1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2b46e756fSKirill A. Shutemov #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 3b46e756fSKirill A. Shutemov 4b46e756fSKirill A. Shutemov #include <linux/mm.h> 5b46e756fSKirill A. Shutemov #include <linux/sched.h> 66e84f315SIngo Molnar #include <linux/sched/mm.h> 7f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 8b46e756fSKirill A. Shutemov #include <linux/mmu_notifier.h> 9b46e756fSKirill A. Shutemov #include <linux/rmap.h> 10b46e756fSKirill A. Shutemov #include <linux/swap.h> 11b46e756fSKirill A. Shutemov #include <linux/mm_inline.h> 12b46e756fSKirill A. Shutemov #include <linux/kthread.h> 13b46e756fSKirill A. Shutemov #include <linux/khugepaged.h> 14b46e756fSKirill A. Shutemov #include <linux/freezer.h> 15b46e756fSKirill A. Shutemov #include <linux/mman.h> 16b46e756fSKirill A. Shutemov #include <linux/hashtable.h> 17b46e756fSKirill A. Shutemov #include <linux/userfaultfd_k.h> 18b46e756fSKirill A. Shutemov #include <linux/page_idle.h> 1980110bbfSPasha Tatashin #include <linux/page_table_check.h> 20b46e756fSKirill A. Shutemov #include <linux/swapops.h> 21f3f0e1d2SKirill A. Shutemov #include <linux/shmem_fs.h> 22b46e756fSKirill A. Shutemov 23b46e756fSKirill A. Shutemov #include <asm/tlb.h> 24b46e756fSKirill A. Shutemov #include <asm/pgalloc.h> 25b46e756fSKirill A. Shutemov #include "internal.h" 26b46e756fSKirill A. Shutemov 27b46e756fSKirill A. Shutemov enum scan_result { 28b46e756fSKirill A. Shutemov SCAN_FAIL, 29b46e756fSKirill A. Shutemov SCAN_SUCCEED, 30b46e756fSKirill A. Shutemov SCAN_PMD_NULL, 31b46e756fSKirill A. Shutemov SCAN_EXCEED_NONE_PTE, 3271a2c112SKirill A. Shutemov SCAN_EXCEED_SWAP_PTE, 3371a2c112SKirill A. Shutemov SCAN_EXCEED_SHARED_PTE, 34b46e756fSKirill A. Shutemov SCAN_PTE_NON_PRESENT, 35e1e267c7SPeter Xu SCAN_PTE_UFFD_WP, 36b46e756fSKirill A. Shutemov SCAN_PAGE_RO, 370db501f7SEbru Akagunduz SCAN_LACK_REFERENCED_PAGE, 38b46e756fSKirill A. Shutemov SCAN_PAGE_NULL, 39b46e756fSKirill A. Shutemov SCAN_SCAN_ABORT, 40b46e756fSKirill A. Shutemov SCAN_PAGE_COUNT, 41b46e756fSKirill A. Shutemov SCAN_PAGE_LRU, 42b46e756fSKirill A. Shutemov SCAN_PAGE_LOCK, 43b46e756fSKirill A. Shutemov SCAN_PAGE_ANON, 44b46e756fSKirill A. Shutemov SCAN_PAGE_COMPOUND, 45b46e756fSKirill A. Shutemov SCAN_ANY_PROCESS, 46b46e756fSKirill A. Shutemov SCAN_VMA_NULL, 47b46e756fSKirill A. Shutemov SCAN_VMA_CHECK, 48b46e756fSKirill A. Shutemov SCAN_ADDRESS_RANGE, 49b46e756fSKirill A. Shutemov SCAN_DEL_PAGE_LRU, 50b46e756fSKirill A. Shutemov SCAN_ALLOC_HUGE_PAGE_FAIL, 51b46e756fSKirill A. Shutemov SCAN_CGROUP_CHARGE_FAIL, 52f3f0e1d2SKirill A. Shutemov SCAN_TRUNCATED, 5399cb0dbdSSong Liu SCAN_PAGE_HAS_PRIVATE, 54b46e756fSKirill A. Shutemov }; 55b46e756fSKirill A. Shutemov 56b46e756fSKirill A. Shutemov #define CREATE_TRACE_POINTS 57b46e756fSKirill A. Shutemov #include <trace/events/huge_memory.h> 58b46e756fSKirill A. Shutemov 594aab2be0SVijay Balakrishna static struct task_struct *khugepaged_thread __read_mostly; 604aab2be0SVijay Balakrishna static DEFINE_MUTEX(khugepaged_mutex); 614aab2be0SVijay Balakrishna 62b46e756fSKirill A. Shutemov /* default scan 8*512 pte (or vmas) every 30 second */ 63b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_to_scan __read_mostly; 64b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_collapsed; 65b46e756fSKirill A. Shutemov static unsigned int khugepaged_full_scans; 66b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 67b46e756fSKirill A. Shutemov /* during fragmentation poll the hugepage allocator once every minute */ 68b46e756fSKirill A. Shutemov static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 69b46e756fSKirill A. Shutemov static unsigned long khugepaged_sleep_expire; 70b46e756fSKirill A. Shutemov static DEFINE_SPINLOCK(khugepaged_mm_lock); 71b46e756fSKirill A. Shutemov static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 72b46e756fSKirill A. Shutemov /* 73b46e756fSKirill A. Shutemov * default collapse hugepages if there is at least one pte mapped like 74b46e756fSKirill A. Shutemov * it would have happened if the vma was large enough during page 75b46e756fSKirill A. Shutemov * fault. 76b46e756fSKirill A. Shutemov */ 77b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_none __read_mostly; 78b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_swap __read_mostly; 7971a2c112SKirill A. Shutemov static unsigned int khugepaged_max_ptes_shared __read_mostly; 80b46e756fSKirill A. Shutemov 81b46e756fSKirill A. Shutemov #define MM_SLOTS_HASH_BITS 10 82b46e756fSKirill A. Shutemov static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 83b46e756fSKirill A. Shutemov 84b46e756fSKirill A. Shutemov static struct kmem_cache *mm_slot_cache __read_mostly; 85b46e756fSKirill A. Shutemov 8627e1f827SSong Liu #define MAX_PTE_MAPPED_THP 8 8727e1f827SSong Liu 8834d6b470SZach O'Keefe struct collapse_control { 8934d6b470SZach O'Keefe /* Num pages scanned per node */ 9034d6b470SZach O'Keefe u32 node_load[MAX_NUMNODES]; 9134d6b470SZach O'Keefe 9234d6b470SZach O'Keefe /* Last target selected in khugepaged_find_target_node() */ 9334d6b470SZach O'Keefe int last_target_node; 9434d6b470SZach O'Keefe }; 9534d6b470SZach O'Keefe 96b46e756fSKirill A. Shutemov /** 97b46e756fSKirill A. Shutemov * struct mm_slot - hash lookup from mm to mm_slot 98b46e756fSKirill A. Shutemov * @hash: hash collision list 99b46e756fSKirill A. Shutemov * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 100b46e756fSKirill A. Shutemov * @mm: the mm that this information is valid for 101336e6b53SAlex Shi * @nr_pte_mapped_thp: number of pte mapped THP 102336e6b53SAlex Shi * @pte_mapped_thp: address array corresponding pte mapped THP 103b46e756fSKirill A. Shutemov */ 104b46e756fSKirill A. Shutemov struct mm_slot { 105b46e756fSKirill A. Shutemov struct hlist_node hash; 106b46e756fSKirill A. Shutemov struct list_head mm_node; 107b46e756fSKirill A. Shutemov struct mm_struct *mm; 10827e1f827SSong Liu 10927e1f827SSong Liu /* pte-mapped THP in this mm */ 11027e1f827SSong Liu int nr_pte_mapped_thp; 11127e1f827SSong Liu unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; 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; 124b46e756fSKirill A. Shutemov struct 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", 388b46e756fSKirill A. Shutemov sizeof(struct mm_slot), 389b46e756fSKirill A. Shutemov __alignof__(struct mm_slot), 0, NULL); 390b46e756fSKirill A. Shutemov if (!mm_slot_cache) 391b46e756fSKirill A. Shutemov return -ENOMEM; 392b46e756fSKirill A. Shutemov 393b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 394b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 395b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 39671a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; 397b46e756fSKirill A. Shutemov 398b46e756fSKirill A. Shutemov return 0; 399b46e756fSKirill A. Shutemov } 400b46e756fSKirill A. Shutemov 401b46e756fSKirill A. Shutemov void __init khugepaged_destroy(void) 402b46e756fSKirill A. Shutemov { 403b46e756fSKirill A. Shutemov kmem_cache_destroy(mm_slot_cache); 404b46e756fSKirill A. Shutemov } 405b46e756fSKirill A. Shutemov 406b46e756fSKirill A. Shutemov static inline struct mm_slot *alloc_mm_slot(void) 407b46e756fSKirill A. Shutemov { 408b46e756fSKirill A. Shutemov if (!mm_slot_cache) /* initialization failed */ 409b46e756fSKirill A. Shutemov return NULL; 410b46e756fSKirill A. Shutemov return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 411b46e756fSKirill A. Shutemov } 412b46e756fSKirill A. Shutemov 413b46e756fSKirill A. Shutemov static inline void free_mm_slot(struct mm_slot *mm_slot) 414b46e756fSKirill A. Shutemov { 415b46e756fSKirill A. Shutemov kmem_cache_free(mm_slot_cache, mm_slot); 416b46e756fSKirill A. Shutemov } 417b46e756fSKirill A. Shutemov 418b46e756fSKirill A. Shutemov static struct mm_slot *get_mm_slot(struct mm_struct *mm) 419b46e756fSKirill A. Shutemov { 420b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 421b46e756fSKirill A. Shutemov 422b46e756fSKirill A. Shutemov hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 423b46e756fSKirill A. Shutemov if (mm == mm_slot->mm) 424b46e756fSKirill A. Shutemov return mm_slot; 425b46e756fSKirill A. Shutemov 426b46e756fSKirill A. Shutemov return NULL; 427b46e756fSKirill A. Shutemov } 428b46e756fSKirill A. Shutemov 429b46e756fSKirill A. Shutemov static void insert_to_mm_slots_hash(struct mm_struct *mm, 430b46e756fSKirill A. Shutemov struct mm_slot *mm_slot) 431b46e756fSKirill A. Shutemov { 432b46e756fSKirill A. Shutemov mm_slot->mm = mm; 433b46e756fSKirill A. Shutemov hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 434b46e756fSKirill A. Shutemov } 435b46e756fSKirill A. Shutemov 436b46e756fSKirill A. Shutemov static inline int khugepaged_test_exit(struct mm_struct *mm) 437b46e756fSKirill A. Shutemov { 4384d45e75aSJann Horn return atomic_read(&mm->mm_users) == 0; 439b46e756fSKirill A. Shutemov } 440b46e756fSKirill A. Shutemov 441d2081b2bSYang Shi void __khugepaged_enter(struct mm_struct *mm) 442b46e756fSKirill A. Shutemov { 443b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 444b46e756fSKirill A. Shutemov int wakeup; 445b46e756fSKirill A. Shutemov 446b46e756fSKirill A. Shutemov mm_slot = alloc_mm_slot(); 447b46e756fSKirill A. Shutemov if (!mm_slot) 448d2081b2bSYang Shi return; 449b46e756fSKirill A. Shutemov 450b46e756fSKirill A. Shutemov /* __khugepaged_exit() must not run from under us */ 45128ff0a3cSMiaohe Lin VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); 452b46e756fSKirill A. Shutemov if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 453b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 454d2081b2bSYang Shi return; 455b46e756fSKirill A. Shutemov } 456b46e756fSKirill A. Shutemov 457b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 458b46e756fSKirill A. Shutemov insert_to_mm_slots_hash(mm, mm_slot); 459b46e756fSKirill A. Shutemov /* 460b46e756fSKirill A. Shutemov * Insert just behind the scanning cursor, to let the area settle 461b46e756fSKirill A. Shutemov * down a little. 462b46e756fSKirill A. Shutemov */ 463b46e756fSKirill A. Shutemov wakeup = list_empty(&khugepaged_scan.mm_head); 464b46e756fSKirill A. Shutemov list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 465b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 466b46e756fSKirill A. Shutemov 467f1f10076SVegard Nossum mmgrab(mm); 468b46e756fSKirill A. Shutemov if (wakeup) 469b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 470b46e756fSKirill A. Shutemov } 471b46e756fSKirill A. Shutemov 472c791576cSYang Shi void khugepaged_enter_vma(struct vm_area_struct *vma, 473b46e756fSKirill A. Shutemov unsigned long vm_flags) 474b46e756fSKirill A. Shutemov { 4752647d11bSYang Shi if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && 4761064026bSYang Shi hugepage_flags_enabled()) { 4777da4e2cbSYang Shi if (hugepage_vma_check(vma, vm_flags, false, false)) 4782647d11bSYang Shi __khugepaged_enter(vma->vm_mm); 4792647d11bSYang Shi } 480b46e756fSKirill A. Shutemov } 481b46e756fSKirill A. Shutemov 482b46e756fSKirill A. Shutemov void __khugepaged_exit(struct mm_struct *mm) 483b46e756fSKirill A. Shutemov { 484b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 485b46e756fSKirill A. Shutemov int free = 0; 486b46e756fSKirill A. Shutemov 487b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 488b46e756fSKirill A. Shutemov mm_slot = get_mm_slot(mm); 489b46e756fSKirill A. Shutemov if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 490b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 491b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 492b46e756fSKirill A. Shutemov free = 1; 493b46e756fSKirill A. Shutemov } 494b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 495b46e756fSKirill A. Shutemov 496b46e756fSKirill A. Shutemov if (free) { 497b46e756fSKirill A. Shutemov clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 498b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 499b46e756fSKirill A. Shutemov mmdrop(mm); 500b46e756fSKirill A. Shutemov } else if (mm_slot) { 501b46e756fSKirill A. Shutemov /* 502b46e756fSKirill A. Shutemov * This is required to serialize against 503b46e756fSKirill A. Shutemov * khugepaged_test_exit() (which is guaranteed to run 504b46e756fSKirill A. Shutemov * under mmap sem read mode). Stop here (after we 505b46e756fSKirill A. Shutemov * return all pagetables will be destroyed) until 506b46e756fSKirill A. Shutemov * khugepaged has finished working on the pagetables 507c1e8d7c6SMichel Lespinasse * under the mmap_lock. 508b46e756fSKirill A. Shutemov */ 509d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 