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; 56171a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0, result = 0, 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); 675b46e756fSKirill A. Shutemov return 1; 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); 681b46e756fSKirill A. Shutemov return 0; 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 816*9710a78aSZach 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); 821b46e756fSKirill A. Shutemov *hpage = ERR_PTR(-ENOMEM); 822*9710a78aSZach O'Keefe return false; 823b46e756fSKirill A. Shutemov } 824b46e756fSKirill A. Shutemov 825b46e756fSKirill A. Shutemov prep_transhuge_page(*hpage); 826b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 827*9710a78aSZach O'Keefe return true; 828b46e756fSKirill A. Shutemov } 829b46e756fSKirill A. Shutemov 830b46e756fSKirill A. Shutemov /* 831c1e8d7c6SMichel Lespinasse * If mmap_lock temporarily dropped, revalidate vma 832c1e8d7c6SMichel Lespinasse * before taking mmap_lock. 833b46e756fSKirill A. Shutemov * Return 0 if succeeds, otherwise return none-zero 834b46e756fSKirill A. Shutemov * value (scan code). 835b46e756fSKirill A. Shutemov */ 836b46e756fSKirill A. Shutemov 837c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 838c131f751SKirill A. Shutemov struct vm_area_struct **vmap) 839b46e756fSKirill A. Shutemov { 840b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 841b46e756fSKirill A. Shutemov 842b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 843b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 844b46e756fSKirill A. Shutemov 845c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 846b46e756fSKirill A. Shutemov if (!vma) 847b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 848b46e756fSKirill A. Shutemov 8494fa6893fSYang Shi if (!transhuge_vma_suitable(vma, address)) 850b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 8517da4e2cbSYang Shi if (!hugepage_vma_check(vma, vma->vm_flags, false, false)) 852b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 853f707fa49SYang Shi /* 854f707fa49SYang Shi * Anon VMA expected, the address may be unmapped then 855f707fa49SYang Shi * remapped to file after khugepaged reaquired the mmap_lock. 856f707fa49SYang Shi * 857f707fa49SYang Shi * hugepage_vma_check may return true for qualified file 858f707fa49SYang Shi * vmas. 859f707fa49SYang Shi */ 86025fa414aSxu xin if (!vma->anon_vma || !vma_is_anonymous(vma)) 861594cced1SKirill A. Shutemov return SCAN_VMA_CHECK; 862b46e756fSKirill A. Shutemov return 0; 863b46e756fSKirill A. Shutemov } 864b46e756fSKirill A. Shutemov 865b46e756fSKirill A. Shutemov /* 866b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 867b46e756fSKirill A. Shutemov * Only done if khugepaged_scan_pmd believes it is worthwhile. 868b46e756fSKirill A. Shutemov * 8694d928e20SMiaohe Lin * Called and returns without pte mapped or spinlocks held. 8704d928e20SMiaohe Lin * Note that if false is returned, mmap_lock will be released. 871b46e756fSKirill A. Shutemov */ 872b46e756fSKirill A. Shutemov 873b46e756fSKirill A. Shutemov static bool __collapse_huge_page_swapin(struct mm_struct *mm, 874b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 8752b635dd3SWill Deacon unsigned long haddr, pmd_t *pmd, 8760db501f7SEbru Akagunduz int referenced) 877b46e756fSKirill A. Shutemov { 8782b740303SSouptick Joarder int swapped_in = 0; 8792b740303SSouptick Joarder vm_fault_t ret = 0; 8802b635dd3SWill Deacon unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); 8812b635dd3SWill Deacon 8822b635dd3SWill Deacon for (address = haddr; address < end; address += PAGE_SIZE) { 88382b0f8c3SJan Kara struct vm_fault vmf = { 884b46e756fSKirill A. Shutemov .vma = vma, 885b46e756fSKirill A. Shutemov .address = address, 8862b635dd3SWill Deacon .pgoff = linear_page_index(vma, haddr), 887b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 888b46e756fSKirill A. Shutemov .pmd = pmd, 889b46e756fSKirill A. Shutemov }; 890b46e756fSKirill A. Shutemov 89182b0f8c3SJan Kara vmf.pte = pte_offset_map(pmd, address); 8922994302bSJan Kara vmf.orig_pte = *vmf.pte; 8932b635dd3SWill Deacon if (!is_swap_pte(vmf.orig_pte)) { 8942b635dd3SWill Deacon pte_unmap(vmf.pte); 895b46e756fSKirill A. Shutemov continue; 8962b635dd3SWill Deacon } 8972994302bSJan Kara ret = do_swap_page(&vmf); 8980db501f7SEbru Akagunduz 8994d928e20SMiaohe Lin /* 9004d928e20SMiaohe Lin * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. 9014d928e20SMiaohe Lin * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because 9024d928e20SMiaohe Lin * we do not retry here and swap entry will remain in pagetable 9034d928e20SMiaohe Lin * resulting in later failure. 