Shutemov { 1724baa355fdSKirill A. Shutemov enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 1725baa355fdSKirill A. Shutemov TTU_RMAP_LOCKED; 1726fec89c10SKirill A. Shutemov int i, ret; 1727e9b61f19SKirill A. Shutemov 1728e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageHead(page), page); 1729e9b61f19SKirill A. Shutemov 1730baa355fdSKirill A. Shutemov if (PageAnon(page)) 1731baa355fdSKirill A. Shutemov ttu_flags |= TTU_MIGRATION; 1732baa355fdSKirill A. Shutemov 1733fec89c10SKirill A. Shutemov /* We only need TTU_SPLIT_HUGE_PMD once */ 1734fec89c10SKirill A. Shutemov ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD); 1735fec89c10SKirill A. Shutemov for (i = 1; !ret && i < HPAGE_PMD_NR; i++) { 1736fec89c10SKirill A. Shutemov /* Cut short if the page is unmapped */ 1737fec89c10SKirill A. Shutemov if (page_count(page) == 1) 1738e9b61f19SKirill A. Shutemov return; 1739bd56086fSKirill A. Shutemov 1740fec89c10SKirill A. Shutemov ret = try_to_unmap(page + i, ttu_flags); 1741fec89c10SKirill A. Shutemov } 1742baa355fdSKirill A. Shutemov VM_BUG_ON_PAGE(ret, page + i - 1); 1743bd56086fSKirill A. Shutemov } 1744bd56086fSKirill A. Shutemov 1745fec89c10SKirill A. Shutemov static void unfreeze_page(struct page *page) 1746e9b61f19SKirill A. Shutemov { 1747fec89c10SKirill A. Shutemov int i; 1748e9b61f19SKirill A. Shutemov 1749fec89c10SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 1750fec89c10SKirill A. Shutemov remove_migration_ptes(page + i, page + i, true); 1751e9b61f19SKirill A. Shutemov } 1752e9b61f19SKirill A. Shutemov 17538df651c7SKirill A. Shutemov static void __split_huge_page_tail(struct page *head, int tail, 1754e9b61f19SKirill A. Shutemov struct lruvec *lruvec, struct list_head *list) 1755e9b61f19SKirill A. Shutemov { 1756e9b61f19SKirill A. Shutemov struct page *page_tail = head + tail; 1757e9b61f19SKirill A. Shutemov 17588df651c7SKirill A. Shutemov VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 1759fe896d18SJoonsoo Kim VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail); 1760e9b61f19SKirill A. Shutemov 1761e9b61f19SKirill A. Shutemov /* 17620139aa7bSJoonsoo Kim * tail_page->_refcount is zero and not changing from under us. But 1763e9b61f19SKirill A. Shutemov * get_page_unless_zero() may be running from under us on the 1764baa355fdSKirill A. Shutemov * tail_page. If we used atomic_set() below instead of atomic_inc() or 1765baa355fdSKirill A. Shutemov * atomic_add(), we would then run atomic_set() concurrently with 1766e9b61f19SKirill A. Shutemov * get_page_unless_zero(), and atomic_set() is implemented in C not 1767e9b61f19SKirill A. Shutemov * using locked ops. spin_unlock on x86 sometime uses locked ops 1768e9b61f19SKirill A. Shutemov * because of PPro errata 66, 92, so unless somebody can guarantee 1769e9b61f19SKirill A. Shutemov * atomic_set() here would be safe on all archs (and not only on x86), 1770baa355fdSKirill A. Shutemov * it's safer to use atomic_inc()/atomic_add(). 1771e9b61f19SKirill A. Shutemov */ 1772baa355fdSKirill A. Shutemov if (PageAnon(head)) { 1773fe896d18SJoonsoo Kim page_ref_inc(page_tail); 1774baa355fdSKirill A. Shutemov } else { 1775baa355fdSKirill A. Shutemov /* Additional pin to radix tree */ 1776baa355fdSKirill A. Shutemov page_ref_add(page_tail, 2); 1777baa355fdSKirill A. Shutemov } 1778e9b61f19SKirill A. Shutemov 1779e9b61f19SKirill A. Shutemov page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 1780e9b61f19SKirill A. Shutemov page_tail->flags |= (head->flags & 1781e9b61f19SKirill A. Shutemov ((1L << PG_referenced) | 1782e9b61f19SKirill A. Shutemov (1L << PG_swapbacked) | 1783e9b61f19SKirill A. Shutemov (1L << PG_mlocked) | 1784e9b61f19SKirill A. Shutemov (1L << PG_uptodate) | 1785e9b61f19SKirill A. Shutemov (1L << PG_active) | 1786e9b61f19SKirill A. Shutemov (1L << PG_locked) | 1787b8d3c4c3SMinchan Kim (1L << PG_unevictable) | 1788b8d3c4c3SMinchan Kim (1L << PG_dirty))); 1789e9b61f19SKirill A. Shutemov 1790e9b61f19SKirill A. Shutemov /* 1791e9b61f19SKirill A. Shutemov * After clearing PageTail the gup refcount can be released. 