1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2b46e756fSKirill A. Shutemov #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 3b46e756fSKirill A. Shutemov 4b46e756fSKirill A. Shutemov #include <linux/mm.h> 5b46e756fSKirill A. Shutemov #include <linux/sched.h> 66e84f315SIngo Molnar #include <linux/sched/mm.h> 7f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 8b46e756fSKirill A. Shutemov #include <linux/mmu_notifier.h> 9b46e756fSKirill A. Shutemov #include <linux/rmap.h> 10b46e756fSKirill A. Shutemov #include <linux/swap.h> 11b46e756fSKirill A. Shutemov #include <linux/mm_inline.h> 12b46e756fSKirill A. Shutemov #include <linux/kthread.h> 13b46e756fSKirill A. Shutemov #include <linux/khugepaged.h> 14b46e756fSKirill A. Shutemov #include <linux/freezer.h> 15b46e756fSKirill A. Shutemov #include <linux/mman.h> 16b46e756fSKirill A. Shutemov #include <linux/hashtable.h> 17b46e756fSKirill A. Shutemov #include <linux/userfaultfd_k.h> 18b46e756fSKirill A. Shutemov #include <linux/page_idle.h> 1980110bbfSPasha Tatashin #include <linux/page_table_check.h> 20b46e756fSKirill A. Shutemov #include <linux/swapops.h> 21f3f0e1d2SKirill A. Shutemov #include <linux/shmem_fs.h> 22b46e756fSKirill A. Shutemov 23b46e756fSKirill A. Shutemov #include <asm/tlb.h> 24b46e756fSKirill A. Shutemov #include <asm/pgalloc.h> 25b46e756fSKirill A. Shutemov #include "internal.h" 26b46e756fSKirill A. Shutemov 27b46e756fSKirill A. Shutemov enum scan_result { 28b46e756fSKirill A. Shutemov SCAN_FAIL, 29b46e756fSKirill A. Shutemov SCAN_SUCCEED, 30b46e756fSKirill A. Shutemov SCAN_PMD_NULL, 3150722804SZach O'Keefe SCAN_PMD_MAPPED, 32b46e756fSKirill A. Shutemov SCAN_EXCEED_NONE_PTE, 3371a2c112SKirill A. Shutemov SCAN_EXCEED_SWAP_PTE, 3471a2c112SKirill A. Shutemov SCAN_EXCEED_SHARED_PTE, 35b46e756fSKirill A. Shutemov SCAN_PTE_NON_PRESENT, 36e1e267c7SPeter Xu SCAN_PTE_UFFD_WP, 37b46e756fSKirill A. Shutemov SCAN_PAGE_RO, 380db501f7SEbru Akagunduz SCAN_LACK_REFERENCED_PAGE, 39b46e756fSKirill A. Shutemov SCAN_PAGE_NULL, 40b46e756fSKirill A. Shutemov SCAN_SCAN_ABORT, 41b46e756fSKirill A. Shutemov SCAN_PAGE_COUNT, 42b46e756fSKirill A. Shutemov SCAN_PAGE_LRU, 43b46e756fSKirill A. Shutemov SCAN_PAGE_LOCK, 44b46e756fSKirill A. Shutemov SCAN_PAGE_ANON, 45b46e756fSKirill A. Shutemov SCAN_PAGE_COMPOUND, 46b46e756fSKirill A. Shutemov SCAN_ANY_PROCESS, 47b46e756fSKirill A. Shutemov SCAN_VMA_NULL, 48b46e756fSKirill A. Shutemov SCAN_VMA_CHECK, 49b46e756fSKirill A. Shutemov SCAN_ADDRESS_RANGE, 50b46e756fSKirill A. Shutemov SCAN_DEL_PAGE_LRU, 51b46e756fSKirill A. Shutemov SCAN_ALLOC_HUGE_PAGE_FAIL, 52b46e756fSKirill A. Shutemov SCAN_CGROUP_CHARGE_FAIL, 53f3f0e1d2SKirill A. Shutemov SCAN_TRUNCATED, 5499cb0dbdSSong Liu SCAN_PAGE_HAS_PRIVATE, 55b46e756fSKirill A. Shutemov }; 56b46e756fSKirill A. Shutemov 57b46e756fSKirill A. Shutemov #define CREATE_TRACE_POINTS 58b46e756fSKirill A. Shutemov #include <trace/events/huge_memory.h> 59b46e756fSKirill A. Shutemov 604aab2be0SVijay Balakrishna static struct task_struct *khugepaged_thread __read_mostly; 614aab2be0SVijay Balakrishna static DEFINE_MUTEX(khugepaged_mutex); 624aab2be0SVijay Balakrishna 63b46e756fSKirill A. Shutemov /* default scan 8*512 pte (or vmas) every 30 second */ 64b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_to_scan __read_mostly; 65b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_collapsed; 66b46e756fSKirill A. Shutemov static unsigned int khugepaged_full_scans; 67b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 68b46e756fSKirill A. Shutemov /* during fragmentation poll the hugepage allocator once every minute */ 69b46e756fSKirill A. Shutemov static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 70b46e756fSKirill A. Shutemov static unsigned long khugepaged_sleep_expire; 71b46e756fSKirill A. Shutemov static DEFINE_SPINLOCK(khugepaged_mm_lock); 72b46e756fSKirill A. Shutemov static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 73b46e756fSKirill A. Shutemov /* 74b46e756fSKirill A. Shutemov * default collapse hugepages if there is at least one pte mapped like 75b46e756fSKirill A. Shutemov * it would have happened if the vma was large enough during page 76b46e756fSKirill A. Shutemov * fault. 77d8ea7cc8SZach O'Keefe * 78d8ea7cc8SZach O'Keefe * Note that these are only respected if collapse was initiated by khugepaged. 79b46e756fSKirill A. Shutemov */ 80b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_none __read_mostly; 81b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_swap __read_mostly; 8271a2c112SKirill A. Shutemov static unsigned int khugepaged_max_ptes_shared __read_mostly; 83b46e756fSKirill A. Shutemov 84b46e756fSKirill A. Shutemov #define MM_SLOTS_HASH_BITS 10 85b46e756fSKirill A. Shutemov static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 86b46e756fSKirill A. Shutemov 87b46e756fSKirill A. Shutemov static struct kmem_cache *mm_slot_cache __read_mostly; 88b46e756fSKirill A. Shutemov 8927e1f827SSong Liu #define MAX_PTE_MAPPED_THP 8 9027e1f827SSong Liu 9134d6b470SZach O'Keefe struct collapse_control { 92d8ea7cc8SZach O'Keefe bool is_khugepaged; 93d8ea7cc8SZach O'Keefe 9434d6b470SZach O'Keefe /* Num pages scanned per node */ 9534d6b470SZach O'Keefe u32 node_load[MAX_NUMNODES]; 9634d6b470SZach O'Keefe 97*7d2c4385SZach O'Keefe /* Last target selected in hpage_collapse_find_target_node() */ 9834d6b470SZach O'Keefe int last_target_node; 9934d6b470SZach O'Keefe }; 10034d6b470SZach O'Keefe 101b46e756fSKirill A. Shutemov /** 102b46e756fSKirill A. Shutemov * struct mm_slot - hash lookup from mm to mm_slot 103b46e756fSKirill A. Shutemov * @hash: hash collision list 104b46e756fSKirill A. Shutemov * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 105b46e756fSKirill A. Shutemov * @mm: the mm that this information is valid for 106336e6b53SAlex Shi * @nr_pte_mapped_thp: number of pte mapped THP 107336e6b53SAlex Shi * @pte_mapped_thp: address array corresponding pte mapped THP 108b46e756fSKirill A. Shutemov */ 109b46e756fSKirill A. Shutemov struct mm_slot { 110b46e756fSKirill A. Shutemov struct hlist_node hash; 111b46e756fSKirill A. Shutemov struct list_head mm_node; 112b46e756fSKirill A. Shutemov struct mm_struct *mm; 11327e1f827SSong Liu 11427e1f827SSong Liu /* pte-mapped THP in this mm */ 11527e1f827SSong Liu int nr_pte_mapped_thp; 11627e1f827SSong Liu unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; 117b46e756fSKirill A. Shutemov }; 118b46e756fSKirill A. Shutemov 119b46e756fSKirill A. Shutemov /** 120b46e756fSKirill A. Shutemov * struct khugepaged_scan - cursor for scanning 121b46e756fSKirill A. Shutemov * @mm_head: the head of the mm list to scan 122b46e756fSKirill A. Shutemov * @mm_slot: the current mm_slot we are scanning 123b46e756fSKirill A. Shutemov * @address: the next address inside that to be scanned 124b46e756fSKirill A. Shutemov * 125b46e756fSKirill A. Shutemov * There is only the one khugepaged_scan instance of this cursor structure. 126b46e756fSKirill A. Shutemov */ 127b46e756fSKirill A. Shutemov struct khugepaged_scan { 128b46e756fSKirill A. Shutemov struct list_head mm_head; 129b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 130b46e756fSKirill A. Shutemov unsigned long address; 131b46e756fSKirill A. Shutemov }; 132b46e756fSKirill A. Shutemov 133b46e756fSKirill A. Shutemov static struct khugepaged_scan khugepaged_scan = { 134b46e756fSKirill A. Shutemov .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 135b46e756fSKirill A. Shutemov }; 136b46e756fSKirill A. Shutemov 137e1465d12SJérémy Lefaure #ifdef CONFIG_SYSFS 138b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 139b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 140b46e756fSKirill A. Shutemov char *buf) 141b46e756fSKirill A. Shutemov { 142ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); 143b46e756fSKirill A. Shutemov } 144b46e756fSKirill A. Shutemov 145b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 146b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 147b46e756fSKirill A. Shutemov const char *buf, size_t count) 148b46e756fSKirill A. Shutemov { 149dfefd226SAlexey Dobriyan unsigned int msecs; 150b46e756fSKirill A. Shutemov int err; 151b46e756fSKirill A. Shutemov 152dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 153dfefd226SAlexey Dobriyan if (err) 154b46e756fSKirill A. Shutemov return -EINVAL; 155b46e756fSKirill A. Shutemov 156b46e756fSKirill A. Shutemov khugepaged_scan_sleep_millisecs = msecs; 157b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 158b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 159b46e756fSKirill A. Shutemov 160b46e756fSKirill A. Shutemov return count; 161b46e756fSKirill A. Shutemov } 162b46e756fSKirill A. Shutemov static struct kobj_attribute scan_sleep_millisecs_attr = 1636dcdc94dSMiaohe Lin __ATTR_RW(scan_sleep_millisecs); 164b46e756fSKirill A. Shutemov 165b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 166b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 167b46e756fSKirill A. Shutemov char *buf) 168b46e756fSKirill A. Shutemov { 169ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 170b46e756fSKirill A. Shutemov } 171b46e756fSKirill A. Shutemov 172b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 173b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 174b46e756fSKirill A. Shutemov const char *buf, size_t count) 175b46e756fSKirill A. Shutemov { 176dfefd226SAlexey Dobriyan unsigned int msecs; 177b46e756fSKirill A. Shutemov int err; 178b46e756fSKirill A. Shutemov 179dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 180dfefd226SAlexey Dobriyan if (err) 181b46e756fSKirill A. Shutemov return -EINVAL; 182b46e756fSKirill A. Shutemov 183b46e756fSKirill A. Shutemov khugepaged_alloc_sleep_millisecs = msecs; 184b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 185b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 186b46e756fSKirill A. Shutemov 187b46e756fSKirill A. Shutemov return count; 188b46e756fSKirill A. Shutemov } 189b46e756fSKirill A. Shutemov static struct kobj_attribute alloc_sleep_millisecs_attr = 1906dcdc94dSMiaohe Lin __ATTR_RW(alloc_sleep_millisecs); 191b46e756fSKirill A. Shutemov 192b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_show(struct kobject *kobj, 193b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 194b46e756fSKirill A. Shutemov char *buf) 195b46e756fSKirill A. Shutemov { 196ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); 197b46e756fSKirill A. Shutemov } 198b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_store(struct kobject *kobj, 199b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 200b46e756fSKirill A. Shutemov const char *buf, size_t count) 201b46e756fSKirill A. Shutemov { 202dfefd226SAlexey Dobriyan unsigned int pages; 203b46e756fSKirill A. Shutemov int err; 204b46e756fSKirill A. Shutemov 205dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &pages); 206dfefd226SAlexey Dobriyan if (err || !pages) 207b46e756fSKirill A. Shutemov return -EINVAL; 208b46e756fSKirill A. Shutemov 209b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = pages; 210b46e756fSKirill A. Shutemov 211b46e756fSKirill A. Shutemov return count; 212b46e756fSKirill A. Shutemov } 213b46e756fSKirill A. Shutemov static struct kobj_attribute pages_to_scan_attr = 2146dcdc94dSMiaohe Lin __ATTR_RW(pages_to_scan); 215b46e756fSKirill A. Shutemov 216b46e756fSKirill A. Shutemov static ssize_t pages_collapsed_show(struct kobject *kobj, 217b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 218b46e756fSKirill A. Shutemov char *buf) 219b46e756fSKirill A. Shutemov { 220ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); 221b46e756fSKirill A. Shutemov } 222b46e756fSKirill A. Shutemov static struct kobj_attribute pages_collapsed_attr = 223b46e756fSKirill A. Shutemov __ATTR_RO(pages_collapsed); 224b46e756fSKirill A. Shutemov 225b46e756fSKirill A. Shutemov static ssize_t full_scans_show(struct kobject *kobj, 226b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 227b46e756fSKirill A. Shutemov char *buf) 228b46e756fSKirill A. Shutemov { 229ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_full_scans); 230b46e756fSKirill A. Shutemov } 231b46e756fSKirill A. Shutemov static struct kobj_attribute full_scans_attr = 232b46e756fSKirill A. Shutemov __ATTR_RO(full_scans); 233b46e756fSKirill A. Shutemov 2346dcdc94dSMiaohe Lin static ssize_t defrag_show(struct kobject *kobj, 235b46e756fSKirill A. Shutemov struct kobj_attribute *attr, char *buf) 236b46e756fSKirill A. Shutemov { 237b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 238b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 239b46e756fSKirill A. Shutemov } 2406dcdc94dSMiaohe Lin static ssize_t defrag_store(struct kobject *kobj, 241b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 242b46e756fSKirill A. Shutemov const char *buf, size_t count) 243b46e756fSKirill A. Shutemov { 244b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 245b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 246b46e756fSKirill A. Shutemov } 247b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_defrag_attr = 2486dcdc94dSMiaohe Lin __ATTR_RW(defrag); 249b46e756fSKirill A. Shutemov 250b46e756fSKirill A. Shutemov /* 251b46e756fSKirill A. Shutemov * max_ptes_none controls if khugepaged should collapse hugepages over 252b46e756fSKirill A. Shutemov * any unmapped ptes in turn potentially increasing the memory 253b46e756fSKirill A. Shutemov * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 254b46e756fSKirill A. Shutemov * reduce the available free memory in the system as it 255b46e756fSKirill A. Shutemov * runs. Increasing max_ptes_none will instead potentially reduce the 256b46e756fSKirill A. Shutemov * free memory in the system during the khugepaged scan. 257b46e756fSKirill A. Shutemov */ 2586dcdc94dSMiaohe Lin static ssize_t max_ptes_none_show(struct kobject *kobj, 259b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 260b46e756fSKirill A. Shutemov char *buf) 261b46e756fSKirill A. Shutemov { 262ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); 263b46e756fSKirill A. Shutemov } 2646dcdc94dSMiaohe Lin static ssize_t max_ptes_none_store(struct kobject *kobj, 265b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 266b46e756fSKirill A. Shutemov const char *buf, size_t count) 267b46e756fSKirill A. Shutemov { 268b46e756fSKirill A. Shutemov int err; 269b46e756fSKirill A. Shutemov unsigned long max_ptes_none; 270b46e756fSKirill A. Shutemov 271b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_none); 272b46e756fSKirill A. Shutemov if (err || max_ptes_none > HPAGE_PMD_NR - 1) 273b46e756fSKirill A. Shutemov return -EINVAL; 274b46e756fSKirill A. Shutemov 275b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = max_ptes_none; 276b46e756fSKirill A. Shutemov 277b46e756fSKirill A. Shutemov return count; 278b46e756fSKirill A. Shutemov } 279b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_none_attr = 2806dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_none); 281b46e756fSKirill A. Shutemov 2826dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_show(struct kobject *kobj, 283b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 284b46e756fSKirill A. Shutemov char *buf) 285b46e756fSKirill A. Shutemov { 286ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); 287b46e756fSKirill A. Shutemov } 288b46e756fSKirill A. Shutemov 2896dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_store(struct kobject *kobj, 290b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 291b46e756fSKirill A. Shutemov const char *buf, size_t count) 292b46e756fSKirill A. Shutemov { 293b46e756fSKirill A. Shutemov int err; 294b46e756fSKirill A. Shutemov unsigned long max_ptes_swap; 295b46e756fSKirill A. Shutemov 296b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_swap); 