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> 22e2942062Sxu xin #include <linux/ksm.h> 23b46e756fSKirill A. Shutemov 24b46e756fSKirill A. Shutemov #include <asm/tlb.h> 25b46e756fSKirill A. Shutemov #include <asm/pgalloc.h> 26b46e756fSKirill A. Shutemov #include "internal.h" 27b26e2701SQi Zheng #include "mm_slot.h" 28b46e756fSKirill A. Shutemov 29b46e756fSKirill A. Shutemov enum scan_result { 30b46e756fSKirill A. Shutemov SCAN_FAIL, 31b46e756fSKirill A. Shutemov SCAN_SUCCEED, 32b46e756fSKirill A. Shutemov SCAN_PMD_NULL, 3334488399SZach O'Keefe SCAN_PMD_NONE, 3450722804SZach O'Keefe SCAN_PMD_MAPPED, 35b46e756fSKirill A. Shutemov SCAN_EXCEED_NONE_PTE, 3671a2c112SKirill A. Shutemov SCAN_EXCEED_SWAP_PTE, 3771a2c112SKirill A. Shutemov SCAN_EXCEED_SHARED_PTE, 38b46e756fSKirill A. Shutemov SCAN_PTE_NON_PRESENT, 39e1e267c7SPeter Xu SCAN_PTE_UFFD_WP, 4058ac9a89SZach O'Keefe SCAN_PTE_MAPPED_HUGEPAGE, 41b46e756fSKirill A. Shutemov SCAN_PAGE_RO, 420db501f7SEbru Akagunduz SCAN_LACK_REFERENCED_PAGE, 43b46e756fSKirill A. Shutemov SCAN_PAGE_NULL, 44b46e756fSKirill A. Shutemov SCAN_SCAN_ABORT, 45b46e756fSKirill A. Shutemov SCAN_PAGE_COUNT, 46b46e756fSKirill A. Shutemov SCAN_PAGE_LRU, 47b46e756fSKirill A. Shutemov SCAN_PAGE_LOCK, 48b46e756fSKirill A. Shutemov SCAN_PAGE_ANON, 49b46e756fSKirill A. Shutemov SCAN_PAGE_COMPOUND, 50b46e756fSKirill A. Shutemov SCAN_ANY_PROCESS, 51b46e756fSKirill A. Shutemov SCAN_VMA_NULL, 52b46e756fSKirill A. Shutemov SCAN_VMA_CHECK, 53b46e756fSKirill A. Shutemov SCAN_ADDRESS_RANGE, 54b46e756fSKirill A. Shutemov SCAN_DEL_PAGE_LRU, 55b46e756fSKirill A. Shutemov SCAN_ALLOC_HUGE_PAGE_FAIL, 56b46e756fSKirill A. Shutemov SCAN_CGROUP_CHARGE_FAIL, 57f3f0e1d2SKirill A. Shutemov SCAN_TRUNCATED, 5899cb0dbdSSong Liu SCAN_PAGE_HAS_PRIVATE, 592ce0bdfeSIvan Orlov SCAN_STORE_FAILED, 6098c76c9fSJiaqi Yan SCAN_COPY_MC, 61ac492b9cSDavid Stevens SCAN_PAGE_FILLED, 62b46e756fSKirill A. Shutemov }; 63b46e756fSKirill A. Shutemov 64b46e756fSKirill A. Shutemov #define CREATE_TRACE_POINTS 65b46e756fSKirill A. Shutemov #include <trace/events/huge_memory.h> 66b46e756fSKirill A. Shutemov 674aab2be0SVijay Balakrishna static struct task_struct *khugepaged_thread __read_mostly; 684aab2be0SVijay Balakrishna static DEFINE_MUTEX(khugepaged_mutex); 694aab2be0SVijay Balakrishna 70b46e756fSKirill A. Shutemov /* default scan 8*512 pte (or vmas) every 30 second */ 71b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_to_scan __read_mostly; 72b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_collapsed; 73b46e756fSKirill A. Shutemov static unsigned int khugepaged_full_scans; 74b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 75b46e756fSKirill A. Shutemov /* during fragmentation poll the hugepage allocator once every minute */ 76b46e756fSKirill A. Shutemov static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 77b46e756fSKirill A. Shutemov static unsigned long khugepaged_sleep_expire; 78b46e756fSKirill A. Shutemov static DEFINE_SPINLOCK(khugepaged_mm_lock); 79b46e756fSKirill A. Shutemov static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 80b46e756fSKirill A. Shutemov /* 81b46e756fSKirill A. Shutemov * default collapse hugepages if there is at least one pte mapped like 82b46e756fSKirill A. Shutemov * it would have happened if the vma was large enough during page 83b46e756fSKirill A. Shutemov * fault. 84d8ea7cc8SZach O'Keefe * 85d8ea7cc8SZach O'Keefe * Note that these are only respected if collapse was initiated by khugepaged. 86b46e756fSKirill A. Shutemov */ 87b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_none __read_mostly; 88b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_swap __read_mostly; 8971a2c112SKirill A. Shutemov static unsigned int khugepaged_max_ptes_shared __read_mostly; 90b46e756fSKirill A. Shutemov 91b46e756fSKirill A. Shutemov #define MM_SLOTS_HASH_BITS 10 92e1ad3e66SNick Desaulniers static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 93b46e756fSKirill A. Shutemov 94b46e756fSKirill A. Shutemov static struct kmem_cache *mm_slot_cache __read_mostly; 95b46e756fSKirill A. Shutemov 9634d6b470SZach O'Keefe struct collapse_control { 97d8ea7cc8SZach O'Keefe bool is_khugepaged; 98d8ea7cc8SZach O'Keefe 9934d6b470SZach O'Keefe /* Num pages scanned per node */ 10034d6b470SZach O'Keefe u32 node_load[MAX_NUMNODES]; 10134d6b470SZach O'Keefe 102e031ff96SYang Shi /* nodemask for allocation fallback */ 103e031ff96SYang Shi nodemask_t alloc_nmask; 10434d6b470SZach O'Keefe }; 10534d6b470SZach O'Keefe 106b46e756fSKirill A. Shutemov /** 107b26e2701SQi Zheng * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned 108b26e2701SQi Zheng * @slot: hash lookup from mm to mm_slot 109b46e756fSKirill A. Shutemov */ 110b26e2701SQi Zheng struct khugepaged_mm_slot { 111b26e2701SQi Zheng struct mm_slot slot; 112b46e756fSKirill A. Shutemov }; 113b46e756fSKirill A. Shutemov 114b46e756fSKirill A. Shutemov /** 115b46e756fSKirill A. Shutemov * struct khugepaged_scan - cursor for scanning 116b46e756fSKirill A. Shutemov * @mm_head: the head of the mm list to scan 117b46e756fSKirill A. Shutemov * @mm_slot: the current mm_slot we are scanning 118b46e756fSKirill A. Shutemov * @address: the next address inside that to be scanned 119b46e756fSKirill A. Shutemov * 120b46e756fSKirill A. Shutemov * There is only the one khugepaged_scan instance of this cursor structure. 121b46e756fSKirill A. Shutemov */ 122b46e756fSKirill A. Shutemov struct khugepaged_scan { 123b46e756fSKirill A. Shutemov struct list_head mm_head; 124b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 125b46e756fSKirill A. Shutemov unsigned long address; 126b46e756fSKirill A. Shutemov }; 127b46e756fSKirill A. Shutemov 128b46e756fSKirill A. Shutemov static struct khugepaged_scan khugepaged_scan = { 129b46e756fSKirill A. Shutemov .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 130b46e756fSKirill A. Shutemov }; 131b46e756fSKirill A. Shutemov 132e1465d12SJérémy Lefaure #ifdef CONFIG_SYSFS 133b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 134b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 135b46e756fSKirill A. Shutemov char *buf) 136b46e756fSKirill A. Shutemov { 137ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); 138b46e756fSKirill A. Shutemov } 139b46e756fSKirill A. Shutemov 140b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 141b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 142b46e756fSKirill A. Shutemov const char *buf, size_t count) 143b46e756fSKirill A. Shutemov { 144dfefd226SAlexey Dobriyan unsigned int msecs; 145b46e756fSKirill A. Shutemov int err; 146b46e756fSKirill A. Shutemov 147dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 148dfefd226SAlexey Dobriyan if (err) 149b46e756fSKirill A. Shutemov return -EINVAL; 150b46e756fSKirill A. Shutemov 151b46e756fSKirill A. Shutemov khugepaged_scan_sleep_millisecs = msecs; 152b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 153b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 154b46e756fSKirill A. Shutemov 155b46e756fSKirill A. Shutemov return count; 156b46e756fSKirill A. Shutemov } 157b46e756fSKirill A. Shutemov static struct kobj_attribute scan_sleep_millisecs_attr = 1586dcdc94dSMiaohe Lin __ATTR_RW(scan_sleep_millisecs); 159b46e756fSKirill A. Shutemov 160b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 161b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 162b46e756fSKirill A. Shutemov char *buf) 163b46e756fSKirill A. Shutemov { 164ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 165b46e756fSKirill A. Shutemov } 166b46e756fSKirill A. Shutemov 167b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 168b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 169b46e756fSKirill A. Shutemov const char *buf, size_t count) 170b46e756fSKirill A. Shutemov { 171dfefd226SAlexey Dobriyan unsigned int msecs; 172b46e756fSKirill A. Shutemov int err; 173b46e756fSKirill A. Shutemov 174dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 175dfefd226SAlexey Dobriyan if (err) 176b46e756fSKirill A. Shutemov return -EINVAL; 177b46e756fSKirill A. Shutemov 178b46e756fSKirill A. Shutemov khugepaged_alloc_sleep_millisecs = msecs; 179b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 180b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 181b46e756fSKirill A. Shutemov 182b46e756fSKirill A. Shutemov return count; 183b46e756fSKirill A. Shutemov } 184b46e756fSKirill A. Shutemov static struct kobj_attribute alloc_sleep_millisecs_attr = 1856dcdc94dSMiaohe Lin __ATTR_RW(alloc_sleep_millisecs); 186b46e756fSKirill A. Shutemov 187b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_show(struct kobject *kobj, 188b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 189b46e756fSKirill A. Shutemov char *buf) 190b46e756fSKirill A. Shutemov { 191ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); 192b46e756fSKirill A. Shutemov } 193b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_store(struct kobject *kobj, 194b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 195b46e756fSKirill A. Shutemov const char *buf, size_t count) 196b46e756fSKirill A. Shutemov { 197dfefd226SAlexey Dobriyan unsigned int pages; 198b46e756fSKirill A. Shutemov int err; 199b46e756fSKirill A. Shutemov 200dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &pages); 201dfefd226SAlexey Dobriyan if (err || !pages) 202b46e756fSKirill A. Shutemov return -EINVAL; 203b46e756fSKirill A. Shutemov 204b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = pages; 205b46e756fSKirill A. Shutemov 206b46e756fSKirill A. Shutemov return count; 207b46e756fSKirill A. Shutemov } 208b46e756fSKirill A. Shutemov static struct kobj_attribute pages_to_scan_attr = 2096dcdc94dSMiaohe Lin __ATTR_RW(pages_to_scan); 210b46e756fSKirill A. Shutemov 211b46e756fSKirill A. Shutemov static ssize_t pages_collapsed_show(struct kobject *kobj, 212b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 213b46e756fSKirill A. Shutemov char *buf) 214b46e756fSKirill A. Shutemov { 215ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); 216b46e756fSKirill A. Shutemov } 217b46e756fSKirill A. Shutemov static struct kobj_attribute pages_collapsed_attr = 218b46e756fSKirill A. Shutemov __ATTR_RO(pages_collapsed); 219b46e756fSKirill A. Shutemov 220b46e756fSKirill A. Shutemov static ssize_t full_scans_show(struct kobject *kobj, 221b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 222b46e756fSKirill A. Shutemov char *buf) 223b46e756fSKirill A. Shutemov { 224ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_full_scans); 225b46e756fSKirill A. Shutemov } 226b46e756fSKirill A. Shutemov static struct kobj_attribute full_scans_attr = 227b46e756fSKirill A. Shutemov __ATTR_RO(full_scans); 228b46e756fSKirill A. Shutemov 2296dcdc94dSMiaohe Lin static ssize_t defrag_show(struct kobject *kobj, 230b46e756fSKirill A. Shutemov struct kobj_attribute *attr, char *buf) 231b46e756fSKirill A. Shutemov { 232b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 233b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 234b46e756fSKirill A. Shutemov } 2356dcdc94dSMiaohe Lin static ssize_t defrag_store(struct kobject *kobj, 236b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 237b46e756fSKirill A. Shutemov const char *buf, size_t count) 238b46e756fSKirill A. Shutemov { 239b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 240b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 241b46e756fSKirill A. Shutemov } 242b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_defrag_attr = 2436dcdc94dSMiaohe Lin __ATTR_RW(defrag); 244b46e756fSKirill A. Shutemov 245b46e756fSKirill A. Shutemov /* 246b46e756fSKirill A. Shutemov * max_ptes_none controls if khugepaged should collapse hugepages over 247b46e756fSKirill A. Shutemov * any unmapped ptes in turn potentially increasing the memory 248b46e756fSKirill A. Shutemov * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 249b46e756fSKirill A. Shutemov * reduce the available free memory in the system as it 250b46e756fSKirill A. Shutemov * runs. Increasing max_ptes_none will instead potentially reduce the 251b46e756fSKirill A. Shutemov * free memory in the system during the khugepaged scan. 252b46e756fSKirill A. Shutemov */ 2536dcdc94dSMiaohe Lin static ssize_t max_ptes_none_show(struct kobject *kobj, 254b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 255b46e756fSKirill A. Shutemov char *buf) 256b46e756fSKirill A. Shutemov { 257ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); 258b46e756fSKirill A. Shutemov } 2596dcdc94dSMiaohe Lin static ssize_t max_ptes_none_store(struct kobject *kobj, 260b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 261b46e756fSKirill A. Shutemov const char *buf, size_t count) 262b46e756fSKirill A. Shutemov { 263b46e756fSKirill A. Shutemov int err; 264b46e756fSKirill A. Shutemov unsigned long max_ptes_none; 265b46e756fSKirill A. Shutemov 266b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_none); 267b46e756fSKirill A. Shutemov if (err || max_ptes_none > HPAGE_PMD_NR - 1) 268b46e756fSKirill A. Shutemov return -EINVAL; 269b46e756fSKirill A. Shutemov 270b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = max_ptes_none; 271b46e756fSKirill A. Shutemov 272b46e756fSKirill A. Shutemov return count; 273b46e756fSKirill A. Shutemov } 274b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_none_attr = 2756dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_none); 276b46e756fSKirill A. Shutemov 2776dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_show(struct kobject *kobj, 278b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 279b46e756fSKirill A. Shutemov char *buf) 280b46e756fSKirill A. Shutemov { 281ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); 282b46e756fSKirill A. Shutemov } 283b46e756fSKirill A. Shutemov 2846dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_store(struct kobject *kobj, 285b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 286b46e756fSKirill A. Shutemov const char *buf, size_t count) 287b46e756fSKirill A. Shutemov { 288b46e756fSKirill A. Shutemov int err; 289b46e756fSKirill A. Shutemov unsigned long max_ptes_swap; 290b46e756fSKirill A. Shutemov 291b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_swap); 292b46e756fSKirill A. Shutemov if (err || max_ptes_swap > HPAGE_PMD_NR - 1) 293b46e756fSKirill A. Shutemov return -EINVAL; 294b46e756fSKirill A. Shutemov 295b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = max_ptes_swap; 296b46e756fSKirill A. Shutemov 297b46e756fSKirill A. Shutemov return count; 298b46e756fSKirill A. Shutemov } 299b46e756fSKirill A. Shutemov 300b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_swap_attr = 3016dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_swap); 302b46e756fSKirill A. Shutemov 3036dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_show(struct kobject *kobj, 30471a2c112SKirill A. Shutemov struct kobj_attribute *attr, 30571a2c112SKirill A. Shutemov char *buf) 30671a2c112SKirill A. Shutemov { 307ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); 30871a2c112SKirill A. Shutemov } 30971a2c112SKirill A. Shutemov 3106dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_store(struct kobject *kobj, 31171a2c112SKirill A. Shutemov struct kobj_attribute *attr, 31271a2c112SKirill A. Shutemov const char *buf, size_t count) 31371a2c112SKirill A. Shutemov { 31471a2c112SKirill A. Shutemov int err; 31571a2c112SKirill A. Shutemov unsigned long max_ptes_shared; 31671a2c112SKirill A. Shutemov 31771a2c112SKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_shared); 31871a2c112SKirill A. Shutemov if (err || max_ptes_shared > HPAGE_PMD_NR - 1) 31971a2c112SKirill A. Shutemov return -EINVAL; 32071a2c112SKirill A. Shutemov 32171a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = max_ptes_shared; 32271a2c112SKirill A. Shutemov 32371a2c112SKirill A. Shutemov return count; 32471a2c112SKirill A. Shutemov } 32571a2c112SKirill A. Shutemov 32671a2c112SKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_shared_attr = 3276dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_shared); 32871a2c112SKirill A. Shutemov 329b46e756fSKirill A. Shutemov static struct attribute *khugepaged_attr[] = { 330b46e756fSKirill A. Shutemov &khugepaged_defrag_attr.attr, 331b46e756fSKirill A. Shutemov &khugepaged_max_ptes_none_attr.attr, 33271a2c112SKirill A. Shutemov &khugepaged_max_ptes_swap_attr.attr, 33371a2c112SKirill A. Shutemov &khugepaged_max_ptes_shared_attr.attr, 334b46e756fSKirill A. Shutemov &pages_to_scan_attr.attr, 335b46e756fSKirill A. Shutemov &pages_collapsed_attr.attr, 336b46e756fSKirill A. Shutemov &full_scans_attr.attr, 337b46e756fSKirill A. Shutemov &scan_sleep_millisecs_attr.attr, 338b46e756fSKirill A. Shutemov &alloc_sleep_millisecs_attr.attr, 339b46e756fSKirill A. Shutemov NULL, 340b46e756fSKirill A. Shutemov }; 341b46e756fSKirill A. Shutemov 342b46e756fSKirill A. Shutemov struct attribute_group khugepaged_attr_group = { 343b46e756fSKirill A. Shutemov .attrs = khugepaged_attr, 344b46e756fSKirill A. Shutemov .name = "khugepaged", 345b46e756fSKirill A. Shutemov }; 346e1465d12SJérémy Lefaure #endif /* CONFIG_SYSFS */ 347b46e756fSKirill A. Shutemov 348b46e756fSKirill A. Shutemov int hugepage_madvise(struct vm_area_struct *vma, 349b46e756fSKirill A. Shutemov unsigned long *vm_flags, int advice) 350b46e756fSKirill A. Shutemov { 351b46e756fSKirill A. Shutemov switch (advice) { 352b46e756fSKirill A. Shutemov case MADV_HUGEPAGE: 353b46e756fSKirill A. Shutemov #ifdef CONFIG_S390 354b46e756fSKirill A. Shutemov /* 355b46e756fSKirill A. Shutemov * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 356b46e756fSKirill A. Shutemov * can't handle this properly after s390_enable_sie, so we simply 357b46e756fSKirill A. Shutemov * ignore the madvise to prevent qemu from causing a SIGSEGV. 