1792e9b61f19SKirill A. Shutemov * Page flags also must be visible before we make the page non-compound. 1793e9b61f19SKirill A. Shutemov */ 1794e9b61f19SKirill A. Shutemov smp_wmb(); 1795e9b61f19SKirill A. Shutemov 1796e9b61f19SKirill A. Shutemov clear_compound_head(page_tail); 1797e9b61f19SKirill A. Shutemov 1798e9b61f19SKirill A. Shutemov if (page_is_young(head)) 1799e9b61f19SKirill A. Shutemov set_page_young(page_tail); 1800e9b61f19SKirill A. Shutemov if (page_is_idle(head)) 1801e9b61f19SKirill A. Shutemov set_page_idle(page_tail); 1802e9b61f19SKirill A. Shutemov 1803e9b61f19SKirill A. Shutemov /* ->mapping in first tail page is compound_mapcount */ 18049a982250SKirill A. Shutemov VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 1805e9b61f19SKirill A. Shutemov page_tail); 1806e9b61f19SKirill A. Shutemov page_tail->mapping = head->mapping; 1807e9b61f19SKirill A. Shutemov 1808e9b61f19SKirill A. Shutemov page_tail->index = head->index + tail; 1809e9b61f19SKirill A. Shutemov page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 1810e9b61f19SKirill A. Shutemov lru_add_page_tail(head, page_tail, lruvec, list); 1811e9b61f19SKirill A. Shutemov } 1812e9b61f19SKirill A. Shutemov 1813baa355fdSKirill A. Shutemov static void __split_huge_page(struct page *page, struct list_head *list, 1814baa355fdSKirill A. Shutemov unsigned long flags) 1815e9b61f19SKirill A. Shutemov { 1816e9b61f19SKirill A. Shutemov struct page *head = compound_head(page); 1817e9b61f19SKirill A. Shutemov struct zone *zone = page_zone(head); 1818e9b61f19SKirill A. Shutemov struct lruvec *lruvec; 1819baa355fdSKirill A. Shutemov pgoff_t end = -1; 18208df651c7SKirill A. Shutemov int i; 1821e9b61f19SKirill A. Shutemov 1822599d0c95SMel Gorman lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); 1823e9b61f19SKirill A. Shutemov 1824e9b61f19SKirill A. Shutemov /* complete memcg works before add pages to LRU */ 1825e9b61f19SKirill A. Shutemov mem_cgroup_split_huge_fixup(head); 1826e9b61f19SKirill A. Shutemov 1827baa355fdSKirill A. Shutemov if (!PageAnon(page)) 1828baa355fdSKirill A. Shutemov end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE); 1829baa355fdSKirill A. Shutemov 1830baa355fdSKirill A. Shutemov for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 18318df651c7SKirill A. Shutemov __split_huge_page_tail(head, i, lruvec, list); 1832baa355fdSKirill A. Shutemov /* Some pages can be beyond i_size: drop them from page cache */ 1833baa355fdSKirill A. Shutemov if (head[i].index >= end) { 1834baa355fdSKirill A. Shutemov __ClearPageDirty(head + i); 1835baa355fdSKirill A. Shutemov __delete_from_page_cache(head + i, NULL); 1836800d8c63SKirill A. Shutemov if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 1837800d8c63SKirill A. Shutemov shmem_uncharge(head->mapping->host, 1); 1838baa355fdSKirill A. Shutemov put_page(head + i); 1839baa355fdSKirill A. Shutemov } 1840baa355fdSKirill A. Shutemov } 1841e9b61f19SKirill A. Shutemov 1842e9b61f19SKirill A. Shutemov ClearPageCompound(head); 1843baa355fdSKirill A. Shutemov /* See comment in __split_huge_page_tail() */ 1844baa355fdSKirill A. Shutemov if (PageAnon(head)) { 1845baa355fdSKirill A. Shutemov page_ref_inc(head); 1846baa355fdSKirill A. Shutemov } else { 1847baa355fdSKirill A. Shutemov /* Additional pin to radix tree */ 1848baa355fdSKirill A. Shutemov page_ref_add(head, 2); 1849baa355fdSKirill A. Shutemov spin_unlock(&head->mapping->tree_lock); 1850baa355fdSKirill