9044d928e20SMiaohe Lin */ 905b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 9060db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 907b46e756fSKirill A. Shutemov return false; 90847f863eaSEbru Akagunduz } 909b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 9104d928e20SMiaohe Lin mmap_read_unlock(mm); 9110db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 912b46e756fSKirill A. Shutemov return false; 913b46e756fSKirill A. Shutemov } 9144d928e20SMiaohe Lin swapped_in++; 915b46e756fSKirill A. Shutemov } 916ae2c5d80SKirill A. Shutemov 917ae2c5d80SKirill A. Shutemov /* Drain LRU add pagevec to remove extra pin on the swapped in pages */ 918ae2c5d80SKirill A. Shutemov if (swapped_in) 919ae2c5d80SKirill A. Shutemov lru_add_drain(); 920ae2c5d80SKirill A. Shutemov 9210db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); 922b46e756fSKirill A. Shutemov return true; 923b46e756fSKirill A. Shutemov } 924b46e756fSKirill A. Shutemov 925*9710a78aSZach O'Keefe static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, 926*9710a78aSZach O'Keefe struct collapse_control *cc) 927*9710a78aSZach O'Keefe { 928*9710a78aSZach O'Keefe /* Only allocate from the target node */ 929*9710a78aSZach O'Keefe gfp_t gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 930*9710a78aSZach O'Keefe int node = khugepaged_find_target_node(cc); 931*9710a78aSZach O'Keefe 932*9710a78aSZach O'Keefe if (!khugepaged_alloc_page(hpage, gfp, node)) 933*9710a78aSZach O'Keefe return SCAN_ALLOC_HUGE_PAGE_FAIL; 934*9710a78aSZach O'Keefe if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp))) 935*9710a78aSZach O'Keefe return SCAN_CGROUP_CHARGE_FAIL; 936*9710a78aSZach O'Keefe count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC); 937*9710a78aSZach O'Keefe return SCAN_SUCCEED; 938*9710a78aSZach O'Keefe } 939*9710a78aSZach O'Keefe 940*9710a78aSZach O'Keefe static void collapse_huge_page(struct mm_struct *mm, unsigned long address, 941*9710a78aSZach O'Keefe struct page **hpage, int referenced, 942*9710a78aSZach O'Keefe int unmapped, struct collapse_control *cc) 943b46e756fSKirill A. Shutemov { 9445503fbf2SKirill A. Shutemov LIST_HEAD(compound_pagelist); 945b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 946b46e756fSKirill A. Shutemov pte_t *pte; 947b46e756fSKirill A. Shutemov pgtable_t pgtable; 948b46e756fSKirill A. Shutemov struct page *new_page; 949b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 950b46e756fSKirill A. Shutemov int isolated = 0, result = 0; 951c131f751SKirill A. Shutemov struct vm_area_struct *vma; 952ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 953b46e756fSKirill A. Shutemov 954b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 955b46e756fSKirill A. Shutemov 956988ddb71SKirill A. Shutemov /* 957c1e8d7c6SMichel Lespinasse * Before allocating the hugepage, release the mmap_lock read lock. 958988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 959c1e8d7c6SMichel Lespinasse * sync compaction, and we do not need to hold the mmap_lock during 960988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 961988ddb71SKirill A. Shutemov */ 962d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 963b46e756fSKirill A. Shutemov 964*9710a78aSZach O'Keefe result = alloc_charge_hpage(hpage, mm, cc); 965*9710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 966b46e756fSKirill A. Shutemov goto out_nolock; 967*9710a78aSZach O'Keefe 968*9710a78aSZach O'Keefe new_page = *hpage; 969b46e756fSKirill A. Shutemov 970d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 971c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 972b46e756fSKirill A. Shutemov if (result) { 973d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 974b46e756fSKirill A. Shutemov goto out_nolock; 975b46e756fSKirill A. Shutemov } 976b46e756fSKirill A. Shutemov 977b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 978b46e756fSKirill A. Shutemov if (!pmd) { 979b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 980d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 981b46e756fSKirill A. Shutemov goto out_nolock; 982b46e756fSKirill A. Shutemov } 983b46e756fSKirill A. Shutemov 984b46e756fSKirill A. Shutemov /* 9854d928e20SMiaohe Lin * __collapse_huge_page_swapin will return with mmap_lock released 9864d928e20SMiaohe Lin * when it fails. So we jump out_nolock directly in that case. 987b46e756fSKirill A. Shutemov * Continuing to collapse causes inconsistency. 988b46e756fSKirill A. Shutemov */ 989ffe945e6SKirill A. Shutemov if (unmapped && !__collapse_huge_page_swapin(mm, vma, address, 990ffe945e6SKirill A. Shutemov pmd, referenced)) { 991b46e756fSKirill A. Shutemov goto out_nolock; 992b46e756fSKirill A. Shutemov } 993b46e756fSKirill A. Shutemov 994d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 995b46e756fSKirill A. Shutemov /* 996b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 997b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 998b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 999b46e756fSKirill A. Shutemov */ 1000d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 1001c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 1002b46e756fSKirill A. Shutemov if (result) 100318d24a7cSMiaohe Lin goto out_up_write; 1004b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 1005b46e756fSKirill A. Shutemov if (mm_find_pmd(mm, address) != pmd) 100618d24a7cSMiaohe Lin goto out_up_write; 1007b46e756fSKirill A. Shutemov 