358b46e756fSKirill A. Shutemov */ 359b46e756fSKirill A. Shutemov if (mm_has_pgste(vma->vm_mm)) 360b46e756fSKirill A. Shutemov return 0; 361b46e756fSKirill A. Shutemov #endif 362b46e756fSKirill A. Shutemov *vm_flags &= ~VM_NOHUGEPAGE; 363b46e756fSKirill A. Shutemov *vm_flags |= VM_HUGEPAGE; 364b46e756fSKirill A. Shutemov /* 365b46e756fSKirill A. Shutemov * If the vma become good for khugepaged to scan, 366b46e756fSKirill A. Shutemov * register it here without waiting a page fault that 367b46e756fSKirill A. Shutemov * may not happen any time soon. 368b46e756fSKirill A. Shutemov */ 369c791576cSYang Shi khugepaged_enter_vma(vma, *vm_flags); 370b46e756fSKirill A. Shutemov break; 371b46e756fSKirill A. Shutemov case MADV_NOHUGEPAGE: 372b46e756fSKirill A. Shutemov *vm_flags &= ~VM_HUGEPAGE; 373b46e756fSKirill A. Shutemov *vm_flags |= VM_NOHUGEPAGE; 374b46e756fSKirill A. Shutemov /* 375b46e756fSKirill A. Shutemov * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 376b46e756fSKirill A. Shutemov * this vma even if we leave the mm registered in khugepaged if 377b46e756fSKirill A. Shutemov * it got registered before VM_NOHUGEPAGE was set. 378b46e756fSKirill A. Shutemov */ 379b46e756fSKirill A. Shutemov break; 380b46e756fSKirill A. Shutemov } 381b46e756fSKirill A. Shutemov 382b46e756fSKirill A. Shutemov return 0; 383b46e756fSKirill A. Shutemov } 384b46e756fSKirill A. Shutemov 385b46e756fSKirill A. Shutemov int __init khugepaged_init(void) 386b46e756fSKirill A. Shutemov { 387b46e756fSKirill A. Shutemov mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 388b26e2701SQi Zheng sizeof(struct khugepaged_mm_slot), 389b26e2701SQi Zheng __alignof__(struct khugepaged_mm_slot), 390b26e2701SQi Zheng 0, NULL); 391b46e756fSKirill A. Shutemov if (!mm_slot_cache) 392b46e756fSKirill A. Shutemov return -ENOMEM; 393b46e756fSKirill A. Shutemov 394b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 395b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 396b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 39771a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; 398b46e756fSKirill A. Shutemov 399b46e756fSKirill A. Shutemov return 0; 400b46e756fSKirill A. Shutemov } 401b46e756fSKirill A. Shutemov 402b46e756fSKirill A. Shutemov void __init khugepaged_destroy(void) 403b46e756fSKirill A. Shutemov { 404b46e756fSKirill A. Shutemov kmem_cache_destroy(mm_slot_cache); 405b46e756fSKirill A. Shutemov } 406b46e756fSKirill A. Shutemov 4077d2c4385SZach O'Keefe static inline int hpage_collapse_test_exit(struct mm_struct *mm) 408b46e756fSKirill A. Shutemov { 4094d45e75aSJann Horn return atomic_read(&mm->mm_users) == 0; 410b46e756fSKirill A. Shutemov } 411b46e756fSKirill A. Shutemov 412d2081b2bSYang Shi void __khugepaged_enter(struct mm_struct *mm) 413b46e756fSKirill A. Shutemov { 414b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 415b26e2701SQi Zheng struct mm_slot *slot; 416b46e756fSKirill A. Shutemov int wakeup; 417b46e756fSKirill A. Shutemov 41816618670SXin Hao /* __khugepaged_exit() must not run from under us */ 41916618670SXin Hao VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm); 42016618670SXin Hao if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) 42116618670SXin Hao return; 42216618670SXin Hao 423b26e2701SQi Zheng mm_slot = mm_slot_alloc(mm_slot_cache); 424b46e756fSKirill A. Shutemov if (!mm_slot) 425d2081b2bSYang Shi return; 426b46e756fSKirill A. Shutemov 427b26e2701SQi Zheng slot = &mm_slot->slot; 428b26e2701SQi Zheng 429b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 430b26e2701SQi Zheng mm_slot_insert(mm_slots_hash, mm, slot); 431b46e756fSKirill A. Shutemov /* 432b46e756fSKirill A. Shutemov * Insert just behind the scanning cursor, to let the area settle 433b46e756fSKirill A. Shutemov * down a little. 434b46e756fSKirill A. Shutemov */ 435b46e756fSKirill A. Shutemov wakeup = list_empty(&khugepaged_scan.mm_head); 436b26e2701SQi Zheng list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head); 437b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 438b46e756fSKirill A. Shutemov 439f1f10076SVegard Nossum mmgrab(mm); 440b46e756fSKirill A. Shutemov if (wakeup) 441b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 442b46e756fSKirill A. Shutemov } 443b46e756fSKirill A. Shutemov 444c791576cSYang Shi void khugepaged_enter_vma(struct vm_area_struct *vma, 445b46e756fSKirill A. Shutemov unsigned long vm_flags) 446b46e756fSKirill A. Shutemov { 4472647d11bSYang Shi if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && 4481064026bSYang Shi hugepage_flags_enabled()) { 449a7f4e6e4SZach O'Keefe if (hugepage_vma_check(vma, vm_flags, false, false, true)) 4502647d11bSYang Shi __khugepaged_enter(vma->vm_mm); 4512647d11bSYang Shi } 452b46e756fSKirill A. Shutemov } 453b46e756fSKirill A. Shutemov 454b46e756fSKirill A. Shutemov void __khugepaged_exit(struct mm_struct *mm) 455b46e756fSKirill A. Shutemov { 456b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 457b26e2701SQi Zheng struct mm_slot *slot; 458b46e756fSKirill A. Shutemov int free = 0; 459b46e756fSKirill A. Shutemov 460b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 461b26e2701SQi Zheng slot = mm_slot_lookup(mm_slots_hash, mm); 462b26e2701SQi Zheng mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 463b46e756fSKirill A. Shutemov if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 464b26e2701SQi Zheng hash_del(&slot->hash); 465b26e2701SQi Zheng list_del(&slot->mm_node); 466b46e756fSKirill A. Shutemov free = 1; 467b46e756fSKirill A. Shutemov } 468b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 469b46e756fSKirill A. Shutemov 470b46e756fSKirill A. Shutemov if (free) { 471b46e756fSKirill A. Shutemov clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 472b26e2701SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 473b46e756fSKirill A. Shutemov mmdrop(mm); 474b46e756fSKirill A. Shutemov } else if (mm_slot) { 475b46e756fSKirill A. Shutemov /* 476b46e756fSKirill A. Shutemov * This is required to serialize against 4777d2c4385SZach O'Keefe * hpage_collapse_test_exit() (which is guaranteed to run 4787d2c4385SZach O'Keefe * under mmap sem read mode). Stop here (after we return all 4797d2c4385SZach O'Keefe * pagetables will be destroyed) until khugepaged has finished 4807d2c4385SZach O'Keefe * working on the pagetables under the mmap_lock. 481b46e756fSKirill A. Shutemov */ 482d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 483d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 484b46e756fSKirill A. Shutemov } 485b46e756fSKirill A. Shutemov } 486b46e756fSKirill A. Shutemov 48792644f58SVishal Moola (Oracle) static void release_pte_folio(struct folio *folio) 48892644f58SVishal Moola (Oracle) { 48992644f58SVishal Moola (Oracle) node_stat_mod_folio(folio, 49092644f58SVishal Moola (Oracle) NR_ISOLATED_ANON + folio_is_file_lru(folio), 49192644f58SVishal Moola (Oracle) -folio_nr_pages(folio)); 49292644f58SVishal Moola (Oracle) folio_unlock(folio); 49392644f58SVishal Moola (Oracle) folio_putback_lru(folio); 49492644f58SVishal Moola (Oracle) } 49592644f58SVishal Moola (Oracle) 496b46e756fSKirill A. Shutemov static void release_pte_page(struct page *page) 497b46e756fSKirill A. Shutemov { 49892644f58SVishal Moola (Oracle) release_pte_folio(page_folio(page)); 499b46e756fSKirill A. Shutemov } 500b46e756fSKirill A. Shutemov 5015503fbf2SKirill A. Shutemov static void release_pte_pages(pte_t *pte, pte_t *_pte, 5025503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 503b46e756fSKirill A. Shutemov { 5049bdfeea4SVishal Moola (Oracle) struct folio *folio, *tmp; 5055503fbf2SKirill A. Shutemov 506b46e756fSKirill A. Shutemov while (--_pte >= pte) { 507c33c7948SRyan Roberts pte_t pteval = ptep_get(_pte); 508f528260bSVishal Moola (Oracle) unsigned long pfn; 5095503fbf2SKirill A. Shutemov 510f528260bSVishal Moola (Oracle) if (pte_none(pteval)) 511f528260bSVishal Moola (Oracle) continue; 512f528260bSVishal Moola (Oracle) pfn = pte_pfn(pteval); 513f528260bSVishal Moola (Oracle) if (is_zero_pfn(pfn)) 514f528260bSVishal Moola (Oracle) continue; 515f528260bSVishal Moola (Oracle) folio = pfn_folio(pfn); 516f528260bSVishal Moola (Oracle) if (folio_test_large(folio)) 517f528260bSVishal Moola (Oracle) continue; 5189bdfeea4SVishal Moola (Oracle) release_pte_folio(folio); 5195503fbf2SKirill A. Shutemov } 5205503fbf2SKirill A. Shutemov 5219bdfeea4SVishal Moola (Oracle) list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) { 5229bdfeea4SVishal Moola (Oracle) list_del(&folio->lru); 5239bdfeea4SVishal Moola (Oracle) release_pte_folio(folio); 524b46e756fSKirill A. Shutemov } 525b46e756fSKirill A. Shutemov } 526b46e756fSKirill A. Shutemov 527dbf85c21SVishal Moola (Oracle) static bool is_refcount_suitable(struct folio *folio) 5289445689fSKirill A. Shutemov { 5299445689fSKirill A. Shutemov int expected_refcount; 5309445689fSKirill A. Shutemov 531dbf85c21SVishal Moola (Oracle) expected_refcount = folio_mapcount(folio); 532dbf85c21SVishal Moola (Oracle) if (folio_test_swapcache(folio)) 533dbf85c21SVishal Moola (Oracle) expected_refcount += folio_nr_pages(folio); 5349445689fSKirill A. Shutemov 535dbf85c21SVishal Moola (Oracle) return folio_ref_count(folio) == expected_refcount; 5369445689fSKirill A. Shutemov } 5379445689fSKirill A. Shutemov 538b46e756fSKirill A. Shutemov static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 539b46e756fSKirill A. Shutemov unsigned long address, 5405503fbf2SKirill A. Shutemov pte_t *pte, 541d8ea7cc8SZach O'Keefe struct collapse_control *cc, 5425503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 543b46e756fSKirill A. Shutemov { 544b46e756fSKirill A. Shutemov struct page *page = NULL; 5458dd1e896SVishal Moola (Oracle) struct folio *folio = NULL; 546b46e756fSKirill A. Shutemov pte_t *_pte; 54750ad2f24SZach O'Keefe int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0; 5480db501f7SEbru Akagunduz bool writable = false; 549b46e756fSKirill A. Shutemov 550b46e756fSKirill A. Shutemov for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 551b46e756fSKirill A. Shutemov _pte++, address += PAGE_SIZE) { 552c33c7948SRyan Roberts pte_t pteval = ptep_get(_pte); 553b46e756fSKirill A. Shutemov if (pte_none(pteval) || (pte_present(pteval) && 554b46e756fSKirill A. Shutemov is_zero_pfn(pte_pfn(pteval)))) { 555d8ea7cc8SZach O'Keefe ++none_or_zero; 556b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 557d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 558d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 559b46e756fSKirill A. Shutemov continue; 560b46e756fSKirill A. Shutemov } else { 561b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 562e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 563b46e756fSKirill A. Shutemov goto out; 564b46e756fSKirill A. Shutemov } 565b46e756fSKirill A. Shutemov } 566b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 567b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 568b46e756fSKirill A. Shutemov goto out; 569b46e756fSKirill A. Shutemov } 570dd47ac42SPeter Xu if (pte_uffd_wp(pteval)) { 571dd47ac42SPeter Xu result = SCAN_PTE_UFFD_WP; 572dd47ac42SPeter Xu goto out; 573dd47ac42SPeter Xu } 574b46e756fSKirill A. Shutemov page = vm_normal_page(vma, address, pteval); 5753218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 576b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 577b46e756fSKirill A. Shutemov goto out; 578b46e756fSKirill A. Shutemov } 579b46e756fSKirill A. Shutemov 5808dd1e896SVishal Moola (Oracle) folio = page_folio(page); 5818dd1e896SVishal Moola (Oracle) VM_BUG_ON_FOLIO(!folio_test_anon(folio), folio); 582b46e756fSKirill A. Shutemov 583d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 584d8ea7cc8SZach O'Keefe ++shared; 585d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 586d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 58771a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 588e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 58971a2c112SKirill A. Shutemov goto out; 59071a2c112SKirill A. Shutemov } 591d8ea7cc8SZach O'Keefe } 59271a2c112SKirill A. Shutemov 5938dd1e896SVishal Moola (Oracle) if (folio_test_large(folio)) { 5948dd1e896SVishal Moola (Oracle) struct folio *f; 5955503fbf2SKirill A. Shutemov 5965503fbf2SKirill A. Shutemov /* 5975503fbf2SKirill A. Shutemov * Check if we have dealt with the compound page 5985503fbf2SKirill A. Shutemov * already 5995503fbf2SKirill A. Shutemov */ 6008dd1e896SVishal Moola (Oracle) list_for_each_entry(f, compound_pagelist, lru) { 6018dd1e896SVishal Moola (Oracle) if (folio == f) 6025503fbf2SKirill A. Shutemov goto next; 6035503fbf2SKirill A. Shutemov } 6045503fbf2SKirill A. Shutemov } 6055503fbf2SKirill A. Shutemov 606b46e756fSKirill A. Shutemov /* 607b46e756fSKirill A. Shutemov * We can do it before isolate_lru_page because the 608b46e756fSKirill A. Shutemov * page can't be freed from under us. NOTE: PG_lock 609b46e756fSKirill A. Shutemov * is needed to serialize against split_huge_page 610b46e756fSKirill A. Shutemov * when invoked from the VM. 611b46e756fSKirill A. Shutemov */ 6128dd1e896SVishal Moola (Oracle) if (!folio_trylock(folio)) { 613b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 614b46e756fSKirill A. Shutemov goto out; 615b46e756fSKirill A. Shutemov } 616b46e756fSKirill A. Shutemov 617b46e756fSKirill A. Shutemov /* 6189445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 6199445689fSKirill A. Shutemov * 6209445689fSKirill A. Shutemov * The page table that maps the page has been already unlinked 6219445689fSKirill A. Shutemov * from the page table tree and this process cannot get 622f0953a1bSIngo Molnar * an additional pin on the page. 6239445689fSKirill A. Shutemov * 6249445689fSKirill A. Shutemov * New pins can come later if the page is shared across fork, 6259445689fSKirill A. Shutemov * but not from this process. The other process cannot write to 6269445689fSKirill A. Shutemov * the page, only trigger CoW. 627b46e756fSKirill A. Shutemov */ 628dbf85c21SVishal Moola (Oracle) if (!is_refcount_suitable(folio)) { 6298dd1e896SVishal Moola (Oracle) folio_unlock(folio); 630b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 631b46e756fSKirill A. Shutemov goto out; 632b46e756fSKirill A. Shutemov } 633b46e756fSKirill A. Shutemov 634b46e756fSKirill A. Shutemov /* 635b46e756fSKirill A. Shutemov * Isolate the page to avoid collapsing an hugepage 636b46e756fSKirill A. Shutemov * currently in use by the VM. 637b46e756fSKirill A. Shutemov */ 6388dd1e896SVishal Moola (Oracle) if (!folio_isolate_lru(folio)) { 6398dd1e896SVishal Moola (Oracle) folio_unlock(folio); 640b46e756fSKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 641b46e756fSKirill A. Shutemov goto out; 642b46e756fSKirill A. Shutemov } 6438dd1e896SVishal Moola (Oracle) node_stat_mod_folio(folio, 6448dd1e896SVishal Moola (Oracle) NR_ISOLATED_ANON + folio_is_file_lru(folio), 6458dd1e896SVishal Moola (Oracle) folio_nr_pages(folio)); 6468dd1e896SVishal Moola (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 6478dd1e896SVishal Moola (Oracle) VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); 648b46e756fSKirill A. Shutemov 6498dd1e896SVishal Moola (Oracle) if (folio_test_large(folio)) 6508dd1e896SVishal Moola (Oracle) list_add_tail(&folio->lru, compound_pagelist); 6515503fbf2SKirill A. Shutemov next: 652d8ea7cc8SZach O'Keefe /* 653d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 654d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 655d8ea7cc8SZach O'Keefe */ 656d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 6578dd1e896SVishal Moola (Oracle) (pte_young(pteval) || folio_test_young(folio) || 6588dd1e896SVishal Moola (Oracle) folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm, 659d8ea7cc8SZach O'Keefe address))) 6600db501f7SEbru Akagunduz referenced++; 6615503fbf2SKirill A. Shutemov 6625503fbf2SKirill A. Shutemov if (pte_write(pteval)) 6635503fbf2SKirill A. Shutemov writable = true; 664b46e756fSKirill A. Shutemov } 66574e579bfSMiaohe Lin 66674e579bfSMiaohe Lin if (unlikely(!writable)) { 66774e579bfSMiaohe Lin result = SCAN_PAGE_RO; 668d8ea7cc8SZach O'Keefe } else if (unlikely(cc->is_khugepaged && !referenced)) { 66974e579bfSMiaohe Lin result = SCAN_LACK_REFERENCED_PAGE; 67074e579bfSMiaohe Lin } else { 671b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 6728dd1e896SVishal Moola (Oracle) trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero, 673b46e756fSKirill A. Shutemov referenced, writable, result); 67450ad2f24SZach O'Keefe return result; 675b46e756fSKirill A. Shutemov } 676b46e756fSKirill A. Shutemov out: 6775503fbf2SKirill A. Shutemov release_pte_pages(pte, _pte, compound_pagelist); 6788dd1e896SVishal Moola (Oracle) trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero, 679b46e756fSKirill A. Shutemov referenced, writable, result); 68050ad2f24SZach O'Keefe return result; 681b46e756fSKirill A. Shutemov } 682b46e756fSKirill A. Shutemov 68398c76c9fSJiaqi Yan static void __collapse_huge_page_copy_succeeded(pte_t *pte, 684b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 685b46e756fSKirill A. Shutemov unsigned long address, 6865503fbf2SKirill A. Shutemov spinlock_t *ptl, 6875503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 688b46e756fSKirill A. Shutemov { 68998c76c9fSJiaqi Yan struct page *src_page; 69098c76c9fSJiaqi Yan struct page *tmp; 691b46e756fSKirill A. Shutemov pte_t *_pte; 69298c76c9fSJiaqi Yan pte_t pteval; 693b46e756fSKirill A. Shutemov 69498c76c9fSJiaqi Yan for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 69598c76c9fSJiaqi Yan _pte++, address += PAGE_SIZE) { 696c33c7948SRyan Roberts pteval = ptep_get(_pte); 697b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 698b46e756fSKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 699b46e756fSKirill A. Shutemov if (is_zero_pfn(pte_pfn(pteval))) { 700b46e756fSKirill A. Shutemov /* 701b46e756fSKirill A. Shutemov * ptl mostly unnecessary. 