510d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 511b46e756fSKirill A. Shutemov } 512b46e756fSKirill A. Shutemov } 513b46e756fSKirill A. Shutemov 514b46e756fSKirill A. Shutemov static void release_pte_page(struct page *page) 515b46e756fSKirill A. Shutemov { 5165503fbf2SKirill A. Shutemov mod_node_page_state(page_pgdat(page), 5175503fbf2SKirill A. Shutemov NR_ISOLATED_ANON + page_is_file_lru(page), 5185503fbf2SKirill A. Shutemov -compound_nr(page)); 519b46e756fSKirill A. Shutemov unlock_page(page); 520b46e756fSKirill A. Shutemov putback_lru_page(page); 521b46e756fSKirill A. Shutemov } 522b46e756fSKirill A. Shutemov 5235503fbf2SKirill A. Shutemov static void release_pte_pages(pte_t *pte, pte_t *_pte, 5245503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 525b46e756fSKirill A. Shutemov { 5265503fbf2SKirill A. Shutemov struct page *page, *tmp; 5275503fbf2SKirill A. Shutemov 528b46e756fSKirill A. Shutemov while (--_pte >= pte) { 529b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 5305503fbf2SKirill A. Shutemov 5315503fbf2SKirill A. Shutemov page = pte_page(pteval); 5325503fbf2SKirill A. Shutemov if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) && 5335503fbf2SKirill A. Shutemov !PageCompound(page)) 5345503fbf2SKirill A. Shutemov release_pte_page(page); 5355503fbf2SKirill A. Shutemov } 5365503fbf2SKirill A. Shutemov 5375503fbf2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, compound_pagelist, lru) { 5385503fbf2SKirill A. Shutemov list_del(&page->lru); 5395503fbf2SKirill A. Shutemov release_pte_page(page); 540b46e756fSKirill A. Shutemov } 541b46e756fSKirill A. Shutemov } 542b46e756fSKirill A. Shutemov 5439445689fSKirill A. Shutemov static bool is_refcount_suitable(struct page *page) 5449445689fSKirill A. Shutemov { 5459445689fSKirill A. Shutemov int expected_refcount; 5469445689fSKirill A. Shutemov 5479445689fSKirill A. Shutemov expected_refcount = total_mapcount(page); 5489445689fSKirill A. Shutemov if (PageSwapCache(page)) 5499445689fSKirill A. Shutemov expected_refcount += compound_nr(page); 5509445689fSKirill A. Shutemov 5519445689fSKirill A. Shutemov return page_count(page) == expected_refcount; 5529445689fSKirill A. Shutemov } 5539445689fSKirill A. Shutemov 554b46e756fSKirill A. Shutemov static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 555b46e756fSKirill A. Shutemov unsigned long address, 5565503fbf2SKirill A. Shutemov pte_t *pte, 5575503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 558b46e756fSKirill A. Shutemov { 559b46e756fSKirill A. Shutemov struct page *page = NULL; 560b46e756fSKirill A. Shutemov pte_t *_pte; 561*50ad2f24SZach O'Keefe int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0; 5620db501f7SEbru Akagunduz bool writable = false; 563b46e756fSKirill A. Shutemov 564b46e756fSKirill A. Shutemov for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 565b46e756fSKirill A. Shutemov _pte++, address += PAGE_SIZE) { 566b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 567b46e756fSKirill A. Shutemov if (pte_none(pteval) || (pte_present(pteval) && 568b46e756fSKirill A. Shutemov is_zero_pfn(pte_pfn(pteval)))) { 569b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 570b46e756fSKirill A. Shutemov ++none_or_zero <= khugepaged_max_ptes_none) { 571b46e756fSKirill A. Shutemov continue; 572b46e756fSKirill A. Shutemov } else { 573b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 574e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 575b46e756fSKirill A. Shutemov goto out; 576b46e756fSKirill A. Shutemov } 577b46e756fSKirill A. Shutemov } 578b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 579b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 580b46e756fSKirill A. Shutemov goto out; 581b46e756fSKirill A. Shutemov } 582b46e756fSKirill A. Shutemov page = vm_normal_page(vma, address, pteval); 5833218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 584b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 585b46e756fSKirill A. Shutemov goto out; 586b46e756fSKirill A. Shutemov } 587b46e756fSKirill A. Shutemov 588b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageAnon(page), page); 589b46e756fSKirill A. Shutemov 59071a2c112SKirill A. Shutemov if (page_mapcount(page) > 1 && 59171a2c112SKirill A. Shutemov ++shared > khugepaged_max_ptes_shared) { 59271a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 593e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 59471a2c112SKirill A. Shutemov goto out; 59571a2c112SKirill A. Shutemov } 59671a2c112SKirill A. Shutemov 5975503fbf2SKirill A. Shutemov if (PageCompound(page)) { 5985503fbf2SKirill A. Shutemov struct page *p; 5995503fbf2SKirill A. Shutemov page = compound_head(page); 6005503fbf2SKirill A. Shutemov 6015503fbf2SKirill A. Shutemov /* 6025503fbf2SKirill A. Shutemov * Check if we have dealt with the compound page 6035503fbf2SKirill A. Shutemov * already 6045503fbf2SKirill A. Shutemov */ 6055503fbf2SKirill A. Shutemov list_for_each_entry(p, compound_pagelist, lru) { 6065503fbf2SKirill A. Shutemov if (page == p) 6075503fbf2SKirill A. Shutemov goto next; 6085503fbf2SKirill A. Shutemov } 6095503fbf2SKirill A. Shutemov } 6105503fbf2SKirill A. Shutemov 611b46e756fSKirill A. Shutemov /* 612b46e756fSKirill A. Shutemov * We can do it before isolate_lru_page because the 613b46e756fSKirill A. Shutemov * page can't be freed from under us. NOTE: PG_lock 614b46e756fSKirill A. Shutemov * is needed to serialize against split_huge_page 615b46e756fSKirill A. Shutemov * when invoked from the VM. 616b46e756fSKirill A. Shutemov */ 617b46e756fSKirill A. Shutemov if (!trylock_page(page)) { 618b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 619b46e756fSKirill A. Shutemov goto out; 620b46e756fSKirill A. Shutemov } 621b46e756fSKirill A. Shutemov 622b46e756fSKirill A. Shutemov /* 6239445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 6249445689fSKirill A. Shutemov * 6259445689fSKirill A. Shutemov * The page table that maps the page has been already unlinked 6269445689fSKirill A. Shutemov * from the page table tree and this process cannot get 627f0953a1bSIngo Molnar * an additional pin on the page. 6289445689fSKirill A. Shutemov * 6299445689fSKirill A. Shutemov * New pins can come later if the page is shared across fork, 6309445689fSKirill A. Shutemov * but not from this process. The other process cannot write to 6319445689fSKirill A. Shutemov * the page, only trigger CoW. 632b46e756fSKirill A. Shutemov */ 6339445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 634b46e756fSKirill A. Shutemov unlock_page(page); 635b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 636b46e756fSKirill A. Shutemov goto out; 637b46e756fSKirill A. Shutemov } 638b46e756fSKirill A. Shutemov 639b46e756fSKirill A. Shutemov /* 640b46e756fSKirill A. Shutemov * Isolate the page to avoid collapsing an hugepage 641b46e756fSKirill A. Shutemov * currently in use by the VM. 642b46e756fSKirill A. Shutemov */ 643b46e756fSKirill A. Shutemov if (isolate_lru_page(page)) { 644b46e756fSKirill A. Shutemov unlock_page(page); 645b46e756fSKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 646b46e756fSKirill A. Shutemov goto out; 647b46e756fSKirill A. Shutemov } 6485503fbf2SKirill A. Shutemov mod_node_page_state(page_pgdat(page), 6495503fbf2SKirill A. Shutemov NR_ISOLATED_ANON + page_is_file_lru(page), 6505503fbf2SKirill A. Shutemov compound_nr(page)); 651b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 652b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(PageLRU(page), page); 653b46e756fSKirill A. Shutemov 6545503fbf2SKirill A. Shutemov if (PageCompound(page)) 6555503fbf2SKirill A. Shutemov list_add_tail(&page->lru, compound_pagelist); 6565503fbf2SKirill A. Shutemov next: 6570db501f7SEbru Akagunduz /* There should be enough young pte to collapse the page */ 658b46e756fSKirill A. Shutemov if (pte_young(pteval) || 659b46e756fSKirill A. Shutemov page_is_young(page) || PageReferenced(page) || 660b46e756fSKirill A. Shutemov mmu_notifier_test_young(vma->vm_mm, address)) 6610db501f7SEbru Akagunduz referenced++; 6625503fbf2SKirill A. Shutemov 6635503fbf2SKirill A. Shutemov if (pte_write(pteval)) 6645503fbf2SKirill A. Shutemov writable = true; 665b46e756fSKirill A. Shutemov } 66674e579bfSMiaohe Lin 66774e579bfSMiaohe Lin if (unlikely(!writable)) { 66874e579bfSMiaohe Lin result = SCAN_PAGE_RO; 66974e579bfSMiaohe Lin } else if (unlikely(!referenced)) { 67074e579bfSMiaohe Lin result = SCAN_LACK_REFERENCED_PAGE; 67174e579bfSMiaohe Lin } else { 672b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 673b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 674b46e756fSKirill A. Shutemov referenced, writable, result); 675*50ad2f24SZach O'Keefe return result; 676b46e756fSKirill A. Shutemov } 677b46e756fSKirill A. Shutemov out: 6785503fbf2SKirill A. Shutemov release_pte_pages(pte, _pte, compound_pagelist); 679b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 680b46e756fSKirill A. Shutemov referenced, writable, result); 681*50ad2f24SZach O'Keefe return result; 682b46e756fSKirill A. Shutemov } 683b46e756fSKirill A. Shutemov 684b46e756fSKirill A. Shutemov static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 685b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 686b46e756fSKirill A. Shutemov unsigned long address, 6875503fbf2SKirill A. Shutemov spinlock_t *ptl, 6885503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 689b46e756fSKirill A. Shutemov { 6905503fbf2SKirill A. Shutemov struct page *src_page, *tmp; 691b46e756fSKirill A. Shutemov pte_t *_pte; 692338a16baSDavid Rientjes for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 693338a16baSDavid Rientjes _pte++, page++, address += PAGE_SIZE) { 694b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 695b46e756fSKirill A. Shutemov 696b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 697b46e756fSKirill A. Shutemov clear_user_highpage(page, address); 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); 706b46e756fSKirill A. Shutemov } 707b46e756fSKirill A. Shutemov } else { 708b46e756fSKirill A. Shutemov src_page = pte_page(pteval); 709b46e756fSKirill A. Shutemov copy_user_highpage(page, src_page, address, vma); 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 736b46e756fSKirill A. Shutemov static void khugepaged_alloc_sleep(void) 737b46e756fSKirill A. Shutemov { 738b46e756fSKirill A. Shutemov DEFINE_WAIT(wait); 739b46e756fSKirill A. Shutemov 740b46e756fSKirill A. Shutemov add_wait_queue(&khugepaged_wait, &wait); 741b46e756fSKirill A. Shutemov freezable_schedule_timeout_interruptible( 742b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 743b46e756fSKirill A. Shutemov remove_wait_queue(&khugepaged_wait, &wait); 744b46e756fSKirill A. Shutemov } 745b46e756fSKirill A. Shutemov 746b46e756fSKirill A. Shutemov 74734d6b470SZach O'Keefe struct collapse_control khugepaged_collapse_control = { 74834d6b470SZach O'Keefe .last_target_node = NUMA_NO_NODE, 74934d6b470SZach O'Keefe }; 75034d6b470SZach O'Keefe 75134d6b470SZach O'Keefe static bool khugepaged_scan_abort(int nid, struct collapse_control *cc) 752b46e756fSKirill A. Shutemov { 753b46e756fSKirill A. Shutemov int i; 754b46e756fSKirill A. Shutemov 755b46e756fSKirill A. Shutemov /* 756a5f5f91dSMel Gorman * If node_reclaim_mode is disabled, then no extra effort is made to 757b46e756fSKirill A. Shutemov * allocate memory locally. 