297b46e756fSKirill A. Shutemov if (err || max_ptes_swap > HPAGE_PMD_NR - 1) 298b46e756fSKirill A. Shutemov return -EINVAL; 299b46e756fSKirill A. Shutemov 300b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = max_ptes_swap; 301b46e756fSKirill A. Shutemov 302b46e756fSKirill A. Shutemov return count; 303b46e756fSKirill A. Shutemov } 304b46e756fSKirill A. Shutemov 305b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_swap_attr = 3066dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_swap); 307b46e756fSKirill A. Shutemov 3086dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_show(struct kobject *kobj, 30971a2c112SKirill A. Shutemov struct kobj_attribute *attr, 31071a2c112SKirill A. Shutemov char *buf) 31171a2c112SKirill A. Shutemov { 312ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); 31371a2c112SKirill A. Shutemov } 31471a2c112SKirill A. Shutemov 3156dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_store(struct kobject *kobj, 31671a2c112SKirill A. Shutemov struct kobj_attribute *attr, 31771a2c112SKirill A. Shutemov const char *buf, size_t count) 31871a2c112SKirill A. Shutemov { 31971a2c112SKirill A. Shutemov int err; 32071a2c112SKirill A. Shutemov unsigned long max_ptes_shared; 32171a2c112SKirill A. Shutemov 32271a2c112SKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_shared); 32371a2c112SKirill A. Shutemov if (err || max_ptes_shared > HPAGE_PMD_NR - 1) 32471a2c112SKirill A. Shutemov return -EINVAL; 32571a2c112SKirill A. Shutemov 32671a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = max_ptes_shared; 32771a2c112SKirill A. Shutemov 32871a2c112SKirill A. Shutemov return count; 32971a2c112SKirill A. Shutemov } 33071a2c112SKirill A. Shutemov 33171a2c112SKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_shared_attr = 3326dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_shared); 33371a2c112SKirill A. Shutemov 334b46e756fSKirill A. Shutemov static struct attribute *khugepaged_attr[] = { 335b46e756fSKirill A. Shutemov &khugepaged_defrag_attr.attr, 336b46e756fSKirill A. Shutemov &khugepaged_max_ptes_none_attr.attr, 33771a2c112SKirill A. Shutemov &khugepaged_max_ptes_swap_attr.attr, 33871a2c112SKirill A. Shutemov &khugepaged_max_ptes_shared_attr.attr, 339b46e756fSKirill A. Shutemov &pages_to_scan_attr.attr, 340b46e756fSKirill A. Shutemov &pages_collapsed_attr.attr, 341b46e756fSKirill A. Shutemov &full_scans_attr.attr, 342b46e756fSKirill A. Shutemov &scan_sleep_millisecs_attr.attr, 343b46e756fSKirill A. Shutemov &alloc_sleep_millisecs_attr.attr, 344b46e756fSKirill A. Shutemov NULL, 345b46e756fSKirill A. Shutemov }; 346b46e756fSKirill A. Shutemov 347b46e756fSKirill A. Shutemov struct attribute_group khugepaged_attr_group = { 348b46e756fSKirill A. Shutemov .attrs = khugepaged_attr, 349b46e756fSKirill A. Shutemov .name = "khugepaged", 350b46e756fSKirill A. Shutemov }; 351e1465d12SJérémy Lefaure #endif /* CONFIG_SYSFS */ 352b46e756fSKirill A. Shutemov 353b46e756fSKirill A. Shutemov int hugepage_madvise(struct vm_area_struct *vma, 354b46e756fSKirill A. Shutemov unsigned long *vm_flags, int advice) 355b46e756fSKirill A. Shutemov { 356b46e756fSKirill A. Shutemov switch (advice) { 357b46e756fSKirill A. Shutemov case MADV_HUGEPAGE: 358b46e756fSKirill A. Shutemov #ifdef CONFIG_S390 359b46e756fSKirill A. Shutemov /* 360b46e756fSKirill A. Shutemov * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 361b46e756fSKirill A. Shutemov * can't handle this properly after s390_enable_sie, so we simply 362b46e756fSKirill A. Shutemov * ignore the madvise to prevent qemu from causing a SIGSEGV. 363b46e756fSKirill A. Shutemov */ 364b46e756fSKirill A. Shutemov if (mm_has_pgste(vma->vm_mm)) 365b46e756fSKirill A. Shutemov return 0; 366b46e756fSKirill A. Shutemov #endif 367b46e756fSKirill A. Shutemov *vm_flags &= ~VM_NOHUGEPAGE; 368b46e756fSKirill A. Shutemov *vm_flags |= VM_HUGEPAGE; 369b46e756fSKirill A. Shutemov /* 370b46e756fSKirill A. Shutemov * If the vma become good for khugepaged to scan, 371b46e756fSKirill A. Shutemov * register it here without waiting a page fault that 372b46e756fSKirill A. Shutemov * may not happen any time soon. 373b46e756fSKirill A. Shutemov */ 374c791576cSYang Shi khugepaged_enter_vma(vma, *vm_flags); 375b46e756fSKirill A. Shutemov break; 376b46e756fSKirill A. Shutemov case MADV_NOHUGEPAGE: 377b46e756fSKirill A. Shutemov *vm_flags &= ~VM_HUGEPAGE; 378b46e756fSKirill A. Shutemov *vm_flags |= VM_NOHUGEPAGE; 379b46e756fSKirill A. Shutemov /* 380b46e756fSKirill A. Shutemov * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 381b46e756fSKirill A. Shutemov * this vma even if we leave the mm registered in khugepaged if 382b46e756fSKirill A. Shutemov * it got registered before VM_NOHUGEPAGE was set. 383b46e756fSKirill A. Shutemov */ 384b46e756fSKirill A. Shutemov break; 385b46e756fSKirill A. Shutemov } 386b46e756fSKirill A. Shutemov 387b46e756fSKirill A. Shutemov return 0; 388b46e756fSKirill A. Shutemov } 389b46e756fSKirill A. Shutemov 390b46e756fSKirill A. Shutemov int __init khugepaged_init(void) 391b46e756fSKirill A. Shutemov { 392b46e756fSKirill A. Shutemov mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 393b46e756fSKirill A. Shutemov sizeof(struct mm_slot), 394b46e756fSKirill A. Shutemov __alignof__(struct mm_slot), 0, NULL); 395b46e756fSKirill A. Shutemov if (!mm_slot_cache) 396b46e756fSKirill A. Shutemov return -ENOMEM; 397b46e756fSKirill A. Shutemov 398b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 399b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 400b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 40171a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; 402b46e756fSKirill A. Shutemov 403b46e756fSKirill A. Shutemov return 0; 404b46e756fSKirill A. Shutemov } 405b46e756fSKirill A. Shutemov 406b46e756fSKirill A. Shutemov void __init khugepaged_destroy(void) 407b46e756fSKirill A. Shutemov { 408b46e756fSKirill A. Shutemov kmem_cache_destroy(mm_slot_cache); 409b46e756fSKirill A. Shutemov } 410b46e756fSKirill A. Shutemov 411b46e756fSKirill A. Shutemov static inline struct mm_slot *alloc_mm_slot(void) 412b46e756fSKirill A. Shutemov { 413b46e756fSKirill A. Shutemov if (!mm_slot_cache) /* initialization failed */ 414b46e756fSKirill A. Shutemov return NULL; 415b46e756fSKirill A. Shutemov return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 416b46e756fSKirill A. Shutemov } 417b46e756fSKirill A. Shutemov 418b46e756fSKirill A. Shutemov static inline void free_mm_slot(struct mm_slot *mm_slot) 419b46e756fSKirill A. Shutemov { 420b46e756fSKirill A. Shutemov kmem_cache_free(mm_slot_cache, mm_slot); 421b46e756fSKirill A. Shutemov } 422b46e756fSKirill A. Shutemov 423b46e756fSKirill A. Shutemov static struct mm_slot *get_mm_slot(struct mm_struct *mm) 424b46e756fSKirill A. Shutemov { 425b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 426b46e756fSKirill A. Shutemov 427b46e756fSKirill A. Shutemov hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 428b46e756fSKirill A. Shutemov if (mm == mm_slot->mm) 429b46e756fSKirill A. Shutemov return mm_slot; 430b46e756fSKirill A. Shutemov 431b46e756fSKirill A. Shutemov return NULL; 432b46e756fSKirill A. Shutemov } 433b46e756fSKirill A. Shutemov 434b46e756fSKirill A. Shutemov static void insert_to_mm_slots_hash(struct mm_struct *mm, 435b46e756fSKirill A. Shutemov struct mm_slot *mm_slot) 436b46e756fSKirill A. Shutemov { 437b46e756fSKirill A. Shutemov mm_slot->mm = mm; 438b46e756fSKirill A. Shutemov hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 439b46e756fSKirill A. Shutemov } 440b46e756fSKirill A. Shutemov 441*7d2c4385SZach O'Keefe static inline int hpage_collapse_test_exit(struct mm_struct *mm) 442b46e756fSKirill A. Shutemov { 4434d45e75aSJann Horn return atomic_read(&mm->mm_users) == 0; 444b46e756fSKirill A. Shutemov } 445b46e756fSKirill A. Shutemov 446d2081b2bSYang Shi void __khugepaged_enter(struct mm_struct *mm) 447b46e756fSKirill A. Shutemov { 448b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 449b46e756fSKirill A. Shutemov int wakeup; 450b46e756fSKirill A. Shutemov 451b46e756fSKirill A. Shutemov mm_slot = alloc_mm_slot(); 452b46e756fSKirill A. Shutemov if (!mm_slot) 453d2081b2bSYang Shi return; 454b46e756fSKirill A. Shutemov 455b46e756fSKirill A. Shutemov /* __khugepaged_exit() must not run from under us */ 456*7d2c4385SZach O'Keefe VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm); 457b46e756fSKirill A. Shutemov if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 458b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 459d2081b2bSYang Shi return; 460b46e756fSKirill A. Shutemov } 461b46e756fSKirill A. Shutemov 462b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 463b46e756fSKirill A. Shutemov insert_to_mm_slots_hash(mm, mm_slot); 464b46e756fSKirill A. Shutemov /* 465b46e756fSKirill A. Shutemov * Insert just behind the scanning cursor, to let the area settle 466b46e756fSKirill A. Shutemov * down a little. 467b46e756fSKirill A. Shutemov */ 468b46e756fSKirill A. Shutemov wakeup = list_empty(&khugepaged_scan.mm_head); 469b46e756fSKirill A. Shutemov list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 470b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 471b46e756fSKirill A. Shutemov 472f1f10076SVegard Nossum mmgrab(mm); 473b46e756fSKirill A. Shutemov if (wakeup) 474b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 475b46e756fSKirill A. Shutemov } 476b46e756fSKirill A. Shutemov 477c791576cSYang Shi void khugepaged_enter_vma(struct vm_area_struct *vma, 478b46e756fSKirill A. Shutemov unsigned long vm_flags) 479b46e756fSKirill A. Shutemov { 4802647d11bSYang Shi if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && 4811064026bSYang Shi hugepage_flags_enabled()) { 482a7f4e6e4SZach O'Keefe if (hugepage_vma_check(vma, vm_flags, false, false, true)) 4832647d11bSYang Shi __khugepaged_enter(vma->vm_mm); 4842647d11bSYang Shi } 485b46e756fSKirill A. Shutemov } 486b46e756fSKirill A. Shutemov 487b46e756fSKirill A. Shutemov void __khugepaged_exit(struct mm_struct *mm) 488b46e756fSKirill A. Shutemov { 489b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 490b46e756fSKirill A. Shutemov int free = 0; 491b46e756fSKirill A. Shutemov 492b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 493b46e756fSKirill A. Shutemov mm_slot = get_mm_slot(mm); 494b46e756fSKirill A. Shutemov if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 495b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 496b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 497b46e756fSKirill A. Shutemov free = 1; 498b46e756fSKirill A. Shutemov } 499b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 500b46e756fSKirill A. Shutemov 501b46e756fSKirill A. Shutemov if (free) { 502b46e756fSKirill A. Shutemov clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 503b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 504b46e756fSKirill A. Shutemov mmdrop(mm); 505b46e756fSKirill A. Shutemov } else if (mm_slot) { 506b46e756fSKirill A. Shutemov /* 507b46e756fSKirill A. Shutemov * This is required to serialize against 508*7d2c4385SZach O'Keefe * hpage_collapse_test_exit() (which is guaranteed to run 509*7d2c4385SZach O'Keefe * under mmap sem read mode). Stop here (after we return all 510*7d2c4385SZach O'Keefe * pagetables will be destroyed) until khugepaged has finished 511*7d2c4385SZach O'Keefe * working on the pagetables under the mmap_lock. 512b46e756fSKirill A. Shutemov */ 513d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 514d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 515b46e756fSKirill A. Shutemov } 516b46e756fSKirill A. Shutemov } 517b46e756fSKirill A. Shutemov 518b46e756fSKirill A. Shutemov static void release_pte_page(struct page *page) 519b46e756fSKirill A. Shutemov { 5205503fbf2SKirill A. Shutemov mod_node_page_state(page_pgdat(page), 5215503fbf2SKirill A. Shutemov NR_ISOLATED_ANON + page_is_file_lru(page), 5225503fbf2SKirill A. Shutemov -compound_nr(page)); 523b46e756fSKirill A. Shutemov unlock_page(page); 524b46e756fSKirill A. Shutemov putback_lru_page(page); 525b46e756fSKirill A. Shutemov } 526b46e756fSKirill A. Shutemov 5275503fbf2SKirill A. Shutemov static void release_pte_pages(pte_t *pte, pte_t *_pte, 5285503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 529b46e756fSKirill A. Shutemov { 5305503fbf2SKirill A. Shutemov struct page *page, *tmp; 5315503fbf2SKirill A. Shutemov 532b46e756fSKirill A. Shutemov while (--_pte >= pte) { 533b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 5345503fbf2SKirill A. Shutemov 5355503fbf2SKirill A. Shutemov page = pte_page(pteval); 5365503fbf2SKirill A. Shutemov if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) && 5375503fbf2SKirill A. Shutemov !PageCompound(page)) 5385503fbf2SKirill A. Shutemov release_pte_page(page); 5395503fbf2SKirill A. Shutemov } 5405503fbf2SKirill A. Shutemov 5415503fbf2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, compound_pagelist, lru) { 5425503fbf2SKirill A. Shutemov list_del(&page->lru); 5435503fbf2SKirill A. Shutemov release_pte_page(page); 544b46e756fSKirill A. Shutemov } 545b46e756fSKirill A. Shutemov } 546b46e756fSKirill A. Shutemov 5479445689fSKirill A. Shutemov static bool is_refcount_suitable(struct page *page) 5489445689fSKirill A. Shutemov { 5499445689fSKirill A. Shutemov int expected_refcount; 5509445689fSKirill A. Shutemov 5519445689fSKirill A. Shutemov expected_refcount = total_mapcount(page); 5529445689fSKirill A. Shutemov if (PageSwapCache(page)) 5539445689fSKirill A. Shutemov expected_refcount += compound_nr(page); 5549445689fSKirill A. Shutemov 5559445689fSKirill A. Shutemov return page_count(page) == expected_refcount; 5569445689fSKirill A. Shutemov } 5579445689fSKirill A. Shutemov 558b46e756fSKirill A. Shutemov static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 559b46e756fSKirill A. Shutemov unsigned long address, 5605503fbf2SKirill A. Shutemov pte_t *pte, 561d8ea7cc8SZach O'Keefe struct collapse_control *cc, 5625503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 563b46e756fSKirill A. Shutemov { 564b46e756fSKirill A. Shutemov struct page *page = NULL; 565b46e756fSKirill A. Shutemov pte_t *_pte; 56650ad2f24SZach O'Keefe int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0; 5670db501f7SEbru Akagunduz bool writable = false; 568b46e756fSKirill A. Shutemov 569b46e756fSKirill A. Shutemov for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 570b46e756fSKirill A. Shutemov _pte++, address += PAGE_SIZE) { 571b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 572b46e756fSKirill A. Shutemov if (pte_none(pteval) || (pte_present(pteval) && 573b46e756fSKirill A. Shutemov is_zero_pfn(pte_pfn(pteval)))) { 574d8ea7cc8SZach O'Keefe ++none_or_zero; 575b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 576d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 577d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 578b46e756fSKirill A. Shutemov continue; 579b46e756fSKirill A. Shutemov } else { 580b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 581e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 582b46e756fSKirill A. Shutemov goto out; 583b46e756fSKirill A. Shutemov } 584b46e756fSKirill A. Shutemov } 585b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 586b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 587b46e756fSKirill A. Shutemov goto out; 588b46e756fSKirill A. Shutemov } 589b46e756fSKirill A. Shutemov page = vm_normal_page(vma, address, pteval); 5903218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 591b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 592b46e756fSKirill A. Shutemov goto out; 593b46e756fSKirill A. Shutemov } 594b46e756fSKirill A. Shutemov 595b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageAnon(page), page); 596b46e756fSKirill