1008b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1009b46e756fSKirill A. Shutemov 10107269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 10116f4f13e8SJérôme Glisse address, address + HPAGE_PMD_SIZE); 1012ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1013ec649c9dSVille Syrjälä 1014ec649c9dSVille Syrjälä pte = pte_offset_map(pmd, address); 1015ec649c9dSVille Syrjälä pte_ptl = pte_lockptr(mm, pmd); 1016ec649c9dSVille Syrjälä 1017b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1018b46e756fSKirill A. Shutemov /* 1019b46e756fSKirill A. Shutemov * After this gup_fast can't run anymore. This also removes 1020b46e756fSKirill A. Shutemov * any huge TLB entry from the CPU so we won't allow 1021b46e756fSKirill A. Shutemov * huge and small TLB entries for the same virtual address 1022b46e756fSKirill A. Shutemov * to avoid the risk of CPU bugs in that area. 1023b46e756fSKirill A. Shutemov */ 1024b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1025b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1026ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1027b46e756fSKirill A. Shutemov 1028b46e756fSKirill A. Shutemov spin_lock(pte_ptl); 10295503fbf2SKirill A. Shutemov isolated = __collapse_huge_page_isolate(vma, address, pte, 10305503fbf2SKirill A. Shutemov &compound_pagelist); 1031b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1032b46e756fSKirill A. Shutemov 1033b46e756fSKirill A. Shutemov if (unlikely(!isolated)) { 1034b46e756fSKirill A. Shutemov pte_unmap(pte); 1035b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1036b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1037b46e756fSKirill A. Shutemov /* 1038b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1039b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1040b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1041b46e756fSKirill A. Shutemov */ 1042b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1043b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1044b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1045b46e756fSKirill A. Shutemov result = SCAN_FAIL; 104618d24a7cSMiaohe Lin goto out_up_write; 1047b46e756fSKirill A. Shutemov } 1048b46e756fSKirill A. Shutemov 1049b46e756fSKirill A. Shutemov /* 1050b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1051b46e756fSKirill A. Shutemov * can't run anymore. 1052b46e756fSKirill A. Shutemov */ 1053b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1054b46e756fSKirill A. Shutemov 10555503fbf2SKirill A. Shutemov __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl, 10565503fbf2SKirill A. Shutemov &compound_pagelist); 1057b46e756fSKirill A. Shutemov pte_unmap(pte); 1058588d01f9SMiaohe Lin /* 1059588d01f9SMiaohe Lin * spin_lock() below is not the equivalent of smp_wmb(), but 1060588d01f9SMiaohe Lin * the smp_wmb() inside __SetPageUptodate() can be reused to 1061588d01f9SMiaohe Lin * avoid the copy_huge_page writes to become visible after 1062588d01f9SMiaohe Lin * the set_pmd_at() write. 1063588d01f9SMiaohe Lin */ 1064b46e756fSKirill A. Shutemov __SetPageUptodate(new_page); 1065b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1066b46e756fSKirill A. Shutemov 1067b46e756fSKirill A. Shutemov _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); 1068f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1069b46e756fSKirill A. Shutemov 1070b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1071b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 107240f2bbf7SDavid Hildenbrand page_add_new_anon_rmap(new_page, vma, address); 1073b518154eSJoonsoo Kim lru_cache_add_inactive_or_unevictable(new_page, vma); 1074b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1075b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1076b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1077b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1078b46e756fSKirill A. Shutemov 1079b46e756fSKirill A. Shutemov *hpage = NULL; 1080b46e756fSKirill A. Shutemov 1081b46e756fSKirill A. Shutemov khugepaged_pages_collapsed++; 1082b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1083b46e756fSKirill A. Shutemov out_up_write: 1084d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 1085b46e756fSKirill A. Shutemov out_nolock: 1086c6a7f445SYang Shi if (!IS_ERR_OR_NULL(*hpage)) { 1087bbc6b703SMatthew Wilcox (Oracle) mem_cgroup_uncharge(page_folio(*hpage)); 1088c6a7f445SYang Shi put_page(*hpage); 1089c6a7f445SYang Shi } 1090b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page(mm, isolated, result); 1091b46e756fSKirill A. Shutemov return; 1092b46e756fSKirill A. Shutemov } 1093b46e756fSKirill A. Shutemov 109434d6b470SZach O'Keefe static int khugepaged_scan_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 109534d6b470SZach O'Keefe unsigned long address, struct page **hpage, 109634d6b470SZach O'Keefe struct collapse_control *cc) 1097b46e756fSKirill A. Shutemov { 1098b46e756fSKirill A. Shutemov pmd_t *pmd; 1099b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 110071a2c112SKirill A. Shutemov int ret = 0, result = 0, referenced = 0; 110171a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0; 1102b46e756fSKirill A. Shutemov struct page *page = NULL; 1103b46e756fSKirill A. Shutemov unsigned long _address; 1104b46e756fSKirill A. Shutemov spinlock_t *ptl; 1105b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 11060db501f7SEbru Akagunduz bool writable = false; 1107b46e756fSKirill A. Shutemov 1108b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1109b46e756fSKirill A. Shutemov 1110b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 1111b46e756fSKirill A. Shutemov if (!pmd) { 1112b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 1113b46e756fSKirill A. Shutemov goto out; 1114b46e756fSKirill A. Shutemov } 1115b46e756fSKirill A. Shutemov 111634d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1117b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1118b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; 1119b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1120b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 1121b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1122b46e756fSKirill A. Shutemov if (++unmapped <= khugepaged_max_ptes_swap) { 1123e1e267c7SPeter Xu /* 1124e1e267c7SPeter Xu * Always be strict with uffd-wp 1125e1e267c7SPeter Xu * enabled swap entries. Please see 1126e1e267c7SPeter Xu * comment below for pte_uffd_wp(). 1127e1e267c7SPeter Xu */ 1128e1e267c7SPeter Xu if (pte_swp_uffd_wp(pteval)) { 1129e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1130e1e267c7SPeter Xu goto out_unmap; 1131e1e267c7SPeter Xu } 1132b46e756fSKirill A. Shutemov continue; 1133b46e756fSKirill A. Shutemov } else { 1134b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1135e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1136b46e756fSKirill A. Shutemov goto out_unmap; 1137b46e756fSKirill A. Shutemov } 1138b46e756fSKirill A. Shutemov } 1139b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1140b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1141b46e756fSKirill A. Shutemov ++none_or_zero <= khugepaged_max_ptes_none) { 1142b46e756fSKirill A. Shutemov continue; 1143b46e756fSKirill A. Shutemov } else { 1144b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1145e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1146b46e756fSKirill A. Shutemov goto out_unmap; 1147b46e756fSKirill A. Shutemov } 1148b46e756fSKirill A. Shutemov } 1149e1e267c7SPeter Xu if (pte_uffd_wp(pteval)) { 1150e1e267c7SPeter Xu /* 1151e1e267c7SPeter Xu * Don't collapse the page if any of the small 1152e1e267c7SPeter Xu * PTEs are armed with uffd write protection. 1153e1e267c7SPeter Xu * Here we can also mark the new huge pmd as 1154e1e267c7SPeter Xu * write protected if any of the small ones is 11558958b249SHaitao Shi * marked but that could bring unknown 1156e1e267c7SPeter Xu * userfault messages that falls outside of 1157e1e267c7SPeter Xu * the registered range. So, just be simple. 1158e1e267c7SPeter Xu */ 1159e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1160e1e267c7SPeter Xu goto out_unmap; 1161e1e267c7SPeter Xu } 1162b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1163b46e756fSKirill A. Shutemov writable = true; 1164b46e756fSKirill A. Shutemov 1165b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 11663218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 1167b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1168b46e756fSKirill A. Shutemov goto out_unmap; 1169b46e756fSKirill A. Shutemov } 1170b46e756fSKirill A. Shutemov 117171a2c112SKirill A. Shutemov if (page_mapcount(page) > 1 && 117271a2c112SKirill A. Shutemov ++shared > khugepaged_max_ptes_shared) { 117371a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 1174e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 117571a2c112SKirill A. Shutemov goto out_unmap; 117671a2c112SKirill A. Shutemov } 117771a2c112SKirill A. Shutemov 11785503fbf2SKirill A. Shutemov page = compound_head(page); 1179b46e756fSKirill A. Shutemov 1180b46e756fSKirill A. Shutemov /* 1181b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 118234d6b470SZach O'Keefe * information to cc->node_load[]. 11830b8f0d87SQuanfa Fu * Khugepaged will allocate hugepage from the node has the max 1184b46e756fSKirill A. Shutemov * hit record. 1185b46e756fSKirill A. Shutemov */ 1186b46e756fSKirill A. Shutemov node = page_to_nid(page); 118734d6b470SZach O'Keefe if (khugepaged_scan_abort(node, cc)) { 1188b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1189b46e756fSKirill A. Shutemov goto out_unmap; 1190b46e756fSKirill A. Shutemov } 119134d6b470SZach O'Keefe cc->node_load[node]++; 1192b46e756fSKirill A. Shutemov if (!PageLRU(page)) { 1193b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1194b46e756fSKirill A. Shutemov goto out_unmap; 1195b46e756fSKirill A. Shutemov } 1196b46e756fSKirill A. Shutemov if (PageLocked(page)) { 1197b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1198b46e756fSKirill A. Shutemov goto out_unmap; 1199b46e756fSKirill A. Shutemov } 1200b46e756fSKirill A. Shutemov if (!PageAnon(page)) { 1201b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1202b46e756fSKirill A. Shutemov goto out_unmap; 1203b46e756fSKirill A. Shutemov } 1204b46e756fSKirill A. Shutemov 1205b46e756fSKirill A. Shutemov /* 12069445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 12079445689fSKirill A. Shutemov * 120836ee2c78SMiaohe Lin * Here the check is racy it may see total_mapcount > refcount 12099445689fSKirill A. Shutemov * in some cases. 