A. Shutemov } 1851baa355fdSKirill A. Shutemov 1852a52633d8SMel Gorman spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 1853e9b61f19SKirill A. Shutemov 1854fec89c10SKirill A. Shutemov unfreeze_page(head); 1855e9b61f19SKirill A. Shutemov 1856e9b61f19SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) { 1857e9b61f19SKirill A. Shutemov struct page *subpage = head + i; 1858e9b61f19SKirill A. Shutemov if (subpage == page) 1859e9b61f19SKirill A. Shutemov continue; 1860e9b61f19SKirill A. Shutemov unlock_page(subpage); 1861e9b61f19SKirill A. Shutemov 1862e9b61f19SKirill A. Shutemov /* 1863e9b61f19SKirill A. Shutemov * Subpages may be freed if there wasn't any mapping 1864e9b61f19SKirill A. Shutemov * like if add_to_swap() is running on a lru page that 1865e9b61f19SKirill A. Shutemov * had its mapping zapped. And freeing these pages 1866e9b61f19SKirill A. Shutemov * requires taking the lru_lock so we do the put_page 1867e9b61f19SKirill A. Shutemov * of the tail pages after the split is complete. 1868e9b61f19SKirill A. Shutemov */ 1869e9b61f19SKirill A. Shutemov put_page(subpage); 1870e9b61f19SKirill A. Shutemov } 1871e9b61f19SKirill A. Shutemov } 1872e9b61f19SKirill A. Shutemov 1873b20ce5e0SKirill A. Shutemov int total_mapcount(struct page *page) 1874b20ce5e0SKirill A. Shutemov { 1875dd78feddSKirill A. Shutemov int i, compound, ret; 1876b20ce5e0SKirill A. Shutemov 1877b20ce5e0SKirill A. Shutemov VM_BUG_ON_PAGE(PageTail(page), page); 1878b20ce5e0SKirill A. Shutemov 1879b20ce5e0SKirill A. Shutemov if (likely(!PageCompound(page))) 1880b20ce5e0SKirill A. Shutemov return atomic_read(&page->_mapcount) + 1; 1881b20ce5e0SKirill A. Shutemov 1882dd78feddSKirill A. Shutemov compound = compound_mapcount(page); 1883b20ce5e0SKirill A. Shutemov if (PageHuge(page)) 1884dd78feddSKirill A. Shutemov return compound; 1885dd78feddSKirill A. Shutemov ret = compound; 1886b20ce5e0SKirill A. Shutemov for (i = 0; i < HPAGE_PMD_NR; i++) 1887b20ce5e0SKirill A. Shutemov ret += atomic_read(&page[i]._mapcount) + 1; 1888dd78feddSKirill A. Shutemov /* File pages has compound_mapcount included in _mapcount */ 1889dd78feddSKirill A. Shutemov if (!PageAnon(page)) 1890dd78feddSKirill A. Shutemov return ret - compound * HPAGE_PMD_NR; 1891b20ce5e0SKirill A. Shutemov if (PageDoubleMap(page)) 1892b20ce5e0SKirill A. Shutemov ret -= HPAGE_PMD_NR; 1893b20ce5e0SKirill A. Shutemov return ret; 1894b20ce5e0SKirill A. Shutemov } 1895b20ce5e0SKirill A. Shutemov 1896e9b61f19SKirill A. Shutemov /* 18976d0a07edSAndrea Arcangeli * This calculates accurately how many mappings a transparent hugepage 18986d0a07edSAndrea Arcangeli * has (unlike page_mapcount() which isn't fully accurate). This full 18996d0a07edSAndrea Arcangeli * accuracy is primarily needed to know if copy-on-write faults can 19006d0a07edSAndrea Arcangeli * reuse the page and change the mapping to read-write instead of 19016d0a07edSAndrea Arcangeli * copying them. At the same time this returns the total_mapcount too. 19026d0a07edSAndrea Arcangeli * 19036d0a07edSAndrea Arcangeli * The function returns the highest mapcount any one of the subpages 19046d0a07edSAndrea Arcangeli * has. If the return value is one, even if different processes are 19056d0a07edSAndrea Arcangeli * mapping different subpages of the transparent hugepage, they can 19066d0a07edSAndrea Arcangeli * all reuse it, because each process is reusing a different subpage. 19076d0a07edSAndrea Arcangeli * 19086d0a07edSAndrea Arcangeli * The total_mapcount is instead counting all virtual mappings of the 19096d0a07edSAndrea Arcangeli * subpages. If the total_mapcount is equal to "one", it tells the 19106d0a07edSAndrea Arcangeli * caller all mappings belong to the same "mm" and in turn the 19116d0a07edSAndrea Arcangeli * anon_vma of the transparent hugepage can become the vma->anon_vma 19126d0a07edSAndrea Arcangeli * local one as no other process may be mapping any of the subpages. 