702b46e756fSKirill A. Shutemov */ 703b46e756fSKirill A. Shutemov spin_lock(ptl); 70408d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 705b46e756fSKirill A. Shutemov spin_unlock(ptl); 7066080d19fSxu xin ksm_might_unmap_zero_page(vma->vm_mm, pteval); 707b46e756fSKirill A. Shutemov } 708b46e756fSKirill A. Shutemov } else { 709b46e756fSKirill A. Shutemov src_page = pte_page(pteval); 7105503fbf2SKirill A. Shutemov if (!PageCompound(src_page)) 711b46e756fSKirill A. Shutemov release_pte_page(src_page); 712b46e756fSKirill A. Shutemov /* 713b46e756fSKirill A. Shutemov * ptl mostly unnecessary, but preempt has to 714b46e756fSKirill A. Shutemov * be disabled to update the per-cpu stats 715b46e756fSKirill A. Shutemov * inside page_remove_rmap(). 716b46e756fSKirill A. Shutemov */ 717b46e756fSKirill A. Shutemov spin_lock(ptl); 71808d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 719cea86fe2SHugh Dickins page_remove_rmap(src_page, vma, false); 720b46e756fSKirill A. Shutemov spin_unlock(ptl); 721b46e756fSKirill A. Shutemov free_page_and_swap_cache(src_page); 722b46e756fSKirill A. Shutemov } 723b46e756fSKirill A. Shutemov } 7245503fbf2SKirill A. Shutemov 7255503fbf2SKirill A. Shutemov list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { 7265503fbf2SKirill A. Shutemov list_del(&src_page->lru); 7271baec203SMiaohe Lin mod_node_page_state(page_pgdat(src_page), 7281baec203SMiaohe Lin NR_ISOLATED_ANON + page_is_file_lru(src_page), 7291baec203SMiaohe Lin -compound_nr(src_page)); 7301baec203SMiaohe Lin unlock_page(src_page); 7311baec203SMiaohe Lin free_swap_cache(src_page); 7321baec203SMiaohe Lin putback_lru_page(src_page); 7335503fbf2SKirill A. Shutemov } 734b46e756fSKirill A. Shutemov } 735b46e756fSKirill A. Shutemov 73698c76c9fSJiaqi Yan static void __collapse_huge_page_copy_failed(pte_t *pte, 73798c76c9fSJiaqi Yan pmd_t *pmd, 73898c76c9fSJiaqi Yan pmd_t orig_pmd, 73998c76c9fSJiaqi Yan struct vm_area_struct *vma, 74098c76c9fSJiaqi Yan struct list_head *compound_pagelist) 74198c76c9fSJiaqi Yan { 74298c76c9fSJiaqi Yan spinlock_t *pmd_ptl; 74398c76c9fSJiaqi Yan 74498c76c9fSJiaqi Yan /* 74598c76c9fSJiaqi Yan * Re-establish the PMD to point to the original page table 74698c76c9fSJiaqi Yan * entry. Restoring PMD needs to be done prior to releasing 74798c76c9fSJiaqi Yan * pages. Since pages are still isolated and locked here, 74898c76c9fSJiaqi Yan * acquiring anon_vma_lock_write is unnecessary. 74998c76c9fSJiaqi Yan */ 75098c76c9fSJiaqi Yan pmd_ptl = pmd_lock(vma->vm_mm, pmd); 75198c76c9fSJiaqi Yan pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd)); 75298c76c9fSJiaqi Yan spin_unlock(pmd_ptl); 75398c76c9fSJiaqi Yan /* 75498c76c9fSJiaqi Yan * Release both raw and compound pages isolated 75598c76c9fSJiaqi Yan * in __collapse_huge_page_isolate. 75698c76c9fSJiaqi Yan */ 75798c76c9fSJiaqi Yan release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist); 75898c76c9fSJiaqi Yan } 75998c76c9fSJiaqi Yan 76098c76c9fSJiaqi Yan /* 76198c76c9fSJiaqi Yan * __collapse_huge_page_copy - attempts to copy memory contents from raw 76298c76c9fSJiaqi Yan * pages to a hugepage. Cleans up the raw pages if copying succeeds; 76398c76c9fSJiaqi Yan * otherwise restores the original page table and releases isolated raw pages. 76498c76c9fSJiaqi Yan * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC. 76598c76c9fSJiaqi Yan * 76698c76c9fSJiaqi Yan * @pte: starting of the PTEs to copy from 76798c76c9fSJiaqi Yan * @page: the new hugepage to copy contents to 76898c76c9fSJiaqi Yan * @pmd: pointer to the new hugepage's PMD 76998c76c9fSJiaqi Yan * @orig_pmd: the original raw pages' PMD 77098c76c9fSJiaqi Yan * @vma: the original raw pages' virtual memory area 77198c76c9fSJiaqi Yan * @address: starting address to copy 77298c76c9fSJiaqi Yan * @ptl: lock on raw pages' PTEs 77398c76c9fSJiaqi Yan * @compound_pagelist: list that stores compound pages 77498c76c9fSJiaqi Yan */ 77598c76c9fSJiaqi Yan static int __collapse_huge_page_copy(pte_t *pte, 77698c76c9fSJiaqi Yan struct page *page, 77798c76c9fSJiaqi Yan pmd_t *pmd, 77898c76c9fSJiaqi Yan pmd_t orig_pmd, 77998c76c9fSJiaqi Yan struct vm_area_struct *vma, 78098c76c9fSJiaqi Yan unsigned long address, 78198c76c9fSJiaqi Yan spinlock_t *ptl, 78298c76c9fSJiaqi Yan struct list_head *compound_pagelist) 78398c76c9fSJiaqi Yan { 78498c76c9fSJiaqi Yan struct page *src_page; 78598c76c9fSJiaqi Yan pte_t *_pte; 78698c76c9fSJiaqi Yan pte_t pteval; 78798c76c9fSJiaqi Yan unsigned long _address; 78898c76c9fSJiaqi Yan int result = SCAN_SUCCEED; 78998c76c9fSJiaqi Yan 79098c76c9fSJiaqi Yan /* 79198c76c9fSJiaqi Yan * Copying pages' contents is subject to memory poison at any iteration. 79298c76c9fSJiaqi Yan */ 79398c76c9fSJiaqi Yan for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; 79498c76c9fSJiaqi Yan _pte++, page++, _address += PAGE_SIZE) { 795c33c7948SRyan Roberts pteval = ptep_get(_pte); 79698c76c9fSJiaqi Yan if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 79798c76c9fSJiaqi Yan clear_user_highpage(page, _address); 79898c76c9fSJiaqi Yan continue; 79998c76c9fSJiaqi Yan } 80098c76c9fSJiaqi Yan src_page = pte_page(pteval); 80198c76c9fSJiaqi Yan if (copy_mc_user_highpage(page, src_page, _address, vma) > 0) { 80298c76c9fSJiaqi Yan result = SCAN_COPY_MC; 80398c76c9fSJiaqi Yan break; 80498c76c9fSJiaqi Yan } 80598c76c9fSJiaqi Yan } 80698c76c9fSJiaqi Yan 80798c76c9fSJiaqi Yan if (likely(result == SCAN_SUCCEED)) 80898c76c9fSJiaqi Yan __collapse_huge_page_copy_succeeded(pte, vma, address, ptl, 80998c76c9fSJiaqi Yan compound_pagelist); 81098c76c9fSJiaqi Yan else 81198c76c9fSJiaqi Yan __collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma, 81298c76c9fSJiaqi Yan compound_pagelist); 81398c76c9fSJiaqi Yan 81498c76c9fSJiaqi Yan return result; 81598c76c9fSJiaqi Yan } 81698c76c9fSJiaqi Yan 817b46e756fSKirill A. Shutemov static void khugepaged_alloc_sleep(void) 818b46e756fSKirill A. Shutemov { 819b46e756fSKirill A. Shutemov DEFINE_WAIT(wait); 820b46e756fSKirill A. Shutemov 821b46e756fSKirill A. Shutemov add_wait_queue(&khugepaged_wait, &wait); 822f5d39b02SPeter Zijlstra __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE); 823f5d39b02SPeter Zijlstra schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 824b46e756fSKirill A. Shutemov remove_wait_queue(&khugepaged_wait, &wait); 825b46e756fSKirill A. Shutemov } 826b46e756fSKirill A. Shutemov 82734d6b470SZach O'Keefe struct collapse_control khugepaged_collapse_control = { 828d8ea7cc8SZach O'Keefe .is_khugepaged = true, 82934d6b470SZach O'Keefe }; 830b46e756fSKirill A. Shutemov 8317d2c4385SZach O'Keefe static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc) 832b46e756fSKirill A. Shutemov { 833b46e756fSKirill A. Shutemov int i; 834b46e756fSKirill A. Shutemov 835b46e756fSKirill A. Shutemov /* 836a5f5f91dSMel Gorman * If node_reclaim_mode is disabled, then no extra effort is made to 837b46e756fSKirill A. Shutemov * allocate memory locally. 838b46e756fSKirill A. Shutemov */ 839202e35dbSDave Hansen if (!node_reclaim_enabled()) 840b46e756fSKirill A. Shutemov return false; 841b46e756fSKirill A. Shutemov 842b46e756fSKirill A. Shutemov /* If there is a count for this node already, it must be acceptable */ 84334d6b470SZach O'Keefe if (cc->node_load[nid]) 844b46e756fSKirill A. Shutemov return false; 845b46e756fSKirill A. Shutemov 846b46e756fSKirill A. Shutemov for (i = 0; i < MAX_NUMNODES; i++) { 84734d6b470SZach O'Keefe if (!cc->node_load[i]) 848b46e756fSKirill A. Shutemov continue; 849a55c7454SMatt Fleming if (node_distance(nid, i) > node_reclaim_distance) 850b46e756fSKirill A. Shutemov return true; 851b46e756fSKirill A. Shutemov } 852b46e756fSKirill A. Shutemov return false; 853b46e756fSKirill A. Shutemov } 854b46e756fSKirill A. Shutemov 8551064026bSYang Shi #define khugepaged_defrag() \ 8561064026bSYang Shi (transparent_hugepage_flags & \ 8571064026bSYang Shi (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)) 8581064026bSYang Shi 859b46e756fSKirill A. Shutemov /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 860b46e756fSKirill A. Shutemov static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 861b46e756fSKirill A. Shutemov { 86225160354SVlastimil Babka return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 863b46e756fSKirill A. Shutemov } 864b46e756fSKirill A. Shutemov 865b46e756fSKirill A. Shutemov #ifdef CONFIG_NUMA 8667d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 867b46e756fSKirill A. Shutemov { 868b46e756fSKirill A. Shutemov int nid, target_node = 0, max_value = 0; 869b46e756fSKirill A. Shutemov 870b46e756fSKirill A. Shutemov /* find first node with max normal pages hit */ 871b46e756fSKirill A. Shutemov for (nid = 0; nid < MAX_NUMNODES; nid++) 87234d6b470SZach O'Keefe if (cc->node_load[nid] > max_value) { 87334d6b470SZach O'Keefe max_value = cc->node_load[nid]; 874b46e756fSKirill A. Shutemov target_node = nid; 875b46e756fSKirill A. Shutemov } 876b46e756fSKirill A. Shutemov 877e031ff96SYang Shi for_each_online_node(nid) { 878e031ff96SYang Shi if (max_value == cc->node_load[nid]) 879e031ff96SYang Shi node_set(nid, cc->alloc_nmask); 880b46e756fSKirill A. Shutemov } 881b46e756fSKirill A. Shutemov 882b46e756fSKirill A. Shutemov return target_node; 883b46e756fSKirill A. Shutemov } 884c6a7f445SYang Shi #else 8857d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 886b46e756fSKirill A. Shutemov { 887c6a7f445SYang Shi return 0; 888b46e756fSKirill A. Shutemov } 889c6a7f445SYang Shi #endif 890b46e756fSKirill A. Shutemov 891b455f39dSVishal Moola (Oracle) static bool hpage_collapse_alloc_folio(struct folio **folio, gfp_t gfp, int node, 892e031ff96SYang Shi nodemask_t *nmask) 893b46e756fSKirill A. Shutemov { 894b455f39dSVishal Moola (Oracle) *folio = __folio_alloc(gfp, HPAGE_PMD_ORDER, node, nmask); 895b455f39dSVishal Moola (Oracle) 896b455f39dSVishal Moola (Oracle) if (unlikely(!*folio)) { 897b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 8989710a78aSZach O'Keefe return false; 899b46e756fSKirill A. Shutemov } 900b46e756fSKirill A. Shutemov 901b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 902b46e756fSKirill A. Shutemov return true; 903b46e756fSKirill A. Shutemov } 904b46e756fSKirill A. Shutemov 905b46e756fSKirill A. Shutemov /* 906c1e8d7c6SMichel Lespinasse * If mmap_lock temporarily dropped, revalidate vma 907c1e8d7c6SMichel Lespinasse * before taking mmap_lock. 90850ad2f24SZach O'Keefe * Returns enum scan_result value. 909b46e756fSKirill A. Shutemov */ 910b46e756fSKirill A. Shutemov 911c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 91234488399SZach O'Keefe bool expect_anon, 913a7f4e6e4SZach O'Keefe struct vm_area_struct **vmap, 914a7f4e6e4SZach O'Keefe struct collapse_control *cc) 915b46e756fSKirill A. Shutemov { 916b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 917b46e756fSKirill A. Shutemov 9187d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 919b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 920b46e756fSKirill A. Shutemov 921c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 922b46e756fSKirill A. Shutemov if (!vma) 923b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 924b46e756fSKirill A. Shutemov 9254fa6893fSYang Shi if (!transhuge_vma_suitable(vma, address)) 926b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 927a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, 928a7f4e6e4SZach O'Keefe cc->is_khugepaged)) 929b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 930f707fa49SYang Shi /* 931f707fa49SYang Shi * Anon VMA expected, the address may be unmapped then 932f707fa49SYang Shi * remapped to file after khugepaged reaquired the mmap_lock. 933f707fa49SYang Shi * 934f707fa49SYang Shi * hugepage_vma_check may return true for qualified file 935f707fa49SYang Shi * vmas. 936f707fa49SYang Shi */ 93734488399SZach O'Keefe if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) 93834488399SZach O'Keefe return SCAN_PAGE_ANON; 93950ad2f24SZach O'Keefe return SCAN_SUCCEED; 940b46e756fSKirill A. Shutemov } 941b46e756fSKirill A. Shutemov 94250722804SZach O'Keefe static int find_pmd_or_thp_or_none(struct mm_struct *mm, 94350722804SZach O'Keefe unsigned long address, 94450722804SZach O'Keefe pmd_t **pmd) 94550722804SZach O'Keefe { 94650722804SZach O'Keefe pmd_t pmde; 94750722804SZach O'Keefe 94850722804SZach O'Keefe *pmd = mm_find_pmd(mm, address); 94950722804SZach O'Keefe if (!*pmd) 95050722804SZach O'Keefe return SCAN_PMD_NULL; 95150722804SZach O'Keefe 952dab6e717SPeter Zijlstra pmde = pmdp_get_lockless(*pmd); 95334488399SZach O'Keefe if (pmd_none(pmde)) 95434488399SZach O'Keefe return SCAN_PMD_NONE; 955edb5d0cfSZach O'Keefe if (!pmd_present(pmde)) 956edb5d0cfSZach O'Keefe return SCAN_PMD_NULL; 95750722804SZach O'Keefe if (pmd_trans_huge(pmde)) 95850722804SZach O'Keefe return SCAN_PMD_MAPPED; 959edb5d0cfSZach O'Keefe if (pmd_devmap(pmde)) 960edb5d0cfSZach O'Keefe return SCAN_PMD_NULL; 96150722804SZach O'Keefe if (pmd_bad(pmde)) 96250722804SZach O'Keefe return SCAN_PMD_NULL; 96350722804SZach O'Keefe return SCAN_SUCCEED; 96450722804SZach O'Keefe } 96550722804SZach O'Keefe 96650722804SZach O'Keefe static int check_pmd_still_valid(struct mm_struct *mm, 96750722804SZach O'Keefe unsigned long address, 96850722804SZach O'Keefe pmd_t *pmd) 96950722804SZach O'Keefe { 97050722804SZach O'Keefe pmd_t *new_pmd; 97150722804SZach O'Keefe int result = find_pmd_or_thp_or_none(mm, address, &new_pmd); 97250722804SZach O'Keefe 97350722804SZach O'Keefe if (result != SCAN_SUCCEED) 97450722804SZach O'Keefe return result; 97550722804SZach O'Keefe if (new_pmd != pmd) 97650722804SZach O'Keefe return SCAN_FAIL; 97750722804SZach O'Keefe return SCAN_SUCCEED; 978b46e756fSKirill A. Shutemov } 979b46e756fSKirill A. Shutemov 980b46e756fSKirill A. Shutemov /* 981b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 9827d2c4385SZach O'Keefe * Only done if hpage_collapse_scan_pmd believes it is worthwhile. 983b46e756fSKirill A. Shutemov * 9844d928e20SMiaohe Lin * Called and returns without pte mapped or spinlocks held. 985895f5ee4SHugh Dickins * Returns result: if not SCAN_SUCCEED, mmap_lock has been released. 986b46e756fSKirill A. Shutemov */ 98750ad2f24SZach O'Keefe static int __collapse_huge_page_swapin(struct mm_struct *mm, 988b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 9892b635dd3SWill Deacon unsigned long haddr, pmd_t *pmd, 9900db501f7SEbru Akagunduz int referenced) 991b46e756fSKirill A. Shutemov { 9922b740303SSouptick Joarder int swapped_in = 0; 9932b740303SSouptick Joarder vm_fault_t ret = 0; 9942b635dd3SWill Deacon unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); 995895f5ee4SHugh Dickins int result; 996895f5ee4SHugh Dickins pte_t *pte = NULL; 997c7ad0880SHugh Dickins spinlock_t *ptl; 9982b635dd3SWill Deacon 9992b635dd3SWill Deacon for (address = haddr; address < end; address += PAGE_SIZE) { 100082b0f8c3SJan Kara struct vm_fault vmf = { 1001b46e756fSKirill A. Shutemov .vma = vma, 1002b46e756fSKirill A. Shutemov .address = address, 1003895f5ee4SHugh Dickins .pgoff = linear_page_index(vma, address), 1004b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 1005b46e756fSKirill A. Shutemov .pmd = pmd, 1006b46e756fSKirill A. Shutemov }; 1007b46e756fSKirill A. Shutemov 1008895f5ee4SHugh Dickins if (!pte++) { 1009c7ad0880SHugh Dickins pte = pte_offset_map_nolock(mm, pmd, address, &ptl); 1010895f5ee4SHugh Dickins if (!pte) { 1011895f5ee4SHugh Dickins mmap_read_unlock(mm); 1012895f5ee4SHugh Dickins result = SCAN_PMD_NULL; 1013895f5ee4SHugh Dickins goto out; 10142b635dd3SWill Deacon } 1015895f5ee4SHugh Dickins } 1016895f5ee4SHugh Dickins 1017c7ad0880SHugh Dickins vmf.orig_pte = ptep_get_lockless(pte); 1018895f5ee4SHugh Dickins if (!is_swap_pte(vmf.orig_pte)) 1019895f5ee4SHugh Dickins continue; 1020895f5ee4SHugh Dickins 1021895f5ee4SHugh Dickins vmf.pte = pte; 1022c7ad0880SHugh Dickins vmf.ptl = ptl; 10232994302bSJan Kara ret = do_swap_page(&vmf); 1024895f5ee4SHugh Dickins /* Which unmaps pte (after perhaps re-checking the entry) */ 1025895f5ee4SHugh Dickins pte = NULL; 10260db501f7SEbru Akagunduz 10274d928e20SMiaohe Lin /* 10284d928e20SMiaohe Lin * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. 10294d928e20SMiaohe Lin * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because 10304d928e20SMiaohe Lin * we do not retry here and swap entry will remain in pagetable 10314d928e20SMiaohe Lin * resulting in later failure. 