A. Shutemov 597d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 598d8ea7cc8SZach O'Keefe ++shared; 599d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 600d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 60171a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 602e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 60371a2c112SKirill A. Shutemov goto out; 60471a2c112SKirill A. Shutemov } 605d8ea7cc8SZach O'Keefe } 60671a2c112SKirill A. Shutemov 6075503fbf2SKirill A. Shutemov if (PageCompound(page)) { 6085503fbf2SKirill A. Shutemov struct page *p; 6095503fbf2SKirill A. Shutemov page = compound_head(page); 6105503fbf2SKirill A. Shutemov 6115503fbf2SKirill A. Shutemov /* 6125503fbf2SKirill A. Shutemov * Check if we have dealt with the compound page 6135503fbf2SKirill A. Shutemov * already 6145503fbf2SKirill A. Shutemov */ 6155503fbf2SKirill A. Shutemov list_for_each_entry(p, compound_pagelist, lru) { 6165503fbf2SKirill A. Shutemov if (page == p) 6175503fbf2SKirill A. Shutemov goto next; 6185503fbf2SKirill A. Shutemov } 6195503fbf2SKirill A. Shutemov } 6205503fbf2SKirill A. Shutemov 621b46e756fSKirill A. Shutemov /* 622b46e756fSKirill A. Shutemov * We can do it before isolate_lru_page because the 623b46e756fSKirill A. Shutemov * page can't be freed from under us. NOTE: PG_lock 624b46e756fSKirill A. Shutemov * is needed to serialize against split_huge_page 625b46e756fSKirill A. Shutemov * when invoked from the VM. 626b46e756fSKirill A. Shutemov */ 627b46e756fSKirill A. Shutemov if (!trylock_page(page)) { 628b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 629b46e756fSKirill A. Shutemov goto out; 630b46e756fSKirill A. Shutemov } 631b46e756fSKirill A. Shutemov 632b46e756fSKirill A. Shutemov /* 6339445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 6349445689fSKirill A. Shutemov * 6359445689fSKirill A. Shutemov * The page table that maps the page has been already unlinked 6369445689fSKirill A. Shutemov * from the page table tree and this process cannot get 637f0953a1bSIngo Molnar * an additional pin on the page. 6389445689fSKirill A. Shutemov * 6399445689fSKirill A. Shutemov * New pins can come later if the page is shared across fork, 6409445689fSKirill A. Shutemov * but not from this process. The other process cannot write to 6419445689fSKirill A. Shutemov * the page, only trigger CoW. 642b46e756fSKirill A. Shutemov */ 6439445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 644b46e756fSKirill A. Shutemov unlock_page(page); 645b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 646b46e756fSKirill A. Shutemov goto out; 647b46e756fSKirill A. Shutemov } 648b46e756fSKirill A. Shutemov 649b46e756fSKirill A. Shutemov /* 650b46e756fSKirill A. Shutemov * Isolate the page to avoid collapsing an hugepage 651b46e756fSKirill A. Shutemov * currently in use by the VM. 652b46e756fSKirill A. Shutemov */ 653b46e756fSKirill A. Shutemov if (isolate_lru_page(page)) { 654b46e756fSKirill A. Shutemov unlock_page(page); 655b46e756fSKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 656b46e756fSKirill A. Shutemov goto out; 657b46e756fSKirill A. Shutemov } 6585503fbf2SKirill A. Shutemov mod_node_page_state(page_pgdat(page), 6595503fbf2SKirill A. Shutemov NR_ISOLATED_ANON + page_is_file_lru(page), 6605503fbf2SKirill A. Shutemov compound_nr(page)); 661b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 662b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(PageLRU(page), page); 663b46e756fSKirill A. Shutemov 6645503fbf2SKirill A. Shutemov if (PageCompound(page)) 6655503fbf2SKirill A. Shutemov list_add_tail(&page->lru, compound_pagelist); 6665503fbf2SKirill A. Shutemov next: 667d8ea7cc8SZach O'Keefe /* 668d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 669d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 670d8ea7cc8SZach O'Keefe */ 671d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 672d8ea7cc8SZach O'Keefe (pte_young(pteval) || page_is_young(page) || 673d8ea7cc8SZach O'Keefe PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, 674d8ea7cc8SZach O'Keefe address))) 6750db501f7SEbru Akagunduz referenced++; 6765503fbf2SKirill A. Shutemov 6775503fbf2SKirill A. Shutemov if (pte_write(pteval)) 6785503fbf2SKirill A. Shutemov writable = true; 679b46e756fSKirill A. Shutemov } 68074e579bfSMiaohe Lin 68174e579bfSMiaohe Lin if (unlikely(!writable)) { 68274e579bfSMiaohe Lin result = SCAN_PAGE_RO; 683d8ea7cc8SZach O'Keefe } else if (unlikely(cc->is_khugepaged && !referenced)) { 68474e579bfSMiaohe Lin result = SCAN_LACK_REFERENCED_PAGE; 68574e579bfSMiaohe Lin } else { 686b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 687b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 688b46e756fSKirill A. Shutemov referenced, writable, result); 68950ad2f24SZach O'Keefe return result; 690b46e756fSKirill A. Shutemov } 691b46e756fSKirill A. Shutemov out: 6925503fbf2SKirill A. Shutemov release_pte_pages(pte, _pte, compound_pagelist); 693b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 694b46e756fSKirill A. Shutemov referenced, writable, result); 69550ad2f24SZach O'Keefe return result; 696b46e756fSKirill A. Shutemov } 697b46e756fSKirill A. Shutemov 698b46e756fSKirill A. Shutemov static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 699b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 700b46e756fSKirill A. Shutemov unsigned long address, 7015503fbf2SKirill A. Shutemov spinlock_t *ptl, 7025503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 703b46e756fSKirill A. Shutemov { 7045503fbf2SKirill A. Shutemov struct page *src_page, *tmp; 705b46e756fSKirill A. Shutemov pte_t *_pte; 706338a16baSDavid Rientjes for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 707338a16baSDavid Rientjes _pte++, page++, address += PAGE_SIZE) { 708b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 709b46e756fSKirill A. Shutemov 710b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 711b46e756fSKirill A. Shutemov clear_user_highpage(page, address); 712b46e756fSKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 713b46e756fSKirill A. Shutemov if (is_zero_pfn(pte_pfn(pteval))) { 714b46e756fSKirill A. Shutemov /* 715b46e756fSKirill A. Shutemov * ptl mostly unnecessary. 716b46e756fSKirill A. Shutemov */ 717b46e756fSKirill A. Shutemov spin_lock(ptl); 71808d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 719b46e756fSKirill A. Shutemov spin_unlock(ptl); 720b46e756fSKirill A. Shutemov } 721b46e756fSKirill A. Shutemov } else { 722b46e756fSKirill A. Shutemov src_page = pte_page(pteval); 723b46e756fSKirill A. Shutemov copy_user_highpage(page, src_page, address, vma); 7245503fbf2SKirill A. Shutemov if (!PageCompound(src_page)) 725b46e756fSKirill A. Shutemov release_pte_page(src_page); 726b46e756fSKirill A. Shutemov /* 727b46e756fSKirill A. Shutemov * ptl mostly unnecessary, but preempt has to 728b46e756fSKirill A. Shutemov * be disabled to update the per-cpu stats 729b46e756fSKirill A. Shutemov * inside page_remove_rmap(). 730b46e756fSKirill A. Shutemov */ 731b46e756fSKirill A. Shutemov spin_lock(ptl); 73208d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 733cea86fe2SHugh Dickins page_remove_rmap(src_page, vma, false); 734b46e756fSKirill A. Shutemov spin_unlock(ptl); 735b46e756fSKirill A. Shutemov free_page_and_swap_cache(src_page); 736b46e756fSKirill A. Shutemov } 737b46e756fSKirill A. Shutemov } 7385503fbf2SKirill A. Shutemov 7395503fbf2SKirill A. Shutemov list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { 7405503fbf2SKirill A. Shutemov list_del(&src_page->lru); 7411baec203SMiaohe Lin mod_node_page_state(page_pgdat(src_page), 7421baec203SMiaohe Lin NR_ISOLATED_ANON + page_is_file_lru(src_page), 7431baec203SMiaohe Lin -compound_nr(src_page)); 7441baec203SMiaohe Lin unlock_page(src_page); 7451baec203SMiaohe Lin free_swap_cache(src_page); 7461baec203SMiaohe Lin putback_lru_page(src_page); 7475503fbf2SKirill A. Shutemov } 748b46e756fSKirill A. Shutemov } 749b46e756fSKirill A. Shutemov 750b46e756fSKirill A. Shutemov static void khugepaged_alloc_sleep(void) 751b46e756fSKirill A. Shutemov { 752b46e756fSKirill A. Shutemov DEFINE_WAIT(wait); 753b46e756fSKirill A. Shutemov 754b46e756fSKirill A. Shutemov add_wait_queue(&khugepaged_wait, &wait); 755b46e756fSKirill A. Shutemov freezable_schedule_timeout_interruptible( 756b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 757b46e756fSKirill A. Shutemov remove_wait_queue(&khugepaged_wait, &wait); 758b46e756fSKirill A. Shutemov } 759b46e756fSKirill A. Shutemov 76034d6b470SZach O'Keefe struct collapse_control khugepaged_collapse_control = { 761d8ea7cc8SZach O'Keefe .is_khugepaged = true, 76234d6b470SZach O'Keefe .last_target_node = NUMA_NO_NODE, 76334d6b470SZach O'Keefe }; 76434d6b470SZach O'Keefe 765*7d2c4385SZach O'Keefe static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc) 766b46e756fSKirill A. Shutemov { 767b46e756fSKirill A. Shutemov int i; 768b46e756fSKirill A. Shutemov 769b46e756fSKirill A. Shutemov /* 770a5f5f91dSMel Gorman * If node_reclaim_mode is disabled, then no extra effort is made to 771b46e756fSKirill A. Shutemov * allocate memory locally. 772b46e756fSKirill A. Shutemov */ 773202e35dbSDave Hansen if (!node_reclaim_enabled()) 774b46e756fSKirill A. Shutemov return false; 775b46e756fSKirill A. Shutemov 776b46e756fSKirill A. Shutemov /* If there is a count for this node already, it must be acceptable */ 77734d6b470SZach O'Keefe if (cc->node_load[nid]) 778b46e756fSKirill A. Shutemov return false; 779b46e756fSKirill A. Shutemov 780b46e756fSKirill A. Shutemov for (i = 0; i < MAX_NUMNODES; i++) { 78134d6b470SZach O'Keefe if (!cc->node_load[i]) 782b46e756fSKirill A. Shutemov continue; 783a55c7454SMatt Fleming if (node_distance(nid, i) > node_reclaim_distance) 784b46e756fSKirill A. Shutemov return true; 785b46e756fSKirill A. Shutemov } 786b46e756fSKirill A. Shutemov return false; 787b46e756fSKirill A. Shutemov } 788b46e756fSKirill A. Shutemov 7891064026bSYang Shi #define khugepaged_defrag() \ 7901064026bSYang Shi (transparent_hugepage_flags & \ 7911064026bSYang Shi (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)) 7921064026bSYang Shi 793b46e756fSKirill A. Shutemov /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 794b46e756fSKirill A. Shutemov static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 795b46e756fSKirill A. Shutemov { 79625160354SVlastimil Babka return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 797b46e756fSKirill A. Shutemov } 798b46e756fSKirill A. Shutemov 799b46e756fSKirill A. Shutemov #ifdef CONFIG_NUMA 800*7d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 801b46e756fSKirill A. Shutemov { 802b46e756fSKirill A. Shutemov int nid, target_node = 0, max_value = 0; 803b46e756fSKirill A. Shutemov 804b46e756fSKirill A. Shutemov /* find first node with max normal pages hit */ 805b46e756fSKirill A. Shutemov for (nid = 0; nid < MAX_NUMNODES; nid++) 80634d6b470SZach O'Keefe if (cc->node_load[nid] > max_value) { 80734d6b470SZach O'Keefe max_value = cc->node_load[nid]; 808b46e756fSKirill A. Shutemov target_node = nid; 809b46e756fSKirill A. Shutemov } 810b46e756fSKirill A. Shutemov 811b46e756fSKirill A. Shutemov /* do some balance if several nodes have the same hit record */ 81234d6b470SZach O'Keefe if (target_node <= cc->last_target_node) 81334d6b470SZach O'Keefe for (nid = cc->last_target_node + 1; nid < MAX_NUMNODES; 814b46e756fSKirill A. Shutemov nid++) 81534d6b470SZach O'Keefe if (max_value == cc->node_load[nid]) { 816b46e756fSKirill A. Shutemov target_node = nid; 817b46e756fSKirill A. Shutemov break; 818b46e756fSKirill A. Shutemov } 819b46e756fSKirill A. Shutemov 82034d6b470SZach O'Keefe cc->last_target_node = target_node; 821b46e756fSKirill A. Shutemov return target_node; 822b46e756fSKirill A. Shutemov } 823c6a7f445SYang Shi #else 824*7d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 825b46e756fSKirill A. Shutemov { 826c6a7f445SYang Shi return 0; 827b46e756fSKirill A. Shutemov } 828c6a7f445SYang Shi #endif 829b46e756fSKirill A. Shutemov 830*7d2c4385SZach O'Keefe static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node) 831b46e756fSKirill A. Shutemov { 832b46e756fSKirill A. Shutemov *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); 833b46e756fSKirill A. Shutemov if (unlikely(!*hpage)) { 834b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 8359710a78aSZach O'Keefe return false; 836b46e756fSKirill A. Shutemov } 837b46e756fSKirill A. Shutemov 838b46e756fSKirill A. Shutemov prep_transhuge_page(*hpage); 839b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 8409710a78aSZach O'Keefe return true; 841b46e756fSKirill A. Shutemov } 842b46e756fSKirill A. Shutemov 843b46e756fSKirill A. Shutemov /* 844c1e8d7c6SMichel Lespinasse * If mmap_lock temporarily dropped, revalidate vma 845c1e8d7c6SMichel Lespinasse * before taking mmap_lock. 84650ad2f24SZach O'Keefe * Returns enum scan_result value. 847b46e756fSKirill A. Shutemov */ 848b46e756fSKirill A. Shutemov 849c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 850a7f4e6e4SZach O'Keefe struct vm_area_struct **vmap, 851a7f4e6e4SZach O'Keefe struct collapse_control *cc) 852b46e756fSKirill A. Shutemov { 853b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 854b46e756fSKirill A. Shutemov 855*7d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 856b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 857b46e756fSKirill A. Shutemov 858c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 859b46e756fSKirill A. Shutemov if (!vma) 860b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 861b46e756fSKirill A. Shutemov 8624fa6893fSYang Shi if (!transhuge_vma_suitable(vma, address)) 863b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 864a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, 865a7f4e6e4SZach O'Keefe cc->is_khugepaged)) 866b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 867f707fa49SYang Shi /* 868f707fa49SYang Shi * Anon VMA expected, the address may be unmapped then 869f707fa49SYang Shi * remapped to file after khugepaged reaquired the mmap_lock. 870f707fa49SYang Shi * 871f707fa49SYang Shi * hugepage_vma_check may return true for qualified file 872f707fa49SYang Shi * vmas. 873f707fa49SYang Shi */ 87425fa414aSxu xin if (!vma->anon_vma || !vma_is_anonymous(vma)) 875594cced1SKirill A. Shutemov return SCAN_VMA_CHECK; 87650ad2f24SZach O'Keefe return SCAN_SUCCEED; 877b46e756fSKirill A. Shutemov } 878b46e756fSKirill A. Shutemov 87950722804SZach O'Keefe static int find_pmd_or_thp_or_none(struct mm_struct *mm, 88050722804SZach O'Keefe unsigned long address, 88150722804SZach O'Keefe pmd_t **pmd) 88250722804SZach O'Keefe { 88350722804SZach O'Keefe pmd_t pmde; 88450722804SZach O'Keefe 88550722804SZach O'Keefe *pmd = mm_find_pmd(mm, address); 88650722804SZach O'Keefe if (!*pmd) 88750722804SZach O'Keefe return SCAN_PMD_NULL; 88850722804SZach O'Keefe 88950722804SZach O'Keefe pmde = pmd_read_atomic(*pmd); 89050722804SZach O'Keefe 89150722804SZach O'Keefe #ifdef CONFIG_TRANSPARENT_HUGEPAGE 89250722804SZach O'Keefe /* See comments in pmd_none_or_trans_huge_or_clear_bad() */ 89350722804SZach O'Keefe barrier(); 89450722804SZach O'Keefe #endif 89550722804SZach O'Keefe if (!pmd_present(pmde)) 89650722804SZach O'Keefe return SCAN_PMD_NULL; 89750722804SZach O'Keefe if (pmd_trans_huge(pmde)) 89850722804SZach O'Keefe return SCAN_PMD_MAPPED; 89950722804SZach O'Keefe if (pmd_bad(pmde)) 90050722804SZach O'Keefe return SCAN_PMD_NULL; 90150722804SZach O'Keefe return SCAN_SUCCEED; 90250722804SZach O'Keefe } 90350722804SZach O'Keefe 90450722804SZach O'Keefe static int check_pmd_still_valid(struct mm_struct *mm, 90550722804SZach O'Keefe unsigned long address, 90650722804SZach O'Keefe pmd_t *pmd) 90750722804SZach O'Keefe { 90850722804SZach O'Keefe pmd_t *new_pmd; 90950722804SZach O'Keefe int result = find_pmd_or_thp_or_none(mm, address, &new_pmd); 91050722804SZach O'Keefe 91150722804SZach O'Keefe if (result != SCAN_SUCCEED) 91250722804SZach O'Keefe return result; 91350722804SZach O'Keefe if (new_pmd != pmd) 91450722804SZach O'Keefe return SCAN_FAIL; 91550722804SZach O'Keefe return SCAN_SUCCEED; 91650722804SZach O'Keefe } 91750722804SZach O'Keefe 918b46e756fSKirill A. Shutemov /* 919b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 920*7d2c4385SZach O'Keefe * Only done if hpage_collapse_scan_pmd believes it is worthwhile. 