19136d0a07edSAndrea Arcangeli * 19146d0a07edSAndrea Arcangeli * It would be more accurate to replace page_mapcount() with 19156d0a07edSAndrea Arcangeli * page_trans_huge_mapcount(), however we only use 19166d0a07edSAndrea Arcangeli * page_trans_huge_mapcount() in the copy-on-write faults where we 19176d0a07edSAndrea Arcangeli * need full accuracy to avoid breaking page pinning, because 19186d0a07edSAndrea Arcangeli * page_trans_huge_mapcount() is slower than page_mapcount(). 19196d0a07edSAndrea Arcangeli */ 19206d0a07edSAndrea Arcangeli int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 19216d0a07edSAndrea Arcangeli { 19226d0a07edSAndrea Arcangeli int i, ret, _total_mapcount, mapcount; 19236d0a07edSAndrea Arcangeli 19246d0a07edSAndrea Arcangeli /* hugetlbfs shouldn't call it */ 19256d0a07edSAndrea Arcangeli VM_BUG_ON_PAGE(PageHuge(page), page); 19266d0a07edSAndrea Arcangeli 19276d0a07edSAndrea Arcangeli if (likely(!PageTransCompound(page))) { 19286d0a07edSAndrea Arcangeli mapcount = atomic_read(&page->_mapcount) + 1; 19296d0a07edSAndrea Arcangeli if (total_mapcount) 19306d0a07edSAndrea Arcangeli *total_mapcount = mapcount; 19316d0a07edSAndrea Arcangeli return mapcount; 19326d0a07edSAndrea Arcangeli } 19336d0a07edSAndrea Arcangeli 19346d0a07edSAndrea Arcangeli page = compound_head(page); 19356d0a07edSAndrea Arcangeli 19366d0a07edSAndrea Arcangeli _total_mapcount = ret = 0; 19376d0a07edSAndrea Arcangeli for (i = 0; i < HPAGE_PMD_NR; i++) { 19386d0a07edSAndrea Arcangeli mapcount = atomic_read(&page[i]._mapcount) + 1; 19396d0a07edSAndrea Arcangeli ret = max(ret, mapcount); 19406d0a07edSAndrea Arcangeli _total_mapcount += mapcount; 19416d0a07edSAndrea Arcangeli } 19426d0a07edSAndrea Arcangeli if (PageDoubleMap(page)) { 19436d0a07edSAndrea Arcangeli ret -= 1; 19446d0a07edSAndrea Arcangeli _total_mapcount -= HPAGE_PMD_NR; 19456d0a07edSAndrea Arcangeli } 19466d0a07edSAndrea Arcangeli mapcount = compound_mapcount(page); 19476d0a07edSAndrea Arcangeli ret += mapcount; 19486d0a07edSAndrea Arcangeli _total_mapcount += mapcount; 19496d0a07edSAndrea Arcangeli if (total_mapcount) 19506d0a07edSAndrea Arcangeli *total_mapcount = _total_mapcount; 19516d0a07edSAndrea Arcangeli return ret; 19526d0a07edSAndrea Arcangeli } 19536d0a07edSAndrea Arcangeli 19546d0a07edSAndrea Arcangeli /* 1955e9b61f19SKirill A. Shutemov * This function splits huge page into normal pages. @page can point to any 1956e9b61f19SKirill A. Shutemov * subpage of huge page to split. Split doesn't change the position of @page. 1957e9b61f19SKirill A. Shutemov * 1958e9b61f19SKirill A. Shutemov * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 1959e9b61f19SKirill A. Shutemov * The huge page must be locked. 1960e9b61f19SKirill A. Shutemov * 1961e9b61f19SKirill A. Shutemov * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 1962e9b61f19SKirill A. Shutemov * 1963e9b61f19SKirill A. Shutemov * Both head page and tail pages will inherit mapping, flags, and so on from 1964e9b61f19SKirill A. Shutemov * the hugepage. 1965e9b61f19SKirill A. Shutemov * 1966e9b61f19SKirill A. Shutemov * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 1967e9b61f19SKirill A. Shutemov * they are not mapped. 1968e9b61f19SKirill A. Shutemov * 1969e9b61f19SKirill A. Shutemov * Returns 0 if the hugepage is split successfully. 