758b46e756fSKirill A. Shutemov */ 759202e35dbSDave Hansen if (!node_reclaim_enabled()) 760b46e756fSKirill A. Shutemov return false; 761b46e756fSKirill A. Shutemov 762b46e756fSKirill A. Shutemov /* If there is a count for this node already, it must be acceptable */ 76334d6b470SZach O'Keefe if (cc->node_load[nid]) 764b46e756fSKirill A. Shutemov return false; 765b46e756fSKirill A. Shutemov 766b46e756fSKirill A. Shutemov for (i = 0; i < MAX_NUMNODES; i++) { 76734d6b470SZach O'Keefe if (!cc->node_load[i]) 768b46e756fSKirill A. Shutemov continue; 769a55c7454SMatt Fleming if (node_distance(nid, i) > node_reclaim_distance) 770b46e756fSKirill A. Shutemov return true; 771b46e756fSKirill A. Shutemov } 772b46e756fSKirill A. Shutemov return false; 773b46e756fSKirill A. Shutemov } 774b46e756fSKirill A. Shutemov 7751064026bSYang Shi #define khugepaged_defrag() \ 7761064026bSYang Shi (transparent_hugepage_flags & \ 7771064026bSYang Shi (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)) 7781064026bSYang Shi 779b46e756fSKirill A. Shutemov /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 780b46e756fSKirill A. Shutemov static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 781b46e756fSKirill A. Shutemov { 78225160354SVlastimil Babka return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 783b46e756fSKirill A. Shutemov } 784b46e756fSKirill A. Shutemov 785b46e756fSKirill A. Shutemov #ifdef CONFIG_NUMA 78634d6b470SZach O'Keefe static int khugepaged_find_target_node(struct collapse_control *cc) 787b46e756fSKirill A. Shutemov { 788b46e756fSKirill A. Shutemov int nid, target_node = 0, max_value = 0; 789b46e756fSKirill A. Shutemov 790b46e756fSKirill A. Shutemov /* find first node with max normal pages hit */ 791b46e756fSKirill A. Shutemov for (nid = 0; nid < MAX_NUMNODES; nid++) 79234d6b470SZach O'Keefe if (cc->node_load[nid] > max_value) { 79334d6b470SZach O'Keefe max_value = cc->node_load[nid]; 794b46e756fSKirill A. Shutemov target_node = nid; 795b46e756fSKirill A. Shutemov } 796b46e756fSKirill A. Shutemov 797b46e756fSKirill A. Shutemov /* do some balance if several nodes have the same hit record */ 79834d6b470SZach O'Keefe if (target_node <= cc->last_target_node) 79934d6b470SZach O'Keefe for (nid = cc->last_target_node + 1; nid < MAX_NUMNODES; 800b46e756fSKirill A. Shutemov nid++) 80134d6b470SZach O'Keefe if (max_value == cc->node_load[nid]) { 802b46e756fSKirill A. Shutemov target_node = nid; 803b46e756fSKirill A. Shutemov break; 804b46e756fSKirill A. Shutemov } 805b46e756fSKirill A. Shutemov 80634d6b470SZach O'Keefe cc->last_target_node = target_node; 807b46e756fSKirill A. Shutemov return target_node; 808b46e756fSKirill A. Shutemov } 809c6a7f445SYang Shi #else 81034d6b470SZach O'Keefe static int khugepaged_find_target_node(struct collapse_control *cc) 811b46e756fSKirill A. Shutemov { 812c6a7f445SYang Shi return 0; 813b46e756fSKirill A. Shutemov } 814c6a7f445SYang Shi #endif 815b46e756fSKirill A. Shutemov 8169710a78aSZach O'Keefe static bool khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) 817b46e756fSKirill A. Shutemov { 818b46e756fSKirill A. Shutemov *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); 819b46e756fSKirill A. Shutemov if (unlikely(!*hpage)) { 820b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 8219710a78aSZach O'Keefe return false; 822b46e756fSKirill A. Shutemov } 823b46e756fSKirill A. Shutemov 824b46e756fSKirill A. Shutemov prep_transhuge_page(*hpage); 825b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 8269710a78aSZach O'Keefe return true; 827b46e756fSKirill A. Shutemov } 828b46e756fSKirill A. Shutemov 829b46e756fSKirill A. Shutemov /* 830c1e8d7c6SMichel Lespinasse * If mmap_lock temporarily dropped, revalidate vma 831c1e8d7c6SMichel Lespinasse * before taking mmap_lock. 832*50ad2f24SZach O'Keefe * Returns enum scan_result value. 833b46e756fSKirill A. Shutemov */ 834b46e756fSKirill A. Shutemov 835c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 836c131f751SKirill A. Shutemov struct vm_area_struct **vmap) 837b46e756fSKirill A. Shutemov { 838b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 839b46e756fSKirill A. Shutemov 840b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 841b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 842b46e756fSKirill A. Shutemov 843c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 844b46e756fSKirill A. Shutemov if (!vma) 845b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 846b46e756fSKirill A. Shutemov 8474fa6893fSYang Shi if (!transhuge_vma_suitable(vma, address)) 848b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 8497da4e2cbSYang Shi if (!hugepage_vma_check(vma, vma->vm_flags, false, false)) 850b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 851f707fa49SYang Shi /* 852f707fa49SYang Shi * Anon VMA expected, the address may be unmapped then 853f707fa49SYang Shi * remapped to file after khugepaged reaquired the mmap_lock. 854f707fa49SYang Shi * 855f707fa49SYang Shi * hugepage_vma_check may return true for qualified file 856f707fa49SYang Shi * vmas. 857f707fa49SYang Shi */ 85825fa414aSxu xin if (!vma->anon_vma || !vma_is_anonymous(vma)) 859594cced1SKirill A. Shutemov return SCAN_VMA_CHECK; 860*50ad2f24SZach O'Keefe return SCAN_SUCCEED; 861b46e756fSKirill A. Shutemov } 862b46e756fSKirill A. Shutemov 863b46e756fSKirill A. Shutemov /* 864b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 865b46e756fSKirill A. Shutemov * Only done if khugepaged_scan_pmd believes it is worthwhile. 866b46e756fSKirill A. Shutemov * 8674d928e20SMiaohe Lin * Called and returns without pte mapped or spinlocks held. 8684d928e20SMiaohe Lin * Note that if false is returned, mmap_lock will be released. 869b46e756fSKirill A. Shutemov */ 870b46e756fSKirill A. Shutemov 871*50ad2f24SZach O'Keefe static int __collapse_huge_page_swapin(struct mm_struct *mm, 872b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 8732b635dd3SWill Deacon unsigned long haddr, pmd_t *pmd, 8740db501f7SEbru Akagunduz int referenced) 875b46e756fSKirill A. Shutemov { 8762b740303SSouptick Joarder int swapped_in = 0; 8772b740303SSouptick Joarder vm_fault_t ret = 0; 8782b635dd3SWill Deacon unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); 8792b635dd3SWill Deacon 8802b635dd3SWill Deacon for (address = haddr; address < end; address += PAGE_SIZE) { 88182b0f8c3SJan Kara struct vm_fault vmf = { 882b46e756fSKirill A. Shutemov .vma = vma, 883b46e756fSKirill A. Shutemov .address = address, 8842b635dd3SWill Deacon .pgoff = linear_page_index(vma, haddr), 885b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 886b46e756fSKirill A. Shutemov .pmd = pmd, 887b46e756fSKirill A. Shutemov }; 888b46e756fSKirill A. Shutemov 88982b0f8c3SJan Kara vmf.pte = pte_offset_map(pmd, address); 8902994302bSJan Kara vmf.orig_pte = *vmf.pte; 8912b635dd3SWill Deacon if (!is_swap_pte(vmf.orig_pte)) { 8922b635dd3SWill Deacon pte_unmap(vmf.pte); 893b46e756fSKirill A. Shutemov continue; 8942b635dd3SWill Deacon } 8952994302bSJan Kara ret = do_swap_page(&vmf); 8960db501f7SEbru Akagunduz 8974d928e20SMiaohe Lin /* 8984d928e20SMiaohe Lin * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. 8994d928e20SMiaohe Lin * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because 9004d928e20SMiaohe Lin * we do not retry here and swap entry will remain in pagetable 9014d928e20SMiaohe Lin * resulting in later failure. 9024d928e20SMiaohe Lin */ 903b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 9040db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 905*50ad2f24SZach O'Keefe /* Likely, but not guaranteed, that page lock failed */ 906*50ad2f24SZach O'Keefe return SCAN_PAGE_LOCK; 90747f863eaSEbru Akagunduz } 908b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 9094d928e20SMiaohe Lin mmap_read_unlock(mm); 9100db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 911*50ad2f24SZach O'Keefe return SCAN_FAIL; 912b46e756fSKirill A. Shutemov } 9134d928e20SMiaohe Lin swapped_in++; 914b46e756fSKirill A. Shutemov } 915ae2c5d80SKirill A. Shutemov 916ae2c5d80SKirill A. Shutemov /* Drain LRU add pagevec to remove extra pin on the swapped in pages */ 917ae2c5d80SKirill A. Shutemov if (swapped_in) 918ae2c5d80SKirill A. Shutemov lru_add_drain(); 919ae2c5d80SKirill A. Shutemov 9200db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); 921*50ad2f24SZach O'Keefe return SCAN_SUCCEED; 922b46e756fSKirill A. Shutemov } 923b46e756fSKirill A. Shutemov 9249710a78aSZach O'Keefe static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, 9259710a78aSZach O'Keefe struct collapse_control *cc) 9269710a78aSZach O'Keefe { 9279710a78aSZach O'Keefe /* Only allocate from the target node */ 9289710a78aSZach O'Keefe gfp_t gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 9299710a78aSZach O'Keefe int node = khugepaged_find_target_node(cc); 9309710a78aSZach O'Keefe 9319710a78aSZach O'Keefe if (!khugepaged_alloc_page(hpage, gfp, node)) 9329710a78aSZach O'Keefe return SCAN_ALLOC_HUGE_PAGE_FAIL; 9339710a78aSZach O'Keefe if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp))) 9349710a78aSZach O'Keefe return SCAN_CGROUP_CHARGE_FAIL; 9359710a78aSZach O'Keefe count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC); 9369710a78aSZach O'Keefe return SCAN_SUCCEED; 9379710a78aSZach O'Keefe } 9389710a78aSZach O'Keefe 939*50ad2f24SZach O'Keefe static int collapse_huge_page(struct mm_struct *mm, unsigned long address, 940*50ad2f24SZach O'Keefe int referenced, int unmapped, 941*50ad2f24SZach O'Keefe struct collapse_control *cc) 942b46e756fSKirill A. Shutemov { 9435503fbf2SKirill A. Shutemov LIST_HEAD(compound_pagelist); 944b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 945b46e756fSKirill A. Shutemov pte_t *pte; 946b46e756fSKirill A. Shutemov pgtable_t pgtable; 947*50ad2f24SZach O'Keefe struct page *hpage; 948b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 949*50ad2f24SZach O'Keefe int result = SCAN_FAIL; 950c131f751SKirill A. Shutemov struct vm_area_struct *vma; 951ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 952b46e756fSKirill A. Shutemov 953b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 954b46e756fSKirill A. Shutemov 955988ddb71SKirill A. Shutemov /* 956c1e8d7c6SMichel Lespinasse * Before allocating the hugepage, release the mmap_lock read lock. 957988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 958c1e8d7c6SMichel Lespinasse * sync compaction, and we do not need to hold the mmap_lock during 959988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 960988ddb71SKirill A. Shutemov */ 961d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 962b46e756fSKirill A. Shutemov 963*50ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 9649710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 965b46e756fSKirill A. Shutemov goto out_nolock; 9669710a78aSZach O'Keefe 967d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 968c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 969*50ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) { 970d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 971b46e756fSKirill A. Shutemov goto out_nolock; 972b46e756fSKirill A. Shutemov } 973b46e756fSKirill A. Shutemov 974b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 975b46e756fSKirill A. Shutemov if (!pmd) { 976b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 977d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 978b46e756fSKirill A. Shutemov goto out_nolock; 979b46e756fSKirill A. Shutemov } 980b46e756fSKirill A. Shutemov 981*50ad2f24SZach O'Keefe if (unmapped) { 982b46e756fSKirill A. Shutemov /* 983*50ad2f24SZach O'Keefe * __collapse_huge_page_swapin will return with mmap_lock 984*50ad2f24SZach O'Keefe * released when it fails. So we jump out_nolock directly in 985*50ad2f24SZach O'Keefe * that case. Continuing to collapse causes inconsistency. 986b46e756fSKirill A. Shutemov */ 987*50ad2f24SZach O'Keefe result = __collapse_huge_page_swapin(mm, vma, address, pmd, 988*50ad2f24SZach O'Keefe referenced); 989*50ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 990b46e756fSKirill A. Shutemov goto out_nolock; 991b46e756fSKirill A. Shutemov } 992b46e756fSKirill A. Shutemov 993d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 994b46e756fSKirill A. Shutemov /* 995b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 996b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 997b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 998b46e756fSKirill A. Shutemov */ 999d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 1000c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 1001*50ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 100218d24a7cSMiaohe Lin goto out_up_write; 1003b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 1004b46e756fSKirill A. Shutemov if (mm_find_pmd(mm, address) != pmd) 100518d24a7cSMiaohe Lin goto out_up_write; 1006b46e756fSKirill A. Shutemov 1007b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1008b46e756fSKirill A. Shutemov 10097269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 10106f4f13e8SJérôme Glisse address, address + HPAGE_PMD_SIZE); 1011ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1012ec649c9dSVille Syrjälä 1013ec649c9dSVille Syrjälä pte = pte_offset_map(pmd, address); 1014ec649c9dSVille Syrjälä pte_ptl = pte_lockptr(mm, pmd); 1015ec649c9dSVille Syrjälä 1016b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1017b46e756fSKirill A. Shutemov /* 1018b46e756fSKirill A. Shutemov * After this gup_fast can't run anymore. This also removes 1019b46e756fSKirill A. Shutemov * any huge TLB entry from the CPU so we won't allow 1020b46e756fSKirill A. Shutemov * huge and small TLB entries for the same virtual address 1021b46e756fSKirill A. Shutemov * to avoid the risk of CPU bugs in that area. 1022b46e756fSKirill A. Shutemov */ 1023b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1024b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1025ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1026b46e756fSKirill A. Shutemov 1027b46e756fSKirill A. Shutemov spin_lock(pte_ptl); 1028*50ad2f24SZach O'Keefe result = __collapse_huge_page_isolate(vma, address, pte, 10295503fbf2SKirill A. Shutemov &compound_pagelist); 1030b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1031b46e756fSKirill A. Shutemov 1032*50ad2f24SZach O'Keefe if (unlikely(result != SCAN_SUCCEED)) { 1033b46e756fSKirill A. Shutemov pte_unmap(pte); 1034b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1035b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1036b46e756fSKirill A. Shutemov /* 1037b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1038b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1039b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1040b46e756fSKirill A. Shutemov */ 1041b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1042b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1043b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 104418d24a7cSMiaohe Lin goto out_up_write; 1045b46e756fSKirill A. Shutemov } 1046b46e756fSKirill A. Shutemov 1047b46e756fSKirill A. Shutemov /* 1048b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1049b46e756fSKirill A. Shutemov * can't run anymore. 