12109445689fSKirill A. Shutemov * For example, one process with one forked child process. 12119445689fSKirill A. Shutemov * The parent has the PMD split due to MADV_DONTNEED, then 12129445689fSKirill A. Shutemov * the child is trying unmap the whole PMD, but khugepaged 12139445689fSKirill A. Shutemov * may be scanning the parent between the child has 12149445689fSKirill A. Shutemov * PageDoubleMap flag cleared and dec the mapcount. So 12159445689fSKirill A. Shutemov * khugepaged may see total_mapcount > refcount. 12169445689fSKirill A. Shutemov * 12179445689fSKirill A. Shutemov * But such case is ephemeral we could always retry collapse 12189445689fSKirill A. Shutemov * later. However it may report false positive if the page 12199445689fSKirill A. Shutemov * has excessive GUP pins (i.e. 512). Anyway the same check 12209445689fSKirill A. Shutemov * will be done again later the risk seems low. 1221b46e756fSKirill A. Shutemov */ 12229445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 1223b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1224b46e756fSKirill A. Shutemov goto out_unmap; 1225b46e756fSKirill A. Shutemov } 1226b46e756fSKirill A. Shutemov if (pte_young(pteval) || 1227b46e756fSKirill A. Shutemov page_is_young(page) || PageReferenced(page) || 1228b46e756fSKirill A. Shutemov mmu_notifier_test_young(vma->vm_mm, address)) 12290db501f7SEbru Akagunduz referenced++; 1230b46e756fSKirill A. Shutemov } 1231ffe945e6SKirill A. Shutemov if (!writable) { 1232ffe945e6SKirill A. Shutemov result = SCAN_PAGE_RO; 1233ffe945e6SKirill A. Shutemov } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) { 1234ffe945e6SKirill A. Shutemov result = SCAN_LACK_REFERENCED_PAGE; 1235ffe945e6SKirill A. Shutemov } else { 1236b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1237b46e756fSKirill A. Shutemov ret = 1; 1238b46e756fSKirill A. Shutemov } 1239b46e756fSKirill A. Shutemov out_unmap: 1240b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 1241b46e756fSKirill A. Shutemov if (ret) { 1242c1e8d7c6SMichel Lespinasse /* collapse_huge_page will return with the mmap_lock released */ 1243*9710a78aSZach O'Keefe collapse_huge_page(mm, address, hpage, referenced, unmapped, 1244*9710a78aSZach O'Keefe cc); 1245b46e756fSKirill A. Shutemov } 1246b46e756fSKirill A. Shutemov out: 1247b46e756fSKirill A. Shutemov trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, 1248b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 1249b46e756fSKirill A. Shutemov return ret; 1250b46e756fSKirill A. Shutemov } 1251b46e756fSKirill A. Shutemov 1252b46e756fSKirill A. Shutemov static void collect_mm_slot(struct mm_slot *mm_slot) 1253b46e756fSKirill A. Shutemov { 1254b46e756fSKirill A. Shutemov struct mm_struct *mm = mm_slot->mm; 1255b46e756fSKirill A. Shutemov 125635f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1257b46e756fSKirill A. Shutemov 1258b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm)) { 1259b46e756fSKirill A. Shutemov /* free mm_slot */ 1260b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 1261b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 1262b46e756fSKirill A. Shutemov 1263b46e756fSKirill A. Shutemov /* 1264b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1265b46e756fSKirill A. Shutemov * 1266b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1267b46e756fSKirill A. Shutemov */ 1268b46e756fSKirill A. Shutemov 1269b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1270b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 1271b46e756fSKirill A. Shutemov mmdrop(mm); 1272b46e756fSKirill A. Shutemov } 1273b46e756fSKirill A. Shutemov } 1274b46e756fSKirill A. Shutemov 1275396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 127627e1f827SSong Liu /* 127727e1f827SSong Liu * Notify khugepaged that given addr of the mm is pte-mapped THP. Then 127827e1f827SSong Liu * khugepaged should try to collapse the page table. 