1970e9b61f19SKirill A. Shutemov * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 1971e9b61f19SKirill A. Shutemov * us. 1972e9b61f19SKirill A. Shutemov */ 1973e9b61f19SKirill A. Shutemov int split_huge_page_to_list(struct page *page, struct list_head *list) 1974e9b61f19SKirill A. Shutemov { 1975e9b61f19SKirill A. Shutemov struct page *head = compound_head(page); 1976a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 1977baa355fdSKirill A. Shutemov struct anon_vma *anon_vma = NULL; 1978baa355fdSKirill A. Shutemov struct address_space *mapping = NULL; 1979baa355fdSKirill A. Shutemov int count, mapcount, extra_pins, ret; 1980d9654322SKirill A. Shutemov bool mlocked; 19810b9b6fffSKirill A. Shutemov unsigned long flags; 1982e9b61f19SKirill A. Shutemov 1983e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(is_huge_zero_page(page), page); 1984e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 1985e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageSwapBacked(page), page); 1986e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageCompound(page), page); 1987e9b61f19SKirill A. Shutemov 1988baa355fdSKirill A. Shutemov if (PageAnon(head)) { 1989e9b61f19SKirill A. Shutemov /* 1990baa355fdSKirill A. Shutemov * The caller does not necessarily hold an mmap_sem that would 1991baa355fdSKirill A. Shutemov * prevent the anon_vma disappearing so we first we take a 1992baa355fdSKirill A. Shutemov * reference to it and then lock the anon_vma for write. This 1993baa355fdSKirill A. Shutemov * is similar to page_lock_anon_vma_read except the write lock 1994baa355fdSKirill A. Shutemov * is taken to serialise against parallel split or collapse 1995baa355fdSKirill A. Shutemov * operations. 1996e9b61f19SKirill A. Shutemov */ 1997e9b61f19SKirill A. Shutemov anon_vma = page_get_anon_vma(head); 1998e9b61f19SKirill A. Shutemov if (!anon_vma) { 1999e9b61f19SKirill A. Shutemov ret = -EBUSY; 2000e9b61f19SKirill A. Shutemov goto out; 2001e9b61f19SKirill A. Shutemov } 2002baa355fdSKirill A. Shutemov extra_pins = 0; 2003baa355fdSKirill A. Shutemov mapping = NULL; 2004e9b61f19SKirill A. Shutemov anon_vma_lock_write(anon_vma); 2005baa355fdSKirill A. Shutemov } else { 2006baa355fdSKirill A. Shutemov mapping = head->mapping; 2007baa355fdSKirill A. Shutemov 2008baa355fdSKirill A. Shutemov /* Truncated ? */ 2009baa355fdSKirill A. Shutemov if (!mapping) { 2010baa355fdSKirill A. Shutemov ret = -EBUSY; 2011baa355fdSKirill A. Shutemov goto out; 2012baa355fdSKirill A. Shutemov } 2013baa355fdSKirill A. Shutemov 2014baa355fdSKirill A. Shutemov /* Addidional pins from radix tree */ 2015baa355fdSKirill A. Shutemov extra_pins = HPAGE_PMD_NR; 2016baa355fdSKirill A. Shutemov anon_vma = NULL; 2017baa355fdSKirill A. Shutemov i_mmap_lock_read(mapping); 2018baa355fdSKirill A. Shutemov } 2019e9b61f19SKirill A. Shutemov 2020e9b61f19SKirill A. Shutemov /* 2021e9b61f19SKirill A. Shutemov * Racy check if we can split the page, before freeze_page() will 2022e9b61f19SKirill A. Shutemov * split PMDs 2023e9b61f19SKirill A. Shutemov */ 2024baa355fdSKirill A. Shutemov if (total_mapcount(head) != page_count(head) - extra_pins - 1) { 2025e9b61f19SKirill A. Shutemov ret = -EBUSY; 2026e9b61f19SKirill A. Shutemov goto out_unlock; 2027e9b61f19SKirill A. Shutemov } 2028e9b61f19SKirill A. Shutemov 2029d9654322SKirill A. Shutemov mlocked = PageMlocked(page); 2030fec89c10SKirill A. Shutemov freeze_page(head); 2031e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(compound_mapcount(head), head); 2032e9b61f19SKirill A. Shutemov 2033d9654322SKirill A. Shutemov /* Make sure the page is not on per-CPU pagevec as it takes pin */ 2034d9654322SKirill A. Shutemov if (mlocked) 2035d9654322SKirill A. Shutemov lru_add_drain(); 2036d9654322SKirill A. Shutemov 2037baa355fdSKirill A. Shutemov /* prevent PageLRU to go away from under us, and freeze lru stats */ 2038a52633d8SMel Gorman spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); 2039baa355fdSKirill A. Shutemov 2040baa355fdSKirill A. Shutemov if (mapping) { 2041baa355fdSKirill A. Shutemov void **pslot; 2042baa355fdSKirill A. Shutemov 2043baa355fdSKirill A. Shutemov spin_lock(&mapping->tree_lock); 2044baa355fdSKirill A. Shutemov pslot = radix_tree_lookup_slot(&mapping->page_tree, 2045baa355fdSKirill A. Shutemov page_index(head)); 2046baa355fdSKirill A. Shutemov /* 2047baa355fdSKirill A. Shutemov * Check if the head page is present in radix tree. 2048baa355fdSKirill A. Shutemov * We assume all tail are present too, if head is there. 