1050b46e756fSKirill A. Shutemov */ 1051b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1052b46e756fSKirill A. Shutemov 1053*50ad2f24SZach O'Keefe __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl, 10545503fbf2SKirill A. Shutemov &compound_pagelist); 1055b46e756fSKirill A. Shutemov pte_unmap(pte); 1056588d01f9SMiaohe Lin /* 1057588d01f9SMiaohe Lin * spin_lock() below is not the equivalent of smp_wmb(), but 1058588d01f9SMiaohe Lin * the smp_wmb() inside __SetPageUptodate() can be reused to 1059588d01f9SMiaohe Lin * avoid the copy_huge_page writes to become visible after 1060588d01f9SMiaohe Lin * the set_pmd_at() write. 1061588d01f9SMiaohe Lin */ 1062*50ad2f24SZach O'Keefe __SetPageUptodate(hpage); 1063b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1064b46e756fSKirill A. Shutemov 1065*50ad2f24SZach O'Keefe _pmd = mk_huge_pmd(hpage, vma->vm_page_prot); 1066f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1067b46e756fSKirill A. Shutemov 1068b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1069b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1070*50ad2f24SZach O'Keefe page_add_new_anon_rmap(hpage, vma, address); 1071*50ad2f24SZach O'Keefe lru_cache_add_inactive_or_unevictable(hpage, vma); 1072b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1073b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1074b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1075b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1076b46e756fSKirill A. Shutemov 1077*50ad2f24SZach O'Keefe hpage = NULL; 1078b46e756fSKirill A. Shutemov 1079b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1080b46e756fSKirill A. Shutemov out_up_write: 1081d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 1082b46e756fSKirill A. Shutemov out_nolock: 1083*50ad2f24SZach O'Keefe if (hpage) { 1084*50ad2f24SZach O'Keefe mem_cgroup_uncharge(page_folio(hpage)); 1085*50ad2f24SZach O'Keefe put_page(hpage); 1086c6a7f445SYang Shi } 1087*50ad2f24SZach O'Keefe trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); 1088*50ad2f24SZach O'Keefe return result; 1089b46e756fSKirill A. Shutemov } 1090b46e756fSKirill A. Shutemov 109134d6b470SZach O'Keefe static int khugepaged_scan_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 1092*50ad2f24SZach O'Keefe unsigned long address, bool *mmap_locked, 109334d6b470SZach O'Keefe struct collapse_control *cc) 1094b46e756fSKirill A. Shutemov { 1095b46e756fSKirill A. Shutemov pmd_t *pmd; 1096b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 1097*50ad2f24SZach O'Keefe int result = SCAN_FAIL, referenced = 0; 109871a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0; 1099b46e756fSKirill A. Shutemov struct page *page = NULL; 1100b46e756fSKirill A. Shutemov unsigned long _address; 1101b46e756fSKirill A. Shutemov spinlock_t *ptl; 1102b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 11030db501f7SEbru Akagunduz bool writable = false; 1104b46e756fSKirill A. Shutemov 1105b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1106b46e756fSKirill A. Shutemov 1107b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 1108b46e756fSKirill A. Shutemov if (!pmd) { 1109b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 1110b46e756fSKirill A. Shutemov goto out; 1111b46e756fSKirill A. Shutemov } 1112b46e756fSKirill A. Shutemov 111334d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1114b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1115b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; 1116b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1117b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 1118b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1119b46e756fSKirill A. Shutemov if (++unmapped <= khugepaged_max_ptes_swap) { 1120e1e267c7SPeter Xu /* 1121e1e267c7SPeter Xu * Always be strict with uffd-wp 1122e1e267c7SPeter Xu * enabled swap entries. Please see 1123e1e267c7SPeter Xu * comment below for pte_uffd_wp(). 1124e1e267c7SPeter Xu */ 1125e1e267c7SPeter Xu if (pte_swp_uffd_wp(pteval)) { 1126e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1127e1e267c7SPeter Xu goto out_unmap; 1128e1e267c7SPeter Xu } 1129b46e756fSKirill A. Shutemov continue; 1130b46e756fSKirill A. Shutemov } else { 1131b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1132e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1133b46e756fSKirill A. Shutemov goto out_unmap; 1134b46e756fSKirill A. Shutemov } 1135b46e756fSKirill A. Shutemov } 1136b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1137b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1138b46e756fSKirill A. Shutemov ++none_or_zero <= khugepaged_max_ptes_none) { 1139b46e756fSKirill A. Shutemov continue; 1140b46e756fSKirill A. Shutemov } else { 1141b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1142e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1143b46e756fSKirill A. Shutemov goto out_unmap; 1144b46e756fSKirill A. Shutemov } 1145b46e756fSKirill A. Shutemov } 1146e1e267c7SPeter Xu if (pte_uffd_wp(pteval)) { 1147e1e267c7SPeter Xu /* 1148e1e267c7SPeter Xu * Don't collapse the page if any of the small 1149e1e267c7SPeter Xu * PTEs are armed with uffd write protection. 1150e1e267c7SPeter Xu * Here we can also mark the new huge pmd as 1151e1e267c7SPeter Xu * write protected if any of the small ones is 11528958b249SHaitao Shi * marked but that could bring unknown 1153e1e267c7SPeter Xu * userfault messages that falls outside of 1154e1e267c7SPeter Xu * the registered range. So, just be simple. 1155e1e267c7SPeter Xu */ 1156e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1157e1e267c7SPeter Xu goto out_unmap; 1158e1e267c7SPeter Xu } 1159b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1160b46e756fSKirill A. Shutemov writable = true; 1161b46e756fSKirill A. Shutemov 1162b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 11633218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 1164b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1165b46e756fSKirill A. Shutemov goto out_unmap; 1166b46e756fSKirill A. Shutemov } 1167b46e756fSKirill A. Shutemov 116871a2c112SKirill A. Shutemov if (page_mapcount(page) > 1 && 116971a2c112SKirill A. Shutemov ++shared > khugepaged_max_ptes_shared) { 117071a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 1171e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 117271a2c112SKirill A. Shutemov goto out_unmap; 117371a2c112SKirill A. Shutemov } 117471a2c112SKirill A. Shutemov 11755503fbf2SKirill A. Shutemov page = compound_head(page); 1176b46e756fSKirill A. Shutemov 1177b46e756fSKirill A. Shutemov /* 1178b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 117934d6b470SZach O'Keefe * information to cc->node_load[]. 11800b8f0d87SQuanfa Fu * Khugepaged will allocate hugepage from the node has the max 1181b46e756fSKirill A. Shutemov * hit record. 1182b46e756fSKirill A. Shutemov */ 1183b46e756fSKirill A. Shutemov node = page_to_nid(page); 118434d6b470SZach O'Keefe if (khugepaged_scan_abort(node, cc)) { 1185b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1186b46e756fSKirill A. Shutemov goto out_unmap; 1187b46e756fSKirill A. Shutemov } 118834d6b470SZach O'Keefe cc->node_load[node]++; 1189b46e756fSKirill A. Shutemov if (!PageLRU(page)) { 1190b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1191b46e756fSKirill A. Shutemov goto out_unmap; 1192b46e756fSKirill A. Shutemov } 1193b46e756fSKirill A. Shutemov if (PageLocked(page)) { 1194b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1195b46e756fSKirill A. Shutemov goto out_unmap; 1196b46e756fSKirill A. Shutemov } 1197b46e756fSKirill A. Shutemov if (!PageAnon(page)) { 1198b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1199b46e756fSKirill A. Shutemov goto out_unmap; 1200b46e756fSKirill A. Shutemov } 1201b46e756fSKirill A. Shutemov 1202b46e756fSKirill A. Shutemov /* 12039445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 12049445689fSKirill A. Shutemov * 120536ee2c78SMiaohe Lin * Here the check is racy it may see total_mapcount > refcount 12069445689fSKirill A. Shutemov * in some cases. 12079445689fSKirill A. Shutemov * For example, one process with one forked child process. 12089445689fSKirill A. Shutemov * The parent has the PMD split due to MADV_DONTNEED, then 12099445689fSKirill A. Shutemov * the child is trying unmap the whole PMD, but khugepaged 12109445689fSKirill A. Shutemov * may be scanning the parent between the child has 12119445689fSKirill A. Shutemov * PageDoubleMap flag cleared and dec the mapcount. So 12129445689fSKirill A. Shutemov * khugepaged may see total_mapcount > refcount. 12139445689fSKirill A. Shutemov * 12149445689fSKirill A. Shutemov * But such case is ephemeral we could always retry collapse 12159445689fSKirill A. Shutemov * later. However it may report false positive if the page 12169445689fSKirill A. Shutemov * has excessive GUP pins (i.e. 512). Anyway the same check 12179445689fSKirill A. Shutemov * will be done again later the risk seems low. 1218b46e756fSKirill A. Shutemov */ 12199445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 1220b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1221b46e756fSKirill A. Shutemov goto out_unmap; 1222b46e756fSKirill A. Shutemov } 1223b46e756fSKirill A. Shutemov if (pte_young(pteval) || 1224b46e756fSKirill A. Shutemov page_is_young(page) || PageReferenced(page) || 1225b46e756fSKirill A. Shutemov mmu_notifier_test_young(vma->vm_mm, address)) 12260db501f7SEbru Akagunduz referenced++; 1227b46e756fSKirill A. Shutemov } 1228ffe945e6SKirill A. Shutemov if (!writable) { 1229ffe945e6SKirill A. Shutemov result = SCAN_PAGE_RO; 1230ffe945e6SKirill A. Shutemov } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) { 1231ffe945e6SKirill A. Shutemov result = SCAN_LACK_REFERENCED_PAGE; 1232ffe945e6SKirill A. Shutemov } else { 1233b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1234b46e756fSKirill A. Shutemov } 1235b46e756fSKirill A. Shutemov out_unmap: 1236b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 1237*50ad2f24SZach O'Keefe if (result == SCAN_SUCCEED) { 1238*50ad2f24SZach O'Keefe result = collapse_huge_page(mm, address, referenced, 1239*50ad2f24SZach O'Keefe unmapped, cc); 1240c1e8d7c6SMichel Lespinasse /* collapse_huge_page will return with the mmap_lock released */ 1241*50ad2f24SZach O'Keefe *mmap_locked = false; 1242b46e756fSKirill A. Shutemov } 1243b46e756fSKirill A. Shutemov out: 1244b46e756fSKirill A. Shutemov trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, 1245b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 1246*50ad2f24SZach O'Keefe return result; 1247b46e756fSKirill A. Shutemov } 1248b46e756fSKirill A. Shutemov 1249b46e756fSKirill A. Shutemov static void collect_mm_slot(struct mm_slot *mm_slot) 1250b46e756fSKirill A. Shutemov { 1251b46e756fSKirill A. Shutemov struct mm_struct *mm = mm_slot->mm; 1252b46e756fSKirill A. Shutemov 125335f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1254b46e756fSKirill A. Shutemov 1255b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm)) { 1256b46e756fSKirill A. Shutemov /* free mm_slot */ 1257b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 1258b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 1259b46e756fSKirill A. Shutemov 1260b46e756fSKirill A. Shutemov /* 1261b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1262b46e756fSKirill A. Shutemov * 1263b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1264b46e756fSKirill A. Shutemov */ 1265b46e756fSKirill A. Shutemov 1266b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1267b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 1268b46e756fSKirill A. Shutemov mmdrop(mm); 1269b46e756fSKirill A. Shutemov } 1270b46e756fSKirill A. Shutemov } 1271b46e756fSKirill A. Shutemov 1272396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 127327e1f827SSong Liu /* 127427e1f827SSong Liu * Notify khugepaged that given addr of the mm is pte-mapped THP. Then 127527e1f827SSong Liu * khugepaged should try to collapse the page table. 