10324d928e20SMiaohe Lin */ 1033b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 103450ad2f24SZach O'Keefe /* Likely, but not guaranteed, that page lock failed */ 1035895f5ee4SHugh Dickins result = SCAN_PAGE_LOCK; 1036895f5ee4SHugh Dickins goto out; 103747f863eaSEbru Akagunduz } 1038b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 10394d928e20SMiaohe Lin mmap_read_unlock(mm); 1040895f5ee4SHugh Dickins result = SCAN_FAIL; 1041895f5ee4SHugh Dickins goto out; 1042b46e756fSKirill A. Shutemov } 10434d928e20SMiaohe Lin swapped_in++; 1044b46e756fSKirill A. Shutemov } 1045ae2c5d80SKirill A. Shutemov 1046895f5ee4SHugh Dickins if (pte) 1047895f5ee4SHugh Dickins pte_unmap(pte); 1048895f5ee4SHugh Dickins 10491fec6890SMatthew Wilcox (Oracle) /* Drain LRU cache to remove extra pin on the swapped in pages */ 1050ae2c5d80SKirill A. Shutemov if (swapped_in) 1051ae2c5d80SKirill A. Shutemov lru_add_drain(); 1052ae2c5d80SKirill A. Shutemov 1053895f5ee4SHugh Dickins result = SCAN_SUCCEED; 1054895f5ee4SHugh Dickins out: 1055895f5ee4SHugh Dickins trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result); 1056895f5ee4SHugh Dickins return result; 1057b46e756fSKirill A. Shutemov } 1058b46e756fSKirill A. Shutemov 10599710a78aSZach O'Keefe static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, 10609710a78aSZach O'Keefe struct collapse_control *cc) 10619710a78aSZach O'Keefe { 10627d8faaf1SZach O'Keefe gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() : 1063e031ff96SYang Shi GFP_TRANSHUGE); 10647d2c4385SZach O'Keefe int node = hpage_collapse_find_target_node(cc); 106594c02ad7SPeter Xu struct folio *folio; 10669710a78aSZach O'Keefe 1067b455f39dSVishal Moola (Oracle) if (!hpage_collapse_alloc_folio(&folio, gfp, node, &cc->alloc_nmask)) { 1068b455f39dSVishal Moola (Oracle) *hpage = NULL; 10699710a78aSZach O'Keefe return SCAN_ALLOC_HUGE_PAGE_FAIL; 1070b455f39dSVishal Moola (Oracle) } 107194c02ad7SPeter Xu 107294c02ad7SPeter Xu if (unlikely(mem_cgroup_charge(folio, mm, gfp))) { 107394c02ad7SPeter Xu folio_put(folio); 107494c02ad7SPeter Xu *hpage = NULL; 10759710a78aSZach O'Keefe return SCAN_CGROUP_CHARGE_FAIL; 107694c02ad7SPeter Xu } 107794c02ad7SPeter Xu 1078b455f39dSVishal Moola (Oracle) count_memcg_folio_events(folio, THP_COLLAPSE_ALLOC, 1); 1079b455f39dSVishal Moola (Oracle) 1080b455f39dSVishal Moola (Oracle) *hpage = folio_page(folio, 0); 10819710a78aSZach O'Keefe return SCAN_SUCCEED; 10829710a78aSZach O'Keefe } 10839710a78aSZach O'Keefe 108450ad2f24SZach O'Keefe static int collapse_huge_page(struct mm_struct *mm, unsigned long address, 108550ad2f24SZach O'Keefe int referenced, int unmapped, 108650ad2f24SZach O'Keefe struct collapse_control *cc) 1087b46e756fSKirill A. Shutemov { 10885503fbf2SKirill A. Shutemov LIST_HEAD(compound_pagelist); 1089b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 1090b46e756fSKirill A. Shutemov pte_t *pte; 1091b46e756fSKirill A. Shutemov pgtable_t pgtable; 109250ad2f24SZach O'Keefe struct page *hpage; 1093b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 109450ad2f24SZach O'Keefe int result = SCAN_FAIL; 1095c131f751SKirill A. Shutemov struct vm_area_struct *vma; 1096ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1097b46e756fSKirill A. Shutemov 1098b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1099b46e756fSKirill A. Shutemov 1100988ddb71SKirill A. Shutemov /* 1101c1e8d7c6SMichel Lespinasse * Before allocating the hugepage, release the mmap_lock read lock. 1102988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 1103c1e8d7c6SMichel Lespinasse * sync compaction, and we do not need to hold the mmap_lock during 1104988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 1105988ddb71SKirill A. Shutemov */ 1106d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1107b46e756fSKirill A. Shutemov 110850ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 11099710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1110b46e756fSKirill A. Shutemov goto out_nolock; 1111b46e756fSKirill A. Shutemov 1112d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 111334488399SZach O'Keefe result = hugepage_vma_revalidate(mm, address, true, &vma, cc); 111450ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) { 1115d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1116b46e756fSKirill A. Shutemov goto out_nolock; 1117b46e756fSKirill A. Shutemov } 1118b46e756fSKirill A. Shutemov 111950722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 112050722804SZach O'Keefe if (result != SCAN_SUCCEED) { 1121d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1122b46e756fSKirill A. Shutemov goto out_nolock; 1123b46e756fSKirill A. Shutemov } 1124b46e756fSKirill A. Shutemov 112550ad2f24SZach O'Keefe if (unmapped) { 1126b46e756fSKirill A. Shutemov /* 112750ad2f24SZach O'Keefe * __collapse_huge_page_swapin will return with mmap_lock 112850ad2f24SZach O'Keefe * released when it fails. So we jump out_nolock directly in 112950ad2f24SZach O'Keefe * that case. Continuing to collapse causes inconsistency. 1130b46e756fSKirill A. Shutemov */ 113150ad2f24SZach O'Keefe result = __collapse_huge_page_swapin(mm, vma, address, pmd, 113250ad2f24SZach O'Keefe referenced); 113350ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 1134b46e756fSKirill A. Shutemov goto out_nolock; 1135b46e756fSKirill A. Shutemov } 1136b46e756fSKirill A. Shutemov 1137d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1138b46e756fSKirill A. Shutemov /* 1139b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 1140b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 1141b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 1142b46e756fSKirill A. Shutemov */ 1143d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 114434488399SZach O'Keefe result = hugepage_vma_revalidate(mm, address, true, &vma, cc); 114550ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 114618d24a7cSMiaohe Lin goto out_up_write; 1147b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 114850722804SZach O'Keefe result = check_pmd_still_valid(mm, address, pmd); 114950722804SZach O'Keefe if (result != SCAN_SUCCEED) 115018d24a7cSMiaohe Lin goto out_up_write; 1151b46e756fSKirill A. Shutemov 115255fd6fccSSuren Baghdasaryan vma_start_write(vma); 1153b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1154b46e756fSKirill A. Shutemov 11557d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address, 11567d4a8be0SAlistair Popple address + HPAGE_PMD_SIZE); 1157ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1158ec649c9dSVille Syrjälä 1159b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1160b46e756fSKirill A. Shutemov /* 116170cbc3ccSYang Shi * This removes any huge TLB entry from the CPU so we won't allow 116270cbc3ccSYang Shi * huge and small TLB entries for the same virtual address to 116370cbc3ccSYang Shi * avoid the risk of CPU bugs in that area. 116470cbc3ccSYang Shi * 116570cbc3ccSYang Shi * Parallel fast GUP is fine since fast GUP will back off when 116670cbc3ccSYang Shi * it detects PMD is changed. 1167b46e756fSKirill A. Shutemov */ 1168b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1169b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1170ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 11712ba99c5eSJann Horn tlb_remove_table_sync_one(); 1172b46e756fSKirill A. Shutemov 1173895f5ee4SHugh Dickins pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl); 1174895f5ee4SHugh Dickins if (pte) { 1175d8ea7cc8SZach O'Keefe result = __collapse_huge_page_isolate(vma, address, pte, cc, 11765503fbf2SKirill A. Shutemov &compound_pagelist); 1177b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1178895f5ee4SHugh Dickins } else { 1179895f5ee4SHugh Dickins result = SCAN_PMD_NULL; 1180895f5ee4SHugh Dickins } 1181b46e756fSKirill A. Shutemov 118250ad2f24SZach O'Keefe if (unlikely(result != SCAN_SUCCEED)) { 1183895f5ee4SHugh Dickins if (pte) 1184b46e756fSKirill A. Shutemov pte_unmap(pte); 1185b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1186b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1187b46e756fSKirill A. Shutemov /* 1188b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1189b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1190b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1191b46e756fSKirill A. Shutemov */ 1192b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1193b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1194b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 119518d24a7cSMiaohe Lin goto out_up_write; 1196b46e756fSKirill A. Shutemov } 1197b46e756fSKirill A. Shutemov 1198b46e756fSKirill A. Shutemov /* 1199b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1200b46e756fSKirill A. Shutemov * can't run anymore. 1201b46e756fSKirill A. Shutemov */ 1202b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1203b46e756fSKirill A. Shutemov 120498c76c9fSJiaqi Yan result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd, 120598c76c9fSJiaqi Yan vma, address, pte_ptl, 12065503fbf2SKirill A. Shutemov &compound_pagelist); 1207b46e756fSKirill A. Shutemov pte_unmap(pte); 120898c76c9fSJiaqi Yan if (unlikely(result != SCAN_SUCCEED)) 120998c76c9fSJiaqi Yan goto out_up_write; 121098c76c9fSJiaqi Yan 1211588d01f9SMiaohe Lin /* 1212588d01f9SMiaohe Lin * spin_lock() below is not the equivalent of smp_wmb(), but 1213588d01f9SMiaohe Lin * the smp_wmb() inside __SetPageUptodate() can be reused to 1214588d01f9SMiaohe Lin * avoid the copy_huge_page writes to become visible after 1215588d01f9SMiaohe Lin * the set_pmd_at() write. 1216588d01f9SMiaohe Lin */ 121750ad2f24SZach O'Keefe __SetPageUptodate(hpage); 1218b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1219b46e756fSKirill A. Shutemov 122050ad2f24SZach O'Keefe _pmd = mk_huge_pmd(hpage, vma->vm_page_prot); 1221f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1222b46e756fSKirill A. Shutemov 1223b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1224b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 122550ad2f24SZach O'Keefe page_add_new_anon_rmap(hpage, vma, address); 122650ad2f24SZach O'Keefe lru_cache_add_inactive_or_unevictable(hpage, vma); 1227b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1228b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1229b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1230b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1231b46e756fSKirill A. Shutemov 123250ad2f24SZach O'Keefe hpage = NULL; 1233b46e756fSKirill A. Shutemov 1234b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1235b46e756fSKirill A. Shutemov out_up_write: 1236d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 1237b46e756fSKirill A. Shutemov out_nolock: 12387cb1d7efSPeter Xu if (hpage) 123950ad2f24SZach O'Keefe put_page(hpage); 124050ad2f24SZach O'Keefe trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); 124150ad2f24SZach O'Keefe return result; 1242b46e756fSKirill A. Shutemov } 1243b46e756fSKirill A. Shutemov 12447d2c4385SZach O'Keefe static int hpage_collapse_scan_pmd(struct mm_struct *mm, 1245b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 124650ad2f24SZach O'Keefe unsigned long address, bool *mmap_locked, 124734d6b470SZach O'Keefe struct collapse_control *cc) 1248b46e756fSKirill A. Shutemov { 1249b46e756fSKirill A. Shutemov pmd_t *pmd; 1250b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 125150ad2f24SZach O'Keefe int result = SCAN_FAIL, referenced = 0; 125271a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0; 1253b46e756fSKirill A. Shutemov struct page *page = NULL; 12545c07ebb3SVishal Moola (Oracle) struct folio *folio = NULL; 1255b46e756fSKirill A. Shutemov unsigned long _address; 1256b46e756fSKirill A. Shutemov spinlock_t *ptl; 1257b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 12580db501f7SEbru Akagunduz bool writable = false; 1259b46e756fSKirill A. Shutemov 1260b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1261b46e756fSKirill A. Shutemov 126250722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 126350722804SZach O'Keefe if (result != SCAN_SUCCEED) 1264b46e756fSKirill A. Shutemov goto out; 1265b46e756fSKirill A. Shutemov 126634d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1267e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 1268b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1269895f5ee4SHugh Dickins if (!pte) { 1270895f5ee4SHugh Dickins result = SCAN_PMD_NULL; 1271895f5ee4SHugh Dickins goto out; 1272895f5ee4SHugh Dickins } 1273895f5ee4SHugh Dickins 1274b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; 1275b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1276c33c7948SRyan Roberts pte_t pteval = ptep_get(_pte); 1277b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1278d8ea7cc8SZach O'Keefe ++unmapped; 1279d8ea7cc8SZach O'Keefe if (!cc->is_khugepaged || 1280d8ea7cc8SZach O'Keefe unmapped <= khugepaged_max_ptes_swap) { 1281e1e267c7SPeter Xu /* 1282e1e267c7SPeter Xu * Always be strict with uffd-wp 1283e1e267c7SPeter Xu * enabled swap entries. Please see 1284e1e267c7SPeter Xu * comment below for pte_uffd_wp(). 1285e1e267c7SPeter Xu */ 12862bad466cSPeter Xu if (pte_swp_uffd_wp_any(pteval)) { 1287e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1288e1e267c7SPeter Xu goto out_unmap; 1289e1e267c7SPeter Xu } 1290b46e756fSKirill A. Shutemov continue; 1291b46e756fSKirill A. Shutemov } else { 1292b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1293e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1294b46e756fSKirill A. Shutemov goto out_unmap; 1295b46e756fSKirill A. Shutemov } 1296b46e756fSKirill A. Shutemov } 1297b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1298d8ea7cc8SZach O'Keefe ++none_or_zero; 1299b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1300d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 1301d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 1302b46e756fSKirill A. Shutemov continue; 1303b46e756fSKirill A. Shutemov } else { 1304b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1305e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1306b46e756fSKirill A. Shutemov goto out_unmap; 1307b46e756fSKirill A. Shutemov } 1308b46e756fSKirill A. Shutemov } 1309e1e267c7SPeter Xu if (pte_uffd_wp(pteval)) { 1310e1e267c7SPeter Xu /* 1311e1e267c7SPeter Xu * Don't collapse the page if any of the small 1312e1e267c7SPeter Xu * PTEs are armed with uffd write protection. 1313e1e267c7SPeter Xu * Here we can also mark the new huge pmd as 1314e1e267c7SPeter Xu * write protected if any of the small ones is 13158958b249SHaitao Shi * marked but that could bring unknown 1316e1e267c7SPeter Xu * userfault messages that falls outside of 1317e1e267c7SPeter Xu * the registered range. So, just be simple. 1318e1e267c7SPeter Xu */ 1319e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1320e1e267c7SPeter Xu goto out_unmap; 1321e1e267c7SPeter Xu } 1322b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1323b46e756fSKirill A. Shutemov writable = true; 1324b46e756fSKirill A. Shutemov 1325b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 13263218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 1327b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1328b46e756fSKirill A. Shutemov goto out_unmap; 1329b46e756fSKirill A. Shutemov } 1330b46e756fSKirill A. Shutemov 1331d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 1332d8ea7cc8SZach O'Keefe ++shared; 1333d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 1334d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 133571a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 1336e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 133771a2c112SKirill A. Shutemov goto out_unmap; 133871a2c112SKirill A. Shutemov } 1339d8ea7cc8SZach O'Keefe } 134071a2c112SKirill A. Shutemov 13415c07ebb3SVishal Moola (Oracle) folio = page_folio(page); 1342b46e756fSKirill A. Shutemov /* 1343b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 134434d6b470SZach O'Keefe * information to cc->node_load[]. 13450b8f0d87SQuanfa Fu * Khugepaged will allocate hugepage from the node has the max 1346b46e756fSKirill A. Shutemov * hit record. 1347b46e756fSKirill A. Shutemov */ 13485c07ebb3SVishal Moola (Oracle) node = folio_nid(folio); 13497d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 1350b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1351b46e756fSKirill A. Shutemov goto out_unmap; 1352b46e756fSKirill A. Shutemov } 135334d6b470SZach O'Keefe cc->node_load[node]++; 13545c07ebb3SVishal Moola (Oracle) if (!folio_test_lru(folio)) { 1355b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1356b46e756fSKirill A. Shutemov goto out_unmap; 1357b46e756fSKirill A. Shutemov } 13585c07ebb3SVishal Moola (Oracle) if (folio_test_locked(folio)) { 1359b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1360b46e756fSKirill A. Shutemov goto out_unmap; 1361b46e756fSKirill A. Shutemov } 13625c07ebb3SVishal Moola (Oracle) if (!folio_test_anon(folio)) { 1363b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1364b46e756fSKirill A. Shutemov goto out_unmap; 1365b46e756fSKirill A. Shutemov } 1366b46e756fSKirill A. Shutemov 1367b46e756fSKirill A. Shutemov /* 13689445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 13699445689fSKirill A. Shutemov * 1370cb67f428SHugh Dickins * Here the check may be racy: 1371cb67f428SHugh Dickins * it may see total_mapcount > refcount in some cases? 