921b46e756fSKirill A. Shutemov * 9224d928e20SMiaohe Lin * Called and returns without pte mapped or spinlocks held. 9234d928e20SMiaohe Lin * Note that if false is returned, mmap_lock will be released. 924b46e756fSKirill A. Shutemov */ 925b46e756fSKirill A. Shutemov 92650ad2f24SZach O'Keefe static int __collapse_huge_page_swapin(struct mm_struct *mm, 927b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 9282b635dd3SWill Deacon unsigned long haddr, pmd_t *pmd, 9290db501f7SEbru Akagunduz int referenced) 930b46e756fSKirill A. Shutemov { 9312b740303SSouptick Joarder int swapped_in = 0; 9322b740303SSouptick Joarder vm_fault_t ret = 0; 9332b635dd3SWill Deacon unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); 9342b635dd3SWill Deacon 9352b635dd3SWill Deacon for (address = haddr; address < end; address += PAGE_SIZE) { 93682b0f8c3SJan Kara struct vm_fault vmf = { 937b46e756fSKirill A. Shutemov .vma = vma, 938b46e756fSKirill A. Shutemov .address = address, 9392b635dd3SWill Deacon .pgoff = linear_page_index(vma, haddr), 940b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 941b46e756fSKirill A. Shutemov .pmd = pmd, 942b46e756fSKirill A. Shutemov }; 943b46e756fSKirill A. Shutemov 94482b0f8c3SJan Kara vmf.pte = pte_offset_map(pmd, address); 9452994302bSJan Kara vmf.orig_pte = *vmf.pte; 9462b635dd3SWill Deacon if (!is_swap_pte(vmf.orig_pte)) { 9472b635dd3SWill Deacon pte_unmap(vmf.pte); 948b46e756fSKirill A. Shutemov continue; 9492b635dd3SWill Deacon } 9502994302bSJan Kara ret = do_swap_page(&vmf); 9510db501f7SEbru Akagunduz 9524d928e20SMiaohe Lin /* 9534d928e20SMiaohe Lin * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. 9544d928e20SMiaohe Lin * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because 9554d928e20SMiaohe Lin * we do not retry here and swap entry will remain in pagetable 9564d928e20SMiaohe Lin * resulting in later failure. 9574d928e20SMiaohe Lin */ 958b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 9590db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 96050ad2f24SZach O'Keefe /* Likely, but not guaranteed, that page lock failed */ 96150ad2f24SZach O'Keefe return SCAN_PAGE_LOCK; 96247f863eaSEbru Akagunduz } 963b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 9644d928e20SMiaohe Lin mmap_read_unlock(mm); 9650db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 96650ad2f24SZach O'Keefe return SCAN_FAIL; 967b46e756fSKirill A. Shutemov } 9684d928e20SMiaohe Lin swapped_in++; 969b46e756fSKirill A. Shutemov } 970ae2c5d80SKirill A. Shutemov 971ae2c5d80SKirill A. Shutemov /* Drain LRU add pagevec to remove extra pin on the swapped in pages */ 972ae2c5d80SKirill A. Shutemov if (swapped_in) 973ae2c5d80SKirill A. Shutemov lru_add_drain(); 974ae2c5d80SKirill A. Shutemov 9750db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); 97650ad2f24SZach O'Keefe return SCAN_SUCCEED; 977b46e756fSKirill A. Shutemov } 978b46e756fSKirill A. Shutemov 9799710a78aSZach O'Keefe static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, 9809710a78aSZach O'Keefe struct collapse_control *cc) 9819710a78aSZach O'Keefe { 9829710a78aSZach O'Keefe /* Only allocate from the target node */ 9837d8faaf1SZach O'Keefe gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() : 9847d8faaf1SZach O'Keefe GFP_TRANSHUGE) | __GFP_THISNODE; 985*7d2c4385SZach O'Keefe int node = hpage_collapse_find_target_node(cc); 9869710a78aSZach O'Keefe 987*7d2c4385SZach O'Keefe if (!hpage_collapse_alloc_page(hpage, gfp, node)) 9889710a78aSZach O'Keefe return SCAN_ALLOC_HUGE_PAGE_FAIL; 9899710a78aSZach O'Keefe if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp))) 9909710a78aSZach O'Keefe return SCAN_CGROUP_CHARGE_FAIL; 9919710a78aSZach O'Keefe count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC); 9929710a78aSZach O'Keefe return SCAN_SUCCEED; 9939710a78aSZach O'Keefe } 9949710a78aSZach O'Keefe 99550ad2f24SZach O'Keefe static int collapse_huge_page(struct mm_struct *mm, unsigned long address, 99650ad2f24SZach O'Keefe int referenced, int unmapped, 99750ad2f24SZach O'Keefe struct collapse_control *cc) 998b46e756fSKirill A. Shutemov { 9995503fbf2SKirill A. Shutemov LIST_HEAD(compound_pagelist); 1000b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 1001b46e756fSKirill A. Shutemov pte_t *pte; 1002b46e756fSKirill A. Shutemov pgtable_t pgtable; 100350ad2f24SZach O'Keefe struct page *hpage; 1004b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 100550ad2f24SZach O'Keefe int result = SCAN_FAIL; 1006c131f751SKirill A. Shutemov struct vm_area_struct *vma; 1007ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1008b46e756fSKirill A. Shutemov 1009b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1010b46e756fSKirill A. Shutemov 1011988ddb71SKirill A. Shutemov /* 1012c1e8d7c6SMichel Lespinasse * Before allocating the hugepage, release the mmap_lock read lock. 1013988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 1014c1e8d7c6SMichel Lespinasse * sync compaction, and we do not need to hold the mmap_lock during 1015988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 1016988ddb71SKirill A. Shutemov */ 1017d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1018b46e756fSKirill A. Shutemov 101950ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 10209710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1021b46e756fSKirill A. Shutemov goto out_nolock; 10229710a78aSZach O'Keefe 1023d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 1024a7f4e6e4SZach O'Keefe result = hugepage_vma_revalidate(mm, address, &vma, cc); 102550ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) { 1026d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1027b46e756fSKirill A. Shutemov goto out_nolock; 1028b46e756fSKirill A. Shutemov } 1029b46e756fSKirill A. Shutemov 103050722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 103150722804SZach O'Keefe if (result != SCAN_SUCCEED) { 1032d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1033b46e756fSKirill A. Shutemov goto out_nolock; 1034b46e756fSKirill A. Shutemov } 1035b46e756fSKirill A. Shutemov 103650ad2f24SZach O'Keefe if (unmapped) { 1037b46e756fSKirill A. Shutemov /* 103850ad2f24SZach O'Keefe * __collapse_huge_page_swapin will return with mmap_lock 103950ad2f24SZach O'Keefe * released when it fails. So we jump out_nolock directly in 104050ad2f24SZach O'Keefe * that case. Continuing to collapse causes inconsistency. 1041b46e756fSKirill A. Shutemov */ 104250ad2f24SZach O'Keefe result = __collapse_huge_page_swapin(mm, vma, address, pmd, 104350ad2f24SZach O'Keefe referenced); 104450ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 1045b46e756fSKirill A. Shutemov goto out_nolock; 1046b46e756fSKirill A. Shutemov } 1047b46e756fSKirill A. Shutemov 1048d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1049b46e756fSKirill A. Shutemov /* 1050b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 1051b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 1052b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 1053b46e756fSKirill A. Shutemov */ 1054d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 1055a7f4e6e4SZach O'Keefe result = hugepage_vma_revalidate(mm, address, &vma, cc); 105650ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 105718d24a7cSMiaohe Lin goto out_up_write; 1058b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 105950722804SZach O'Keefe result = check_pmd_still_valid(mm, address, pmd); 106050722804SZach O'Keefe if (result != SCAN_SUCCEED) 106118d24a7cSMiaohe Lin goto out_up_write; 1062b46e756fSKirill A. Shutemov 1063b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1064b46e756fSKirill A. Shutemov 10657269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 10666f4f13e8SJérôme Glisse address, address + HPAGE_PMD_SIZE); 1067ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1068ec649c9dSVille Syrjälä 1069ec649c9dSVille Syrjälä pte = pte_offset_map(pmd, address); 1070ec649c9dSVille Syrjälä pte_ptl = pte_lockptr(mm, pmd); 1071ec649c9dSVille Syrjälä 1072b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1073b46e756fSKirill A. Shutemov /* 1074b46e756fSKirill A. Shutemov * After this gup_fast can't run anymore. This also removes 1075b46e756fSKirill A. Shutemov * any huge TLB entry from the CPU so we won't allow 1076b46e756fSKirill A. Shutemov * huge and small TLB entries for the same virtual address 1077b46e756fSKirill A. Shutemov * to avoid the risk of CPU bugs in that area. 1078b46e756fSKirill A. Shutemov */ 1079b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1080b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1081ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1082b46e756fSKirill A. Shutemov 1083b46e756fSKirill A. Shutemov spin_lock(pte_ptl); 1084d8ea7cc8SZach O'Keefe result = __collapse_huge_page_isolate(vma, address, pte, cc, 10855503fbf2SKirill A. Shutemov &compound_pagelist); 1086b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1087b46e756fSKirill A. Shutemov 108850ad2f24SZach O'Keefe if (unlikely(result != SCAN_SUCCEED)) { 1089b46e756fSKirill A. Shutemov pte_unmap(pte); 1090b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1091b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1092b46e756fSKirill A. Shutemov /* 1093b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1094b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1095b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1096b46e756fSKirill A. Shutemov */ 1097b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1098b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1099b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 110018d24a7cSMiaohe Lin goto out_up_write; 1101b46e756fSKirill A. Shutemov } 1102b46e756fSKirill A. Shutemov 1103b46e756fSKirill A. Shutemov /* 1104b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1105b46e756fSKirill A. Shutemov * can't run anymore. 1106b46e756fSKirill A. Shutemov */ 1107b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1108b46e756fSKirill A. Shutemov 110950ad2f24SZach O'Keefe __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl, 11105503fbf2SKirill A. Shutemov &compound_pagelist); 1111b46e756fSKirill A. Shutemov pte_unmap(pte); 1112588d01f9SMiaohe Lin /* 1113588d01f9SMiaohe Lin * spin_lock() below is not the equivalent of smp_wmb(), but 1114588d01f9SMiaohe Lin * the smp_wmb() inside __SetPageUptodate() can be reused to 1115588d01f9SMiaohe Lin * avoid the copy_huge_page writes to become visible after 1116588d01f9SMiaohe Lin * the set_pmd_at() write. 1117588d01f9SMiaohe Lin */ 111850ad2f24SZach O'Keefe __SetPageUptodate(hpage); 1119b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1120b46e756fSKirill A. Shutemov 112150ad2f24SZach O'Keefe _pmd = mk_huge_pmd(hpage, vma->vm_page_prot); 1122f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1123b46e756fSKirill A. Shutemov 1124b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1125b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 112650ad2f24SZach O'Keefe page_add_new_anon_rmap(hpage, vma, address); 112750ad2f24SZach O'Keefe lru_cache_add_inactive_or_unevictable(hpage, vma); 1128b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1129b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1130b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1131b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1132b46e756fSKirill A. Shutemov 113350ad2f24SZach O'Keefe hpage = NULL; 1134b46e756fSKirill A. Shutemov 1135b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1136b46e756fSKirill A. Shutemov out_up_write: 1137d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 1138b46e756fSKirill A. Shutemov out_nolock: 113950ad2f24SZach O'Keefe if (hpage) { 114050ad2f24SZach O'Keefe mem_cgroup_uncharge(page_folio(hpage)); 114150ad2f24SZach O'Keefe put_page(hpage); 1142c6a7f445SYang Shi } 114350ad2f24SZach O'Keefe trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); 114450ad2f24SZach O'Keefe return result; 1145b46e756fSKirill A. Shutemov } 1146b46e756fSKirill A. Shutemov 1147*7d2c4385SZach O'Keefe static int hpage_collapse_scan_pmd(struct mm_struct *mm, 1148*7d2c4385SZach O'Keefe struct vm_area_struct *vma, 114950ad2f24SZach O'Keefe unsigned long address, bool *mmap_locked, 115034d6b470SZach O'Keefe struct collapse_control *cc) 1151b46e756fSKirill A. Shutemov { 1152b46e756fSKirill A. Shutemov pmd_t *pmd; 1153b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 115450ad2f24SZach O'Keefe int result = SCAN_FAIL, referenced = 0; 115571a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0; 1156b46e756fSKirill A. Shutemov struct page *page = NULL; 1157b46e756fSKirill A. Shutemov unsigned long _address; 1158b46e756fSKirill A. Shutemov spinlock_t *ptl; 1159b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 11600db501f7SEbru Akagunduz bool writable = false; 1161b46e756fSKirill A. Shutemov 1162b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1163b46e756fSKirill A. Shutemov 116450722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 116550722804SZach O'Keefe if (result != SCAN_SUCCEED) 1166b46e756fSKirill A. Shutemov goto out; 1167b46e756fSKirill A. Shutemov 116834d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1169b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1170b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; 1171b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1172b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 1173b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1174d8ea7cc8SZach O'Keefe ++unmapped; 1175d8ea7cc8SZach O'Keefe if (!cc->is_khugepaged || 1176d8ea7cc8SZach O'Keefe unmapped <= khugepaged_max_ptes_swap) { 1177e1e267c7SPeter Xu /* 1178e1e267c7SPeter Xu * Always be strict with uffd-wp 1179e1e267c7SPeter Xu * enabled swap entries. Please see 1180e1e267c7SPeter Xu * comment below for pte_uffd_wp(). 1181e1e267c7SPeter Xu */ 1182e1e267c7SPeter Xu if (pte_swp_uffd_wp(pteval)) { 1183e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1184e1e267c7SPeter Xu goto out_unmap; 1185e1e267c7SPeter Xu } 1186b46e756fSKirill A. Shutemov continue; 1187b46e756fSKirill A. Shutemov } else { 1188b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1189e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1190b46e756fSKirill A. Shutemov goto out_unmap; 1191b46e756fSKirill A. Shutemov } 1192b46e756fSKirill A. Shutemov } 1193b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1194d8ea7cc8SZach O'Keefe ++none_or_zero; 1195b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1196d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 1197d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 1198b46e756fSKirill A. Shutemov continue; 1199b46e756fSKirill A. Shutemov } else { 1200b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1201e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1202b46e756fSKirill A. Shutemov goto out_unmap; 1203b46e756fSKirill A. Shutemov } 1204b46e756fSKirill A. Shutemov } 1205e1e267c7SPeter Xu if (pte_uffd_wp(pteval)) { 1206e1e267c7SPeter Xu /* 1207e1e267c7SPeter Xu * Don't collapse the page if any of the small 1208e1e267c7SPeter Xu * PTEs are armed with uffd write protection. 