127627e1f827SSong Liu */ 1277081c3256SMiaohe Lin static void khugepaged_add_pte_mapped_thp(struct mm_struct *mm, 127827e1f827SSong Liu unsigned long addr) 127927e1f827SSong Liu { 128027e1f827SSong Liu struct mm_slot *mm_slot; 128127e1f827SSong Liu 128227e1f827SSong Liu VM_BUG_ON(addr & ~HPAGE_PMD_MASK); 128327e1f827SSong Liu 128427e1f827SSong Liu spin_lock(&khugepaged_mm_lock); 128527e1f827SSong Liu mm_slot = get_mm_slot(mm); 128627e1f827SSong Liu if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) 128727e1f827SSong Liu mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; 128827e1f827SSong Liu spin_unlock(&khugepaged_mm_lock); 128927e1f827SSong Liu } 129027e1f827SSong Liu 1291e59a47b8SPasha Tatashin static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 1292e59a47b8SPasha Tatashin unsigned long addr, pmd_t *pmdp) 1293e59a47b8SPasha Tatashin { 1294e59a47b8SPasha Tatashin spinlock_t *ptl; 1295e59a47b8SPasha Tatashin pmd_t pmd; 1296e59a47b8SPasha Tatashin 129780110bbfSPasha Tatashin mmap_assert_write_locked(mm); 1298e59a47b8SPasha Tatashin ptl = pmd_lock(vma->vm_mm, pmdp); 1299e59a47b8SPasha Tatashin pmd = pmdp_collapse_flush(vma, addr, pmdp); 1300e59a47b8SPasha Tatashin spin_unlock(ptl); 1301e59a47b8SPasha Tatashin mm_dec_nr_ptes(mm); 130280110bbfSPasha Tatashin page_table_check_pte_clear_range(mm, addr, pmd); 1303e59a47b8SPasha Tatashin pte_free(mm, pmd_pgtable(pmd)); 1304e59a47b8SPasha Tatashin } 1305e59a47b8SPasha Tatashin 130627e1f827SSong Liu /** 1307336e6b53SAlex Shi * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at 1308336e6b53SAlex Shi * address haddr. 1309336e6b53SAlex Shi * 1310336e6b53SAlex Shi * @mm: process address space where collapse happens 1311336e6b53SAlex Shi * @addr: THP collapse address 131227e1f827SSong Liu * 131327e1f827SSong Liu * This function checks whether all the PTEs in the PMD are pointing to the 131427e1f827SSong Liu * right THP. If so, retract the page table so the THP can refault in with 131527e1f827SSong Liu * as pmd-mapped. 131627e1f827SSong Liu */ 131727e1f827SSong Liu void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr) 131827e1f827SSong Liu { 131927e1f827SSong Liu unsigned long haddr = addr & HPAGE_PMD_MASK; 132027e1f827SSong Liu struct vm_area_struct *vma = find_vma(mm, haddr); 1321119a5fc1SHugh Dickins struct page *hpage; 132227e1f827SSong Liu pte_t *start_pte, *pte; 1323e59a47b8SPasha Tatashin pmd_t *pmd; 132427e1f827SSong Liu spinlock_t *ptl; 132527e1f827SSong Liu int count = 0; 132627e1f827SSong Liu int i; 132727e1f827SSong Liu 132827e1f827SSong Liu if (!vma || !vma->vm_file || 1329fef792a4SMiaohe Lin !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) 133027e1f827SSong Liu return; 133127e1f827SSong Liu 133227e1f827SSong Liu /* 133327e1f827SSong Liu * This vm_flags may not have VM_HUGEPAGE if the page was not 133427e1f827SSong Liu * collapsed by this mm. But we can still collapse if the page is 133527e1f827SSong Liu * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check() 133627e1f827SSong Liu * will not fail the vma for missing VM_HUGEPAGE 133727e1f827SSong Liu */ 13387da4e2cbSYang Shi if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE, false, false)) 133927e1f827SSong Liu return; 134027e1f827SSong Liu 1341deb4c93aSPeter Xu /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ 1342deb4c93aSPeter Xu if (userfaultfd_wp(vma)) 1343deb4c93aSPeter Xu return; 1344deb4c93aSPeter Xu 1345119a5fc1SHugh Dickins hpage = find_lock_page(vma->vm_file->f_mapping, 1346119a5fc1SHugh Dickins linear_page_index(vma, haddr)); 1347119a5fc1SHugh Dickins if (!hpage) 1348119a5fc1SHugh Dickins return; 1349119a5fc1SHugh Dickins 1350119a5fc1SHugh Dickins if (!PageHead(hpage)) 1351119a5fc1SHugh Dickins goto drop_hpage; 1352119a5fc1SHugh Dickins 135327e1f827SSong Liu pmd = mm_find_pmd(mm, haddr); 135427e1f827SSong Liu if (!pmd) 1355119a5fc1SHugh Dickins goto drop_hpage; 135627e1f827SSong Liu 135727e1f827SSong Liu start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 135827e1f827SSong Liu 135927e1f827SSong Liu /* step 1: check all mapped PTEs are to the right huge page */ 136027e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 136127e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 136227e1f827SSong Liu struct page *page; 136327e1f827SSong Liu 136427e1f827SSong Liu /* empty pte, skip */ 136527e1f827SSong Liu if (pte_none(*pte)) 136627e1f827SSong Liu continue; 136727e1f827SSong Liu 136827e1f827SSong Liu /* page swapped out, abort */ 136927e1f827SSong Liu if (!pte_present(*pte)) 137027e1f827SSong Liu goto abort; 137127e1f827SSong Liu 137227e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 13733218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 13743218f871SAlex Sierra page = NULL; 137527e1f827SSong Liu /* 1376119a5fc1SHugh Dickins * Note that uprobe, debugger, or MAP_PRIVATE may change the 1377119a5fc1SHugh Dickins * page table, but the new page will not be a subpage of hpage. 137827e1f827SSong Liu */ 1379119a5fc1SHugh Dickins if (hpage + i != page) 138027e1f827SSong Liu goto abort; 138127e1f827SSong Liu count++; 138227e1f827SSong Liu } 138327e1f827SSong Liu 138427e1f827SSong Liu /* step 2: adjust rmap */ 138527e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 138627e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 138727e1f827SSong Liu struct page *page; 138827e1f827SSong Liu 138927e1f827SSong Liu if (pte_none(*pte)) 139027e1f827SSong Liu continue; 139127e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 13923218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 13933218f871SAlex Sierra goto abort; 1394cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 139527e1f827SSong Liu } 139627e1f827SSong Liu 139727e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 139827e1f827SSong Liu 139927e1f827SSong Liu /* step 3: set proper refcount and mm_counters. */ 1400119a5fc1SHugh Dickins if (count) { 140127e1f827SSong Liu page_ref_sub(hpage, count); 140227e1f827SSong Liu add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); 140327e1f827SSong Liu } 140427e1f827SSong Liu 140527e1f827SSong Liu /* step 4: collapse pmd */ 1406e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, haddr, pmd); 1407119a5fc1SHugh Dickins drop_hpage: 1408119a5fc1SHugh Dickins unlock_page(hpage); 1409119a5fc1SHugh Dickins put_page(hpage); 141027e1f827SSong Liu return; 141127e1f827SSong Liu 141227e1f827SSong Liu abort: 141327e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 1414119a5fc1SHugh Dickins goto drop_hpage; 141527e1f827SSong Liu } 141627e1f827SSong Liu 14170edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 141827e1f827SSong Liu { 141927e1f827SSong Liu struct mm_struct *mm = mm_slot->mm; 142027e1f827SSong Liu int i; 142127e1f827SSong Liu 142227e1f827SSong Liu if (likely(mm_slot->nr_pte_mapped_thp == 0)) 14230edf61e5SMiaohe Lin return; 142427e1f827SSong Liu 1425d8ed45c5SMichel Lespinasse if (!mmap_write_trylock(mm)) 14260edf61e5SMiaohe Lin return; 142727e1f827SSong Liu 142827e1f827SSong Liu if (unlikely(khugepaged_test_exit(mm))) 142927e1f827SSong Liu goto out; 143027e1f827SSong Liu 143127e1f827SSong Liu for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) 143227e1f827SSong Liu collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]); 143327e1f827SSong Liu 143427e1f827SSong Liu out: 143527e1f827SSong Liu mm_slot->nr_pte_mapped_thp = 0; 1436d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 143727e1f827SSong Liu } 143827e1f827SSong Liu 1439f3f0e1d2SKirill A. Shutemov static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1440f3f0e1d2SKirill A. Shutemov { 1441f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 144218e77600SHugh Dickins struct mm_struct *mm; 1443f3f0e1d2SKirill A. Shutemov unsigned long addr; 1444e59a47b8SPasha Tatashin pmd_t *pmd; 1445f3f0e1d2SKirill A. Shutemov 1446f3f0e1d2SKirill A. Shutemov i_mmap_lock_write(mapping); 1447f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 144827e1f827SSong Liu /* 144927e1f827SSong Liu * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 145027e1f827SSong Liu * got written to. These VMAs are likely not worth investing 14513e4e28c5SMichel Lespinasse * mmap_write_lock(mm) as PMD-mapping is likely to be split 145227e1f827SSong Liu * later. 145327e1f827SSong Liu * 145436ee2c78SMiaohe Lin * Note that vma->anon_vma check is racy: it can be set up after 1455c1e8d7c6SMichel Lespinasse * the check but before we took mmap_lock by the fault path. 145627e1f827SSong Liu * But page lock would prevent establishing any new ptes of the 145727e1f827SSong Liu * page, so we are safe. 145827e1f827SSong Liu * 145927e1f827SSong Liu * An alternative would be drop the check, but check that page 146027e1f827SSong Liu * table is clear before calling pmdp_collapse_flush() under 146127e1f827SSong Liu * ptl. It has higher chance to recover THP for the VMA, but 146227e1f827SSong Liu * has higher cost too. 146327e1f827SSong Liu */ 1464f3f0e1d2SKirill A. Shutemov if (vma->anon_vma) 1465f3f0e1d2SKirill A. Shutemov continue; 1466f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 1467f3f0e1d2SKirill A. Shutemov if (addr & ~HPAGE_PMD_MASK) 1468f3f0e1d2SKirill A. Shutemov continue; 1469f3f0e1d2SKirill A. Shutemov if (vma->vm_end < addr + HPAGE_PMD_SIZE) 1470f3f0e1d2SKirill A. Shutemov continue; 147118e77600SHugh Dickins mm = vma->vm_mm; 147218e77600SHugh Dickins pmd = mm_find_pmd(mm, addr); 1473f3f0e1d2SKirill A. Shutemov if (!pmd) 1474f3f0e1d2SKirill A. Shutemov continue; 1475f3f0e1d2SKirill A. Shutemov /* 1476c1e8d7c6SMichel Lespinasse * We need exclusive mmap_lock to retract page table. 147727e1f827SSong Liu * 147827e1f827SSong Liu * We use trylock due to lock inversion: we need to acquire 1479c1e8d7c6SMichel Lespinasse * mmap_lock while holding page lock. Fault path does it in 148027e1f827SSong Liu * reverse order. Trylock is a way to avoid deadlock. 1481f3f0e1d2SKirill A. Shutemov */ 148218e77600SHugh Dickins if (mmap_write_trylock(mm)) { 1483deb4c93aSPeter Xu /* 1484deb4c93aSPeter Xu * When a vma is registered with uffd-wp, we can't 1485deb4c93aSPeter Xu * recycle the pmd pgtable because there can be pte 1486deb4c93aSPeter Xu * markers installed. Skip it only, so the rest mm/vma 1487deb4c93aSPeter Xu * can still have the same file mapped hugely, however 1488deb4c93aSPeter Xu * it'll always mapped in small page size for uffd-wp 1489deb4c93aSPeter Xu * registered ranges. 