13729445689fSKirill A. Shutemov * But such case is ephemeral we could always retry collapse 13739445689fSKirill A. Shutemov * later. However it may report false positive if the page 13749445689fSKirill A. Shutemov * has excessive GUP pins (i.e. 512). Anyway the same check 13759445689fSKirill A. Shutemov * will be done again later the risk seems low. 1376b46e756fSKirill A. Shutemov */ 1377dbf85c21SVishal Moola (Oracle) if (!is_refcount_suitable(folio)) { 1378b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1379b46e756fSKirill A. Shutemov goto out_unmap; 1380b46e756fSKirill A. Shutemov } 1381d8ea7cc8SZach O'Keefe 1382d8ea7cc8SZach O'Keefe /* 1383d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 1384d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 1385d8ea7cc8SZach O'Keefe */ 1386d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 13875c07ebb3SVishal Moola (Oracle) (pte_young(pteval) || folio_test_young(folio) || 13885c07ebb3SVishal Moola (Oracle) folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm, 1389d8ea7cc8SZach O'Keefe address))) 13900db501f7SEbru Akagunduz referenced++; 1391b46e756fSKirill A. Shutemov } 1392ffe945e6SKirill A. Shutemov if (!writable) { 1393ffe945e6SKirill A. Shutemov result = SCAN_PAGE_RO; 1394d8ea7cc8SZach O'Keefe } else if (cc->is_khugepaged && 1395d8ea7cc8SZach O'Keefe (!referenced || 1396d8ea7cc8SZach O'Keefe (unmapped && referenced < HPAGE_PMD_NR / 2))) { 1397ffe945e6SKirill A. Shutemov result = SCAN_LACK_REFERENCED_PAGE; 1398ffe945e6SKirill A. Shutemov } else { 1399b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1400b46e756fSKirill A. Shutemov } 1401b46e756fSKirill A. Shutemov out_unmap: 1402b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 140350ad2f24SZach O'Keefe if (result == SCAN_SUCCEED) { 140450ad2f24SZach O'Keefe result = collapse_huge_page(mm, address, referenced, 140550ad2f24SZach O'Keefe unmapped, cc); 1406c1e8d7c6SMichel Lespinasse /* collapse_huge_page will return with the mmap_lock released */ 140750ad2f24SZach O'Keefe *mmap_locked = false; 1408b46e756fSKirill A. Shutemov } 1409b46e756fSKirill A. Shutemov out: 14105c07ebb3SVishal Moola (Oracle) trace_mm_khugepaged_scan_pmd(mm, &folio->page, writable, referenced, 1411b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 141250ad2f24SZach O'Keefe return result; 1413b46e756fSKirill A. Shutemov } 1414b46e756fSKirill A. Shutemov 1415b26e2701SQi Zheng static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot) 1416b46e756fSKirill A. Shutemov { 1417b26e2701SQi Zheng struct mm_slot *slot = &mm_slot->slot; 1418b26e2701SQi Zheng struct mm_struct *mm = slot->mm; 1419b46e756fSKirill A. Shutemov 142035f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1421b46e756fSKirill A. Shutemov 14227d2c4385SZach O'Keefe if (hpage_collapse_test_exit(mm)) { 1423b46e756fSKirill A. Shutemov /* free mm_slot */ 1424b26e2701SQi Zheng hash_del(&slot->hash); 1425b26e2701SQi Zheng list_del(&slot->mm_node); 1426b46e756fSKirill A. Shutemov 1427b46e756fSKirill A. Shutemov /* 1428b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1429b46e756fSKirill A. Shutemov * 1430b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1431b46e756fSKirill A. Shutemov */ 1432b46e756fSKirill A. Shutemov 1433b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1434b26e2701SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 1435b46e756fSKirill A. Shutemov mmdrop(mm); 1436b46e756fSKirill A. Shutemov } 1437b46e756fSKirill A. Shutemov } 1438b46e756fSKirill A. Shutemov 1439396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 14401043173eSHugh Dickins /* hpage must be locked, and mmap_lock must be held */ 144134488399SZach O'Keefe static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, 144234488399SZach O'Keefe pmd_t *pmdp, struct page *hpage) 144334488399SZach O'Keefe { 144434488399SZach O'Keefe struct vm_fault vmf = { 144534488399SZach O'Keefe .vma = vma, 144634488399SZach O'Keefe .address = addr, 144734488399SZach O'Keefe .flags = 0, 144834488399SZach O'Keefe .pmd = pmdp, 144934488399SZach O'Keefe }; 145034488399SZach O'Keefe 145134488399SZach O'Keefe VM_BUG_ON(!PageTransHuge(hpage)); 14521043173eSHugh Dickins mmap_assert_locked(vma->vm_mm); 145334488399SZach O'Keefe 145434488399SZach O'Keefe if (do_set_pmd(&vmf, hpage)) 145534488399SZach O'Keefe return SCAN_FAIL; 145634488399SZach O'Keefe 145734488399SZach O'Keefe get_page(hpage); 145834488399SZach O'Keefe return SCAN_SUCCEED; 145927e1f827SSong Liu } 146027e1f827SSong Liu 146127e1f827SSong Liu /** 1462336e6b53SAlex Shi * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at 1463336e6b53SAlex Shi * address haddr. 1464336e6b53SAlex Shi * 1465336e6b53SAlex Shi * @mm: process address space where collapse happens 1466336e6b53SAlex Shi * @addr: THP collapse address 146734488399SZach O'Keefe * @install_pmd: If a huge PMD should be installed 146827e1f827SSong Liu * 146927e1f827SSong Liu * This function checks whether all the PTEs in the PMD are pointing to the 147027e1f827SSong Liu * right THP. If so, retract the page table so the THP can refault in with 147134488399SZach O'Keefe * as pmd-mapped. Possibly install a huge PMD mapping the THP. 147227e1f827SSong Liu */ 147334488399SZach O'Keefe int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, 147434488399SZach O'Keefe bool install_pmd) 147527e1f827SSong Liu { 14761043173eSHugh Dickins struct mmu_notifier_range range; 14771043173eSHugh Dickins bool notified = false; 147827e1f827SSong Liu unsigned long haddr = addr & HPAGE_PMD_MASK; 147994d815b2SLiam R. Howlett struct vm_area_struct *vma = vma_lookup(mm, haddr); 1480*98b32d29SVishal Moola (Oracle) struct folio *folio; 148127e1f827SSong Liu pte_t *start_pte, *pte; 14821043173eSHugh Dickins pmd_t *pmd, pgt_pmd; 1483a9846049SHugh Dickins spinlock_t *pml = NULL, *ptl; 14841043173eSHugh Dickins int nr_ptes = 0, result = SCAN_FAIL; 148527e1f827SSong Liu int i; 148627e1f827SSong Liu 14871043173eSHugh Dickins mmap_assert_locked(mm); 14881043173eSHugh Dickins 14891043173eSHugh Dickins /* First check VMA found, in case page tables are being torn down */ 14901043173eSHugh Dickins if (!vma || !vma->vm_file || 14911043173eSHugh Dickins !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) 14921043173eSHugh Dickins return SCAN_VMA_CHECK; 149358ac9a89SZach O'Keefe 149434488399SZach O'Keefe /* Fast check before locking page if already PMD-mapped */ 149558ac9a89SZach O'Keefe result = find_pmd_or_thp_or_none(mm, haddr, &pmd); 149634488399SZach O'Keefe if (result == SCAN_PMD_MAPPED) 149734488399SZach O'Keefe return result; 149858ac9a89SZach O'Keefe 149927e1f827SSong Liu /* 1500a7f4e6e4SZach O'Keefe * If we are here, we've succeeded in replacing all the native pages 1501a7f4e6e4SZach O'Keefe * in the page cache with a single hugepage. If a mm were to fault-in 1502a7f4e6e4SZach O'Keefe * this memory (mapped by a suitably aligned VMA), we'd get the hugepage 1503a7f4e6e4SZach O'Keefe * and map it by a PMD, regardless of sysfs THP settings. As such, let's 1504a7f4e6e4SZach O'Keefe * analogously elide sysfs THP settings here. 150527e1f827SSong Liu */ 1506a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) 150734488399SZach O'Keefe return SCAN_VMA_CHECK; 150827e1f827SSong Liu 1509deb4c93aSPeter Xu /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ 1510deb4c93aSPeter Xu if (userfaultfd_wp(vma)) 151134488399SZach O'Keefe return SCAN_PTE_UFFD_WP; 1512deb4c93aSPeter Xu 1513*98b32d29SVishal Moola (Oracle) folio = filemap_lock_folio(vma->vm_file->f_mapping, 1514119a5fc1SHugh Dickins linear_page_index(vma, haddr)); 1515*98b32d29SVishal Moola (Oracle) if (IS_ERR(folio)) 151634488399SZach O'Keefe return SCAN_PAGE_NULL; 1517119a5fc1SHugh Dickins 1518*98b32d29SVishal Moola (Oracle) if (folio_order(folio) != HPAGE_PMD_ORDER) { 151934488399SZach O'Keefe result = SCAN_PAGE_COMPOUND; 1520*98b32d29SVishal Moola (Oracle) goto drop_folio; 152134488399SZach O'Keefe } 1522780a4b6fSZach O'Keefe 15231043173eSHugh Dickins result = find_pmd_or_thp_or_none(mm, haddr, &pmd); 152434488399SZach O'Keefe switch (result) { 152534488399SZach O'Keefe case SCAN_SUCCEED: 152634488399SZach O'Keefe break; 152734488399SZach O'Keefe case SCAN_PMD_NONE: 152834488399SZach O'Keefe /* 15291d65b771SHugh Dickins * All pte entries have been removed and pmd cleared. 15301d65b771SHugh Dickins * Skip all the pte checks and just update the pmd mapping. 153134488399SZach O'Keefe */ 153234488399SZach O'Keefe goto maybe_install_pmd; 153334488399SZach O'Keefe default: 1534*98b32d29SVishal Moola (Oracle) goto drop_folio; 153534488399SZach O'Keefe } 153627e1f827SSong Liu 153734488399SZach O'Keefe result = SCAN_FAIL; 1538895f5ee4SHugh Dickins start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 15391043173eSHugh Dickins if (!start_pte) /* mmap_lock + page lock should prevent this */ 1540*98b32d29SVishal Moola (Oracle) goto drop_folio; 154127e1f827SSong Liu 154227e1f827SSong Liu /* step 1: check all mapped PTEs are to the right huge page */ 154327e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 154427e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 154527e1f827SSong Liu struct page *page; 1546c33c7948SRyan Roberts pte_t ptent = ptep_get(pte); 154727e1f827SSong Liu 154827e1f827SSong Liu /* empty pte, skip */ 1549c33c7948SRyan Roberts if (pte_none(ptent)) 155027e1f827SSong Liu continue; 155127e1f827SSong Liu 155227e1f827SSong Liu /* page swapped out, abort */ 1553c33c7948SRyan Roberts if (!pte_present(ptent)) { 155434488399SZach O'Keefe result = SCAN_PTE_NON_PRESENT; 155527e1f827SSong Liu goto abort; 155634488399SZach O'Keefe } 155727e1f827SSong Liu 1558c33c7948SRyan Roberts page = vm_normal_page(vma, addr, ptent); 15593218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 15603218f871SAlex Sierra page = NULL; 156127e1f827SSong Liu /* 1562119a5fc1SHugh Dickins * Note that uprobe, debugger, or MAP_PRIVATE may change the 1563119a5fc1SHugh Dickins * page table, but the new page will not be a subpage of hpage. 156427e1f827SSong Liu */ 1565*98b32d29SVishal Moola (Oracle) if (folio_page(folio, i) != page) 156627e1f827SSong Liu goto abort; 156727e1f827SSong Liu } 156827e1f827SSong Liu 15691043173eSHugh Dickins pte_unmap_unlock(start_pte, ptl); 15701043173eSHugh Dickins mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 15711043173eSHugh Dickins haddr, haddr + HPAGE_PMD_SIZE); 15721043173eSHugh Dickins mmu_notifier_invalidate_range_start(&range); 15731043173eSHugh Dickins notified = true; 1574a9846049SHugh Dickins 1575a9846049SHugh Dickins /* 1576a9846049SHugh Dickins * pmd_lock covers a wider range than ptl, and (if split from mm's 1577a9846049SHugh Dickins * page_table_lock) ptl nests inside pml. The less time we hold pml, 1578a9846049SHugh Dickins * the better; but userfaultfd's mfill_atomic_pte() on a private VMA 1579a9846049SHugh Dickins * inserts a valid as-if-COWed PTE without even looking up page cache. 1580*98b32d29SVishal Moola (Oracle) * So page lock of folio does not protect from it, so we must not drop 1581a9846049SHugh Dickins * ptl before pgt_pmd is removed, so uffd private needs pml taken now. 1582a9846049SHugh Dickins */ 1583a9846049SHugh Dickins if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED)) 1584a9846049SHugh Dickins pml = pmd_lock(mm, pmd); 1585a9846049SHugh Dickins 1586a9846049SHugh Dickins start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl); 15871043173eSHugh Dickins if (!start_pte) /* mmap_lock + page lock should prevent this */ 15881043173eSHugh Dickins goto abort; 1589a9846049SHugh Dickins if (!pml) 1590a9846049SHugh Dickins spin_lock(ptl); 1591a9846049SHugh Dickins else if (ptl != pml) 1592a9846049SHugh Dickins spin_lock_nested(ptl, SINGLE_DEPTH_NESTING); 15931043173eSHugh Dickins 15941043173eSHugh Dickins /* step 2: clear page table and adjust rmap */ 159527e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 159627e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 159727e1f827SSong Liu struct page *page; 1598c33c7948SRyan Roberts pte_t ptent = ptep_get(pte); 159927e1f827SSong Liu 1600c33c7948SRyan Roberts if (pte_none(ptent)) 160127e1f827SSong Liu continue; 16021043173eSHugh Dickins /* 16031043173eSHugh Dickins * We dropped ptl after the first scan, to do the mmu_notifier: 1604*98b32d29SVishal Moola (Oracle) * page lock stops more PTEs of the folio being faulted in, but 16051043173eSHugh Dickins * does not stop write faults COWing anon copies from existing 16061043173eSHugh Dickins * PTEs; and does not stop those being swapped out or migrated. 16071043173eSHugh Dickins */ 16081043173eSHugh Dickins if (!pte_present(ptent)) { 16091043173eSHugh Dickins result = SCAN_PTE_NON_PRESENT; 16103218f871SAlex Sierra goto abort; 16111043173eSHugh Dickins } 16121043173eSHugh Dickins page = vm_normal_page(vma, addr, ptent); 1613*98b32d29SVishal Moola (Oracle) if (folio_page(folio, i) != page) 16141043173eSHugh Dickins goto abort; 16151043173eSHugh Dickins 16161043173eSHugh Dickins /* 16171043173eSHugh Dickins * Must clear entry, or a racing truncate may re-remove it. 16181043173eSHugh Dickins * TLB flush can be left until pmdp_collapse_flush() does it. 16191043173eSHugh Dickins * PTE dirty? Shmem page is already dirty; file is read-only. 16201043173eSHugh Dickins */ 16211043173eSHugh Dickins ptep_clear(mm, addr, pte); 1622cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 16231043173eSHugh Dickins nr_ptes++; 162427e1f827SSong Liu } 162527e1f827SSong Liu 1626a9846049SHugh Dickins pte_unmap(start_pte); 1627a9846049SHugh Dickins if (!pml) 1628a9846049SHugh Dickins spin_unlock(ptl); 162927e1f827SSong Liu 163027e1f827SSong Liu /* step 3: set proper refcount and mm_counters. */ 16311043173eSHugh Dickins if (nr_ptes) { 1632*98b32d29SVishal Moola (Oracle) folio_ref_sub(folio, nr_ptes); 1633*98b32d29SVishal Moola (Oracle) add_mm_counter(mm, mm_counter_file(&folio->page), -nr_ptes); 163427e1f827SSong Liu } 163527e1f827SSong Liu 1636a9846049SHugh Dickins /* step 4: remove empty page table */ 1637a9846049SHugh Dickins if (!pml) { 16381043173eSHugh Dickins pml = pmd_lock(mm, pmd); 16391043173eSHugh Dickins if (ptl != pml) 16401043173eSHugh Dickins spin_lock_nested(ptl, SINGLE_DEPTH_NESTING); 1641a9846049SHugh Dickins } 16421043173eSHugh Dickins pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd); 16431043173eSHugh Dickins pmdp_get_lockless_sync(); 16441043173eSHugh Dickins if (ptl != pml) 16451043173eSHugh Dickins spin_unlock(ptl); 16461043173eSHugh Dickins spin_unlock(pml); 1647ab0c3f12SHugh Dickins 16481043173eSHugh Dickins mmu_notifier_invalidate_range_end(&range); 164934488399SZach O'Keefe 16501043173eSHugh Dickins mm_dec_nr_ptes(mm); 16511043173eSHugh Dickins page_table_check_pte_clear_range(mm, haddr, pgt_pmd); 16521043173eSHugh Dickins pte_free_defer(mm, pmd_pgtable(pgt_pmd)); 16538d3c106eSJann Horn 165434488399SZach O'Keefe maybe_install_pmd: 165534488399SZach O'Keefe /* step 5: install pmd entry */ 165634488399SZach O'Keefe result = install_pmd 1657*98b32d29SVishal Moola (Oracle) ? set_huge_pmd(vma, haddr, pmd, &folio->page) 165834488399SZach O'Keefe : SCAN_SUCCEED; 1659*98b32d29SVishal Moola (Oracle) goto drop_folio; 16601043173eSHugh Dickins abort: 16611043173eSHugh Dickins if (nr_ptes) { 16621043173eSHugh Dickins flush_tlb_mm(mm); 1663*98b32d29SVishal Moola (Oracle) folio_ref_sub(folio, nr_ptes); 1664*98b32d29SVishal Moola (Oracle) add_mm_counter(mm, mm_counter_file(&folio->page), -nr_ptes); 16651043173eSHugh Dickins } 16661043173eSHugh Dickins if (start_pte) 16671043173eSHugh Dickins pte_unmap_unlock(start_pte, ptl); 1668a9846049SHugh Dickins if (pml && pml != ptl) 1669a9846049SHugh Dickins spin_unlock(pml); 16701043173eSHugh Dickins if (notified) 16711043173eSHugh Dickins mmu_notifier_invalidate_range_end(&range); 1672*98b32d29SVishal Moola (Oracle) drop_folio: 1673*98b32d29SVishal Moola (Oracle) folio_unlock(folio); 1674*98b32d29SVishal Moola (Oracle) folio_put(folio); 167534488399SZach O'Keefe return result; 167627e1f827SSong Liu } 167727e1f827SSong Liu 16781d65b771SHugh Dickins static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1679f3f0e1d2SKirill A. Shutemov { 1680f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 1681f3f0e1d2SKirill A. Shutemov 16821d65b771SHugh Dickins i_mmap_lock_read(mapping); 1683f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 16841d65b771SHugh Dickins struct mmu_notifier_range range; 16851d65b771SHugh Dickins struct mm_struct *mm; 16861d65b771SHugh Dickins unsigned long addr; 16871d65b771SHugh Dickins pmd_t *pmd, pgt_pmd; 16881d65b771SHugh Dickins spinlock_t *pml; 16891d65b771SHugh Dickins spinlock_t *ptl; 16901d65b771SHugh Dickins bool skipped_uffd = false; 169134488399SZach O'Keefe 169227e1f827SSong Liu /* 169327e1f827SSong Liu * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 16941d65b771SHugh Dickins * got written to. These VMAs are likely not worth removing 16951d65b771SHugh Dickins * page tables from, as PMD-mapping is likely to be split later. 169627e1f827SSong Liu */ 16971d65b771SHugh Dickins if (READ_ONCE(vma->anon_vma)) 16981d65b771SHugh Dickins continue; 16991d65b771SHugh Dickins 1700f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 170134488399SZach O'Keefe if (addr & ~HPAGE_PMD_MASK || 17021d65b771SHugh Dickins vma->vm_end < addr + HPAGE_PMD_SIZE) 170334488399SZach O'Keefe continue; 17041d65b771SHugh Dickins 17051d65b771SHugh Dickins mm = vma->vm_mm; 17061d65b771SHugh Dickins if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED) 17071d65b771SHugh Dickins continue; 17081d65b771SHugh Dickins 17091d65b771SHugh Dickins if (hpage_collapse_test_exit(mm)) 17101d65b771SHugh Dickins continue; 17111d65b771SHugh Dickins /* 17121d65b771SHugh Dickins * When a vma is registered with uffd-wp, we cannot recycle 17131d65b771SHugh Dickins * the page table because there may be pte markers installed. 17141d65b771SHugh Dickins * Other vmas can still have the same file mapped hugely, but 17151d65b771SHugh Dickins * skip this one: it will always be mapped in small page size 17161d65b771SHugh Dickins * for uffd-wp registered ranges. 17171d65b771SHugh Dickins */ 17181d65b771SHugh Dickins if (userfaultfd_wp(vma)) 17191d65b771SHugh Dickins continue; 17201d65b771SHugh Dickins 17211d65b771SHugh Dickins /* PTEs were notified when unmapped; but now for the PMD? */ 17221d65b771SHugh Dickins mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 17231d65b771SHugh Dickins addr, addr + HPAGE_PMD_SIZE); 17241d65b771SHugh Dickins mmu_notifier_invalidate_range_start(&range); 17251d65b771SHugh Dickins 17261d65b771SHugh Dickins pml = pmd_lock(mm, pmd); 17271d65b771SHugh Dickins ptl = pte_lockptr(mm, pmd); 17281d65b771SHugh Dickins if (ptl != pml) 17291d65b771SHugh Dickins spin_lock_nested(ptl, SINGLE_DEPTH_NESTING); 17301d65b771SHugh Dickins 17311d65b771SHugh Dickins /* 17321d65b771SHugh Dickins * Huge page lock is still held, so normally the page table 17331d65b771SHugh Dickins * must remain empty; and we have already skipped anon_vma 17341d65b771SHugh Dickins * and userfaultfd_wp() vmas. But since the mmap_lock is not 17351d65b771SHugh Dickins * held, it is still possible for a racing userfaultfd_ioctl() 17361d65b771SHugh Dickins * to have inserted ptes or markers. Now that we hold ptlock, 17371d65b771SHugh Dickins * repeating the anon_vma check protects from one category, 17381d65b771SHugh Dickins * and repeating the userfaultfd_wp() check from another. 