2049baa355fdSKirill A. Shutemov */ 2050baa355fdSKirill A. Shutemov if (radix_tree_deref_slot_protected(pslot, 2051baa355fdSKirill A. Shutemov &mapping->tree_lock) != head) 2052baa355fdSKirill A. Shutemov goto fail; 2053baa355fdSKirill A. Shutemov } 2054baa355fdSKirill A. Shutemov 20550139aa7bSJoonsoo Kim /* Prevent deferred_split_scan() touching ->_refcount */ 2056baa355fdSKirill A. Shutemov spin_lock(&pgdata->split_queue_lock); 2057e9b61f19SKirill A. Shutemov count = page_count(head); 2058e9b61f19SKirill A. Shutemov mapcount = total_mapcount(head); 2059baa355fdSKirill A. Shutemov if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 20609a982250SKirill A. Shutemov if (!list_empty(page_deferred_list(head))) { 2061a3d0a918SKirill A. Shutemov pgdata->split_queue_len--; 20629a982250SKirill A. Shutemov list_del(page_deferred_list(head)); 20639a982250SKirill A. Shutemov } 206465c45377SKirill A. Shutemov if (mapping) 206511fb9989SMel Gorman __dec_node_page_state(page, NR_SHMEM_THPS); 2066baa355fdSKirill A. Shutemov spin_unlock(&pgdata->split_queue_lock); 2067baa355fdSKirill A. Shutemov __split_huge_page(page, list, flags); 2068e9b61f19SKirill A. Shutemov ret = 0; 2069baa355fdSKirill A. Shutemov } else { 2070baa355fdSKirill A. Shutemov if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2071e9b61f19SKirill A. Shutemov pr_alert("total_mapcount: %u, page_count(): %u\n", 2072e9b61f19SKirill A. Shutemov mapcount, count); 2073e9b61f19SKirill A. Shutemov if (PageTail(page)) 2074e9b61f19SKirill A. Shutemov dump_page(head, NULL); 2075bd56086fSKirill A. Shutemov dump_page(page, "total_mapcount(head) > 0"); 2076e9b61f19SKirill A. Shutemov BUG(); 2077baa355fdSKirill A. Shutemov } 2078baa355fdSKirill A. Shutemov spin_unlock(&pgdata->split_queue_lock); 2079baa355fdSKirill A. Shutemov fail: if (mapping) 2080baa355fdSKirill A. Shutemov spin_unlock(&mapping->tree_lock); 2081a52633d8SMel Gorman spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2082fec89c10SKirill A. Shutemov unfreeze_page(head); 2083e9b61f19SKirill A. Shutemov ret = -EBUSY; 2084e9b61f19SKirill A. Shutemov } 2085e9b61f19SKirill A. Shutemov 2086e9b61f19SKirill A. Shutemov out_unlock: 2087baa355fdSKirill A. Shutemov if (anon_vma) { 2088e9b61f19SKirill A. Shutemov anon_vma_unlock_write(anon_vma); 2089e9b61f19SKirill A. Shutemov put_anon_vma(anon_vma); 2090baa355fdSKirill A. Shutemov } 2091baa355fdSKirill A. Shutemov if (mapping) 2092baa355fdSKirill A. Shutemov i_mmap_unlock_read(mapping); 2093e9b61f19SKirill A. Shutemov out: 2094e9b61f19SKirill A. Shutemov count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2095e9b61f19SKirill A. Shutemov return ret; 2096e9b61f19SKirill A. Shutemov } 20979a982250SKirill A. Shutemov 20989a982250SKirill A. Shutemov void free_transhuge_page(struct page *page) 20999a982250SKirill A. Shutemov { 2100a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 21019a982250SKirill A. Shutemov unsigned long flags; 21029a982250SKirill A. Shutemov 2103a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 21049a982250SKirill A. Shutemov if (!list_empty(page_deferred_list(page))) { 2105a3d0a918SKirill A. Shutemov pgdata->split_queue_len--; 21069a982250SKirill A. Shutemov list_del(page_deferred_list(page)); 21079a982250SKirill A. Shutemov } 2108a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 