1490deb4c93aSPeter Xu */ 1491deb4c93aSPeter Xu if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma)) 1492e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, addr, pmd); 149318e77600SHugh Dickins mmap_write_unlock(mm); 149427e1f827SSong Liu } else { 149527e1f827SSong Liu /* Try again later */ 149618e77600SHugh Dickins khugepaged_add_pte_mapped_thp(mm, addr); 1497f3f0e1d2SKirill A. Shutemov } 1498f3f0e1d2SKirill A. Shutemov } 1499f3f0e1d2SKirill A. Shutemov i_mmap_unlock_write(mapping); 1500f3f0e1d2SKirill A. Shutemov } 1501f3f0e1d2SKirill A. Shutemov 1502f3f0e1d2SKirill A. Shutemov /** 150399cb0dbdSSong Liu * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1504f3f0e1d2SKirill A. Shutemov * 1505336e6b53SAlex Shi * @mm: process address space where collapse happens 1506336e6b53SAlex Shi * @file: file that collapse on 1507336e6b53SAlex Shi * @start: collapse start address 15089710a78aSZach O'Keefe * @cc: collapse context and scratchpad 1509336e6b53SAlex Shi * 1510f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 151187c460a0SHugh Dickins * - allocate and lock a new huge page; 151277da9389SMatthew Wilcox * - scan page cache replacing old pages with the new one 151399cb0dbdSSong Liu * + swap/gup in pages if necessary; 1514f3f0e1d2SKirill A. Shutemov * + fill in gaps; 151577da9389SMatthew Wilcox * + keep old pages around in case rollback is required; 151677da9389SMatthew Wilcox * - if replacing succeeds: 1517f3f0e1d2SKirill A. Shutemov * + copy data over; 1518f3f0e1d2SKirill A. Shutemov * + free old pages; 151987c460a0SHugh Dickins * + unlock huge page; 1520f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1521f3f0e1d2SKirill A. Shutemov * + put all pages back and unfreeze them; 152277da9389SMatthew Wilcox * + restore gaps in the page cache; 152387c460a0SHugh Dickins * + unlock and free huge page; 1524f3f0e1d2SKirill A. Shutemov */ 1525*50ad2f24SZach O'Keefe static int collapse_file(struct mm_struct *mm, struct file *file, 1526*50ad2f24SZach O'Keefe pgoff_t start, struct collapse_control *cc) 1527f3f0e1d2SKirill A. Shutemov { 1528579c571eSSong Liu struct address_space *mapping = file->f_mapping; 1529*50ad2f24SZach O'Keefe struct page *hpage; 1530f3f0e1d2SKirill A. Shutemov pgoff_t index, end = start + HPAGE_PMD_NR; 1531f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 153277da9389SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1533f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 153499cb0dbdSSong Liu bool is_shmem = shmem_file(file); 1535bf9eceadSMuchun Song int nr; 1536f3f0e1d2SKirill A. Shutemov 153799cb0dbdSSong Liu VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1538f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1539f3f0e1d2SKirill A. Shutemov 1540*50ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 15419710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1542f3f0e1d2SKirill A. Shutemov goto out; 1543f3f0e1d2SKirill A. Shutemov 15446b24ca4aSMatthew Wilcox (Oracle) /* 15456b24ca4aSMatthew Wilcox (Oracle) * Ensure we have slots for all the pages in the range. This is 15466b24ca4aSMatthew Wilcox (Oracle) * almost certainly a no-op because most of the pages must be present 15476b24ca4aSMatthew Wilcox (Oracle) */ 154895feeabbSHugh Dickins do { 154995feeabbSHugh Dickins xas_lock_irq(&xas); 155095feeabbSHugh Dickins xas_create_range(&xas); 155195feeabbSHugh Dickins if (!xas_error(&xas)) 155295feeabbSHugh Dickins break; 155395feeabbSHugh Dickins xas_unlock_irq(&xas); 155495feeabbSHugh Dickins if (!xas_nomem(&xas, GFP_KERNEL)) { 155595feeabbSHugh Dickins result = SCAN_FAIL; 155695feeabbSHugh Dickins goto out; 155795feeabbSHugh Dickins } 155895feeabbSHugh Dickins } while (1); 155995feeabbSHugh Dickins 1560*50ad2f24SZach O'Keefe __SetPageLocked(hpage); 156199cb0dbdSSong Liu if (is_shmem) 1562*50ad2f24SZach O'Keefe __SetPageSwapBacked(hpage); 1563*50ad2f24SZach O'Keefe hpage->index = start; 1564*50ad2f24SZach O'Keefe hpage->mapping = mapping; 1565f3f0e1d2SKirill A. Shutemov 1566f3f0e1d2SKirill A. Shutemov /* 1567*50ad2f24SZach O'Keefe * At this point the hpage is locked and not up-to-date. 156887c460a0SHugh Dickins * It's safe to insert it into the page cache, because nobody would 156987c460a0SHugh Dickins * be able to map it or use it in another way until we unlock it. 1570f3f0e1d2SKirill A. Shutemov */ 1571f3f0e1d2SKirill A. Shutemov 157277da9389SMatthew Wilcox xas_set(&xas, start); 157377da9389SMatthew Wilcox for (index = start; index < end; index++) { 157477da9389SMatthew Wilcox struct page *page = xas_next(&xas); 157577da9389SMatthew Wilcox 157677da9389SMatthew Wilcox VM_BUG_ON(index != xas.xa_index); 157799cb0dbdSSong Liu if (is_shmem) { 157877da9389SMatthew Wilcox if (!page) { 1579701270faSHugh Dickins /* 158099cb0dbdSSong Liu * Stop if extent has been truncated or 158199cb0dbdSSong Liu * hole-punched, and is now completely 158299cb0dbdSSong Liu * empty. 1583701270faSHugh Dickins */ 1584701270faSHugh Dickins if (index == start) { 1585701270faSHugh Dickins if (!xas_next_entry(&xas, end - 1)) { 1586701270faSHugh Dickins result = SCAN_TRUNCATED; 1587042a3082SHugh Dickins goto xa_locked; 1588701270faSHugh Dickins } 1589701270faSHugh Dickins xas_set(&xas, index); 1590701270faSHugh Dickins } 159177da9389SMatthew Wilcox if (!shmem_charge(mapping->host, 1)) { 1592f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1593042a3082SHugh Dickins goto xa_locked; 1594f3f0e1d2SKirill A. Shutemov } 1595*50ad2f24SZach O'Keefe xas_store(&xas, hpage); 159677da9389SMatthew Wilcox nr_none++; 159777da9389SMatthew Wilcox continue; 1598f3f0e1d2SKirill A. Shutemov } 1599f3f0e1d2SKirill A. Shutemov 16003159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 160177da9389SMatthew Wilcox xas_unlock_irq(&xas); 1602f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 1603f3f0e1d2SKirill A. Shutemov if (shmem_getpage(mapping->host, index, &page, 1604acdd9f8eSHugh Dickins SGP_NOALLOC)) { 1605f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 160677da9389SMatthew Wilcox goto xa_unlocked; 1607f3f0e1d2SKirill A. Shutemov } 1608f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1609f3f0e1d2SKirill A. Shutemov get_page(page); 1610042a3082SHugh Dickins xas_unlock_irq(&xas); 1611f3f0e1d2SKirill A. Shutemov } else { 1612f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1613042a3082SHugh Dickins goto xa_locked; 1614f3f0e1d2SKirill A. Shutemov } 161599cb0dbdSSong Liu } else { /* !is_shmem */ 161699cb0dbdSSong Liu if (!page || xa_is_value(page)) { 161799cb0dbdSSong Liu xas_unlock_irq(&xas); 161899cb0dbdSSong Liu page_cache_sync_readahead(mapping, &file->f_ra, 161999cb0dbdSSong Liu file, index, 1620e5a59d30SDavid Howells end - index); 162199cb0dbdSSong Liu /* drain pagevecs to help isolate_lru_page() */ 162299cb0dbdSSong Liu lru_add_drain(); 162399cb0dbdSSong Liu page = find_lock_page(mapping, index); 162499cb0dbdSSong Liu if (unlikely(page == NULL)) { 162599cb0dbdSSong Liu result = SCAN_FAIL; 162699cb0dbdSSong Liu goto xa_unlocked; 162799cb0dbdSSong Liu } 162875f36069SSong Liu } else if (PageDirty(page)) { 162975f36069SSong Liu /* 163075f36069SSong Liu * khugepaged only works on read-only fd, 163175f36069SSong Liu * so this page is dirty because it hasn't 163275f36069SSong Liu * been flushed since first write. There 163375f36069SSong Liu * won't be new dirty pages. 163475f36069SSong Liu * 163575f36069SSong Liu * Trigger async flush here and hope the 163675f36069SSong Liu * writeback is done when khugepaged 163775f36069SSong Liu * revisits this page. 163875f36069SSong Liu * 163975f36069SSong Liu * This is a one-off situation. We are not 164075f36069SSong Liu * forcing writeback in loop. 164175f36069SSong Liu */ 164275f36069SSong Liu xas_unlock_irq(&xas); 164375f36069SSong Liu filemap_flush(mapping); 164475f36069SSong Liu result = SCAN_FAIL; 164575f36069SSong Liu goto xa_unlocked; 164674c42e1bSRongwei Wang } else if (PageWriteback(page)) { 164774c42e1bSRongwei Wang xas_unlock_irq(&xas); 164874c42e1bSRongwei Wang result = SCAN_FAIL; 164974c42e1bSRongwei Wang goto xa_unlocked; 165099cb0dbdSSong Liu } else if (trylock_page(page)) { 165199cb0dbdSSong Liu get_page(page); 165299cb0dbdSSong Liu xas_unlock_irq(&xas); 165399cb0dbdSSong Liu } else { 165499cb0dbdSSong Liu result = SCAN_PAGE_LOCK; 165599cb0dbdSSong Liu goto xa_locked; 165699cb0dbdSSong Liu } 165799cb0dbdSSong Liu } 1658f3f0e1d2SKirill A. Shutemov 1659f3f0e1d2SKirill A. Shutemov /* 1660b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1661f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1662f3f0e1d2SKirill A. Shutemov */ 1663f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 16644655e5e5SSong Liu 16654655e5e5SSong Liu /* make sure the page is up to date */ 16664655e5e5SSong Liu if (unlikely(!PageUptodate(page))) { 16674655e5e5SSong Liu result = SCAN_FAIL; 16684655e5e5SSong Liu goto out_unlock; 16694655e5e5SSong Liu } 167006a5e126SHugh Dickins 167106a5e126SHugh Dickins /* 167206a5e126SHugh Dickins * If file was truncated then extended, or hole-punched, before 167306a5e126SHugh Dickins * we locked the first page, then a THP might be there already. 167406a5e126SHugh Dickins */ 167506a5e126SHugh Dickins if (PageTransCompound(page)) { 167606a5e126SHugh Dickins result = SCAN_PAGE_COMPOUND; 167706a5e126SHugh Dickins goto out_unlock; 167806a5e126SHugh Dickins } 1679f3f0e1d2SKirill A. Shutemov 1680f3f0e1d2SKirill A. Shutemov if (page_mapping(page) != mapping) { 1681f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1682f3f0e1d2SKirill A. Shutemov goto out_unlock; 1683f3f0e1d2SKirill A. Shutemov } 1684f3f0e1d2SKirill A. Shutemov 168574c42e1bSRongwei Wang if (!is_shmem && (PageDirty(page) || 168674c42e1bSRongwei Wang PageWriteback(page))) { 16874655e5e5SSong Liu /* 16884655e5e5SSong Liu * khugepaged only works on read-only fd, so this 16894655e5e5SSong Liu * page is dirty because it hasn't been flushed 16904655e5e5SSong Liu * since first write. 16914655e5e5SSong Liu */ 16924655e5e5SSong Liu result = SCAN_FAIL; 16934655e5e5SSong Liu goto out_unlock; 16944655e5e5SSong Liu } 16954655e5e5SSong Liu 1696f3f0e1d2SKirill A. Shutemov if (isolate_lru_page(page)) { 1697f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1698042a3082SHugh Dickins goto out_unlock; 1699f3f0e1d2SKirill A. Shutemov } 1700f3f0e1d2SKirill A. Shutemov 170199cb0dbdSSong Liu if (page_has_private(page) && 170299cb0dbdSSong Liu !try_to_release_page(page, GFP_KERNEL)) { 170399cb0dbdSSong Liu result = SCAN_PAGE_HAS_PRIVATE; 17042f33a706SHugh Dickins putback_lru_page(page); 170599cb0dbdSSong Liu goto out_unlock; 170699cb0dbdSSong Liu } 170799cb0dbdSSong Liu 1708f3f0e1d2SKirill A. Shutemov if (page_mapped(page)) 1709869f7ee6SMatthew Wilcox (Oracle) try_to_unmap(page_folio(page), 1710869f7ee6SMatthew Wilcox (Oracle) TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); 1711f3f0e1d2SKirill A. Shutemov 171277da9389SMatthew Wilcox xas_lock_irq(&xas); 171377da9389SMatthew Wilcox xas_set(&xas, index); 1714f3f0e1d2SKirill A. Shutemov 171577da9389SMatthew Wilcox VM_BUG_ON_PAGE(page != xas_load(&xas), page); 1716f3f0e1d2SKirill A. Shutemov 1717f3f0e1d2SKirill A. Shutemov /* 1718f3f0e1d2SKirill A. Shutemov * The page is expected to have page_count() == 3: 1719f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 172077da9389SMatthew Wilcox * - one reference from page cache; 1721f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1722f3f0e1d2SKirill A. Shutemov */ 1723f3f0e1d2SKirill A. Shutemov if (!page_ref_freeze(page, 3)) { 1724f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1725042a3082SHugh Dickins xas_unlock_irq(&xas); 1726042a3082SHugh Dickins putback_lru_page(page); 1727042a3082SHugh Dickins goto out_unlock; 1728f3f0e1d2SKirill A. Shutemov } 1729f3f0e1d2SKirill A. Shutemov 1730f3f0e1d2SKirill A. Shutemov /* 1731f3f0e1d2SKirill A. Shutemov * Add the page to the list to be able to undo the collapse if 1732f3f0e1d2SKirill A. Shutemov * something go wrong. 1733f3f0e1d2SKirill A. Shutemov */ 1734f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 1735f3f0e1d2SKirill A. Shutemov 1736f3f0e1d2SKirill A. Shutemov /* Finally, replace with the new page. */ 1737*50ad2f24SZach O'Keefe xas_store(&xas, hpage); 1738f3f0e1d2SKirill A. Shutemov continue; 1739f3f0e1d2SKirill A. Shutemov out_unlock: 1740f3f0e1d2SKirill A. Shutemov unlock_page(page); 1741f3f0e1d2SKirill A. Shutemov put_page(page); 1742042a3082SHugh Dickins goto xa_unlocked; 1743f3f0e1d2SKirill A. Shutemov } 1744*50ad2f24SZach O'Keefe nr = thp_nr_pages(hpage); 1745f3f0e1d2SKirill A. Shutemov 174699cb0dbdSSong Liu if (is_shmem) 1747*50ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr); 174809d91cdaSSong Liu else { 1749*50ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr); 175009d91cdaSSong Liu filemap_nr_thps_inc(mapping); 1751eb6ecbedSCollin Fijalkovich /* 1752eb6ecbedSCollin Fijalkovich * Paired with smp_mb() in do_dentry_open() to ensure 1753eb6ecbedSCollin Fijalkovich * i_writecount is up to date and the update to nr_thps is 1754eb6ecbedSCollin Fijalkovich * visible. Ensures the page cache will be truncated if the 1755eb6ecbedSCollin Fijalkovich * file is opened writable. 1756eb6ecbedSCollin Fijalkovich */ 1757eb6ecbedSCollin Fijalkovich smp_mb(); 1758eb6ecbedSCollin Fijalkovich if (inode_is_open_for_write(mapping->host)) { 1759eb6ecbedSCollin Fijalkovich result = SCAN_FAIL; 1760*50ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr); 1761eb6ecbedSCollin Fijalkovich filemap_nr_thps_dec(mapping); 1762eb6ecbedSCollin Fijalkovich goto xa_locked; 1763eb6ecbedSCollin Fijalkovich } 176409d91cdaSSong Liu } 176599cb0dbdSSong Liu 1766042a3082SHugh Dickins if (nr_none) { 1767*50ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none); 17682f55f070SMiaohe Lin /* nr_none is always 0 for non-shmem. */ 1769*50ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none); 1770042a3082SHugh Dickins } 1771042a3082SHugh Dickins 17726b24ca4aSMatthew Wilcox (Oracle) /* Join all the small entries into a single multi-index entry */ 17736b24ca4aSMatthew Wilcox (Oracle) xas_set_order(&xas, start, HPAGE_PMD_ORDER); 1774*50ad2f24SZach O'Keefe xas_store(&xas, hpage); 1775042a3082SHugh Dickins xa_locked: 1776042a3082SHugh Dickins xas_unlock_irq(&xas); 177777da9389SMatthew Wilcox xa_unlocked: 1778042a3082SHugh Dickins 17796d9df8a5SHugh Dickins /* 17806d9df8a5SHugh Dickins * If collapse is successful, flush must be done now before copying. 