17391d65b771SHugh Dickins */ 17401d65b771SHugh Dickins if (unlikely(vma->anon_vma || userfaultfd_wp(vma))) { 17411d65b771SHugh Dickins skipped_uffd = true; 17421d65b771SHugh Dickins } else { 17431d65b771SHugh Dickins pgt_pmd = pmdp_collapse_flush(vma, addr, pmd); 17441d65b771SHugh Dickins pmdp_get_lockless_sync(); 174534488399SZach O'Keefe } 17461d65b771SHugh Dickins 17471d65b771SHugh Dickins if (ptl != pml) 17481d65b771SHugh Dickins spin_unlock(ptl); 17491d65b771SHugh Dickins spin_unlock(pml); 17501d65b771SHugh Dickins 17511d65b771SHugh Dickins mmu_notifier_invalidate_range_end(&range); 17521d65b771SHugh Dickins 17531d65b771SHugh Dickins if (!skipped_uffd) { 17541d65b771SHugh Dickins mm_dec_nr_ptes(mm); 17551d65b771SHugh Dickins page_table_check_pte_clear_range(mm, addr, pgt_pmd); 17561d65b771SHugh Dickins pte_free_defer(mm, pmd_pgtable(pgt_pmd)); 1757f3f0e1d2SKirill A. Shutemov } 17581d65b771SHugh Dickins } 17591d65b771SHugh Dickins i_mmap_unlock_read(mapping); 1760f3f0e1d2SKirill A. Shutemov } 1761f3f0e1d2SKirill A. Shutemov 1762f3f0e1d2SKirill A. Shutemov /** 176399cb0dbdSSong Liu * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1764f3f0e1d2SKirill A. Shutemov * 1765336e6b53SAlex Shi * @mm: process address space where collapse happens 176634488399SZach O'Keefe * @addr: virtual collapse start address 1767336e6b53SAlex Shi * @file: file that collapse on 1768336e6b53SAlex Shi * @start: collapse start address 17699710a78aSZach O'Keefe * @cc: collapse context and scratchpad 1770336e6b53SAlex Shi * 1771f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 177287c460a0SHugh Dickins * - allocate and lock a new huge page; 1773a2e17cc2SDavid Stevens * - scan page cache, locking old pages 177499cb0dbdSSong Liu * + swap/gup in pages if necessary; 1775a2e17cc2SDavid Stevens * - copy data to new page 1776a2e17cc2SDavid Stevens * - handle shmem holes 1777a2e17cc2SDavid Stevens * + re-validate that holes weren't filled by someone else 1778a2e17cc2SDavid Stevens * + check for userfaultfd 1779ac492b9cSDavid Stevens * - finalize updates to the page cache; 178077da9389SMatthew Wilcox * - if replacing succeeds: 178187c460a0SHugh Dickins * + unlock huge page; 1782a2e17cc2SDavid Stevens * + free old pages; 1783f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1784a2e17cc2SDavid Stevens * + unlock old pages 178587c460a0SHugh Dickins * + unlock and free huge page; 1786f3f0e1d2SKirill A. Shutemov */ 178734488399SZach O'Keefe static int collapse_file(struct mm_struct *mm, unsigned long addr, 1788579c571eSSong Liu struct file *file, pgoff_t start, 178934488399SZach O'Keefe struct collapse_control *cc) 1790f3f0e1d2SKirill A. Shutemov { 1791579c571eSSong Liu struct address_space *mapping = file->f_mapping; 179250ad2f24SZach O'Keefe struct page *hpage; 179312904d95SJiaqi Yan struct page *page; 179412904d95SJiaqi Yan struct page *tmp; 179512904d95SJiaqi Yan struct folio *folio; 17964c9473e8SGautam Menghani pgoff_t index = 0, end = start + HPAGE_PMD_NR; 1797f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 179877da9389SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1799f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 180099cb0dbdSSong Liu bool is_shmem = shmem_file(file); 18014c9473e8SGautam Menghani int nr = 0; 1802f3f0e1d2SKirill A. Shutemov 180399cb0dbdSSong Liu VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1804f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1805f3f0e1d2SKirill A. Shutemov 180650ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 18079710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1808f3f0e1d2SKirill A. Shutemov goto out; 1809f3f0e1d2SKirill A. Shutemov 1810cae106ddSDavid Stevens __SetPageLocked(hpage); 1811cae106ddSDavid Stevens if (is_shmem) 1812cae106ddSDavid Stevens __SetPageSwapBacked(hpage); 1813cae106ddSDavid Stevens hpage->index = start; 1814cae106ddSDavid Stevens hpage->mapping = mapping; 1815cae106ddSDavid Stevens 18166b24ca4aSMatthew Wilcox (Oracle) /* 18176b24ca4aSMatthew Wilcox (Oracle) * Ensure we have slots for all the pages in the range. This is 18186b24ca4aSMatthew Wilcox (Oracle) * almost certainly a no-op because most of the pages must be present 18196b24ca4aSMatthew Wilcox (Oracle) */ 182095feeabbSHugh Dickins do { 182195feeabbSHugh Dickins xas_lock_irq(&xas); 182295feeabbSHugh Dickins xas_create_range(&xas); 182395feeabbSHugh Dickins if (!xas_error(&xas)) 182495feeabbSHugh Dickins break; 182595feeabbSHugh Dickins xas_unlock_irq(&xas); 182695feeabbSHugh Dickins if (!xas_nomem(&xas, GFP_KERNEL)) { 182795feeabbSHugh Dickins result = SCAN_FAIL; 1828cae106ddSDavid Stevens goto rollback; 182995feeabbSHugh Dickins } 183095feeabbSHugh Dickins } while (1); 183195feeabbSHugh Dickins 183277da9389SMatthew Wilcox for (index = start; index < end; index++) { 1833e8c716bcSHugh Dickins xas_set(&xas, index); 1834e8c716bcSHugh Dickins page = xas_load(&xas); 183577da9389SMatthew Wilcox 183677da9389SMatthew Wilcox VM_BUG_ON(index != xas.xa_index); 183799cb0dbdSSong Liu if (is_shmem) { 183877da9389SMatthew Wilcox if (!page) { 1839701270faSHugh Dickins /* 184099cb0dbdSSong Liu * Stop if extent has been truncated or 184199cb0dbdSSong Liu * hole-punched, and is now completely 184299cb0dbdSSong Liu * empty. 1843701270faSHugh Dickins */ 1844701270faSHugh Dickins if (index == start) { 1845701270faSHugh Dickins if (!xas_next_entry(&xas, end - 1)) { 1846701270faSHugh Dickins result = SCAN_TRUNCATED; 1847042a3082SHugh Dickins goto xa_locked; 1848701270faSHugh Dickins } 1849701270faSHugh Dickins } 185077da9389SMatthew Wilcox nr_none++; 185177da9389SMatthew Wilcox continue; 1852f3f0e1d2SKirill A. Shutemov } 1853f3f0e1d2SKirill A. Shutemov 18543159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 185577da9389SMatthew Wilcox xas_unlock_irq(&xas); 1856f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 18577459c149SMatthew Wilcox (Oracle) if (shmem_get_folio(mapping->host, index, 18587459c149SMatthew Wilcox (Oracle) &folio, SGP_NOALLOC)) { 1859f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 186077da9389SMatthew Wilcox goto xa_unlocked; 1861f3f0e1d2SKirill A. Shutemov } 18621fec6890SMatthew Wilcox (Oracle) /* drain lru cache to help isolate_lru_page() */ 1863efa3d814SDavid Stevens lru_add_drain(); 18647459c149SMatthew Wilcox (Oracle) page = folio_file_page(folio, index); 1865f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1866f3f0e1d2SKirill A. Shutemov get_page(page); 1867042a3082SHugh Dickins xas_unlock_irq(&xas); 1868f3f0e1d2SKirill A. Shutemov } else { 1869f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1870042a3082SHugh Dickins goto xa_locked; 1871f3f0e1d2SKirill A. Shutemov } 187299cb0dbdSSong Liu } else { /* !is_shmem */ 187399cb0dbdSSong Liu if (!page || xa_is_value(page)) { 187499cb0dbdSSong Liu xas_unlock_irq(&xas); 187599cb0dbdSSong Liu page_cache_sync_readahead(mapping, &file->f_ra, 187699cb0dbdSSong Liu file, index, 1877e5a59d30SDavid Howells end - index); 18781fec6890SMatthew Wilcox (Oracle) /* drain lru cache to help isolate_lru_page() */ 187999cb0dbdSSong Liu lru_add_drain(); 188099cb0dbdSSong Liu page = find_lock_page(mapping, index); 188199cb0dbdSSong Liu if (unlikely(page == NULL)) { 188299cb0dbdSSong Liu result = SCAN_FAIL; 188399cb0dbdSSong Liu goto xa_unlocked; 188499cb0dbdSSong Liu } 188575f36069SSong Liu } else if (PageDirty(page)) { 188675f36069SSong Liu /* 188775f36069SSong Liu * khugepaged only works on read-only fd, 188875f36069SSong Liu * so this page is dirty because it hasn't 188975f36069SSong Liu * been flushed since first write. There 189075f36069SSong Liu * won't be new dirty pages. 189175f36069SSong Liu * 189275f36069SSong Liu * Trigger async flush here and hope the 189375f36069SSong Liu * writeback is done when khugepaged 189475f36069SSong Liu * revisits this page. 189575f36069SSong Liu * 189675f36069SSong Liu * This is a one-off situation. We are not 189775f36069SSong Liu * forcing writeback in loop. 189875f36069SSong Liu */ 189975f36069SSong Liu xas_unlock_irq(&xas); 190075f36069SSong Liu filemap_flush(mapping); 190175f36069SSong Liu result = SCAN_FAIL; 190275f36069SSong Liu goto xa_unlocked; 190374c42e1bSRongwei Wang } else if (PageWriteback(page)) { 190474c42e1bSRongwei Wang xas_unlock_irq(&xas); 190574c42e1bSRongwei Wang result = SCAN_FAIL; 190674c42e1bSRongwei Wang goto xa_unlocked; 190799cb0dbdSSong Liu } else if (trylock_page(page)) { 190899cb0dbdSSong Liu get_page(page); 190999cb0dbdSSong Liu xas_unlock_irq(&xas); 191099cb0dbdSSong Liu } else { 191199cb0dbdSSong Liu result = SCAN_PAGE_LOCK; 191299cb0dbdSSong Liu goto xa_locked; 191399cb0dbdSSong Liu } 191499cb0dbdSSong Liu } 1915f3f0e1d2SKirill A. Shutemov 1916f3f0e1d2SKirill A. Shutemov /* 1917b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1918f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1919f3f0e1d2SKirill A. Shutemov */ 1920f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 19214655e5e5SSong Liu 19224655e5e5SSong Liu /* make sure the page is up to date */ 19234655e5e5SSong Liu if (unlikely(!PageUptodate(page))) { 19244655e5e5SSong Liu result = SCAN_FAIL; 19254655e5e5SSong Liu goto out_unlock; 19264655e5e5SSong Liu } 192706a5e126SHugh Dickins 192806a5e126SHugh Dickins /* 192906a5e126SHugh Dickins * If file was truncated then extended, or hole-punched, before 193006a5e126SHugh Dickins * we locked the first page, then a THP might be there already. 193158ac9a89SZach O'Keefe * This will be discovered on the first iteration. 193206a5e126SHugh Dickins */ 193306a5e126SHugh Dickins if (PageTransCompound(page)) { 193458ac9a89SZach O'Keefe struct page *head = compound_head(page); 193558ac9a89SZach O'Keefe 193658ac9a89SZach O'Keefe result = compound_order(head) == HPAGE_PMD_ORDER && 193758ac9a89SZach O'Keefe head->index == start 193858ac9a89SZach O'Keefe /* Maybe PMD-mapped */ 193958ac9a89SZach O'Keefe ? SCAN_PTE_MAPPED_HUGEPAGE 194058ac9a89SZach O'Keefe : SCAN_PAGE_COMPOUND; 194106a5e126SHugh Dickins goto out_unlock; 194206a5e126SHugh Dickins } 1943f3f0e1d2SKirill A. Shutemov 194464ab3195SVishal Moola (Oracle) folio = page_folio(page); 194564ab3195SVishal Moola (Oracle) 194664ab3195SVishal Moola (Oracle) if (folio_mapping(folio) != mapping) { 1947f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1948f3f0e1d2SKirill A. Shutemov goto out_unlock; 1949f3f0e1d2SKirill A. Shutemov } 1950f3f0e1d2SKirill A. Shutemov 195164ab3195SVishal Moola (Oracle) if (!is_shmem && (folio_test_dirty(folio) || 195264ab3195SVishal Moola (Oracle) folio_test_writeback(folio))) { 19534655e5e5SSong Liu /* 19544655e5e5SSong Liu * khugepaged only works on read-only fd, so this 19554655e5e5SSong Liu * page is dirty because it hasn't been flushed 19564655e5e5SSong Liu * since first write. 19574655e5e5SSong Liu */ 19584655e5e5SSong Liu result = SCAN_FAIL; 19594655e5e5SSong Liu goto out_unlock; 19604655e5e5SSong Liu } 19614655e5e5SSong Liu 1962be2d5756SBaolin Wang if (!folio_isolate_lru(folio)) { 1963f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1964042a3082SHugh Dickins goto out_unlock; 1965f3f0e1d2SKirill A. Shutemov } 1966f3f0e1d2SKirill A. Shutemov 19670201ebf2SDavid Howells if (!filemap_release_folio(folio, GFP_KERNEL)) { 196899cb0dbdSSong Liu result = SCAN_PAGE_HAS_PRIVATE; 196964ab3195SVishal Moola (Oracle) folio_putback_lru(folio); 197099cb0dbdSSong Liu goto out_unlock; 197199cb0dbdSSong Liu } 197299cb0dbdSSong Liu 197364ab3195SVishal Moola (Oracle) if (folio_mapped(folio)) 197464ab3195SVishal Moola (Oracle) try_to_unmap(folio, 1975869f7ee6SMatthew Wilcox (Oracle) TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); 1976f3f0e1d2SKirill A. Shutemov 197777da9389SMatthew Wilcox xas_lock_irq(&xas); 1978f3f0e1d2SKirill A. Shutemov 1979e8c716bcSHugh Dickins VM_BUG_ON_PAGE(page != xa_load(xas.xa, index), page); 1980f3f0e1d2SKirill A. Shutemov 1981f3f0e1d2SKirill A. Shutemov /* 1982a2e17cc2SDavid Stevens * We control three references to the page: 1983f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 198477da9389SMatthew Wilcox * - one reference from page cache; 1985f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1986a2e17cc2SDavid Stevens * If those are the only references, then any new usage of the 1987a2e17cc2SDavid Stevens * page will have to fetch it from the page cache. That requires 1988a2e17cc2SDavid Stevens * locking the page to handle truncate, so any new usage will be 1989a2e17cc2SDavid Stevens * blocked until we unlock page after collapse/during rollback. 1990f3f0e1d2SKirill A. Shutemov */ 1991a2e17cc2SDavid Stevens if (page_count(page) != 3) { 1992f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1993042a3082SHugh Dickins xas_unlock_irq(&xas); 1994042a3082SHugh Dickins putback_lru_page(page); 1995042a3082SHugh Dickins goto out_unlock; 1996f3f0e1d2SKirill A. Shutemov } 1997f3f0e1d2SKirill A. Shutemov 1998f3f0e1d2SKirill A. Shutemov /* 1999a2e17cc2SDavid Stevens * Accumulate the pages that are being collapsed. 2000f3f0e1d2SKirill A. Shutemov */ 2001f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 2002f3f0e1d2SKirill A. Shutemov continue; 2003f3f0e1d2SKirill A. Shutemov out_unlock: 2004f3f0e1d2SKirill A. Shutemov unlock_page(page); 2005f3f0e1d2SKirill A. Shutemov put_page(page); 2006042a3082SHugh Dickins goto xa_unlocked; 2007f3f0e1d2SKirill A. Shutemov } 2008f3f0e1d2SKirill A. Shutemov 200912904d95SJiaqi Yan if (!is_shmem) { 201009d91cdaSSong Liu filemap_nr_thps_inc(mapping); 2011eb6ecbedSCollin Fijalkovich /* 2012eb6ecbedSCollin Fijalkovich * Paired with smp_mb() in do_dentry_open() to ensure 2013eb6ecbedSCollin Fijalkovich * i_writecount is up to date and the update to nr_thps is 2014eb6ecbedSCollin Fijalkovich * visible. Ensures the page cache will be truncated if the 2015eb6ecbedSCollin Fijalkovich * file is opened writable. 2016eb6ecbedSCollin Fijalkovich */ 2017eb6ecbedSCollin Fijalkovich smp_mb(); 2018eb6ecbedSCollin Fijalkovich if (inode_is_open_for_write(mapping->host)) { 2019eb6ecbedSCollin Fijalkovich result = SCAN_FAIL; 2020eb6ecbedSCollin Fijalkovich filemap_nr_thps_dec(mapping); 2021eb6ecbedSCollin Fijalkovich } 202209d91cdaSSong Liu } 202399cb0dbdSSong Liu 2024042a3082SHugh Dickins xa_locked: 2025042a3082SHugh Dickins xas_unlock_irq(&xas); 202677da9389SMatthew Wilcox xa_unlocked: 2027042a3082SHugh Dickins 20286d9df8a5SHugh Dickins /* 20296d9df8a5SHugh Dickins * If collapse is successful, flush must be done now before copying. 20306d9df8a5SHugh Dickins * If collapse is unsuccessful, does flush actually need to be done? 20316d9df8a5SHugh Dickins * Do it anyway, to clear the state. 20326d9df8a5SHugh Dickins */ 20336d9df8a5SHugh Dickins try_to_unmap_flush(); 20346d9df8a5SHugh Dickins 2035509f0069SHugh Dickins if (result == SCAN_SUCCEED && nr_none && 2036509f0069SHugh Dickins !shmem_charge(mapping->host, nr_none)) 2037509f0069SHugh Dickins result = SCAN_FAIL; 2038509f0069SHugh Dickins if (result != SCAN_SUCCEED) { 2039509f0069SHugh Dickins nr_none = 0; 2040cae106ddSDavid Stevens goto rollback; 2041509f0069SHugh Dickins } 2042cae106ddSDavid Stevens 2043f3f0e1d2SKirill A. Shutemov /* 2044a2e17cc2SDavid Stevens * The old pages are locked, so they won't change anymore. 2045f3f0e1d2SKirill A. Shutemov */ 20462af8ff29SHugh Dickins index = start; 204712904d95SJiaqi Yan list_for_each_entry(page, &pagelist, lru) { 20482af8ff29SHugh Dickins while (index < page->index) { 204950ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 20502af8ff29SHugh Dickins index++; 20512af8ff29SHugh Dickins } 2052cae106ddSDavid Stevens if (copy_mc_highpage(hpage + (page->index % HPAGE_PMD_NR), page) > 0) { 205312904d95SJiaqi Yan result = SCAN_COPY_MC; 2054cae106ddSDavid Stevens goto rollback; 205512904d95SJiaqi Yan } 205612904d95SJiaqi Yan index++; 205712904d95SJiaqi Yan } 2058cae106ddSDavid Stevens while (index < end) { 205912904d95SJiaqi Yan clear_highpage(hpage + (index % HPAGE_PMD_NR)); 206012904d95SJiaqi Yan index++; 206112904d95SJiaqi Yan } 206212904d95SJiaqi Yan 2063ac492b9cSDavid Stevens if (nr_none) { 2064ac492b9cSDavid Stevens struct vm_area_struct *vma; 2065ac492b9cSDavid Stevens int nr_none_check = 0; 2066ac492b9cSDavid Stevens 2067ac492b9cSDavid Stevens i_mmap_lock_read(mapping); 2068ac492b9cSDavid Stevens xas_lock_irq(&xas); 2069ac492b9cSDavid Stevens 2070ac492b9cSDavid Stevens xas_set(&xas, start); 2071ac492b9cSDavid Stevens for (index = start; index < end; index++) { 2072ac492b9cSDavid Stevens if (!xas_next(&xas)) { 2073ac492b9cSDavid Stevens xas_store(&xas, XA_RETRY_ENTRY); 2074ac492b9cSDavid Stevens if (xas_error(&xas)) { 2075ac492b9cSDavid Stevens result = SCAN_STORE_FAILED; 2076ac492b9cSDavid Stevens goto immap_locked; 2077ac492b9cSDavid Stevens } 2078ac492b9cSDavid Stevens nr_none_check++; 2079ac492b9cSDavid Stevens } 2080ac492b9cSDavid Stevens } 2081ac492b9cSDavid Stevens 2082ac492b9cSDavid Stevens if (nr_none != nr_none_check) { 2083ac492b9cSDavid Stevens result = SCAN_PAGE_FILLED; 2084ac492b9cSDavid Stevens goto immap_locked; 2085ac492b9cSDavid Stevens } 2086ac492b9cSDavid Stevens 208712904d95SJiaqi Yan /* 2088ac492b9cSDavid Stevens * If userspace observed a missing page in a VMA with a MODE_MISSING 2089ac492b9cSDavid Stevens * userfaultfd, then it might expect a UFFD_EVENT_PAGEFAULT for that 2090ac492b9cSDavid Stevens * page. If so, we need to roll back to avoid suppressing such an 2091ac492b9cSDavid Stevens * event. Since wp/minor userfaultfds don't give userspace any 2092ac492b9cSDavid Stevens * guarantees that the kernel doesn't fill a missing page with a zero 2093ac492b9cSDavid Stevens * page, so they don't matter here. 2094ac492b9cSDavid Stevens * 2095ac492b9cSDavid Stevens * Any userfaultfds registered after this point will not be able to 2096ac492b9cSDavid Stevens * observe any missing pages due to the previously inserted retry 2097ac492b9cSDavid Stevens * entries. 