21099a982250SKirill A. Shutemov free_compound_page(page); 21109a982250SKirill A. Shutemov } 21119a982250SKirill A. Shutemov 21129a982250SKirill A. Shutemov void deferred_split_huge_page(struct page *page) 21139a982250SKirill A. Shutemov { 2114a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 21159a982250SKirill A. Shutemov unsigned long flags; 21169a982250SKirill A. Shutemov 21179a982250SKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 21189a982250SKirill A. Shutemov 2119a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 21209a982250SKirill A. Shutemov if (list_empty(page_deferred_list(page))) { 2121f9719a03SKirill A. Shutemov count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2122a3d0a918SKirill A. Shutemov list_add_tail(page_deferred_list(page), &pgdata->split_queue); 2123a3d0a918SKirill A. Shutemov pgdata->split_queue_len++; 21249a982250SKirill A. Shutemov } 2125a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 21269a982250SKirill A. Shutemov } 21279a982250SKirill A. Shutemov 21289a982250SKirill A. Shutemov static unsigned long deferred_split_count(struct shrinker *shrink, 21299a982250SKirill A. Shutemov struct shrink_control *sc) 21309a982250SKirill A. Shutemov { 2131a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 2132cb8d68ecSKirill A. Shutemov return ACCESS_ONCE(pgdata->split_queue_len); 21339a982250SKirill A. Shutemov } 21349a982250SKirill A. Shutemov 21359a982250SKirill A. Shutemov static unsigned long deferred_split_scan(struct shrinker *shrink, 21369a982250SKirill A. Shutemov struct shrink_control *sc) 21379a982250SKirill A. Shutemov { 2138a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 21399a982250SKirill A. Shutemov unsigned long flags; 21409a982250SKirill A. Shutemov LIST_HEAD(list), *pos, *next; 21419a982250SKirill A. Shutemov struct page *page; 21429a982250SKirill A. Shutemov int split = 0; 21439a982250SKirill A. Shutemov 2144a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 21459a982250SKirill A. Shutemov /* Take pin on all head pages to avoid freeing them under us */ 2146ae026204SKirill A. Shutemov list_for_each_safe(pos, next, &pgdata->split_queue) { 21479a982250SKirill A. Shutemov page = list_entry((void *)pos, struct page, mapping); 21489a982250SKirill A. Shutemov page = compound_head(page); 2149e3ae1953SKirill A. Shutemov if (get_page_unless_zero(page)) { 2150e3ae1953SKirill A. Shutemov list_move(page_deferred_list(page), &list); 2151e3ae1953SKirill A. Shutemov } else { 2152e3ae1953SKirill A. Shutemov /* We lost race with put_compound_page() */ 21539a982250SKirill A. Shutemov list_del_init(page_deferred_list(page)); 2154a3d0a918SKirill A. Shutemov pgdata->split_queue_len--; 21559a982250SKirill A. Shutemov } 2156e3ae1953SKirill A. Shutemov if (!--sc->nr_to_scan) 2157e3ae1953SKirill A. Shutemov break; 21589a982250SKirill A. Shutemov } 2159a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 21609a982250SKirill A. Shutemov 21619a982250SKirill A. Shutemov list_for_each_safe(pos, next, &list) { 21629a982250SKirill A. Shutemov page = list_entry((void *)pos, struct page, mapping); 21639a982250SKirill A. Shutemov lock_page(page); 21649a982250SKirill A. Shutemov /* split_huge_page() removes page from list on success */ 21659a982250SKirill A. Shutemov if (!split_huge_page(page)) 21669a982250SKirill A. Shutemov split++; 21679a982250SKirill A. Shutemov unlock_page(page); 21689a982250SKirill A. Shutemov put_page(page); 21699a982250SKirill A. Shutemov } 21709a982250SKirill A. Shutemov 2171a3d0a918SKirill A. Shutemov spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2172a3d0a918SKirill A. Shutemov list_splice_tail(&list, &pgdata->split_queue); 2173a3d0a918SKirill A. Shutemov spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 21749a982250SKirill A. Shutemov 2175cb8d68ecSKirill A. Shutemov /* 2176cb8d68ecSKirill A. Shutemov * Stop shrinker if we didn't split any page, but the queue is empty. 