1209e1e267c7SPeter Xu * Here we can also mark the new huge pmd as 1210e1e267c7SPeter Xu * write protected if any of the small ones is 12118958b249SHaitao Shi * marked but that could bring unknown 1212e1e267c7SPeter Xu * userfault messages that falls outside of 1213e1e267c7SPeter Xu * the registered range. So, just be simple. 1214e1e267c7SPeter Xu */ 1215e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1216e1e267c7SPeter Xu goto out_unmap; 1217e1e267c7SPeter Xu } 1218b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1219b46e756fSKirill A. Shutemov writable = true; 1220b46e756fSKirill A. Shutemov 1221b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 12223218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 1223b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1224b46e756fSKirill A. Shutemov goto out_unmap; 1225b46e756fSKirill A. Shutemov } 1226b46e756fSKirill A. Shutemov 1227d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 1228d8ea7cc8SZach O'Keefe ++shared; 1229d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 1230d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 123171a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 1232e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 123371a2c112SKirill A. Shutemov goto out_unmap; 123471a2c112SKirill A. Shutemov } 1235d8ea7cc8SZach O'Keefe } 123671a2c112SKirill A. Shutemov 12375503fbf2SKirill A. Shutemov page = compound_head(page); 1238b46e756fSKirill A. Shutemov 1239b46e756fSKirill A. Shutemov /* 1240b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 124134d6b470SZach O'Keefe * information to cc->node_load[]. 12420b8f0d87SQuanfa Fu * Khugepaged will allocate hugepage from the node has the max 1243b46e756fSKirill A. Shutemov * hit record. 1244b46e756fSKirill A. Shutemov */ 1245b46e756fSKirill A. Shutemov node = page_to_nid(page); 1246*7d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 1247b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1248b46e756fSKirill A. Shutemov goto out_unmap; 1249b46e756fSKirill A. Shutemov } 125034d6b470SZach O'Keefe cc->node_load[node]++; 1251b46e756fSKirill A. Shutemov if (!PageLRU(page)) { 1252b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1253b46e756fSKirill A. Shutemov goto out_unmap; 1254b46e756fSKirill A. Shutemov } 1255b46e756fSKirill A. Shutemov if (PageLocked(page)) { 1256b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1257b46e756fSKirill A. Shutemov goto out_unmap; 1258b46e756fSKirill A. Shutemov } 1259b46e756fSKirill A. Shutemov if (!PageAnon(page)) { 1260b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1261b46e756fSKirill A. Shutemov goto out_unmap; 1262b46e756fSKirill A. Shutemov } 1263b46e756fSKirill A. Shutemov 1264b46e756fSKirill A. Shutemov /* 12659445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 12669445689fSKirill A. Shutemov * 126736ee2c78SMiaohe Lin * Here the check is racy it may see total_mapcount > refcount 12689445689fSKirill A. Shutemov * in some cases. 12699445689fSKirill A. Shutemov * For example, one process with one forked child process. 12709445689fSKirill A. Shutemov * The parent has the PMD split due to MADV_DONTNEED, then 12719445689fSKirill A. Shutemov * the child is trying unmap the whole PMD, but khugepaged 12729445689fSKirill A. Shutemov * may be scanning the parent between the child has 12739445689fSKirill A. Shutemov * PageDoubleMap flag cleared and dec the mapcount. So 12749445689fSKirill A. Shutemov * khugepaged may see total_mapcount > refcount. 12759445689fSKirill A. Shutemov * 12769445689fSKirill A. Shutemov * But such case is ephemeral we could always retry collapse 12779445689fSKirill A. Shutemov * later. However it may report false positive if the page 12789445689fSKirill A. Shutemov * has excessive GUP pins (i.e. 512). Anyway the same check 12799445689fSKirill A. Shutemov * will be done again later the risk seems low. 1280b46e756fSKirill A. Shutemov */ 12819445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 1282b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1283b46e756fSKirill A. Shutemov goto out_unmap; 1284b46e756fSKirill A. Shutemov } 1285d8ea7cc8SZach O'Keefe 1286d8ea7cc8SZach O'Keefe /* 1287d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 1288d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 1289d8ea7cc8SZach O'Keefe */ 1290d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 1291d8ea7cc8SZach O'Keefe (pte_young(pteval) || page_is_young(page) || 1292d8ea7cc8SZach O'Keefe PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, 1293d8ea7cc8SZach O'Keefe address))) 12940db501f7SEbru Akagunduz referenced++; 1295b46e756fSKirill A. Shutemov } 1296ffe945e6SKirill A. Shutemov if (!writable) { 1297ffe945e6SKirill A. Shutemov result = SCAN_PAGE_RO; 1298d8ea7cc8SZach O'Keefe } else if (cc->is_khugepaged && 1299d8ea7cc8SZach O'Keefe (!referenced || 1300d8ea7cc8SZach O'Keefe (unmapped && referenced < HPAGE_PMD_NR / 2))) { 1301ffe945e6SKirill A. Shutemov result = SCAN_LACK_REFERENCED_PAGE; 1302ffe945e6SKirill A. Shutemov } else { 1303b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1304b46e756fSKirill A. Shutemov } 1305b46e756fSKirill A. Shutemov out_unmap: 1306b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 130750ad2f24SZach O'Keefe if (result == SCAN_SUCCEED) { 130850ad2f24SZach O'Keefe result = collapse_huge_page(mm, address, referenced, 130950ad2f24SZach O'Keefe unmapped, cc); 1310c1e8d7c6SMichel Lespinasse /* collapse_huge_page will return with the mmap_lock released */ 131150ad2f24SZach O'Keefe *mmap_locked = false; 1312b46e756fSKirill A. Shutemov } 1313b46e756fSKirill A. Shutemov out: 1314b46e756fSKirill A. Shutemov trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, 1315b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 131650ad2f24SZach O'Keefe return result; 1317b46e756fSKirill A. Shutemov } 1318b46e756fSKirill A. Shutemov 1319b46e756fSKirill A. Shutemov static void collect_mm_slot(struct mm_slot *mm_slot) 1320b46e756fSKirill A. Shutemov { 1321b46e756fSKirill A. Shutemov struct mm_struct *mm = mm_slot->mm; 1322b46e756fSKirill A. Shutemov 132335f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1324b46e756fSKirill A. Shutemov 1325*7d2c4385SZach O'Keefe if (hpage_collapse_test_exit(mm)) { 1326b46e756fSKirill A. Shutemov /* free mm_slot */ 1327b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 1328b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 1329b46e756fSKirill A. Shutemov 1330b46e756fSKirill A. Shutemov /* 1331b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1332b46e756fSKirill A. Shutemov * 1333b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1334b46e756fSKirill A. Shutemov */ 1335b46e756fSKirill A. Shutemov 1336b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1337b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 1338b46e756fSKirill A. Shutemov mmdrop(mm); 1339b46e756fSKirill A. Shutemov } 1340b46e756fSKirill A. Shutemov } 1341b46e756fSKirill A. Shutemov 1342396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 134327e1f827SSong Liu /* 134427e1f827SSong Liu * Notify khugepaged that given addr of the mm is pte-mapped THP. Then 134527e1f827SSong Liu * khugepaged should try to collapse the page table. 134627e1f827SSong Liu */ 1347081c3256SMiaohe Lin static void khugepaged_add_pte_mapped_thp(struct mm_struct *mm, 134827e1f827SSong Liu unsigned long addr) 134927e1f827SSong Liu { 135027e1f827SSong Liu struct mm_slot *mm_slot; 135127e1f827SSong Liu 135227e1f827SSong Liu VM_BUG_ON(addr & ~HPAGE_PMD_MASK); 135327e1f827SSong Liu 135427e1f827SSong Liu spin_lock(&khugepaged_mm_lock); 135527e1f827SSong Liu mm_slot = get_mm_slot(mm); 135627e1f827SSong Liu if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) 135727e1f827SSong Liu mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; 135827e1f827SSong Liu spin_unlock(&khugepaged_mm_lock); 135927e1f827SSong Liu } 136027e1f827SSong Liu 1361e59a47b8SPasha Tatashin static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 1362e59a47b8SPasha Tatashin unsigned long addr, pmd_t *pmdp) 1363e59a47b8SPasha Tatashin { 1364e59a47b8SPasha Tatashin spinlock_t *ptl; 1365e59a47b8SPasha Tatashin pmd_t pmd; 1366e59a47b8SPasha Tatashin 136780110bbfSPasha Tatashin mmap_assert_write_locked(mm); 1368e59a47b8SPasha Tatashin ptl = pmd_lock(vma->vm_mm, pmdp); 1369e59a47b8SPasha Tatashin pmd = pmdp_collapse_flush(vma, addr, pmdp); 1370e59a47b8SPasha Tatashin spin_unlock(ptl); 1371e59a47b8SPasha Tatashin mm_dec_nr_ptes(mm); 137280110bbfSPasha Tatashin page_table_check_pte_clear_range(mm, addr, pmd); 1373e59a47b8SPasha Tatashin pte_free(mm, pmd_pgtable(pmd)); 1374e59a47b8SPasha Tatashin } 1375e59a47b8SPasha Tatashin 137627e1f827SSong Liu /** 1377336e6b53SAlex Shi * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at 1378336e6b53SAlex Shi * address haddr. 1379336e6b53SAlex Shi * 1380336e6b53SAlex Shi * @mm: process address space where collapse happens 1381336e6b53SAlex Shi * @addr: THP collapse address 138227e1f827SSong Liu * 138327e1f827SSong Liu * This function checks whether all the PTEs in the PMD are pointing to the 138427e1f827SSong Liu * right THP. If so, retract the page table so the THP can refault in with 138527e1f827SSong Liu * as pmd-mapped. 138627e1f827SSong Liu */ 138727e1f827SSong Liu void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr) 138827e1f827SSong Liu { 138927e1f827SSong Liu unsigned long haddr = addr & HPAGE_PMD_MASK; 139027e1f827SSong Liu struct vm_area_struct *vma = find_vma(mm, haddr); 1391119a5fc1SHugh Dickins struct page *hpage; 139227e1f827SSong Liu pte_t *start_pte, *pte; 1393e59a47b8SPasha Tatashin pmd_t *pmd; 139427e1f827SSong Liu spinlock_t *ptl; 139527e1f827SSong Liu int count = 0; 139627e1f827SSong Liu int i; 139727e1f827SSong Liu 139827e1f827SSong Liu if (!vma || !vma->vm_file || 1399fef792a4SMiaohe Lin !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) 140027e1f827SSong Liu return; 140127e1f827SSong Liu 140227e1f827SSong Liu /* 1403a7f4e6e4SZach O'Keefe * If we are here, we've succeeded in replacing all the native pages 1404a7f4e6e4SZach O'Keefe * in the page cache with a single hugepage. If a mm were to fault-in 1405a7f4e6e4SZach O'Keefe * this memory (mapped by a suitably aligned VMA), we'd get the hugepage 1406a7f4e6e4SZach O'Keefe * and map it by a PMD, regardless of sysfs THP settings. As such, let's 1407a7f4e6e4SZach O'Keefe * analogously elide sysfs THP settings here. 140827e1f827SSong Liu */ 1409a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) 141027e1f827SSong Liu return; 141127e1f827SSong Liu 1412deb4c93aSPeter Xu /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ 1413deb4c93aSPeter Xu if (userfaultfd_wp(vma)) 1414deb4c93aSPeter Xu return; 1415deb4c93aSPeter Xu 1416119a5fc1SHugh Dickins hpage = find_lock_page(vma->vm_file->f_mapping, 1417119a5fc1SHugh Dickins linear_page_index(vma, haddr)); 1418119a5fc1SHugh Dickins if (!hpage) 1419119a5fc1SHugh Dickins return; 1420119a5fc1SHugh Dickins 1421119a5fc1SHugh Dickins if (!PageHead(hpage)) 1422119a5fc1SHugh Dickins goto drop_hpage; 1423119a5fc1SHugh Dickins 142450722804SZach O'Keefe if (find_pmd_or_thp_or_none(mm, haddr, &pmd) != SCAN_SUCCEED) 1425119a5fc1SHugh Dickins goto drop_hpage; 142627e1f827SSong Liu 142727e1f827SSong Liu start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 142827e1f827SSong Liu 142927e1f827SSong Liu /* step 1: check all mapped PTEs are to the right huge page */ 143027e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 143127e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 143227e1f827SSong Liu struct page *page; 143327e1f827SSong Liu 143427e1f827SSong Liu /* empty pte, skip */ 143527e1f827SSong Liu if (pte_none(*pte)) 143627e1f827SSong Liu continue; 143727e1f827SSong Liu 143827e1f827SSong Liu /* page swapped out, abort */ 143927e1f827SSong Liu if (!pte_present(*pte)) 144027e1f827SSong Liu goto abort; 144127e1f827SSong Liu 144227e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 14433218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 14443218f871SAlex Sierra page = NULL; 144527e1f827SSong Liu /* 1446119a5fc1SHugh Dickins * Note that uprobe, debugger, or MAP_PRIVATE may change the 1447119a5fc1SHugh Dickins * page table, but the new page will not be a subpage of hpage. 144827e1f827SSong Liu */ 1449119a5fc1SHugh Dickins if (hpage + i != page) 145027e1f827SSong Liu goto abort; 145127e1f827SSong Liu count++; 145227e1f827SSong Liu } 145327e1f827SSong Liu 145427e1f827SSong Liu /* step 2: adjust rmap */ 145527e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 145627e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 145727e1f827SSong Liu struct page *page; 145827e1f827SSong Liu 145927e1f827SSong Liu if (pte_none(*pte)) 146027e1f827SSong Liu continue; 146127e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 14623218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 14633218f871SAlex Sierra goto abort; 1464cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 146527e1f827SSong Liu } 146627e1f827SSong Liu 146727e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 146827e1f827SSong Liu 146927e1f827SSong Liu /* step 3: set proper refcount and mm_counters. */ 1470119a5fc1SHugh Dickins if (count) { 147127e1f827SSong Liu page_ref_sub(hpage, count); 147227e1f827SSong Liu add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); 147327e1f827SSong Liu } 147427e1f827SSong Liu 147527e1f827SSong Liu /* step 4: collapse pmd */ 1476e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, haddr, pmd); 1477119a5fc1SHugh Dickins drop_hpage: 1478119a5fc1SHugh Dickins unlock_page(hpage); 1479119a5fc1SHugh Dickins put_page(hpage); 148027e1f827SSong Liu return; 148127e1f827SSong Liu 148227e1f827SSong Liu abort: 148327e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 1484119a5fc1SHugh Dickins goto drop_hpage; 148527e1f827SSong Liu } 148627e1f827SSong Liu 14870edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 148827e1f827SSong Liu { 148927e1f827SSong Liu struct mm_struct *mm = mm_slot->mm; 149027e1f827SSong Liu int i; 149127e1f827SSong Liu 149227e1f827SSong Liu if (likely(mm_slot->nr_pte_mapped_thp == 0)) 14930edf61e5SMiaohe Lin return; 149427e1f827SSong Liu 1495d8ed45c5SMichel Lespinasse if (!mmap_write_trylock(mm)) 14960edf61e5SMiaohe Lin return; 149727e1f827SSong Liu 1498*7d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 149927e1f827SSong Liu goto out; 150027e1f827SSong Liu 150127e1f827SSong Liu for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) 150227e1f827SSong Liu collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]); 150327e1f827SSong Liu 150427e1f827SSong Liu out: 150527e1f827SSong Liu mm_slot->nr_pte_mapped_thp = 0; 1506d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 150727e1f827SSong Liu } 150827e1f827SSong Liu 1509f3f0e1d2SKirill A. Shutemov static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1510f3f0e1d2SKirill A. Shutemov { 1511f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 151218e77600SHugh Dickins struct mm_struct *mm; 1513f3f0e1d2SKirill A. Shutemov unsigned long addr; 1514e59a47b8SPasha Tatashin pmd_t *pmd; 1515f3f0e1d2SKirill A. Shutemov 1516f3f0e1d2SKirill A. Shutemov i_mmap_lock_write(mapping); 1517f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 151827e1f827SSong Liu /* 151927e1f827SSong Liu * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 152027e1f827SSong Liu * got written to. These VMAs are likely not worth investing 15213e4e28c5SMichel Lespinasse * mmap_write_lock(mm) as PMD-mapping is likely to be split 152227e1f827SSong Liu * later. 