209812904d95SJiaqi Yan */ 2099ac492b9cSDavid Stevens vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) { 2100ac492b9cSDavid Stevens if (userfaultfd_missing(vma)) { 2101ac492b9cSDavid Stevens result = SCAN_EXCEED_NONE_PTE; 2102ac492b9cSDavid Stevens goto immap_locked; 2103ac492b9cSDavid Stevens } 2104ac492b9cSDavid Stevens } 2105ac492b9cSDavid Stevens 2106ac492b9cSDavid Stevens immap_locked: 2107ac492b9cSDavid Stevens i_mmap_unlock_read(mapping); 2108ac492b9cSDavid Stevens if (result != SCAN_SUCCEED) { 2109ac492b9cSDavid Stevens xas_set(&xas, start); 2110ac492b9cSDavid Stevens for (index = start; index < end; index++) { 2111ac492b9cSDavid Stevens if (xas_next(&xas) == XA_RETRY_ENTRY) 2112ac492b9cSDavid Stevens xas_store(&xas, NULL); 2113ac492b9cSDavid Stevens } 2114ac492b9cSDavid Stevens 2115ac492b9cSDavid Stevens xas_unlock_irq(&xas); 2116ac492b9cSDavid Stevens goto rollback; 2117ac492b9cSDavid Stevens } 2118ac492b9cSDavid Stevens } else { 2119ac492b9cSDavid Stevens xas_lock_irq(&xas); 21202af8ff29SHugh Dickins } 212112904d95SJiaqi Yan 2122cae106ddSDavid Stevens nr = thp_nr_pages(hpage); 212312904d95SJiaqi Yan if (is_shmem) 212412904d95SJiaqi Yan __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr); 212512904d95SJiaqi Yan else 212612904d95SJiaqi Yan __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr); 212712904d95SJiaqi Yan 212812904d95SJiaqi Yan if (nr_none) { 212912904d95SJiaqi Yan __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none); 213012904d95SJiaqi Yan /* nr_none is always 0 for non-shmem. */ 213112904d95SJiaqi Yan __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none); 2132f3f0e1d2SKirill A. Shutemov } 2133f3f0e1d2SKirill A. Shutemov 2134a2e17cc2SDavid Stevens /* 2135a2e17cc2SDavid Stevens * Mark hpage as uptodate before inserting it into the page cache so 2136a2e17cc2SDavid Stevens * that it isn't mistaken for an fallocated but unwritten page. 2137a2e17cc2SDavid Stevens */ 2138284a344eSVishal Moola (Oracle) folio = page_folio(hpage); 2139284a344eSVishal Moola (Oracle) folio_mark_uptodate(folio); 2140284a344eSVishal Moola (Oracle) folio_ref_add(folio, HPAGE_PMD_NR - 1); 2141284a344eSVishal Moola (Oracle) 21426058eaecSJohannes Weiner if (is_shmem) 2143284a344eSVishal Moola (Oracle) folio_mark_dirty(folio); 2144284a344eSVishal Moola (Oracle) folio_add_lru(folio); 2145f3f0e1d2SKirill A. Shutemov 2146a2e17cc2SDavid Stevens /* Join all the small entries into a single multi-index entry. */ 2147a2e17cc2SDavid Stevens xas_set_order(&xas, start, HPAGE_PMD_ORDER); 2148a2e17cc2SDavid Stevens xas_store(&xas, hpage); 21490175ab61SHugh Dickins WARN_ON_ONCE(xas_error(&xas)); 2150a2e17cc2SDavid Stevens xas_unlock_irq(&xas); 2151a2e17cc2SDavid Stevens 2152042a3082SHugh Dickins /* 2153042a3082SHugh Dickins * Remove pte page tables, so we can re-fault the page as huge. 21541d65b771SHugh Dickins * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp(). 2155042a3082SHugh Dickins */ 21561d65b771SHugh Dickins retract_page_tables(mapping, start); 21571d65b771SHugh Dickins if (cc && !cc->is_khugepaged) 21581d65b771SHugh Dickins result = SCAN_PTE_MAPPED_HUGEPAGE; 215950ad2f24SZach O'Keefe unlock_page(hpage); 2160ac492b9cSDavid Stevens 2161ac492b9cSDavid Stevens /* 2162ac492b9cSDavid Stevens * The collapse has succeeded, so free the old pages. 2163ac492b9cSDavid Stevens */ 2164ac492b9cSDavid Stevens list_for_each_entry_safe(page, tmp, &pagelist, lru) { 2165ac492b9cSDavid Stevens list_del(&page->lru); 2166ac492b9cSDavid Stevens page->mapping = NULL; 2167ac492b9cSDavid Stevens ClearPageActive(page); 2168ac492b9cSDavid Stevens ClearPageUnevictable(page); 2169ac492b9cSDavid Stevens unlock_page(page); 2170a2e17cc2SDavid Stevens folio_put_refs(page_folio(page), 3); 2171ac492b9cSDavid Stevens } 2172ac492b9cSDavid Stevens 2173cae106ddSDavid Stevens goto out; 2174cae106ddSDavid Stevens 2175cae106ddSDavid Stevens rollback: 217677da9389SMatthew Wilcox /* Something went wrong: roll back page cache changes */ 21772f55f070SMiaohe Lin if (nr_none) { 2178a2e17cc2SDavid Stevens xas_lock_irq(&xas); 2179aaa52e34SHugh Dickins mapping->nrpages -= nr_none; 218077da9389SMatthew Wilcox xas_unlock_irq(&xas); 2181509f0069SHugh Dickins shmem_uncharge(mapping->host, nr_none); 2182a2e17cc2SDavid Stevens } 2183a2e17cc2SDavid Stevens 2184a2e17cc2SDavid Stevens list_for_each_entry_safe(page, tmp, &pagelist, lru) { 2185a2e17cc2SDavid Stevens list_del(&page->lru); 2186f3f0e1d2SKirill A. Shutemov unlock_page(page); 2187042a3082SHugh Dickins putback_lru_page(page); 2188a2e17cc2SDavid Stevens put_page(page); 2189f3f0e1d2SKirill A. Shutemov } 219012904d95SJiaqi Yan /* 219112904d95SJiaqi Yan * Undo the updates of filemap_nr_thps_inc for non-SHMEM 219212904d95SJiaqi Yan * file only. This undo is not needed unless failure is 219312904d95SJiaqi Yan * due to SCAN_COPY_MC. 219412904d95SJiaqi Yan */ 219512904d95SJiaqi Yan if (!is_shmem && result == SCAN_COPY_MC) { 219612904d95SJiaqi Yan filemap_nr_thps_dec(mapping); 219712904d95SJiaqi Yan /* 219812904d95SJiaqi Yan * Paired with smp_mb() in do_dentry_open() to 219912904d95SJiaqi Yan * ensure the update to nr_thps is visible. 220012904d95SJiaqi Yan */ 220112904d95SJiaqi Yan smp_mb(); 220212904d95SJiaqi Yan } 220312904d95SJiaqi Yan 220450ad2f24SZach O'Keefe hpage->mapping = NULL; 2205042a3082SHugh Dickins 220650ad2f24SZach O'Keefe unlock_page(hpage); 2207cae106ddSDavid Stevens put_page(hpage); 2208f3f0e1d2SKirill A. Shutemov out: 2209f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 22104c9473e8SGautam Menghani trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result); 221150ad2f24SZach O'Keefe return result; 2212f3f0e1d2SKirill A. Shutemov } 2213f3f0e1d2SKirill A. Shutemov 221434488399SZach O'Keefe static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, 221534488399SZach O'Keefe struct file *file, pgoff_t start, 221634488399SZach O'Keefe struct collapse_control *cc) 2217f3f0e1d2SKirill A. Shutemov { 2218f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 2219579c571eSSong Liu struct address_space *mapping = file->f_mapping; 222085b392dbSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, start); 2221f3f0e1d2SKirill A. Shutemov int present, swap; 2222f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 2223f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 2224f3f0e1d2SKirill A. Shutemov 2225f3f0e1d2SKirill A. Shutemov present = 0; 2226f3f0e1d2SKirill A. Shutemov swap = 0; 222734d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 2228e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 2229f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 223085b392dbSMatthew Wilcox xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 223185b392dbSMatthew Wilcox if (xas_retry(&xas, page)) 2232f3f0e1d2SKirill A. Shutemov continue; 2233f3f0e1d2SKirill A. Shutemov 223485b392dbSMatthew Wilcox if (xa_is_value(page)) { 2235d8ea7cc8SZach O'Keefe ++swap; 2236d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 2237d8ea7cc8SZach O'Keefe swap > khugepaged_max_ptes_swap) { 2238f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 2239e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 2240f3f0e1d2SKirill A. Shutemov break; 2241f3f0e1d2SKirill A. Shutemov } 2242f3f0e1d2SKirill A. Shutemov continue; 2243f3f0e1d2SKirill A. Shutemov } 2244f3f0e1d2SKirill A. Shutemov 22456b24ca4aSMatthew Wilcox (Oracle) /* 224658ac9a89SZach O'Keefe * TODO: khugepaged should compact smaller compound pages 22476b24ca4aSMatthew Wilcox (Oracle) * into a PMD sized page 22486b24ca4aSMatthew Wilcox (Oracle) */ 2249f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 225058ac9a89SZach O'Keefe struct page *head = compound_head(page); 225158ac9a89SZach O'Keefe 225258ac9a89SZach O'Keefe result = compound_order(head) == HPAGE_PMD_ORDER && 225358ac9a89SZach O'Keefe head->index == start 225458ac9a89SZach O'Keefe /* Maybe PMD-mapped */ 225558ac9a89SZach O'Keefe ? SCAN_PTE_MAPPED_HUGEPAGE 225658ac9a89SZach O'Keefe : SCAN_PAGE_COMPOUND; 225758ac9a89SZach O'Keefe /* 225858ac9a89SZach O'Keefe * For SCAN_PTE_MAPPED_HUGEPAGE, further processing 225958ac9a89SZach O'Keefe * by the caller won't touch the page cache, and so 226058ac9a89SZach O'Keefe * it's safe to skip LRU and refcount checks before 226158ac9a89SZach O'Keefe * returning. 226258ac9a89SZach O'Keefe */ 2263f3f0e1d2SKirill A. Shutemov break; 2264f3f0e1d2SKirill A. Shutemov } 2265f3f0e1d2SKirill A. Shutemov 2266f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 22677d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 2268f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 2269f3f0e1d2SKirill A. Shutemov break; 2270f3f0e1d2SKirill A. Shutemov } 227134d6b470SZach O'Keefe cc->node_load[node]++; 2272f3f0e1d2SKirill A. Shutemov 2273f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 2274f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 2275f3f0e1d2SKirill A. Shutemov break; 2276f3f0e1d2SKirill A. Shutemov } 2277f3f0e1d2SKirill A. Shutemov 227899cb0dbdSSong Liu if (page_count(page) != 227999cb0dbdSSong Liu 1 + page_mapcount(page) + page_has_private(page)) { 2280f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 2281f3f0e1d2SKirill A. Shutemov break; 2282f3f0e1d2SKirill A. Shutemov } 2283f3f0e1d2SKirill A. Shutemov 2284f3f0e1d2SKirill A. Shutemov /* 2285f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 2286f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 2287f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 2288f3f0e1d2SKirill A. Shutemov */ 2289f3f0e1d2SKirill A. Shutemov 2290f3f0e1d2SKirill A. Shutemov present++; 2291f3f0e1d2SKirill A. Shutemov 2292f3f0e1d2SKirill A. Shutemov if (need_resched()) { 229385b392dbSMatthew Wilcox xas_pause(&xas); 2294f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 2295f3f0e1d2SKirill A. Shutemov } 2296f3f0e1d2SKirill A. Shutemov } 2297f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 2298f3f0e1d2SKirill A. Shutemov 2299f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 2300d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 2301d8ea7cc8SZach O'Keefe present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 2302f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 2303e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 2304f3f0e1d2SKirill A. Shutemov } else { 230534488399SZach O'Keefe result = collapse_file(mm, addr, file, start, cc); 2306f3f0e1d2SKirill A. Shutemov } 2307f3f0e1d2SKirill A. Shutemov } 2308f3f0e1d2SKirill A. Shutemov 2309045634ffSGautam Menghani trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result); 231050ad2f24SZach O'Keefe return result; 2311f3f0e1d2SKirill A. Shutemov } 2312f3f0e1d2SKirill A. Shutemov #else 231334488399SZach O'Keefe static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, 231434488399SZach O'Keefe struct file *file, pgoff_t start, 231534488399SZach O'Keefe struct collapse_control *cc) 2316f3f0e1d2SKirill A. Shutemov { 2317f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 2318f3f0e1d2SKirill A. Shutemov } 2319f3f0e1d2SKirill A. Shutemov #endif 2320f3f0e1d2SKirill A. Shutemov 232150ad2f24SZach O'Keefe static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, 232234d6b470SZach O'Keefe struct collapse_control *cc) 2323b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 2324b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 2325b46e756fSKirill A. Shutemov { 232668540502SMatthew Wilcox (Oracle) struct vma_iterator vmi; 2327b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 2328b26e2701SQi Zheng struct mm_slot *slot; 2329b46e756fSKirill A. Shutemov struct mm_struct *mm; 2330b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 2331b46e756fSKirill A. Shutemov int progress = 0; 2332b46e756fSKirill A. Shutemov 2333b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 233435f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 233550ad2f24SZach O'Keefe *result = SCAN_FAIL; 2336b46e756fSKirill A. Shutemov 2337b26e2701SQi Zheng if (khugepaged_scan.mm_slot) { 2338b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2339b26e2701SQi Zheng slot = &mm_slot->slot; 2340b26e2701SQi Zheng } else { 2341b26e2701SQi Zheng slot = list_entry(khugepaged_scan.mm_head.next, 2342b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2343b26e2701SQi Zheng mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 2344b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2345b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 2346b46e756fSKirill A. Shutemov } 2347b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2348b46e756fSKirill A. Shutemov 2349b26e2701SQi Zheng mm = slot->mm; 23503b454ad3SYang Shi /* 23513b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 23523b454ad3SYang Shi * the next mm on the list. 23533b454ad3SYang Shi */ 2354b46e756fSKirill A. Shutemov vma = NULL; 2355d8ed45c5SMichel Lespinasse if (unlikely(!mmap_read_trylock(mm))) 2356c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2357b46e756fSKirill A. Shutemov 2358b46e756fSKirill A. Shutemov progress++; 235968540502SMatthew Wilcox (Oracle) if (unlikely(hpage_collapse_test_exit(mm))) 236068540502SMatthew Wilcox (Oracle) goto breakouterloop; 236168540502SMatthew Wilcox (Oracle) 236268540502SMatthew Wilcox (Oracle) vma_iter_init(&vmi, mm, khugepaged_scan.address); 236368540502SMatthew Wilcox (Oracle) for_each_vma(vmi, vma) { 2364b46e756fSKirill A. Shutemov unsigned long hstart, hend; 2365b46e756fSKirill A. Shutemov 2366b46e756fSKirill A. Shutemov cond_resched(); 23677d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) { 2368b46e756fSKirill A. Shutemov progress++; 2369b46e756fSKirill A. Shutemov break; 2370b46e756fSKirill A. Shutemov } 2371a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) { 2372b46e756fSKirill A. Shutemov skip: 2373b46e756fSKirill A. Shutemov progress++; 2374b46e756fSKirill A. Shutemov continue; 2375b46e756fSKirill A. Shutemov } 23764fa6893fSYang Shi hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); 23774fa6893fSYang Shi hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); 2378b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 2379b46e756fSKirill A. Shutemov goto skip; 2380b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 2381b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 2382b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2383b46e756fSKirill A. Shutemov 2384b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 238550ad2f24SZach O'Keefe bool mmap_locked = true; 238650ad2f24SZach O'Keefe 2387b46e756fSKirill A. Shutemov cond_resched(); 23887d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 2389b46e756fSKirill A. Shutemov goto breakouterloop; 2390b46e756fSKirill A. Shutemov 2391b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 2392b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 2393b46e756fSKirill A. Shutemov hend); 239499cb0dbdSSong Liu if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 2395396bcc52SMatthew Wilcox (Oracle) struct file *file = get_file(vma->vm_file); 2396f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 2397f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 239899cb0dbdSSong Liu 2399d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 240050ad2f24SZach O'Keefe mmap_locked = false; 2401d50791c2SHugh Dickins *result = hpage_collapse_scan_file(mm, 2402d50791c2SHugh Dickins khugepaged_scan.address, file, pgoff, cc); 2403f3f0e1d2SKirill A. Shutemov fput(file); 2404d50791c2SHugh Dickins if (*result == SCAN_PTE_MAPPED_HUGEPAGE) { 2405d50791c2SHugh Dickins mmap_read_lock(mm); 2406d50791c2SHugh Dickins if (hpage_collapse_test_exit(mm)) 2407d50791c2SHugh Dickins goto breakouterloop; 2408d50791c2SHugh Dickins *result = collapse_pte_mapped_thp(mm, 2409d50791c2SHugh Dickins khugepaged_scan.address, false); 2410d50791c2SHugh Dickins if (*result == SCAN_PMD_MAPPED) 2411d50791c2SHugh Dickins *result = SCAN_SUCCEED; 2412d50791c2SHugh Dickins mmap_read_unlock(mm); 2413d50791c2SHugh Dickins } 2414f3f0e1d2SKirill A. Shutemov } else { 24157d2c4385SZach O'Keefe *result = hpage_collapse_scan_pmd(mm, vma, 2416d50791c2SHugh Dickins khugepaged_scan.address, &mmap_locked, cc); 2417f3f0e1d2SKirill A. Shutemov } 241858ac9a89SZach O'Keefe 2419d50791c2SHugh Dickins if (*result == SCAN_SUCCEED) 242050ad2f24SZach O'Keefe ++khugepaged_pages_collapsed; 242158ac9a89SZach O'Keefe 2422b46e756fSKirill A. Shutemov /* move to next address */ 2423b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 2424b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 242550ad2f24SZach O'Keefe if (!mmap_locked) 242650ad2f24SZach O'Keefe /* 242750ad2f24SZach O'Keefe * We released mmap_lock so break loop. Note 242850ad2f24SZach O'Keefe * that we drop mmap_lock before all hugepage 242950ad2f24SZach O'Keefe * allocations, so if allocation fails, we are 243050ad2f24SZach O'Keefe * guaranteed to break here and report the 243150ad2f24SZach O'Keefe * correct result back to caller. 