17816d9df8a5SHugh Dickins * If collapse is unsuccessful, does flush actually need to be done? 17826d9df8a5SHugh Dickins * Do it anyway, to clear the state. 17836d9df8a5SHugh Dickins */ 17846d9df8a5SHugh Dickins try_to_unmap_flush(); 17856d9df8a5SHugh Dickins 1786f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 178777da9389SMatthew Wilcox struct page *page, *tmp; 1788f3f0e1d2SKirill A. Shutemov 1789f3f0e1d2SKirill A. Shutemov /* 179077da9389SMatthew Wilcox * Replacing old pages with new one has succeeded, now we 179177da9389SMatthew Wilcox * need to copy the content and free the old pages. 1792f3f0e1d2SKirill A. Shutemov */ 17932af8ff29SHugh Dickins index = start; 1794f3f0e1d2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, &pagelist, lru) { 17952af8ff29SHugh Dickins while (index < page->index) { 1796*50ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 17972af8ff29SHugh Dickins index++; 17982af8ff29SHugh Dickins } 1799*50ad2f24SZach O'Keefe copy_highpage(hpage + (page->index % HPAGE_PMD_NR), 1800f3f0e1d2SKirill A. Shutemov page); 1801f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1802f3f0e1d2SKirill A. Shutemov page->mapping = NULL; 1803042a3082SHugh Dickins page_ref_unfreeze(page, 1); 1804f3f0e1d2SKirill A. Shutemov ClearPageActive(page); 1805f3f0e1d2SKirill A. Shutemov ClearPageUnevictable(page); 1806042a3082SHugh Dickins unlock_page(page); 1807f3f0e1d2SKirill A. Shutemov put_page(page); 18082af8ff29SHugh Dickins index++; 18092af8ff29SHugh Dickins } 18102af8ff29SHugh Dickins while (index < end) { 1811*50ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 18122af8ff29SHugh Dickins index++; 1813f3f0e1d2SKirill A. Shutemov } 1814f3f0e1d2SKirill A. Shutemov 1815*50ad2f24SZach O'Keefe SetPageUptodate(hpage); 1816*50ad2f24SZach O'Keefe page_ref_add(hpage, HPAGE_PMD_NR - 1); 18176058eaecSJohannes Weiner if (is_shmem) 1818*50ad2f24SZach O'Keefe set_page_dirty(hpage); 1819*50ad2f24SZach O'Keefe lru_cache_add(hpage); 1820f3f0e1d2SKirill A. Shutemov 1821042a3082SHugh Dickins /* 1822042a3082SHugh Dickins * Remove pte page tables, so we can re-fault the page as huge. 1823042a3082SHugh Dickins */ 1824042a3082SHugh Dickins retract_page_tables(mapping, start); 1825*50ad2f24SZach O'Keefe unlock_page(hpage); 1826*50ad2f24SZach O'Keefe hpage = NULL; 1827f3f0e1d2SKirill A. Shutemov } else { 182877da9389SMatthew Wilcox struct page *page; 1829aaa52e34SHugh Dickins 183077da9389SMatthew Wilcox /* Something went wrong: roll back page cache changes */ 183177da9389SMatthew Wilcox xas_lock_irq(&xas); 18322f55f070SMiaohe Lin if (nr_none) { 1833aaa52e34SHugh Dickins mapping->nrpages -= nr_none; 1834aaa52e34SHugh Dickins shmem_uncharge(mapping->host, nr_none); 18352f55f070SMiaohe Lin } 1836aaa52e34SHugh Dickins 183777da9389SMatthew Wilcox xas_set(&xas, start); 183877da9389SMatthew Wilcox xas_for_each(&xas, page, end - 1) { 1839f3f0e1d2SKirill A. Shutemov page = list_first_entry_or_null(&pagelist, 1840f3f0e1d2SKirill A. Shutemov struct page, lru); 184177da9389SMatthew Wilcox if (!page || xas.xa_index < page->index) { 1842f3f0e1d2SKirill A. Shutemov if (!nr_none) 1843f3f0e1d2SKirill A. Shutemov break; 1844f3f0e1d2SKirill A. Shutemov nr_none--; 184559749e6cSJohannes Weiner /* Put holes back where they were */ 184677da9389SMatthew Wilcox xas_store(&xas, NULL); 1847f3f0e1d2SKirill A. Shutemov continue; 1848f3f0e1d2SKirill A. Shutemov } 1849f3f0e1d2SKirill A. Shutemov 185077da9389SMatthew Wilcox VM_BUG_ON_PAGE(page->index != xas.xa_index, page); 1851f3f0e1d2SKirill A. Shutemov 1852f3f0e1d2SKirill A. Shutemov /* Unfreeze the page. */ 1853f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1854f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(page, 2); 185577da9389SMatthew Wilcox xas_store(&xas, page); 185677da9389SMatthew Wilcox xas_pause(&xas); 185777da9389SMatthew Wilcox xas_unlock_irq(&xas); 1858f3f0e1d2SKirill A. Shutemov unlock_page(page); 1859042a3082SHugh Dickins putback_lru_page(page); 186077da9389SMatthew Wilcox xas_lock_irq(&xas); 1861f3f0e1d2SKirill A. Shutemov } 1862f3f0e1d2SKirill A. Shutemov VM_BUG_ON(nr_none); 186377da9389SMatthew Wilcox xas_unlock_irq(&xas); 1864f3f0e1d2SKirill A. Shutemov 1865*50ad2f24SZach O'Keefe hpage->mapping = NULL; 1866f3f0e1d2SKirill A. Shutemov } 1867042a3082SHugh Dickins 1868*50ad2f24SZach O'Keefe if (hpage) 1869*50ad2f24SZach O'Keefe unlock_page(hpage); 1870f3f0e1d2SKirill A. Shutemov out: 1871f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 1872*50ad2f24SZach O'Keefe if (hpage) { 1873*50ad2f24SZach O'Keefe mem_cgroup_uncharge(page_folio(hpage)); 1874*50ad2f24SZach O'Keefe put_page(hpage); 1875c6a7f445SYang Shi } 1876f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 1877*50ad2f24SZach O'Keefe return result; 1878f3f0e1d2SKirill A. Shutemov } 1879f3f0e1d2SKirill A. Shutemov 1880*50ad2f24SZach O'Keefe static int khugepaged_scan_file(struct mm_struct *mm, struct file *file, 1881*50ad2f24SZach O'Keefe pgoff_t start, struct collapse_control *cc) 1882f3f0e1d2SKirill A. Shutemov { 1883f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 1884579c571eSSong Liu struct address_space *mapping = file->f_mapping; 188585b392dbSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, start); 1886f3f0e1d2SKirill A. Shutemov int present, swap; 1887f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 1888f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 1889f3f0e1d2SKirill A. Shutemov 1890f3f0e1d2SKirill A. Shutemov present = 0; 1891f3f0e1d2SKirill A. Shutemov swap = 0; 189234d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1893f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 189485b392dbSMatthew Wilcox xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 189585b392dbSMatthew Wilcox if (xas_retry(&xas, page)) 1896f3f0e1d2SKirill A. Shutemov continue; 1897f3f0e1d2SKirill A. Shutemov 189885b392dbSMatthew Wilcox if (xa_is_value(page)) { 1899f3f0e1d2SKirill A. Shutemov if (++swap > khugepaged_max_ptes_swap) { 1900f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1901e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1902f3f0e1d2SKirill A. Shutemov break; 1903f3f0e1d2SKirill A. Shutemov } 1904f3f0e1d2SKirill A. Shutemov continue; 1905f3f0e1d2SKirill A. Shutemov } 1906f3f0e1d2SKirill A. Shutemov 19076b24ca4aSMatthew Wilcox (Oracle) /* 19086b24ca4aSMatthew Wilcox (Oracle) * XXX: khugepaged should compact smaller compound pages 19096b24ca4aSMatthew Wilcox (Oracle) * into a PMD sized page 19106b24ca4aSMatthew Wilcox (Oracle) */ 1911f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 1912f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COMPOUND; 1913f3f0e1d2SKirill A. Shutemov break; 1914f3f0e1d2SKirill A. Shutemov } 1915f3f0e1d2SKirill A. Shutemov 1916f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 191734d6b470SZach O'Keefe if (khugepaged_scan_abort(node, cc)) { 1918f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 1919f3f0e1d2SKirill A. Shutemov break; 1920f3f0e1d2SKirill A. Shutemov } 192134d6b470SZach O'Keefe cc->node_load[node]++; 1922f3f0e1d2SKirill A. Shutemov 1923f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 1924f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 1925f3f0e1d2SKirill A. Shutemov break; 1926f3f0e1d2SKirill A. Shutemov } 1927f3f0e1d2SKirill A. Shutemov 192899cb0dbdSSong Liu if (page_count(page) != 192999cb0dbdSSong Liu 1 + page_mapcount(page) + page_has_private(page)) { 1930f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1931f3f0e1d2SKirill A. Shutemov break; 1932f3f0e1d2SKirill A. Shutemov } 1933f3f0e1d2SKirill A. Shutemov 1934f3f0e1d2SKirill A. Shutemov /* 1935f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 1936f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 1937f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 1938f3f0e1d2SKirill A. Shutemov */ 1939f3f0e1d2SKirill A. Shutemov 1940f3f0e1d2SKirill A. Shutemov present++; 1941f3f0e1d2SKirill A. Shutemov 1942f3f0e1d2SKirill A. Shutemov if (need_resched()) { 194385b392dbSMatthew Wilcox xas_pause(&xas); 1944f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 1945f3f0e1d2SKirill A. Shutemov } 1946f3f0e1d2SKirill A. Shutemov } 1947f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 1948f3f0e1d2SKirill A. Shutemov 1949f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 1950f3f0e1d2SKirill A. Shutemov if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 1951f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1952e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1953f3f0e1d2SKirill A. Shutemov } else { 1954*50ad2f24SZach O'Keefe result = collapse_file(mm, file, start, cc); 1955f3f0e1d2SKirill A. Shutemov } 1956f3f0e1d2SKirill A. Shutemov } 1957f3f0e1d2SKirill A. Shutemov 1958f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 1959*50ad2f24SZach O'Keefe return result; 1960f3f0e1d2SKirill A. Shutemov } 1961f3f0e1d2SKirill A. Shutemov #else 1962*50ad2f24SZach O'Keefe static int khugepaged_scan_file(struct mm_struct *mm, struct file *file, 1963*50ad2f24SZach O'Keefe pgoff_t start, struct collapse_control *cc) 1964f3f0e1d2SKirill A. Shutemov { 1965f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 1966f3f0e1d2SKirill A. Shutemov } 196727e1f827SSong Liu 19680edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 196927e1f827SSong Liu { 197027e1f827SSong Liu } 1971f3f0e1d2SKirill A. Shutemov #endif 1972f3f0e1d2SKirill A. Shutemov 1973*50ad2f24SZach O'Keefe static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, 197434d6b470SZach O'Keefe struct collapse_control *cc) 1975b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 1976b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 1977b46e756fSKirill A. Shutemov { 1978b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 1979b46e756fSKirill A. Shutemov struct mm_struct *mm; 1980b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 1981b46e756fSKirill A. Shutemov int progress = 0; 1982b46e756fSKirill A. Shutemov 1983b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 198435f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1985*50ad2f24SZach O'Keefe *result = SCAN_FAIL; 1986b46e756fSKirill A. Shutemov 1987b46e756fSKirill A. Shutemov if (khugepaged_scan.mm_slot) 1988b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 1989b46e756fSKirill A. Shutemov else { 1990b46e756fSKirill A. Shutemov mm_slot = list_entry(khugepaged_scan.mm_head.next, 1991b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 1992b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 1993b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 1994b46e756fSKirill A. Shutemov } 1995b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 199627e1f827SSong Liu khugepaged_collapse_pte_mapped_thps(mm_slot); 1997b46e756fSKirill A. Shutemov 1998b46e756fSKirill A. Shutemov mm = mm_slot->mm; 19993b454ad3SYang Shi /* 20003b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 20013b454ad3SYang Shi * the next mm on the list. 20023b454ad3SYang Shi */ 2003b46e756fSKirill A. Shutemov vma = NULL; 2004d8ed45c5SMichel Lespinasse if (unlikely(!mmap_read_trylock(mm))) 2005c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 20063b454ad3SYang Shi if (likely(!khugepaged_test_exit(mm))) 2007b46e756fSKirill A. Shutemov vma = find_vma(mm, khugepaged_scan.address); 2008b46e756fSKirill A. Shutemov 2009b46e756fSKirill A. Shutemov progress++; 2010b46e756fSKirill A. Shutemov for (; vma; vma = vma->vm_next) { 2011b46e756fSKirill A. Shutemov unsigned long hstart, hend; 2012b46e756fSKirill A. Shutemov 2013b46e756fSKirill A. Shutemov cond_resched(); 2014b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) { 2015b46e756fSKirill A. Shutemov progress++; 2016b46e756fSKirill A. Shutemov break; 2017b46e756fSKirill A. Shutemov } 20187da4e2cbSYang Shi if (!hugepage_vma_check(vma, vma->vm_flags, false, false)) { 2019b46e756fSKirill A. Shutemov skip: 2020b46e756fSKirill A. Shutemov progress++; 2021b46e756fSKirill A. Shutemov continue; 2022b46e756fSKirill A. Shutemov } 20234fa6893fSYang Shi hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); 20244fa6893fSYang Shi hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); 2025b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 2026b46e756fSKirill A. Shutemov goto skip; 2027b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 2028b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 2029b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2030b46e756fSKirill A. Shutemov 2031b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 2032*50ad2f24SZach O'Keefe bool mmap_locked = true; 2033*50ad2f24SZach O'Keefe 2034b46e756fSKirill A. Shutemov cond_resched(); 2035b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 2036b46e756fSKirill A. Shutemov goto breakouterloop; 2037b46e756fSKirill A. Shutemov 2038b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 2039b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 2040b46e756fSKirill A. Shutemov hend); 204199cb0dbdSSong Liu if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 2042396bcc52SMatthew Wilcox (Oracle) struct file *file = get_file(vma->vm_file); 2043f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 2044f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 204599cb0dbdSSong Liu 2046d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 2047*50ad2f24SZach O'Keefe *result = khugepaged_scan_file(mm, file, pgoff, 204834d6b470SZach O'Keefe cc); 2049*50ad2f24SZach O'Keefe mmap_locked = false; 2050f3f0e1d2SKirill A. Shutemov fput(file); 2051f3f0e1d2SKirill A. Shutemov } else { 2052*50ad2f24SZach O'Keefe *result = khugepaged_scan_pmd(mm, vma, 2053b46e756fSKirill A. Shutemov khugepaged_scan.address, 2054*50ad2f24SZach O'Keefe &mmap_locked, cc); 2055f3f0e1d2SKirill A. Shutemov } 2056*50ad2f24SZach O'Keefe if (*result == SCAN_SUCCEED) 2057*50ad2f24SZach O'Keefe ++khugepaged_pages_collapsed; 2058b46e756fSKirill A. Shutemov /* move to next address */ 2059b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 2060b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 2061*50ad2f24SZach O'Keefe if (!mmap_locked) 2062*50ad2f24SZach O'Keefe /* 2063*50ad2f24SZach O'Keefe * We released mmap_lock so break loop. Note 2064*50ad2f24SZach O'Keefe * that we drop mmap_lock before all hugepage 2065*50ad2f24SZach O'Keefe * allocations, so if allocation fails, we are 2066*50ad2f24SZach O'Keefe * guaranteed to break here and report the 2067*50ad2f24SZach O'Keefe * correct result back to caller. 