152327e1f827SSong Liu * 152436ee2c78SMiaohe Lin * Note that vma->anon_vma check is racy: it can be set up after 1525c1e8d7c6SMichel Lespinasse * the check but before we took mmap_lock by the fault path. 152627e1f827SSong Liu * But page lock would prevent establishing any new ptes of the 152727e1f827SSong Liu * page, so we are safe. 152827e1f827SSong Liu * 152927e1f827SSong Liu * An alternative would be drop the check, but check that page 153027e1f827SSong Liu * table is clear before calling pmdp_collapse_flush() under 153127e1f827SSong Liu * ptl. It has higher chance to recover THP for the VMA, but 153227e1f827SSong Liu * has higher cost too. 153327e1f827SSong Liu */ 1534f3f0e1d2SKirill A. Shutemov if (vma->anon_vma) 1535f3f0e1d2SKirill A. Shutemov continue; 1536f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 1537f3f0e1d2SKirill A. Shutemov if (addr & ~HPAGE_PMD_MASK) 1538f3f0e1d2SKirill A. Shutemov continue; 1539f3f0e1d2SKirill A. Shutemov if (vma->vm_end < addr + HPAGE_PMD_SIZE) 1540f3f0e1d2SKirill A. Shutemov continue; 154118e77600SHugh Dickins mm = vma->vm_mm; 154250722804SZach O'Keefe if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED) 1543f3f0e1d2SKirill A. Shutemov continue; 1544f3f0e1d2SKirill A. Shutemov /* 1545c1e8d7c6SMichel Lespinasse * We need exclusive mmap_lock to retract page table. 154627e1f827SSong Liu * 154727e1f827SSong Liu * We use trylock due to lock inversion: we need to acquire 1548c1e8d7c6SMichel Lespinasse * mmap_lock while holding page lock. Fault path does it in 154927e1f827SSong Liu * reverse order. Trylock is a way to avoid deadlock. 1550f3f0e1d2SKirill A. Shutemov */ 155118e77600SHugh Dickins if (mmap_write_trylock(mm)) { 1552deb4c93aSPeter Xu /* 1553deb4c93aSPeter Xu * When a vma is registered with uffd-wp, we can't 1554deb4c93aSPeter Xu * recycle the pmd pgtable because there can be pte 1555deb4c93aSPeter Xu * markers installed. Skip it only, so the rest mm/vma 1556deb4c93aSPeter Xu * can still have the same file mapped hugely, however 1557deb4c93aSPeter Xu * it'll always mapped in small page size for uffd-wp 1558deb4c93aSPeter Xu * registered ranges. 1559deb4c93aSPeter Xu */ 1560*7d2c4385SZach O'Keefe if (!hpage_collapse_test_exit(mm) && 1561*7d2c4385SZach O'Keefe !userfaultfd_wp(vma)) 1562e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, addr, pmd); 156318e77600SHugh Dickins mmap_write_unlock(mm); 156427e1f827SSong Liu } else { 156527e1f827SSong Liu /* Try again later */ 156618e77600SHugh Dickins khugepaged_add_pte_mapped_thp(mm, addr); 1567f3f0e1d2SKirill A. Shutemov } 1568f3f0e1d2SKirill A. Shutemov } 1569f3f0e1d2SKirill A. Shutemov i_mmap_unlock_write(mapping); 1570f3f0e1d2SKirill A. Shutemov } 1571f3f0e1d2SKirill A. Shutemov 1572f3f0e1d2SKirill A. Shutemov /** 157399cb0dbdSSong Liu * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1574f3f0e1d2SKirill A. Shutemov * 1575336e6b53SAlex Shi * @mm: process address space where collapse happens 1576336e6b53SAlex Shi * @file: file that collapse on 1577336e6b53SAlex Shi * @start: collapse start address 15789710a78aSZach O'Keefe * @cc: collapse context and scratchpad 1579336e6b53SAlex Shi * 1580f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 158187c460a0SHugh Dickins * - allocate and lock a new huge page; 158277da9389SMatthew Wilcox * - scan page cache replacing old pages with the new one 158399cb0dbdSSong Liu * + swap/gup in pages if necessary; 1584f3f0e1d2SKirill A. Shutemov * + fill in gaps; 158577da9389SMatthew Wilcox * + keep old pages around in case rollback is required; 158677da9389SMatthew Wilcox * - if replacing succeeds: 1587f3f0e1d2SKirill A. Shutemov * + copy data over; 1588f3f0e1d2SKirill A. Shutemov * + free old pages; 158987c460a0SHugh Dickins * + unlock huge page; 1590f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1591f3f0e1d2SKirill A. Shutemov * + put all pages back and unfreeze them; 159277da9389SMatthew Wilcox * + restore gaps in the page cache; 159387c460a0SHugh Dickins * + unlock and free huge page; 1594f3f0e1d2SKirill A. Shutemov */ 159550ad2f24SZach O'Keefe static int collapse_file(struct mm_struct *mm, struct file *file, 159650ad2f24SZach O'Keefe pgoff_t start, struct collapse_control *cc) 1597f3f0e1d2SKirill A. Shutemov { 1598579c571eSSong Liu struct address_space *mapping = file->f_mapping; 159950ad2f24SZach O'Keefe struct page *hpage; 1600f3f0e1d2SKirill A. Shutemov pgoff_t index, end = start + HPAGE_PMD_NR; 1601f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 160277da9389SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1603f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 160499cb0dbdSSong Liu bool is_shmem = shmem_file(file); 1605bf9eceadSMuchun Song int nr; 1606f3f0e1d2SKirill A. Shutemov 160799cb0dbdSSong Liu VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1608f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1609f3f0e1d2SKirill A. Shutemov 161050ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 16119710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1612f3f0e1d2SKirill A. Shutemov goto out; 1613f3f0e1d2SKirill A. Shutemov 16146b24ca4aSMatthew Wilcox (Oracle) /* 16156b24ca4aSMatthew Wilcox (Oracle) * Ensure we have slots for all the pages in the range. This is 16166b24ca4aSMatthew Wilcox (Oracle) * almost certainly a no-op because most of the pages must be present 16176b24ca4aSMatthew Wilcox (Oracle) */ 161895feeabbSHugh Dickins do { 161995feeabbSHugh Dickins xas_lock_irq(&xas); 162095feeabbSHugh Dickins xas_create_range(&xas); 162195feeabbSHugh Dickins if (!xas_error(&xas)) 162295feeabbSHugh Dickins break; 162395feeabbSHugh Dickins xas_unlock_irq(&xas); 162495feeabbSHugh Dickins if (!xas_nomem(&xas, GFP_KERNEL)) { 162595feeabbSHugh Dickins result = SCAN_FAIL; 162695feeabbSHugh Dickins goto out; 162795feeabbSHugh Dickins } 162895feeabbSHugh Dickins } while (1); 162995feeabbSHugh Dickins 163050ad2f24SZach O'Keefe __SetPageLocked(hpage); 163199cb0dbdSSong Liu if (is_shmem) 163250ad2f24SZach O'Keefe __SetPageSwapBacked(hpage); 163350ad2f24SZach O'Keefe hpage->index = start; 163450ad2f24SZach O'Keefe hpage->mapping = mapping; 1635f3f0e1d2SKirill A. Shutemov 1636f3f0e1d2SKirill A. Shutemov /* 163750ad2f24SZach O'Keefe * At this point the hpage is locked and not up-to-date. 163887c460a0SHugh Dickins * It's safe to insert it into the page cache, because nobody would 163987c460a0SHugh Dickins * be able to map it or use it in another way until we unlock it. 1640f3f0e1d2SKirill A. Shutemov */ 1641f3f0e1d2SKirill A. Shutemov 164277da9389SMatthew Wilcox xas_set(&xas, start); 164377da9389SMatthew Wilcox for (index = start; index < end; index++) { 164477da9389SMatthew Wilcox struct page *page = xas_next(&xas); 164577da9389SMatthew Wilcox 164677da9389SMatthew Wilcox VM_BUG_ON(index != xas.xa_index); 164799cb0dbdSSong Liu if (is_shmem) { 164877da9389SMatthew Wilcox if (!page) { 1649701270faSHugh Dickins /* 165099cb0dbdSSong Liu * Stop if extent has been truncated or 165199cb0dbdSSong Liu * hole-punched, and is now completely 165299cb0dbdSSong Liu * empty. 1653701270faSHugh Dickins */ 1654701270faSHugh Dickins if (index == start) { 1655701270faSHugh Dickins if (!xas_next_entry(&xas, end - 1)) { 1656701270faSHugh Dickins result = SCAN_TRUNCATED; 1657042a3082SHugh Dickins goto xa_locked; 1658701270faSHugh Dickins } 1659701270faSHugh Dickins xas_set(&xas, index); 1660701270faSHugh Dickins } 166177da9389SMatthew Wilcox if (!shmem_charge(mapping->host, 1)) { 1662f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1663042a3082SHugh Dickins goto xa_locked; 1664f3f0e1d2SKirill A. Shutemov } 166550ad2f24SZach O'Keefe xas_store(&xas, hpage); 166677da9389SMatthew Wilcox nr_none++; 166777da9389SMatthew Wilcox continue; 1668f3f0e1d2SKirill A. Shutemov } 1669f3f0e1d2SKirill A. Shutemov 16703159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 167177da9389SMatthew Wilcox xas_unlock_irq(&xas); 1672f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 1673f3f0e1d2SKirill A. Shutemov if (shmem_getpage(mapping->host, index, &page, 1674acdd9f8eSHugh Dickins SGP_NOALLOC)) { 1675f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 167677da9389SMatthew Wilcox goto xa_unlocked; 1677f3f0e1d2SKirill A. Shutemov } 1678f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1679f3f0e1d2SKirill A. Shutemov get_page(page); 1680042a3082SHugh Dickins xas_unlock_irq(&xas); 1681f3f0e1d2SKirill A. Shutemov } else { 1682f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1683042a3082SHugh Dickins goto xa_locked; 1684f3f0e1d2SKirill A. Shutemov } 168599cb0dbdSSong Liu } else { /* !is_shmem */ 168699cb0dbdSSong Liu if (!page || xa_is_value(page)) { 168799cb0dbdSSong Liu xas_unlock_irq(&xas); 168899cb0dbdSSong Liu page_cache_sync_readahead(mapping, &file->f_ra, 168999cb0dbdSSong Liu file, index, 1690e5a59d30SDavid Howells end - index); 169199cb0dbdSSong Liu /* drain pagevecs to help isolate_lru_page() */ 169299cb0dbdSSong Liu lru_add_drain(); 169399cb0dbdSSong Liu page = find_lock_page(mapping, index); 169499cb0dbdSSong Liu if (unlikely(page == NULL)) { 169599cb0dbdSSong Liu result = SCAN_FAIL; 169699cb0dbdSSong Liu goto xa_unlocked; 169799cb0dbdSSong Liu } 169875f36069SSong Liu } else if (PageDirty(page)) { 169975f36069SSong Liu /* 170075f36069SSong Liu * khugepaged only works on read-only fd, 170175f36069SSong Liu * so this page is dirty because it hasn't 170275f36069SSong Liu * been flushed since first write. There 170375f36069SSong Liu * won't be new dirty pages. 170475f36069SSong Liu * 170575f36069SSong Liu * Trigger async flush here and hope the 170675f36069SSong Liu * writeback is done when khugepaged 170775f36069SSong Liu * revisits this page. 170875f36069SSong Liu * 170975f36069SSong Liu * This is a one-off situation. We are not 171075f36069SSong Liu * forcing writeback in loop. 171175f36069SSong Liu */ 171275f36069SSong Liu xas_unlock_irq(&xas); 171375f36069SSong Liu filemap_flush(mapping); 171475f36069SSong Liu result = SCAN_FAIL; 171575f36069SSong Liu goto xa_unlocked; 171674c42e1bSRongwei Wang } else if (PageWriteback(page)) { 171774c42e1bSRongwei Wang xas_unlock_irq(&xas); 171874c42e1bSRongwei Wang result = SCAN_FAIL; 171974c42e1bSRongwei Wang goto xa_unlocked; 172099cb0dbdSSong Liu } else if (trylock_page(page)) { 172199cb0dbdSSong Liu get_page(page); 172299cb0dbdSSong Liu xas_unlock_irq(&xas); 172399cb0dbdSSong Liu } else { 172499cb0dbdSSong Liu result = SCAN_PAGE_LOCK; 172599cb0dbdSSong Liu goto xa_locked; 172699cb0dbdSSong Liu } 172799cb0dbdSSong Liu } 1728f3f0e1d2SKirill A. Shutemov 1729f3f0e1d2SKirill A. Shutemov /* 1730b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1731f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1732f3f0e1d2SKirill A. Shutemov */ 1733f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 17344655e5e5SSong Liu 17354655e5e5SSong Liu /* make sure the page is up to date */ 17364655e5e5SSong Liu if (unlikely(!PageUptodate(page))) { 17374655e5e5SSong Liu result = SCAN_FAIL; 17384655e5e5SSong Liu goto out_unlock; 17394655e5e5SSong Liu } 174006a5e126SHugh Dickins 174106a5e126SHugh Dickins /* 174206a5e126SHugh Dickins * If file was truncated then extended, or hole-punched, before 174306a5e126SHugh Dickins * we locked the first page, then a THP might be there already. 174406a5e126SHugh Dickins */ 174506a5e126SHugh Dickins if (PageTransCompound(page)) { 174606a5e126SHugh Dickins result = SCAN_PAGE_COMPOUND; 174706a5e126SHugh Dickins goto out_unlock; 174806a5e126SHugh Dickins } 1749f3f0e1d2SKirill A. Shutemov 1750f3f0e1d2SKirill A. Shutemov if (page_mapping(page) != mapping) { 1751f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1752f3f0e1d2SKirill A. Shutemov goto out_unlock; 1753f3f0e1d2SKirill A. Shutemov } 1754f3f0e1d2SKirill A. Shutemov 175574c42e1bSRongwei Wang if (!is_shmem && (PageDirty(page) || 175674c42e1bSRongwei Wang PageWriteback(page))) { 17574655e5e5SSong Liu /* 17584655e5e5SSong Liu * khugepaged only works on read-only fd, so this 17594655e5e5SSong Liu * page is dirty because it hasn't been flushed 17604655e5e5SSong Liu * since first write. 17614655e5e5SSong Liu */ 17624655e5e5SSong Liu result = SCAN_FAIL; 17634655e5e5SSong Liu goto out_unlock; 17644655e5e5SSong Liu } 17654655e5e5SSong Liu 1766f3f0e1d2SKirill A. Shutemov if (isolate_lru_page(page)) { 1767f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1768042a3082SHugh Dickins goto out_unlock; 1769f3f0e1d2SKirill A. Shutemov } 1770f3f0e1d2SKirill A. Shutemov 177199cb0dbdSSong Liu if (page_has_private(page) && 177299cb0dbdSSong Liu !try_to_release_page(page, GFP_KERNEL)) { 177399cb0dbdSSong Liu result = SCAN_PAGE_HAS_PRIVATE; 17742f33a706SHugh Dickins putback_lru_page(page); 177599cb0dbdSSong Liu goto out_unlock; 177699cb0dbdSSong Liu } 177799cb0dbdSSong Liu 1778f3f0e1d2SKirill A. Shutemov if (page_mapped(page)) 1779869f7ee6SMatthew Wilcox (Oracle) try_to_unmap(page_folio(page), 1780869f7ee6SMatthew Wilcox (Oracle) TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); 1781f3f0e1d2SKirill A. Shutemov 178277da9389SMatthew Wilcox xas_lock_irq(&xas); 178377da9389SMatthew Wilcox xas_set(&xas, index); 1784f3f0e1d2SKirill A. Shutemov 178577da9389SMatthew Wilcox VM_BUG_ON_PAGE(page != xas_load(&xas), page); 1786f3f0e1d2SKirill A. Shutemov 1787f3f0e1d2SKirill A. Shutemov /* 1788f3f0e1d2SKirill A. Shutemov * The page is expected to have page_count() == 3: 1789f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 179077da9389SMatthew Wilcox * - one reference from page cache; 1791f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1792f3f0e1d2SKirill A. Shutemov */ 1793f3f0e1d2SKirill A. Shutemov if (!page_ref_freeze(page, 3)) { 1794f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1795042a3082SHugh Dickins xas_unlock_irq(&xas); 1796042a3082SHugh Dickins putback_lru_page(page); 1797042a3082SHugh Dickins goto out_unlock; 1798f3f0e1d2SKirill A. Shutemov } 1799f3f0e1d2SKirill A. Shutemov 1800f3f0e1d2SKirill A. Shutemov /* 1801f3f0e1d2SKirill A. Shutemov * Add the page to the list to be able to undo the collapse if 1802f3f0e1d2SKirill A. Shutemov * something go wrong. 1803f3f0e1d2SKirill A. Shutemov */ 1804f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 1805f3f0e1d2SKirill A. Shutemov 1806f3f0e1d2SKirill A. Shutemov /* Finally, replace with the new page. */ 180750ad2f24SZach O'Keefe xas_store(&xas, hpage); 1808f3f0e1d2SKirill A. Shutemov continue; 1809f3f0e1d2SKirill A. Shutemov out_unlock: 1810f3f0e1d2SKirill A. Shutemov unlock_page(page); 1811f3f0e1d2SKirill A. Shutemov put_page(page); 1812042a3082SHugh Dickins goto xa_unlocked; 1813f3f0e1d2SKirill A. Shutemov } 181450ad2f24SZach O'Keefe nr = thp_nr_pages(hpage); 1815f3f0e1d2SKirill A. Shutemov 181699cb0dbdSSong Liu if (is_shmem) 181750ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr); 181809d91cdaSSong Liu else { 181950ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr); 182009d91cdaSSong Liu filemap_nr_thps_inc(mapping); 1821eb6ecbedSCollin Fijalkovich /* 1822eb6ecbedSCollin Fijalkovich * Paired with smp_mb() in do_dentry_open() to ensure 1823eb6ecbedSCollin Fijalkovich * i_writecount is up to date and the update to nr_thps is 1824eb6ecbedSCollin Fijalkovich * visible. Ensures the page cache will be truncated if the 1825eb6ecbedSCollin Fijalkovich * file is opened writable. 1826eb6ecbedSCollin Fijalkovich */ 1827eb6ecbedSCollin Fijalkovich smp_mb(); 1828eb6ecbedSCollin Fijalkovich if (inode_is_open_for_write(mapping->host)) { 1829eb6ecbedSCollin Fijalkovich result = SCAN_FAIL; 183050ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr); 1831eb6ecbedSCollin Fijalkovich filemap_nr_thps_dec(mapping); 1832eb6ecbedSCollin Fijalkovich goto xa_locked; 1833eb6ecbedSCollin Fijalkovich } 183409d91cdaSSong Liu } 183599cb0dbdSSong Liu 1836042a3082SHugh Dickins if (nr_none) { 183750ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none); 18382f55f070SMiaohe Lin /* nr_none is always 0 for non-shmem. */ 183950ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none); 1840042a3082SHugh Dickins } 1841042a3082SHugh Dickins 18426b24ca4aSMatthew Wilcox (Oracle) /* Join all the small entries into a single multi-index entry */ 18436b24ca4aSMatthew Wilcox (Oracle) xas_set_order(&xas, start, HPAGE_PMD_ORDER); 184450ad2f24SZach O'Keefe xas_store(&xas, hpage); 1845042a3082SHugh Dickins xa_locked: 1846042a3082SHugh Dickins xas_unlock_irq(&xas); 184777da9389SMatthew Wilcox xa_unlocked: 1848042a3082SHugh Dickins 18496d9df8a5SHugh Dickins /* 18506d9df8a5SHugh Dickins * If collapse is successful, flush must be done now before copying. 