127927e1f827SSong Liu */ 1280081c3256SMiaohe Lin static void khugepaged_add_pte_mapped_thp(struct mm_struct *mm, 128127e1f827SSong Liu unsigned long addr) 128227e1f827SSong Liu { 128327e1f827SSong Liu struct mm_slot *mm_slot; 128427e1f827SSong Liu 128527e1f827SSong Liu VM_BUG_ON(addr & ~HPAGE_PMD_MASK); 128627e1f827SSong Liu 128727e1f827SSong Liu spin_lock(&khugepaged_mm_lock); 128827e1f827SSong Liu mm_slot = get_mm_slot(mm); 128927e1f827SSong Liu if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) 129027e1f827SSong Liu mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; 129127e1f827SSong Liu spin_unlock(&khugepaged_mm_lock); 129227e1f827SSong Liu } 129327e1f827SSong Liu 1294e59a47b8SPasha Tatashin static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 1295e59a47b8SPasha Tatashin unsigned long addr, pmd_t *pmdp) 1296e59a47b8SPasha Tatashin { 1297e59a47b8SPasha Tatashin spinlock_t *ptl; 1298e59a47b8SPasha Tatashin pmd_t pmd; 1299e59a47b8SPasha Tatashin 130080110bbfSPasha Tatashin mmap_assert_write_locked(mm); 1301e59a47b8SPasha Tatashin ptl = pmd_lock(vma->vm_mm, pmdp); 1302e59a47b8SPasha Tatashin pmd = pmdp_collapse_flush(vma, addr, pmdp); 1303e59a47b8SPasha Tatashin spin_unlock(ptl); 1304e59a47b8SPasha Tatashin mm_dec_nr_ptes(mm); 130580110bbfSPasha Tatashin page_table_check_pte_clear_range(mm, addr, pmd); 1306e59a47b8SPasha Tatashin pte_free(mm, pmd_pgtable(pmd)); 1307e59a47b8SPasha Tatashin } 1308e59a47b8SPasha Tatashin 130927e1f827SSong Liu /** 1310336e6b53SAlex Shi * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at 1311336e6b53SAlex Shi * address haddr. 1312336e6b53SAlex Shi * 1313336e6b53SAlex Shi * @mm: process address space where collapse happens 1314336e6b53SAlex Shi * @addr: THP collapse address 131527e1f827SSong Liu * 131627e1f827SSong Liu * This function checks whether all the PTEs in the PMD are pointing to the 131727e1f827SSong Liu * right THP. If so, retract the page table so the THP can refault in with 131827e1f827SSong Liu * as pmd-mapped. 131927e1f827SSong Liu */ 132027e1f827SSong Liu void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr) 132127e1f827SSong Liu { 132227e1f827SSong Liu unsigned long haddr = addr & HPAGE_PMD_MASK; 132327e1f827SSong Liu struct vm_area_struct *vma = find_vma(mm, haddr); 1324119a5fc1SHugh Dickins struct page *hpage; 132527e1f827SSong Liu pte_t *start_pte, *pte; 1326e59a47b8SPasha Tatashin pmd_t *pmd; 132727e1f827SSong Liu spinlock_t *ptl; 132827e1f827SSong Liu int count = 0; 132927e1f827SSong Liu int i; 133027e1f827SSong Liu 133127e1f827SSong Liu if (!vma || !vma->vm_file || 1332fef792a4SMiaohe Lin !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) 133327e1f827SSong Liu return; 133427e1f827SSong Liu 133527e1f827SSong Liu /* 133627e1f827SSong Liu * This vm_flags may not have VM_HUGEPAGE if the page was not 133727e1f827SSong Liu * collapsed by this mm. But we can still collapse if the page is 133827e1f827SSong Liu * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check() 133927e1f827SSong Liu * will not fail the vma for missing VM_HUGEPAGE 134027e1f827SSong Liu */ 13417da4e2cbSYang Shi if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE, false, false)) 134227e1f827SSong Liu return; 134327e1f827SSong Liu 1344deb4c93aSPeter Xu /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ 1345deb4c93aSPeter Xu if (userfaultfd_wp(vma)) 1346deb4c93aSPeter Xu return; 1347deb4c93aSPeter Xu 1348119a5fc1SHugh Dickins hpage = find_lock_page(vma->vm_file->f_mapping, 1349119a5fc1SHugh Dickins linear_page_index(vma, haddr)); 1350119a5fc1SHugh Dickins if (!hpage) 1351119a5fc1SHugh Dickins return; 1352119a5fc1SHugh Dickins 1353119a5fc1SHugh Dickins if (!PageHead(hpage)) 1354119a5fc1SHugh Dickins goto drop_hpage; 1355119a5fc1SHugh Dickins 135627e1f827SSong Liu pmd = mm_find_pmd(mm, haddr); 135727e1f827SSong Liu if (!pmd) 1358119a5fc1SHugh Dickins goto drop_hpage; 135927e1f827SSong Liu 136027e1f827SSong Liu start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 136127e1f827SSong Liu 136227e1f827SSong Liu /* step 1: check all mapped PTEs are to the right huge page */ 136327e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 136427e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 136527e1f827SSong Liu struct page *page; 136627e1f827SSong Liu 136727e1f827SSong Liu /* empty pte, skip */ 136827e1f827SSong Liu if (pte_none(*pte)) 136927e1f827SSong Liu continue; 137027e1f827SSong Liu 137127e1f827SSong Liu /* page swapped out, abort */ 137227e1f827SSong Liu if (!pte_present(*pte)) 137327e1f827SSong Liu goto abort; 137427e1f827SSong Liu 137527e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 13763218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 13773218f871SAlex Sierra page = NULL; 137827e1f827SSong Liu /* 1379119a5fc1SHugh Dickins * Note that uprobe, debugger, or MAP_PRIVATE may change the 1380119a5fc1SHugh Dickins * page table, but the new page will not be a subpage of hpage. 138127e1f827SSong Liu */ 1382119a5fc1SHugh Dickins if (hpage + i != page) 138327e1f827SSong Liu goto abort; 138427e1f827SSong Liu count++; 138527e1f827SSong Liu } 138627e1f827SSong Liu 138727e1f827SSong Liu /* step 2: adjust rmap */ 138827e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 138927e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 139027e1f827SSong Liu struct page *page; 139127e1f827SSong Liu 139227e1f827SSong Liu if (pte_none(*pte)) 139327e1f827SSong Liu continue; 139427e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 13953218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 13963218f871SAlex Sierra goto abort; 1397cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 139827e1f827SSong Liu } 139927e1f827SSong Liu 140027e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 140127e1f827SSong Liu 140227e1f827SSong