243250ad2f24SZach O'Keefe */ 2433c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2434b46e756fSKirill A. Shutemov if (progress >= pages) 2435b46e756fSKirill A. Shutemov goto breakouterloop; 2436b46e756fSKirill A. Shutemov } 2437b46e756fSKirill A. Shutemov } 2438b46e756fSKirill A. Shutemov breakouterloop: 2439d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ 2440c1e8d7c6SMichel Lespinasse breakouterloop_mmap_lock: 2441b46e756fSKirill A. Shutemov 2442b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2443b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2444b46e756fSKirill A. Shutemov /* 2445b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 2446b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 2447b46e756fSKirill A. Shutemov */ 24487d2c4385SZach O'Keefe if (hpage_collapse_test_exit(mm) || !vma) { 2449b46e756fSKirill A. Shutemov /* 2450b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 2451b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 2452b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 2453b46e756fSKirill A. Shutemov */ 2454b26e2701SQi Zheng if (slot->mm_node.next != &khugepaged_scan.mm_head) { 2455b26e2701SQi Zheng slot = list_entry(slot->mm_node.next, 2456b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2457b26e2701SQi Zheng khugepaged_scan.mm_slot = 2458b26e2701SQi Zheng mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 2459b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2460b46e756fSKirill A. Shutemov } else { 2461b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2462b46e756fSKirill A. Shutemov khugepaged_full_scans++; 2463b46e756fSKirill A. Shutemov } 2464b46e756fSKirill A. Shutemov 2465b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2466b46e756fSKirill A. Shutemov } 2467b46e756fSKirill A. Shutemov 2468b46e756fSKirill A. Shutemov return progress; 2469b46e756fSKirill A. Shutemov } 2470b46e756fSKirill A. Shutemov 2471b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 2472b46e756fSKirill A. Shutemov { 2473b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 24741064026bSYang Shi hugepage_flags_enabled(); 2475b46e756fSKirill A. Shutemov } 2476b46e756fSKirill A. Shutemov 2477b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 2478b46e756fSKirill A. Shutemov { 2479b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 2480b46e756fSKirill A. Shutemov kthread_should_stop(); 2481b46e756fSKirill A. Shutemov } 2482b46e756fSKirill A. Shutemov 248334d6b470SZach O'Keefe static void khugepaged_do_scan(struct collapse_control *cc) 2484b46e756fSKirill A. Shutemov { 2485b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 248689dc6a96SYanfei Xu unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); 2487b46e756fSKirill A. Shutemov bool wait = true; 248850ad2f24SZach O'Keefe int result = SCAN_SUCCEED; 2489b46e756fSKirill A. Shutemov 2490a980df33SKirill A. Shutemov lru_add_drain_all(); 2491a980df33SKirill A. Shutemov 2492c6a7f445SYang Shi while (true) { 2493b46e756fSKirill A. Shutemov cond_resched(); 2494b46e756fSKirill A. Shutemov 2495b46e756fSKirill A. Shutemov if (unlikely(kthread_should_stop() || try_to_freeze())) 2496b46e756fSKirill A. Shutemov break; 2497b46e756fSKirill A. Shutemov 2498b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2499b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 2500b46e756fSKirill A. Shutemov pass_through_head++; 2501b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 2502b46e756fSKirill A. Shutemov pass_through_head < 2) 2503b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 250450ad2f24SZach O'Keefe &result, cc); 2505b46e756fSKirill A. Shutemov else 2506b46e756fSKirill A. Shutemov progress = pages; 2507b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2508b46e756fSKirill A. Shutemov 2509c6a7f445SYang Shi if (progress >= pages) 2510c6a7f445SYang Shi break; 2511c6a7f445SYang Shi 251250ad2f24SZach O'Keefe if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) { 2513c6a7f445SYang Shi /* 2514c6a7f445SYang Shi * If fail to allocate the first time, try to sleep for 2515c6a7f445SYang Shi * a while. When hit again, cancel the scan. 2516c6a7f445SYang Shi */ 2517c6a7f445SYang Shi if (!wait) 2518c6a7f445SYang Shi break; 2519c6a7f445SYang Shi wait = false; 2520c6a7f445SYang Shi khugepaged_alloc_sleep(); 2521c6a7f445SYang Shi } 2522c6a7f445SYang Shi } 2523b46e756fSKirill A. Shutemov } 2524b46e756fSKirill A. Shutemov 2525b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 2526b46e756fSKirill A. Shutemov { 2527b46e756fSKirill A. Shutemov return kthread_should_stop() || 2528b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 2529b46e756fSKirill A. Shutemov } 2530b46e756fSKirill A. Shutemov 2531b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 2532b46e756fSKirill A. Shutemov { 2533b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 2534b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 2535b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2536b46e756fSKirill A. Shutemov 2537b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 2538b46e756fSKirill A. Shutemov return; 2539b46e756fSKirill A. Shutemov 2540b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2541b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 2542b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 2543b46e756fSKirill A. Shutemov scan_sleep_jiffies); 2544b46e756fSKirill A. Shutemov return; 2545b46e756fSKirill A. Shutemov } 2546b46e756fSKirill A. Shutemov 25471064026bSYang Shi if (hugepage_flags_enabled()) 2548b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2549b46e756fSKirill A. Shutemov } 2550b46e756fSKirill A. Shutemov 2551b46e756fSKirill A. Shutemov static int khugepaged(void *none) 2552b46e756fSKirill A. Shutemov { 2553b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 2554b46e756fSKirill A. Shutemov 2555b46e756fSKirill A. Shutemov set_freezable(); 2556b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 2557b46e756fSKirill A. Shutemov 2558b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 255934d6b470SZach O'Keefe khugepaged_do_scan(&khugepaged_collapse_control); 2560b46e756fSKirill A. Shutemov khugepaged_wait_work(); 2561b46e756fSKirill A. Shutemov } 2562b46e756fSKirill A. Shutemov 2563b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2564b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2565b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2566b46e756fSKirill A. Shutemov if (mm_slot) 2567b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2568b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2569b46e756fSKirill A. Shutemov return 0; 2570b46e756fSKirill A. Shutemov } 2571b46e756fSKirill A. Shutemov 2572b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 2573b46e756fSKirill A. Shutemov { 2574b46e756fSKirill A. Shutemov struct zone *zone; 2575b46e756fSKirill A. Shutemov int nr_zones = 0; 2576b46e756fSKirill A. Shutemov unsigned long recommended_min; 2577b46e756fSKirill A. Shutemov 25781064026bSYang Shi if (!hugepage_flags_enabled()) { 2579bd3400eaSLiangcai Fan calculate_min_free_kbytes(); 2580bd3400eaSLiangcai Fan goto update_wmarks; 2581bd3400eaSLiangcai Fan } 2582bd3400eaSLiangcai Fan 2583b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 2584b7d349c7SJoonsoo Kim /* 2585b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 2586b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 2587b7d349c7SJoonsoo Kim */ 2588b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 2589b7d349c7SJoonsoo Kim continue; 2590b7d349c7SJoonsoo Kim 2591b46e756fSKirill A. Shutemov nr_zones++; 2592b7d349c7SJoonsoo Kim } 2593b46e756fSKirill A. Shutemov 2594b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2595b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 2596b46e756fSKirill A. Shutemov 2597b46e756fSKirill A. Shutemov /* 2598b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 2599b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 2600b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 2601b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 2602b46e756fSKirill A. Shutemov */ 2603b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 2604b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2605b46e756fSKirill A. Shutemov 2606b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 2607b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 2608b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 2609b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 2610b46e756fSKirill A. Shutemov 2611b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 2612b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 2613b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2614b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 2615b46e756fSKirill A. Shutemov 2616b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 2617b46e756fSKirill A. Shutemov } 2618bd3400eaSLiangcai Fan 2619bd3400eaSLiangcai Fan update_wmarks: 2620b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 2621b46e756fSKirill A. Shutemov } 2622b46e756fSKirill A. Shutemov 2623b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 2624b46e756fSKirill A. Shutemov { 2625b46e756fSKirill A. Shutemov int err = 0; 2626b46e756fSKirill A. Shutemov 2627b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 26281064026bSYang Shi if (hugepage_flags_enabled()) { 2629b46e756fSKirill A. Shutemov if (!khugepaged_thread) 2630b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 2631b46e756fSKirill A. Shutemov "khugepaged"); 2632b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 2633b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2634b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 2635b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2636b46e756fSKirill A. Shutemov goto fail; 2637b46e756fSKirill A. Shutemov } 2638b46e756fSKirill A. Shutemov 2639b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 2640b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 2641b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 2642b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 2643b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2644b46e756fSKirill A. Shutemov } 2645bd3400eaSLiangcai Fan set_recommended_min_free_kbytes(); 2646b46e756fSKirill A. Shutemov fail: 2647b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 2648b46e756fSKirill A. Shutemov return err; 2649b46e756fSKirill A. Shutemov } 26504aab2be0SVijay Balakrishna 26514aab2be0SVijay Balakrishna void khugepaged_min_free_kbytes_update(void) 26524aab2be0SVijay Balakrishna { 26534aab2be0SVijay Balakrishna mutex_lock(&khugepaged_mutex); 26541064026bSYang Shi if (hugepage_flags_enabled() && khugepaged_thread) 26554aab2be0SVijay Balakrishna set_recommended_min_free_kbytes(); 26564aab2be0SVijay Balakrishna mutex_unlock(&khugepaged_mutex); 26574aab2be0SVijay Balakrishna } 26587d8faaf1SZach O'Keefe 265957e9cc50SJohannes Weiner bool current_is_khugepaged(void) 266057e9cc50SJohannes Weiner { 266157e9cc50SJohannes Weiner return kthread_func(current) == khugepaged; 266257e9cc50SJohannes Weiner } 266357e9cc50SJohannes Weiner 26647d8faaf1SZach O'Keefe static int madvise_collapse_errno(enum scan_result r) 26657d8faaf1SZach O'Keefe { 26667d8faaf1SZach O'Keefe /* 26677d8faaf1SZach O'Keefe * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide 26687d8faaf1SZach O'Keefe * actionable feedback to caller, so they may take an appropriate 26697d8faaf1SZach O'Keefe * fallback measure depending on the nature of the failure. 26707d8faaf1SZach O'Keefe */ 26717d8faaf1SZach O'Keefe switch (r) { 26727d8faaf1SZach O'Keefe case SCAN_ALLOC_HUGE_PAGE_FAIL: 26737d8faaf1SZach O'Keefe return -ENOMEM; 26747d8faaf1SZach O'Keefe case SCAN_CGROUP_CHARGE_FAIL: 2675ac492b9cSDavid Stevens case SCAN_EXCEED_NONE_PTE: 26767d8faaf1SZach O'Keefe return -EBUSY; 26777d8faaf1SZach O'Keefe /* Resource temporary unavailable - trying again might succeed */ 2678ae63c898SZach O'Keefe case SCAN_PAGE_COUNT: 26797d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 26807d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 26810f3e2a2cSZach O'Keefe case SCAN_DEL_PAGE_LRU: 2682ac492b9cSDavid Stevens case SCAN_PAGE_FILLED: 26837d8faaf1SZach O'Keefe return -EAGAIN; 26847d8faaf1SZach O'Keefe /* 26857d8faaf1SZach O'Keefe * Other: Trying again likely not to succeed / error intrinsic to 26867d8faaf1SZach O'Keefe * specified memory range. khugepaged likely won't be able to collapse 26877d8faaf1SZach O'Keefe * either. 26887d8faaf1SZach O'Keefe */ 26897d8faaf1SZach O'Keefe default: 26907d8faaf1SZach O'Keefe return -EINVAL; 26917d8faaf1SZach O'Keefe } 26927d8faaf1SZach O'Keefe } 26937d8faaf1SZach O'Keefe 26947d8faaf1SZach O'Keefe int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev, 26957d8faaf1SZach O'Keefe unsigned long start, unsigned long end) 26967d8faaf1SZach O'Keefe { 26977d8faaf1SZach O'Keefe struct collapse_control *cc; 26987d8faaf1SZach O'Keefe struct mm_struct *mm = vma->vm_mm; 26997d8faaf1SZach O'Keefe unsigned long hstart, hend, addr; 27007d8faaf1SZach O'Keefe int thps = 0, last_fail = SCAN_FAIL; 27017d8faaf1SZach O'Keefe bool mmap_locked = true; 27027d8faaf1SZach O'Keefe 27037d8faaf1SZach O'Keefe BUG_ON(vma->vm_start > start); 27047d8faaf1SZach O'Keefe BUG_ON(vma->vm_end < end); 27057d8faaf1SZach O'Keefe 27067d8faaf1SZach O'Keefe *prev = vma; 27077d8faaf1SZach O'Keefe 27087d8faaf1SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) 27097d8faaf1SZach O'Keefe return -EINVAL; 27107d8faaf1SZach O'Keefe 27117d8faaf1SZach O'Keefe cc = kmalloc(sizeof(*cc), GFP_KERNEL); 27127d8faaf1SZach O'Keefe if (!cc) 27137d8faaf1SZach O'Keefe return -ENOMEM; 27147d8faaf1SZach O'Keefe cc->is_khugepaged = false; 27157d8faaf1SZach O'Keefe 27167d8faaf1SZach O'Keefe mmgrab(mm); 27177d8faaf1SZach O'Keefe lru_add_drain_all(); 27187d8faaf1SZach O'Keefe 27197d8faaf1SZach O'Keefe hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 27207d8faaf1SZach O'Keefe hend = end & HPAGE_PMD_MASK; 27217d8faaf1SZach O'Keefe 27227d8faaf1SZach O'Keefe for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) { 27237d8faaf1SZach O'Keefe int result = SCAN_FAIL; 27247d8faaf1SZach O'Keefe 27257d8faaf1SZach O'Keefe if (!mmap_locked) { 27267d8faaf1SZach O'Keefe cond_resched(); 27277d8faaf1SZach O'Keefe mmap_read_lock(mm); 27287d8faaf1SZach O'Keefe mmap_locked = true; 272934488399SZach O'Keefe result = hugepage_vma_revalidate(mm, addr, false, &vma, 273034488399SZach O'Keefe cc); 27317d8faaf1SZach O'Keefe if (result != SCAN_SUCCEED) { 27327d8faaf1SZach O'Keefe last_fail = result; 27337d8faaf1SZach O'Keefe goto out_nolock; 27347d8faaf1SZach O'Keefe } 27354d24de94SYang Shi 273652dc0310SZach O'Keefe hend = min(hend, vma->vm_end & HPAGE_PMD_MASK); 27377d8faaf1SZach O'Keefe } 27387d8faaf1SZach O'Keefe mmap_assert_locked(mm); 27397d8faaf1SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 2740e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 274134488399SZach O'Keefe if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 274234488399SZach O'Keefe struct file *file = get_file(vma->vm_file); 274334488399SZach O'Keefe pgoff_t pgoff = linear_page_index(vma, addr); 274434488399SZach O'Keefe 274534488399SZach O'Keefe mmap_read_unlock(mm); 274634488399SZach O'Keefe mmap_locked = false; 274734488399SZach O'Keefe result = hpage_collapse_scan_file(mm, addr, file, pgoff, 27487d2c4385SZach O'Keefe cc); 274934488399SZach O'Keefe fput(file); 275034488399SZach O'Keefe } else { 275134488399SZach O'Keefe result = hpage_collapse_scan_pmd(mm, vma, addr, 275234488399SZach O'Keefe &mmap_locked, cc); 275334488399SZach O'Keefe } 27547d8faaf1SZach O'Keefe if (!mmap_locked) 27557d8faaf1SZach O'Keefe *prev = NULL; /* Tell caller we dropped mmap_lock */ 27567d8faaf1SZach O'Keefe 275734488399SZach O'Keefe handle_result: 27587d8faaf1SZach O'Keefe switch (result) { 27597d8faaf1SZach O'Keefe case SCAN_SUCCEED: 27607d8faaf1SZach O'Keefe case SCAN_PMD_MAPPED: 27617d8faaf1SZach O'Keefe ++thps; 27627d8faaf1SZach O'Keefe break; 276334488399SZach O'Keefe case SCAN_PTE_MAPPED_HUGEPAGE: 276434488399SZach O'Keefe BUG_ON(mmap_locked); 276534488399SZach O'Keefe BUG_ON(*prev); 27661043173eSHugh Dickins mmap_read_lock(mm); 276734488399SZach O'Keefe result = collapse_pte_mapped_thp(mm, addr, true); 27681043173eSHugh Dickins mmap_read_unlock(mm); 276934488399SZach O'Keefe goto handle_result; 27707d8faaf1SZach O'Keefe /* Whitelisted set of results where continuing OK */ 27717d8faaf1SZach O'Keefe case SCAN_PMD_NULL: 27727d8faaf1SZach O'Keefe case SCAN_PTE_NON_PRESENT: 27737d8faaf1SZach O'Keefe case SCAN_PTE_UFFD_WP: 27747d8faaf1SZach O'Keefe case SCAN_PAGE_RO: 27757d8faaf1SZach O'Keefe case SCAN_LACK_REFERENCED_PAGE: 27767d8faaf1SZach O'Keefe case SCAN_PAGE_NULL: 27777d8faaf1SZach O'Keefe case SCAN_PAGE_COUNT: 27787d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 27797d8faaf1SZach O'Keefe case SCAN_PAGE_COMPOUND: 27807d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 27810f3e2a2cSZach O'Keefe case SCAN_DEL_PAGE_LRU: 27827d8faaf1SZach O'Keefe last_fail = result; 27837d8faaf1SZach O'Keefe break; 27847d8faaf1SZach O'Keefe default: 27857d8faaf1SZach O'Keefe last_fail = result; 27867d8faaf1SZach O'Keefe /* Other error, exit */ 27877d8faaf1SZach O'Keefe goto out_maybelock; 27887d8faaf1SZach O'Keefe } 27897d8faaf1SZach O'Keefe } 27907d8faaf1SZach O'Keefe 27917d8faaf1SZach O'Keefe out_maybelock: 27927d8faaf1SZach O'Keefe /* Caller expects us to hold mmap_lock on return */ 27937d8faaf1SZach O'Keefe if (!mmap_locked) 27947d8faaf1SZach O'Keefe mmap_read_lock(mm); 27957d8faaf1SZach O'Keefe out_nolock: 27967d8faaf1SZach O'Keefe mmap_assert_locked(mm); 27977d8faaf1SZach O'Keefe mmdrop(mm); 27987d8faaf1SZach O'Keefe kfree(cc); 27997d8faaf1SZach O'Keefe 28007d8faaf1SZach O'Keefe return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0 28017d8faaf1SZach O'Keefe : madvise_collapse_errno(last_fail); 28027d8faaf1SZach O'Keefe } 2803