18516d9df8a5SHugh Dickins * If collapse is unsuccessful, does flush actually need to be done? 18526d9df8a5SHugh Dickins * Do it anyway, to clear the state. 18536d9df8a5SHugh Dickins */ 18546d9df8a5SHugh Dickins try_to_unmap_flush(); 18556d9df8a5SHugh Dickins 1856f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 185777da9389SMatthew Wilcox struct page *page, *tmp; 1858f3f0e1d2SKirill A. Shutemov 1859f3f0e1d2SKirill A. Shutemov /* 186077da9389SMatthew Wilcox * Replacing old pages with new one has succeeded, now we 186177da9389SMatthew Wilcox * need to copy the content and free the old pages. 1862f3f0e1d2SKirill A. Shutemov */ 18632af8ff29SHugh Dickins index = start; 1864f3f0e1d2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, &pagelist, lru) { 18652af8ff29SHugh Dickins while (index < page->index) { 186650ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 18672af8ff29SHugh Dickins index++; 18682af8ff29SHugh Dickins } 186950ad2f24SZach O'Keefe copy_highpage(hpage + (page->index % HPAGE_PMD_NR), 1870f3f0e1d2SKirill A. Shutemov page); 1871f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1872f3f0e1d2SKirill A. Shutemov page->mapping = NULL; 1873042a3082SHugh Dickins page_ref_unfreeze(page, 1); 1874f3f0e1d2SKirill A. Shutemov ClearPageActive(page); 1875f3f0e1d2SKirill A. Shutemov ClearPageUnevictable(page); 1876042a3082SHugh Dickins unlock_page(page); 1877f3f0e1d2SKirill A. Shutemov put_page(page); 18782af8ff29SHugh Dickins index++; 18792af8ff29SHugh Dickins } 18802af8ff29SHugh Dickins while (index < end) { 188150ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 18822af8ff29SHugh Dickins index++; 1883f3f0e1d2SKirill A. Shutemov } 1884f3f0e1d2SKirill A. Shutemov 188550ad2f24SZach O'Keefe SetPageUptodate(hpage); 188650ad2f24SZach O'Keefe page_ref_add(hpage, HPAGE_PMD_NR - 1); 18876058eaecSJohannes Weiner if (is_shmem) 188850ad2f24SZach O'Keefe set_page_dirty(hpage); 188950ad2f24SZach O'Keefe lru_cache_add(hpage); 1890f3f0e1d2SKirill A. Shutemov 1891042a3082SHugh Dickins /* 1892042a3082SHugh Dickins * Remove pte page tables, so we can re-fault the page as huge. 1893042a3082SHugh Dickins */ 1894042a3082SHugh Dickins retract_page_tables(mapping, start); 189550ad2f24SZach O'Keefe unlock_page(hpage); 189650ad2f24SZach O'Keefe hpage = NULL; 1897f3f0e1d2SKirill A. Shutemov } else { 189877da9389SMatthew Wilcox struct page *page; 1899aaa52e34SHugh Dickins 190077da9389SMatthew Wilcox /* Something went wrong: roll back page cache changes */ 190177da9389SMatthew Wilcox xas_lock_irq(&xas); 19022f55f070SMiaohe Lin if (nr_none) { 1903aaa52e34SHugh Dickins mapping->nrpages -= nr_none; 1904aaa52e34SHugh Dickins shmem_uncharge(mapping->host, nr_none); 19052f55f070SMiaohe Lin } 1906aaa52e34SHugh Dickins 190777da9389SMatthew Wilcox xas_set(&xas, start); 190877da9389SMatthew Wilcox xas_for_each(&xas, page, end - 1) { 1909f3f0e1d2SKirill A. Shutemov page = list_first_entry_or_null(&pagelist, 1910f3f0e1d2SKirill A. Shutemov struct page, lru); 191177da9389SMatthew Wilcox if (!page || xas.xa_index < page->index) { 1912f3f0e1d2SKirill A. Shutemov if (!nr_none) 1913f3f0e1d2SKirill A. Shutemov break; 1914f3f0e1d2SKirill A. Shutemov nr_none--; 191559749e6cSJohannes Weiner /* Put holes back where they were */ 191677da9389SMatthew Wilcox xas_store(&xas, NULL); 1917f3f0e1d2SKirill A. Shutemov continue; 1918f3f0e1d2SKirill A. Shutemov } 1919f3f0e1d2SKirill A. Shutemov 192077da9389SMatthew Wilcox VM_BUG_ON_PAGE(page->index != xas.xa_index, page); 1921f3f0e1d2SKirill A. Shutemov 1922f3f0e1d2SKirill A. Shutemov /* Unfreeze the page. */ 1923f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1924f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(page, 2); 192577da9389SMatthew Wilcox xas_store(&xas, page); 192677da9389SMatthew Wilcox xas_pause(&xas); 192777da9389SMatthew Wilcox xas_unlock_irq(&xas); 1928f3f0e1d2SKirill A. Shutemov unlock_page(page); 1929042a3082SHugh Dickins putback_lru_page(page); 193077da9389SMatthew Wilcox xas_lock_irq(&xas); 1931f3f0e1d2SKirill A. Shutemov } 1932f3f0e1d2SKirill A. Shutemov VM_BUG_ON(nr_none); 193377da9389SMatthew Wilcox xas_unlock_irq(&xas); 1934f3f0e1d2SKirill A. Shutemov 193550ad2f24SZach O'Keefe hpage->mapping = NULL; 1936f3f0e1d2SKirill A. Shutemov } 1937042a3082SHugh Dickins 193850ad2f24SZach O'Keefe if (hpage) 193950ad2f24SZach O'Keefe unlock_page(hpage); 1940f3f0e1d2SKirill A. Shutemov out: 1941f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 194250ad2f24SZach O'Keefe if (hpage) { 194350ad2f24SZach O'Keefe mem_cgroup_uncharge(page_folio(hpage)); 194450ad2f24SZach O'Keefe put_page(hpage); 1945c6a7f445SYang Shi } 1946f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 194750ad2f24SZach O'Keefe return result; 1948f3f0e1d2SKirill A. Shutemov } 1949f3f0e1d2SKirill A. Shutemov 195050ad2f24SZach O'Keefe static int khugepaged_scan_file(struct mm_struct *mm, struct file *file, 195150ad2f24SZach O'Keefe pgoff_t start, struct collapse_control *cc) 1952f3f0e1d2SKirill A. Shutemov { 1953f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 1954579c571eSSong Liu struct address_space *mapping = file->f_mapping; 195585b392dbSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, start); 1956f3f0e1d2SKirill A. Shutemov int present, swap; 1957f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 1958f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 1959f3f0e1d2SKirill A. Shutemov 1960f3f0e1d2SKirill A. Shutemov present = 0; 1961f3f0e1d2SKirill A. Shutemov swap = 0; 196234d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1963f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 196485b392dbSMatthew Wilcox xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 196585b392dbSMatthew Wilcox if (xas_retry(&xas, page)) 1966f3f0e1d2SKirill A. Shutemov continue; 1967f3f0e1d2SKirill A. Shutemov 196885b392dbSMatthew Wilcox if (xa_is_value(page)) { 1969d8ea7cc8SZach O'Keefe ++swap; 1970d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 1971d8ea7cc8SZach O'Keefe swap > khugepaged_max_ptes_swap) { 1972f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1973e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1974f3f0e1d2SKirill A. Shutemov break; 1975f3f0e1d2SKirill A. Shutemov } 1976f3f0e1d2SKirill A. Shutemov continue; 1977f3f0e1d2SKirill A. Shutemov } 1978f3f0e1d2SKirill A. Shutemov 19796b24ca4aSMatthew Wilcox (Oracle) /* 19806b24ca4aSMatthew Wilcox (Oracle) * XXX: khugepaged should compact smaller compound pages 19816b24ca4aSMatthew Wilcox (Oracle) * into a PMD sized page 19826b24ca4aSMatthew Wilcox (Oracle) */ 1983f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 1984f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COMPOUND; 1985f3f0e1d2SKirill A. Shutemov break; 1986f3f0e1d2SKirill A. Shutemov } 1987f3f0e1d2SKirill A. Shutemov 1988f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 1989*7d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 1990f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 1991f3f0e1d2SKirill A. Shutemov break; 1992f3f0e1d2SKirill A. Shutemov } 199334d6b470SZach O'Keefe cc->node_load[node]++; 1994f3f0e1d2SKirill A. Shutemov 1995f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 1996f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 1997f3f0e1d2SKirill A. Shutemov break; 1998f3f0e1d2SKirill A. Shutemov } 1999f3f0e1d2SKirill A. Shutemov 200099cb0dbdSSong Liu if (page_count(page) != 200199cb0dbdSSong Liu 1 + page_mapcount(page) + page_has_private(page)) { 2002f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 2003f3f0e1d2SKirill A. Shutemov break; 2004f3f0e1d2SKirill A. Shutemov } 2005f3f0e1d2SKirill A. Shutemov 2006f3f0e1d2SKirill A. Shutemov /* 2007f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 2008f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 2009f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 2010f3f0e1d2SKirill A. Shutemov */ 2011f3f0e1d2SKirill A. Shutemov 2012f3f0e1d2SKirill A. Shutemov present++; 2013f3f0e1d2SKirill A. Shutemov 2014f3f0e1d2SKirill A. Shutemov if (need_resched()) { 201585b392dbSMatthew Wilcox xas_pause(&xas); 2016f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 2017f3f0e1d2SKirill A. Shutemov } 2018f3f0e1d2SKirill A. Shutemov } 2019f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 2020f3f0e1d2SKirill A. Shutemov 2021f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 2022d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 2023d8ea7cc8SZach O'Keefe present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 2024f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 2025e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 2026f3f0e1d2SKirill A. Shutemov } else { 202750ad2f24SZach O'Keefe result = collapse_file(mm, file, start, cc); 2028f3f0e1d2SKirill A. Shutemov } 2029f3f0e1d2SKirill A. Shutemov } 2030f3f0e1d2SKirill A. Shutemov 2031f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 203250ad2f24SZach O'Keefe return result; 2033f3f0e1d2SKirill A. Shutemov } 2034f3f0e1d2SKirill A. Shutemov #else 203550ad2f24SZach O'Keefe static int khugepaged_scan_file(struct mm_struct *mm, struct file *file, 203650ad2f24SZach O'Keefe pgoff_t start, struct collapse_control *cc) 2037f3f0e1d2SKirill A. Shutemov { 2038f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 2039f3f0e1d2SKirill A. Shutemov } 204027e1f827SSong Liu 20410edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 204227e1f827SSong Liu { 204327e1f827SSong Liu } 2044f3f0e1d2SKirill A. Shutemov #endif 2045f3f0e1d2SKirill A. Shutemov 204650ad2f24SZach O'Keefe static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, 204734d6b470SZach O'Keefe struct collapse_control *cc) 2048b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 2049b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 2050b46e756fSKirill A. Shutemov { 2051b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 2052b46e756fSKirill A. Shutemov struct mm_struct *mm; 2053b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 2054b46e756fSKirill A. Shutemov int progress = 0; 2055b46e756fSKirill A. Shutemov 2056b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 205735f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 205850ad2f24SZach O'Keefe *result = SCAN_FAIL; 2059b46e756fSKirill A. Shutemov 2060b46e756fSKirill A. Shutemov if (khugepaged_scan.mm_slot) 2061b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2062b46e756fSKirill A. Shutemov else { 2063b46e756fSKirill A. Shutemov mm_slot = list_entry(khugepaged_scan.mm_head.next, 2064b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2065b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2066b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 2067b46e756fSKirill A. Shutemov } 2068b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 206927e1f827SSong Liu khugepaged_collapse_pte_mapped_thps(mm_slot); 2070b46e756fSKirill A. Shutemov 2071b46e756fSKirill A. Shutemov mm = mm_slot->mm; 20723b454ad3SYang Shi /* 20733b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 20743b454ad3SYang Shi * the next mm on the list. 20753b454ad3SYang Shi */ 2076b46e756fSKirill A. Shutemov vma = NULL; 2077d8ed45c5SMichel Lespinasse if (unlikely(!mmap_read_trylock(mm))) 2078c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2079*7d2c4385SZach O'Keefe if (likely(!hpage_collapse_test_exit(mm))) 2080b46e756fSKirill A. Shutemov vma = find_vma(mm, khugepaged_scan.address); 2081b46e756fSKirill A. Shutemov 2082b46e756fSKirill A. Shutemov progress++; 2083b46e756fSKirill A. Shutemov for (; vma; vma = vma->vm_next) { 2084b46e756fSKirill A. Shutemov unsigned long hstart, hend; 2085b46e756fSKirill A. Shutemov 2086b46e756fSKirill A. Shutemov cond_resched(); 2087*7d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) { 2088b46e756fSKirill A. Shutemov progress++; 2089b46e756fSKirill A. Shutemov break; 2090b46e756fSKirill A. Shutemov } 2091a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) { 2092b46e756fSKirill A. Shutemov skip: 2093b46e756fSKirill A. Shutemov progress++; 2094b46e756fSKirill A. Shutemov continue; 2095b46e756fSKirill A. Shutemov } 20964fa6893fSYang Shi hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); 20974fa6893fSYang Shi hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); 2098b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 2099b46e756fSKirill A. Shutemov goto skip; 2100b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 2101b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 2102b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2103b46e756fSKirill A. Shutemov 2104b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 210550ad2f24SZach O'Keefe bool mmap_locked = true; 210650ad2f24SZach O'Keefe 2107b46e756fSKirill A. Shutemov cond_resched(); 2108*7d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 2109b46e756fSKirill A. Shutemov goto breakouterloop; 2110b46e756fSKirill A. Shutemov 2111b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 2112b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 2113b46e756fSKirill A. Shutemov hend); 211499cb0dbdSSong Liu if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 2115396bcc52SMatthew Wilcox (Oracle) struct file *file = get_file(vma->vm_file); 2116f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 2117f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 211899cb0dbdSSong Liu 2119d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 212050ad2f24SZach O'Keefe *result = khugepaged_scan_file(mm, file, pgoff, 212134d6b470SZach O'Keefe cc); 212250ad2f24SZach O'Keefe mmap_locked = false; 2123f3f0e1d2SKirill A. Shutemov fput(file); 2124f3f0e1d2SKirill A. Shutemov } else { 2125*7d2c4385SZach O'Keefe *result = hpage_collapse_scan_pmd(mm, vma, 2126b46e756fSKirill A. Shutemov khugepaged_scan.address, 2127*7d2c4385SZach O'Keefe &mmap_locked, 2128*7d2c4385SZach O'Keefe cc); 2129f3f0e1d2SKirill A. Shutemov } 213050ad2f24SZach O'Keefe if (*result == SCAN_SUCCEED) 213150ad2f24SZach O'Keefe ++khugepaged_pages_collapsed; 2132b46e756fSKirill A. Shutemov /* move to next address */ 2133b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 2134b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 213550ad2f24SZach O'Keefe if (!mmap_locked) 213650ad2f24SZach O'Keefe /* 213750ad2f24SZach O'Keefe * We released mmap_lock so break loop. Note 213850ad2f24SZach O'Keefe * that we drop mmap_lock before all hugepage 213950ad2f24SZach O'Keefe * allocations, so if allocation fails, we are 214050ad2f24SZach O'Keefe * guaranteed to break here and report the 214150ad2f24SZach O'Keefe * correct result back to caller. 