2068*50ad2f24SZach O'Keefe */ 2069c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2070b46e756fSKirill A. Shutemov if (progress >= pages) 2071b46e756fSKirill A. Shutemov goto breakouterloop; 2072b46e756fSKirill A. Shutemov } 2073b46e756fSKirill A. Shutemov } 2074b46e756fSKirill A. Shutemov breakouterloop: 2075d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ 2076c1e8d7c6SMichel Lespinasse breakouterloop_mmap_lock: 2077b46e756fSKirill A. Shutemov 2078b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2079b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2080b46e756fSKirill A. Shutemov /* 2081b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 2082b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 2083b46e756fSKirill A. Shutemov */ 2084b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm) || !vma) { 2085b46e756fSKirill A. Shutemov /* 2086b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 2087b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 2088b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 2089b46e756fSKirill A. Shutemov */ 2090b46e756fSKirill A. Shutemov if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2091b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = list_entry( 2092b46e756fSKirill A. Shutemov mm_slot->mm_node.next, 2093b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2094b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2095b46e756fSKirill A. Shutemov } else { 2096b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2097b46e756fSKirill A. Shutemov khugepaged_full_scans++; 2098b46e756fSKirill A. Shutemov } 2099b46e756fSKirill A. Shutemov 2100b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2101b46e756fSKirill A. Shutemov } 2102b46e756fSKirill A. Shutemov 2103b46e756fSKirill A. Shutemov return progress; 2104b46e756fSKirill A. Shutemov } 2105b46e756fSKirill A. Shutemov 2106b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 2107b46e756fSKirill A. Shutemov { 2108b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 21091064026bSYang Shi hugepage_flags_enabled(); 2110b46e756fSKirill A. Shutemov } 2111b46e756fSKirill A. Shutemov 2112b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 2113b46e756fSKirill A. Shutemov { 2114b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 2115b46e756fSKirill A. Shutemov kthread_should_stop(); 2116b46e756fSKirill A. Shutemov } 2117b46e756fSKirill A. Shutemov 211834d6b470SZach O'Keefe static void khugepaged_do_scan(struct collapse_control *cc) 2119b46e756fSKirill A. Shutemov { 2120b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 212189dc6a96SYanfei Xu unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); 2122b46e756fSKirill A. Shutemov bool wait = true; 2123*50ad2f24SZach O'Keefe int result = SCAN_SUCCEED; 2124b46e756fSKirill A. Shutemov 2125a980df33SKirill A. Shutemov lru_add_drain_all(); 2126a980df33SKirill A. Shutemov 2127c6a7f445SYang Shi while (true) { 2128b46e756fSKirill A. Shutemov cond_resched(); 2129b46e756fSKirill A. Shutemov 2130b46e756fSKirill A. Shutemov if (unlikely(kthread_should_stop() || try_to_freeze())) 2131b46e756fSKirill A. Shutemov break; 2132b46e756fSKirill A. Shutemov 2133b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2134b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 2135b46e756fSKirill A. Shutemov pass_through_head++; 2136b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 2137b46e756fSKirill A. Shutemov pass_through_head < 2) 2138b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 2139*50ad2f24SZach O'Keefe &result, cc); 2140b46e756fSKirill A. Shutemov else 2141b46e756fSKirill A. Shutemov progress = pages; 2142b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2143b46e756fSKirill A. Shutemov 2144c6a7f445SYang Shi if (progress >= pages) 2145c6a7f445SYang Shi break; 2146c6a7f445SYang Shi 2147*50ad2f24SZach O'Keefe if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) { 2148c6a7f445SYang Shi /* 2149c6a7f445SYang Shi * If fail to allocate the first time, try to sleep for 2150c6a7f445SYang Shi * a while. When hit again, cancel the scan. 2151c6a7f445SYang Shi */ 2152c6a7f445SYang Shi if (!wait) 2153c6a7f445SYang Shi break; 2154c6a7f445SYang Shi wait = false; 2155c6a7f445SYang Shi khugepaged_alloc_sleep(); 2156c6a7f445SYang Shi } 2157c6a7f445SYang Shi } 2158b46e756fSKirill A. Shutemov } 2159b46e756fSKirill A. Shutemov 2160b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 2161b46e756fSKirill A. Shutemov { 2162b46e756fSKirill A. Shutemov return kthread_should_stop() || 2163b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 2164b46e756fSKirill A. Shutemov } 2165b46e756fSKirill A. Shutemov 2166b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 2167b46e756fSKirill A. Shutemov { 2168b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 2169b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 2170b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2171b46e756fSKirill A. Shutemov 2172b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 2173b46e756fSKirill A. Shutemov return; 2174b46e756fSKirill A. Shutemov 2175b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2176b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 2177b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 2178b46e756fSKirill A. Shutemov scan_sleep_jiffies); 2179b46e756fSKirill A. Shutemov return; 2180b46e756fSKirill A. Shutemov } 2181b46e756fSKirill A. Shutemov 21821064026bSYang Shi if (hugepage_flags_enabled()) 2183b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2184b46e756fSKirill A. Shutemov } 2185b46e756fSKirill A. Shutemov 2186b46e756fSKirill A. Shutemov static int khugepaged(void *none) 2187b46e756fSKirill A. Shutemov { 2188b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 2189b46e756fSKirill A. Shutemov 2190b46e756fSKirill A. Shutemov set_freezable(); 2191b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 2192b46e756fSKirill A. Shutemov 2193b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 219434d6b470SZach O'Keefe khugepaged_do_scan(&khugepaged_collapse_control); 2195b46e756fSKirill A. Shutemov khugepaged_wait_work(); 2196b46e756fSKirill A. Shutemov } 2197b46e756fSKirill A. Shutemov 2198b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2199b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2200b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2201b46e756fSKirill A. Shutemov if (mm_slot) 2202b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2203b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2204b46e756fSKirill A. Shutemov return 0; 2205b46e756fSKirill A. Shutemov } 2206b46e756fSKirill A. Shutemov 2207b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 2208b46e756fSKirill A. Shutemov { 2209b46e756fSKirill A. Shutemov struct zone *zone; 2210b46e756fSKirill A. Shutemov int nr_zones = 0; 2211b46e756fSKirill A. Shutemov unsigned long recommended_min; 2212b46e756fSKirill A. Shutemov 22131064026bSYang Shi if (!hugepage_flags_enabled()) { 2214bd3400eaSLiangcai Fan calculate_min_free_kbytes(); 2215bd3400eaSLiangcai Fan goto update_wmarks; 2216bd3400eaSLiangcai Fan } 2217bd3400eaSLiangcai Fan 2218b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 2219b7d349c7SJoonsoo Kim /* 2220b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 2221b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 2222b7d349c7SJoonsoo Kim */ 2223b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 2224b7d349c7SJoonsoo Kim continue; 2225b7d349c7SJoonsoo Kim 2226b46e756fSKirill A. Shutemov nr_zones++; 2227b7d349c7SJoonsoo Kim } 2228b46e756fSKirill A. Shutemov 2229b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2230b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 2231b46e756fSKirill A. Shutemov 2232b46e756fSKirill A. Shutemov /* 2233b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 2234b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 2235b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 2236b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 2237b46e756fSKirill A. Shutemov */ 2238b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 2239b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2240b46e756fSKirill A. Shutemov 2241b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 2242b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 2243b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 2244b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 2245b46e756fSKirill A. Shutemov 2246b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 2247b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 2248b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2249b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 2250b46e756fSKirill A. Shutemov 2251b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 2252b46e756fSKirill A. Shutemov } 2253bd3400eaSLiangcai Fan 2254bd3400eaSLiangcai Fan update_wmarks: 2255b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 2256b46e756fSKirill A. Shutemov } 2257b46e756fSKirill A. Shutemov 2258b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 2259b46e756fSKirill A. Shutemov { 2260b46e756fSKirill A. Shutemov int err = 0; 2261b46e756fSKirill A. Shutemov 2262b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 22631064026bSYang Shi if (hugepage_flags_enabled()) { 2264b46e756fSKirill A. Shutemov if (!khugepaged_thread) 2265b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 2266b46e756fSKirill A. Shutemov "khugepaged"); 2267b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 2268b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2269b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 2270b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2271b46e756fSKirill A. Shutemov goto fail; 2272b46e756fSKirill A. Shutemov } 2273b46e756fSKirill A. Shutemov 2274b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 2275b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 2276b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 2277b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 2278b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2279b46e756fSKirill A. Shutemov } 2280bd3400eaSLiangcai Fan set_recommended_min_free_kbytes(); 2281b46e756fSKirill A. Shutemov fail: 2282b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 2283b46e756fSKirill A. Shutemov return err; 2284b46e756fSKirill A. Shutemov } 22854aab2be0SVijay Balakrishna 22864aab2be0SVijay Balakrishna void khugepaged_min_free_kbytes_update(void) 22874aab2be0SVijay Balakrishna { 22884aab2be0SVijay Balakrishna mutex_lock(&khugepaged_mutex); 22891064026bSYang Shi if (hugepage_flags_enabled() && khugepaged_thread) 22904aab2be0SVijay Balakrishna set_recommended_min_free_kbytes(); 22914aab2be0SVijay Balakrishna mutex_unlock(&khugepaged_mutex); 22924aab2be0SVijay Balakrishna } 2293