Liu /* step 3: set proper refcount and mm_counters. */ 1403119a5fc1SHugh Dickins if (count) { 140427e1f827SSong Liu page_ref_sub(hpage, count); 140527e1f827SSong Liu add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); 140627e1f827SSong Liu } 140727e1f827SSong Liu 140827e1f827SSong Liu /* step 4: collapse pmd */ 1409e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, haddr, pmd); 1410119a5fc1SHugh Dickins drop_hpage: 1411119a5fc1SHugh Dickins unlock_page(hpage); 1412119a5fc1SHugh Dickins put_page(hpage); 141327e1f827SSong Liu return; 141427e1f827SSong Liu 141527e1f827SSong Liu abort: 141627e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 1417119a5fc1SHugh Dickins goto drop_hpage; 141827e1f827SSong Liu } 141927e1f827SSong Liu 14200edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 142127e1f827SSong Liu { 142227e1f827SSong Liu struct mm_struct *mm = mm_slot->mm; 142327e1f827SSong Liu int i; 142427e1f827SSong Liu 142527e1f827SSong Liu if (likely(mm_slot->nr_pte_mapped_thp == 0)) 14260edf61e5SMiaohe Lin return; 142727e1f827SSong Liu 1428d8ed45c5SMichel Lespinasse if (!mmap_write_trylock(mm)) 14290edf61e5SMiaohe Lin return; 143027e1f827SSong Liu 143127e1f827SSong Liu if (unlikely(khugepaged_test_exit(mm))) 143227e1f827SSong Liu goto out; 143327e1f827SSong Liu 143427e1f827SSong Liu for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) 143527e1f827SSong Liu collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]); 143627e1f827SSong Liu 143727e1f827SSong Liu out: 143827e1f827SSong Liu mm_slot->nr_pte_mapped_thp = 0; 1439d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 144027e1f827SSong Liu } 144127e1f827SSong Liu 1442f3f0e1d2SKirill A. Shutemov static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1443f3f0e1d2SKirill A. Shutemov { 1444f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 144518e77600SHugh Dickins struct mm_struct *mm; 1446f3f0e1d2SKirill A. Shutemov unsigned long addr; 1447e59a47b8SPasha Tatashin pmd_t *pmd; 1448f3f0e1d2SKirill A. Shutemov 1449f3f0e1d2SKirill A. Shutemov i_mmap_lock_write(mapping); 1450f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 145127e1f827SSong Liu /* 145227e1f827SSong Liu * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 145327e1f827SSong Liu * got written to. These VMAs are likely not worth investing 14543e4e28c5SMichel Lespinasse * mmap_write_lock(mm) as PMD-mapping is likely to be split 145527e1f827SSong Liu * later. 145627e1f827SSong Liu * 145736ee2c78SMiaohe Lin * Note that vma->anon_vma check is racy: it can be set up after 1458c1e8d7c6SMichel Lespinasse * the check but before we took mmap_lock by the fault path. 145927e1f827SSong Liu * But page lock would prevent establishing any new ptes of the 146027e1f827SSong Liu * page, so we are safe. 146127e1f827SSong Liu * 146227e1f827SSong Liu * An alternative would be drop the check, but check that page 146327e1f827SSong Liu * table is clear before calling pmdp_collapse_flush() under 146427e1f827SSong Liu * ptl. It has higher chance to recover THP for the VMA, but 146527e1f827SSong Liu * has higher cost too. 146627e1f827SSong Liu */ 1467f3f0e1d2SKirill A. Shutemov if (vma->anon_vma) 1468f3f0e1d2SKirill A. Shutemov continue; 1469f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 1470f3f0e1d2SKirill A. Shutemov if (addr & ~HPAGE_PMD_MASK) 1471f3f0e1d2SKirill A. Shutemov continue; 1472f3f0e1d2SKirill A. Shutemov if (vma->vm_end < addr + HPAGE_PMD_SIZE) 1473f3f0e1d2SKirill A. Shutemov continue; 147418e77600SHugh Dickins mm = vma->vm_mm; 147518e77600SHugh Dickins pmd = mm_find_pmd(mm, addr); 1476f3f0e1d2SKirill A. Shutemov if (!pmd) 1477f3f0e1d2SKirill A. Shutemov continue; 1478f3f0e1d2SKirill A. Shutemov /* 1479c1e8d7c6SMichel Lespinasse * We need exclusive mmap_lock to retract page table. 148027e1f827SSong Liu * 148127e1f827SSong Liu * We use trylock due to lock inversion: we need to acquire 1482c1e8d7c6SMichel Lespinasse * mmap_lock while holding page lock. Fault path does it in 148327e1f827SSong Liu * reverse order. Trylock is a way to avoid deadlock. 1484f3f0e1d2SKirill A. Shutemov */ 148518e77600SHugh Dickins if (mmap_write_trylock(mm)) { 1486deb4c93aSPeter Xu /* 1487deb4c93aSPeter Xu * When a vma is registered with uffd-wp, we can't 1488deb4c93aSPeter Xu * recycle the pmd pgtable because there can be pte 1489deb4c93aSPeter Xu * markers installed. Skip it only, so the rest mm/vma 1490deb4c93aSPeter Xu * can still have the same file mapped hugely, however 1491deb4c93aSPeter Xu * it'll always mapped in small page size for uffd-wp 1492deb4c93aSPeter Xu * registered ranges. 