214250ad2f24SZach O'Keefe */ 2143c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2144b46e756fSKirill A. Shutemov if (progress >= pages) 2145b46e756fSKirill A. Shutemov goto breakouterloop; 2146b46e756fSKirill A. Shutemov } 2147b46e756fSKirill A. Shutemov } 2148b46e756fSKirill A. Shutemov breakouterloop: 2149d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ 2150c1e8d7c6SMichel Lespinasse breakouterloop_mmap_lock: 2151b46e756fSKirill A. Shutemov 2152b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2153b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2154b46e756fSKirill A. Shutemov /* 2155b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 2156b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 2157b46e756fSKirill A. Shutemov */ 2158*7d2c4385SZach O'Keefe if (hpage_collapse_test_exit(mm) || !vma) { 2159b46e756fSKirill A. Shutemov /* 2160b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 2161b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 2162b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 2163b46e756fSKirill A. Shutemov */ 2164b46e756fSKirill A. Shutemov if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2165b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = list_entry( 2166b46e756fSKirill A. Shutemov mm_slot->mm_node.next, 2167b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2168b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2169b46e756fSKirill A. Shutemov } else { 2170b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2171b46e756fSKirill A. Shutemov khugepaged_full_scans++; 2172b46e756fSKirill A. Shutemov } 2173b46e756fSKirill A. Shutemov 2174b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2175b46e756fSKirill A. Shutemov } 2176b46e756fSKirill A. Shutemov 2177b46e756fSKirill A. Shutemov return progress; 2178b46e756fSKirill A. Shutemov } 2179b46e756fSKirill A. Shutemov 2180b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 2181b46e756fSKirill A. Shutemov { 2182b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 21831064026bSYang Shi hugepage_flags_enabled(); 2184b46e756fSKirill A. Shutemov } 2185b46e756fSKirill A. Shutemov 2186b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 2187b46e756fSKirill A. Shutemov { 2188b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 2189b46e756fSKirill A. Shutemov kthread_should_stop(); 2190b46e756fSKirill A. Shutemov } 2191b46e756fSKirill A. Shutemov 219234d6b470SZach O'Keefe static void khugepaged_do_scan(struct collapse_control *cc) 2193b46e756fSKirill A. Shutemov { 2194b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 219589dc6a96SYanfei Xu unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); 2196b46e756fSKirill A. Shutemov bool wait = true; 219750ad2f24SZach O'Keefe int result = SCAN_SUCCEED; 2198b46e756fSKirill A. Shutemov 2199a980df33SKirill A. Shutemov lru_add_drain_all(); 2200a980df33SKirill A. Shutemov 2201c6a7f445SYang Shi while (true) { 2202b46e756fSKirill A. Shutemov cond_resched(); 2203b46e756fSKirill A. Shutemov 2204b46e756fSKirill A. Shutemov if (unlikely(kthread_should_stop() || try_to_freeze())) 2205b46e756fSKirill A. Shutemov break; 2206b46e756fSKirill A. Shutemov 2207b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2208b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 2209b46e756fSKirill A. Shutemov pass_through_head++; 2210b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 2211b46e756fSKirill A. Shutemov pass_through_head < 2) 2212b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 221350ad2f24SZach O'Keefe &result, cc); 2214b46e756fSKirill A. Shutemov else 2215b46e756fSKirill A. Shutemov progress = pages; 2216b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2217b46e756fSKirill A. Shutemov 2218c6a7f445SYang Shi if (progress >= pages) 2219c6a7f445SYang Shi break; 2220c6a7f445SYang Shi 222150ad2f24SZach O'Keefe if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) { 2222c6a7f445SYang Shi /* 2223c6a7f445SYang Shi * If fail to allocate the first time, try to sleep for 2224c6a7f445SYang Shi * a while. When hit again, cancel the scan. 2225c6a7f445SYang Shi */ 2226c6a7f445SYang Shi if (!wait) 2227c6a7f445SYang Shi break; 2228c6a7f445SYang Shi wait = false; 2229c6a7f445SYang Shi khugepaged_alloc_sleep(); 2230c6a7f445SYang Shi } 2231c6a7f445SYang Shi } 2232b46e756fSKirill A. Shutemov } 2233b46e756fSKirill A. Shutemov 2234b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 2235b46e756fSKirill A. Shutemov { 2236b46e756fSKirill A. Shutemov return kthread_should_stop() || 2237b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 2238b46e756fSKirill A. Shutemov } 2239b46e756fSKirill A. Shutemov 2240b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 2241b46e756fSKirill A. Shutemov { 2242b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 2243b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 2244b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2245b46e756fSKirill A. Shutemov 2246b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 2247b46e756fSKirill A. Shutemov return; 2248b46e756fSKirill A. Shutemov 2249b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2250b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 2251b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 2252b46e756fSKirill A. Shutemov scan_sleep_jiffies); 2253b46e756fSKirill A. Shutemov return; 2254b46e756fSKirill A. Shutemov } 2255b46e756fSKirill A. Shutemov 22561064026bSYang Shi if (hugepage_flags_enabled()) 2257b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2258b46e756fSKirill A. Shutemov } 2259b46e756fSKirill A. Shutemov 2260b46e756fSKirill A. Shutemov static int khugepaged(void *none) 2261b46e756fSKirill A. Shutemov { 2262b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 2263b46e756fSKirill A. Shutemov 2264b46e756fSKirill A. Shutemov set_freezable(); 2265b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 2266b46e756fSKirill A. Shutemov 2267b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 226834d6b470SZach O'Keefe khugepaged_do_scan(&khugepaged_collapse_control); 2269b46e756fSKirill A. Shutemov khugepaged_wait_work(); 2270b46e756fSKirill A. Shutemov } 2271b46e756fSKirill A. Shutemov 2272b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2273b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2274b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2275b46e756fSKirill A. Shutemov if (mm_slot) 2276b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2277b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2278b46e756fSKirill A. Shutemov return 0; 2279b46e756fSKirill A. Shutemov } 2280b46e756fSKirill A. Shutemov 2281b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 2282b46e756fSKirill A. Shutemov { 2283b46e756fSKirill A. Shutemov struct zone *zone; 2284b46e756fSKirill A. Shutemov int nr_zones = 0; 2285b46e756fSKirill A. Shutemov unsigned long recommended_min; 2286b46e756fSKirill A. Shutemov 22871064026bSYang Shi if (!hugepage_flags_enabled()) { 2288bd3400eaSLiangcai Fan calculate_min_free_kbytes(); 2289bd3400eaSLiangcai Fan goto update_wmarks; 2290bd3400eaSLiangcai Fan } 2291bd3400eaSLiangcai Fan 2292b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 2293b7d349c7SJoonsoo Kim /* 2294b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 2295b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 2296b7d349c7SJoonsoo Kim */ 2297b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 2298b7d349c7SJoonsoo Kim continue; 2299b7d349c7SJoonsoo Kim 2300b46e756fSKirill A. Shutemov nr_zones++; 2301b7d349c7SJoonsoo Kim } 2302b46e756fSKirill A. Shutemov 2303b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2304b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 2305b46e756fSKirill A. Shutemov 2306b46e756fSKirill A. Shutemov /* 2307b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 2308b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 2309b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 2310b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 2311b46e756fSKirill A. Shutemov */ 2312b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 2313b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2314b46e756fSKirill A. Shutemov 2315b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 2316b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 2317b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 2318b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 2319b46e756fSKirill A. Shutemov 2320b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 2321b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 2322b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2323b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 2324b46e756fSKirill A. Shutemov 2325b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 2326b46e756fSKirill A. Shutemov } 2327bd3400eaSLiangcai Fan 2328bd3400eaSLiangcai Fan update_wmarks: 2329b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 2330b46e756fSKirill A. Shutemov } 2331b46e756fSKirill A. Shutemov 2332b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 2333b46e756fSKirill A. Shutemov { 2334b46e756fSKirill A. Shutemov int err = 0; 2335b46e756fSKirill A. Shutemov 2336b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 23371064026bSYang Shi if (hugepage_flags_enabled()) { 2338b46e756fSKirill A. Shutemov if (!khugepaged_thread) 2339b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 2340b46e756fSKirill A. Shutemov "khugepaged"); 2341b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 2342b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2343b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 2344b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2345b46e756fSKirill A. Shutemov goto fail; 2346b46e756fSKirill A. Shutemov } 2347b46e756fSKirill A. Shutemov 2348b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 2349b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 2350b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 2351b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 2352b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2353b46e756fSKirill A. Shutemov } 2354bd3400eaSLiangcai Fan set_recommended_min_free_kbytes(); 2355b46e756fSKirill A. Shutemov fail: 2356b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 2357b46e756fSKirill A. Shutemov return err; 2358b46e756fSKirill A. Shutemov } 23594aab2be0SVijay Balakrishna 23604aab2be0SVijay Balakrishna void khugepaged_min_free_kbytes_update(void) 23614aab2be0SVijay Balakrishna { 23624aab2be0SVijay Balakrishna mutex_lock(&khugepaged_mutex); 23631064026bSYang Shi if (hugepage_flags_enabled() && khugepaged_thread) 23644aab2be0SVijay Balakrishna set_recommended_min_free_kbytes(); 23654aab2be0SVijay Balakrishna mutex_unlock(&khugepaged_mutex); 23664aab2be0SVijay Balakrishna } 23677d8faaf1SZach O'Keefe 23687d8faaf1SZach O'Keefe static int madvise_collapse_errno(enum scan_result r) 23697d8faaf1SZach O'Keefe { 23707d8faaf1SZach O'Keefe /* 23717d8faaf1SZach O'Keefe * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide 23727d8faaf1SZach O'Keefe * actionable feedback to caller, so they may take an appropriate 23737d8faaf1SZach O'Keefe * fallback measure depending on the nature of the failure. 23747d8faaf1SZach O'Keefe */ 23757d8faaf1SZach O'Keefe switch (r) { 23767d8faaf1SZach O'Keefe case SCAN_ALLOC_HUGE_PAGE_FAIL: 23777d8faaf1SZach O'Keefe return -ENOMEM; 23787d8faaf1SZach O'Keefe case SCAN_CGROUP_CHARGE_FAIL: 23797d8faaf1SZach O'Keefe return -EBUSY; 23807d8faaf1SZach O'Keefe /* Resource temporary unavailable - trying again might succeed */ 23817d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 23827d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 23837d8faaf1SZach O'Keefe return -EAGAIN; 23847d8faaf1SZach O'Keefe /* 23857d8faaf1SZach O'Keefe * Other: Trying again likely not to succeed / error intrinsic to 23867d8faaf1SZach O'Keefe * specified memory range. khugepaged likely won't be able to collapse 23877d8faaf1SZach O'Keefe * either. 23887d8faaf1SZach O'Keefe */ 23897d8faaf1SZach O'Keefe default: 23907d8faaf1SZach O'Keefe return -EINVAL; 23917d8faaf1SZach O'Keefe } 23927d8faaf1SZach O'Keefe } 23937d8faaf1SZach O'Keefe 23947d8faaf1SZach O'Keefe int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev, 23957d8faaf1SZach O'Keefe unsigned long start, unsigned long end) 23967d8faaf1SZach O'Keefe { 23977d8faaf1SZach O'Keefe struct collapse_control *cc; 23987d8faaf1SZach O'Keefe struct mm_struct *mm = vma->vm_mm; 23997d8faaf1SZach O'Keefe unsigned long hstart, hend, addr; 24007d8faaf1SZach O'Keefe int thps = 0, last_fail = SCAN_FAIL; 24017d8faaf1SZach O'Keefe bool mmap_locked = true; 24027d8faaf1SZach O'Keefe 24037d8faaf1SZach O'Keefe BUG_ON(vma->vm_start > start); 24047d8faaf1SZach O'Keefe BUG_ON(vma->vm_end < end); 24057d8faaf1SZach O'Keefe 24067d8faaf1SZach O'Keefe *prev = vma; 24077d8faaf1SZach O'Keefe 24087d8faaf1SZach O'Keefe /* TODO: Support file/shmem */ 24097d8faaf1SZach O'Keefe if (!vma->anon_vma || !vma_is_anonymous(vma)) 24107d8faaf1SZach O'Keefe return -EINVAL; 24117d8faaf1SZach O'Keefe 24127d8faaf1SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) 24137d8faaf1SZach O'Keefe return -EINVAL; 24147d8faaf1SZach O'Keefe 24157d8faaf1SZach O'Keefe cc = kmalloc(sizeof(*cc), GFP_KERNEL); 24167d8faaf1SZach O'Keefe if (!cc) 24177d8faaf1SZach O'Keefe return -ENOMEM; 24187d8faaf1SZach O'Keefe cc->is_khugepaged = false; 24197d8faaf1SZach O'Keefe cc->last_target_node = NUMA_NO_NODE; 24207d8faaf1SZach O'Keefe 24217d8faaf1SZach O'Keefe mmgrab(mm); 24227d8faaf1SZach O'Keefe lru_add_drain_all(); 24237d8faaf1SZach O'Keefe 24247d8faaf1SZach O'Keefe hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 24257d8faaf1SZach O'Keefe hend = end & HPAGE_PMD_MASK; 24267d8faaf1SZach O'Keefe 24277d8faaf1SZach O'Keefe for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) { 24287d8faaf1SZach O'Keefe int result = SCAN_FAIL; 24297d8faaf1SZach O'Keefe 24307d8faaf1SZach O'Keefe if (!mmap_locked) { 24317d8faaf1SZach O'Keefe cond_resched(); 24327d8faaf1SZach O'Keefe mmap_read_lock(mm); 24337d8faaf1SZach O'Keefe mmap_locked = true; 24347d8faaf1SZach O'Keefe result = hugepage_vma_revalidate(mm, addr, &vma, cc); 24357d8faaf1SZach O'Keefe if (result != SCAN_SUCCEED) { 24367d8faaf1SZach O'Keefe last_fail = result; 24377d8faaf1SZach O'Keefe goto out_nolock; 24387d8faaf1SZach O'Keefe } 24397d8faaf1SZach O'Keefe } 24407d8faaf1SZach O'Keefe mmap_assert_locked(mm); 24417d8faaf1SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 2442*7d2c4385SZach O'Keefe result = hpage_collapse_scan_pmd(mm, vma, addr, &mmap_locked, 2443*7d2c4385SZach O'Keefe cc); 24447d8faaf1SZach O'Keefe if (!mmap_locked) 24457d8faaf1SZach O'Keefe *prev = NULL; /* Tell caller we dropped mmap_lock */ 24467d8faaf1SZach O'Keefe 24477d8faaf1SZach O'Keefe switch (result) { 24487d8faaf1SZach O'Keefe case SCAN_SUCCEED: 24497d8faaf1SZach O'Keefe case SCAN_PMD_MAPPED: 24507d8faaf1SZach O'Keefe ++thps; 24517d8faaf1SZach O'Keefe break; 24527d8faaf1SZach O'Keefe /* Whitelisted set of results where continuing OK */ 24537d8faaf1SZach O'Keefe case SCAN_PMD_NULL: 24547d8faaf1SZach O'Keefe case SCAN_PTE_NON_PRESENT: 24557d8faaf1SZach O'Keefe case SCAN_PTE_UFFD_WP: 24567d8faaf1SZach O'Keefe case SCAN_PAGE_RO: 24577d8faaf1SZach O'Keefe case SCAN_LACK_REFERENCED_PAGE: 24587d8faaf1SZach O'Keefe case SCAN_PAGE_NULL: 24597d8faaf1SZach O'Keefe case SCAN_PAGE_COUNT: 24607d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 24617d8faaf1SZach O'Keefe case SCAN_PAGE_COMPOUND: 24627d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 24637d8faaf1SZach O'Keefe last_fail = result; 24647d8faaf1SZach O'Keefe break; 24657d8faaf1SZach O'Keefe default: 24667d8faaf1SZach O'Keefe last_fail = result; 24677d8faaf1SZach O'Keefe /* Other error, exit */ 24687d8faaf1SZach O'Keefe goto out_maybelock; 24697d8faaf1SZach O'Keefe } 24707d8faaf1SZach O'Keefe } 24717d8faaf1SZach O'Keefe 24727d8faaf1SZach O'Keefe out_maybelock: 24737d8faaf1SZach O'Keefe /* Caller expects us to hold mmap_lock on return */ 24747d8faaf1SZach O'Keefe if (!mmap_locked) 24757d8faaf1SZach O'Keefe mmap_read_lock(mm); 24767d8faaf1SZach O'Keefe out_nolock: 24777d8faaf1SZach O'Keefe mmap_assert_locked(mm); 24787d8faaf1SZach O'Keefe mmdrop(mm); 24797d8faaf1SZach O'Keefe kfree(cc); 24807d8faaf1SZach O'Keefe 24817d8faaf1SZach O'Keefe return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0 24827d8faaf1SZach O'Keefe : madvise_collapse_errno(last_fail); 24837d8faaf1SZach O'Keefe } 2484