2177cb8d68ecSKirill A. Shutemov * This can happen if pages were freed under us. 2178cb8d68ecSKirill A. Shutemov */ 2179cb8d68ecSKirill A. Shutemov if (!split && list_empty(&pgdata->split_queue)) 2180cb8d68ecSKirill A. Shutemov return SHRINK_STOP; 2181cb8d68ecSKirill A. Shutemov return split; 21829a982250SKirill A. Shutemov } 21839a982250SKirill A. Shutemov 21849a982250SKirill A. Shutemov static struct shrinker deferred_split_shrinker = { 21859a982250SKirill A. Shutemov .count_objects = deferred_split_count, 21869a982250SKirill A. Shutemov .scan_objects = deferred_split_scan, 21879a982250SKirill A. Shutemov .seeks = DEFAULT_SEEKS, 2188a3d0a918SKirill A. Shutemov .flags = SHRINKER_NUMA_AWARE, 21899a982250SKirill A. Shutemov }; 219049071d43SKirill A. Shutemov 219149071d43SKirill A. Shutemov #ifdef CONFIG_DEBUG_FS 219249071d43SKirill A. Shutemov static int split_huge_pages_set(void *data, u64 val) 219349071d43SKirill A. Shutemov { 219449071d43SKirill A. Shutemov struct zone *zone; 219549071d43SKirill A. Shutemov struct page *page; 219649071d43SKirill A. Shutemov unsigned long pfn, max_zone_pfn; 219749071d43SKirill A. Shutemov unsigned long total = 0, split = 0; 219849071d43SKirill A. Shutemov 219949071d43SKirill A. Shutemov if (val != 1) 220049071d43SKirill A. Shutemov return -EINVAL; 220149071d43SKirill A. Shutemov 220249071d43SKirill A. Shutemov for_each_populated_zone(zone) { 220349071d43SKirill A. Shutemov max_zone_pfn = zone_end_pfn(zone); 220449071d43SKirill A. Shutemov for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 220549071d43SKirill A. Shutemov if (!pfn_valid(pfn)) 220649071d43SKirill A. Shutemov continue; 220749071d43SKirill A. Shutemov 220849071d43SKirill A. Shutemov page = pfn_to_page(pfn); 220949071d43SKirill A. Shutemov if (!get_page_unless_zero(page)) 221049071d43SKirill A. Shutemov continue; 221149071d43SKirill A. Shutemov 221249071d43SKirill A. Shutemov if (zone != page_zone(page)) 221349071d43SKirill A. Shutemov goto next; 221449071d43SKirill A. Shutemov 2215baa355fdSKirill A. Shutemov if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 221649071d43SKirill A. Shutemov goto next; 221749071d43SKirill A. Shutemov 221849071d43SKirill A. Shutemov total++; 221949071d43SKirill A. Shutemov lock_page(page); 222049071d43SKirill A. Shutemov if (!split_huge_page(page)) 222149071d43SKirill A. Shutemov split++; 222249071d43SKirill A. Shutemov unlock_page(page); 222349071d43SKirill A. Shutemov next: 222449071d43SKirill A. Shutemov put_page(page); 222549071d43SKirill A. Shutemov } 222649071d43SKirill A. Shutemov } 222749071d43SKirill A. Shutemov 2228145bdaa1SYang Shi pr_info("%lu of %lu THP split\n", split, total); 222949071d43SKirill A. Shutemov 223049071d43SKirill A. Shutemov return 0; 223149071d43SKirill A. Shutemov } 223249071d43SKirill A. Shutemov DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 223349071d43SKirill A. Shutemov "%llu\n"); 223449071d43SKirill A. Shutemov 223549071d43SKirill A. Shutemov static int __init split_huge_pages_debugfs(void) 223649071d43SKirill A. Shutemov { 223749071d43SKirill A. Shutemov void *ret; 223849071d43SKirill A. Shutemov 2239145bdaa1SYang Shi ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 224049071d43SKirill A. Shutemov &split_huge_pages_fops); 224149071d43SKirill A. Shutemov if (!ret) 224249071d43SKirill A. Shutemov pr_warn("Failed to create split_huge_pages in debugfs"); 224349071d43SKirill A. Shutemov return 0; 224449071d43SKirill A. Shutemov } 224549071d43SKirill A. Shutemov late_initcall(split_huge_pages_debugfs); 224649071d43SKirill A. Shutemov #endif 2247