1493deb4c93aSPeter Xu */ 1494deb4c93aSPeter Xu if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma)) 1495e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, addr, pmd); 149618e77600SHugh Dickins mmap_write_unlock(mm); 149727e1f827SSong Liu } else { 149827e1f827SSong Liu /* Try again later */ 149918e77600SHugh Dickins khugepaged_add_pte_mapped_thp(mm, addr); 1500f3f0e1d2SKirill A. Shutemov } 1501f3f0e1d2SKirill A. Shutemov } 1502f3f0e1d2SKirill A. Shutemov i_mmap_unlock_write(mapping); 1503f3f0e1d2SKirill A. Shutemov } 1504f3f0e1d2SKirill A. Shutemov 1505f3f0e1d2SKirill A. Shutemov /** 150699cb0dbdSSong Liu * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1507f3f0e1d2SKirill A. Shutemov * 1508336e6b53SAlex Shi * @mm: process address space where collapse happens 1509336e6b53SAlex Shi * @file: file that collapse on 1510336e6b53SAlex Shi * @start: collapse start address 1511336e6b53SAlex Shi * @hpage: new allocated huge page for collapse 1512*9710a78aSZach O'Keefe * @cc: collapse context and scratchpad 1513336e6b53SAlex Shi * 1514f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 151587c460a0SHugh Dickins * - allocate and lock a new huge page; 151677da9389SMatthew Wilcox * - scan page cache replacing old pages with the new one 151799cb0dbdSSong Liu * + swap/gup in pages if necessary; 1518f3f0e1d2SKirill A. Shutemov * + fill in gaps; 151977da9389SMatthew Wilcox * + keep old pages around in case rollback is required; 152077da9389SMatthew Wilcox * - if replacing succeeds: 1521f3f0e1d2SKirill A. Shutemov * + copy data over; 1522f3f0e1d2SKirill A. Shutemov * + free old pages; 152387c460a0SHugh Dickins * + unlock huge page; 1524f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1525f3f0e1d2SKirill A. Shutemov * + put all pages back and unfreeze them; 152677da9389SMatthew Wilcox * + restore gaps in the page cache; 152787c460a0SHugh Dickins * + unlock and free huge page; 1528f3f0e1d2SKirill A. Shutemov */ 1529*9710a78aSZach O'Keefe static void collapse_file(struct mm_struct *mm, struct file *file, 1530*9710a78aSZach O'Keefe pgoff_t start, struct page **hpage, 1531*9710a78aSZach O'Keefe struct collapse_control *cc) 1532f3f0e1d2SKirill A. Shutemov { 1533579c571eSSong Liu struct address_space *mapping = file->f_mapping; 153477da9389SMatthew Wilcox struct page *new_page; 1535f3f0e1d2SKirill A. Shutemov pgoff_t index, end = start + HPAGE_PMD_NR; 1536f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 153777da9389SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1538f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 153999cb0dbdSSong Liu bool is_shmem = shmem_file(file); 1540bf9eceadSMuchun Song int nr; 1541f3f0e1d2SKirill A. Shutemov 154299cb0dbdSSong Liu VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1543f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1544f3f0e1d2SKirill A. Shutemov 1545*9710a78aSZach O'Keefe result = alloc_charge_hpage(hpage, mm, cc); 1546*9710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1547f3f0e1d2SKirill A. Shutemov goto out; 1548f3f0e1d2SKirill A. Shutemov 1549*9710a78aSZach O'Keefe new_page = *hpage; 1550f3f0e1d2SKirill A. Shutemov 15516b24ca4aSMatthew Wilcox (Oracle) /* 15526b24ca4aSMatthew Wilcox (Oracle) * Ensure we have slots for all the pages in the range. This is 15536b24ca4aSMatthew Wilcox (Oracle) * almost certainly a no-op because most of the pages must be present 15546b24ca4aSMatthew Wilcox (Oracle) */ 155595feeabbSHugh Dickins do { 155695feeabbSHugh Dickins xas_lock_irq(&xas); 155795feeabbSHugh Dickins xas_create_range(&xas); 155895feeabbSHugh Dickins if (!xas_error(&xas)) 155995feeabbSHugh Dickins break; 156095feeabbSHugh Dickins xas_unlock_irq(&xas); 156195feeabbSHugh Dickins if (!xas_nomem(&xas, GFP_KERNEL)) { 156295feeabbSHugh Dickins result = SCAN_FAIL; 156395feeabbSHugh Dickins goto out; 156495feeabbSHugh Dickins } 156595feeabbSHugh Dickins } while (1); 156695feeabbSHugh Dickins 1567042a3082SHugh Dickins __SetPageLocked(new_page); 156899cb0dbdSSong Liu if (is_shmem) 1569042a3082SHugh Dickins __SetPageSwapBacked(new_page); 1570f3f0e1d2SKirill A. Shutemov new_page->index = start; 1571f3f0e1d2SKirill A. Shutemov new_page->mapping = mapping; 1572f3f0e1d2SKirill A. Shutemov 1573f3f0e1d2SKirill A. Shutemov /* 157487c460a0SHugh Dickins * At this point the new_page is locked and not up-to-date. 157587c460a0SHugh Dickins * It's safe to insert it into the page cache, because nobody would 157687c460a0SHugh Dickins * be able to map it or use it in another way until we unlock it. 1577f3f0e1d2SKirill A. Shutemov */ 1578f3f0e1d2SKirill A. Shutemov 157977da9389SMatthew Wilcox xas_set(&xas, start); 158077da9389SMatthew Wilcox for (index = start; index < end; index++) { 158177da9389SMatthew Wilcox struct page *page = xas_next(&xas); 158277da9389SMatthew Wilcox 158377da9389SMatthew Wilcox VM_BUG_ON(index != xas.xa_index); 158499cb0dbdSSong Liu if (is_shmem) { 158577da9389SMatthew Wilcox if (!page) { 1586701270faSHugh Dickins /* 158799cb0dbdSSong Liu * Stop if extent has been truncated or 158899cb0dbdSSong Liu * hole-punched, and is now completely 158999cb0dbdSSong Liu * empty. 1590701270faSHugh Dickins */ 1591701270faSHugh Dickins if (index == start) { 1592701270faSHugh Dickins if (!xas_next_entry(&xas, end - 1)) { 1593701270faSHugh Dickins result = SCAN_TRUNCATED; 1594042a3082SHugh Dickins goto xa_locked; 1595701270faSHugh Dickins } 1596701270faSHugh Dickins xas_set(&xas, index); 1597701270faSHugh Dickins } 159877da9389SMatthew Wilcox if (!shmem_charge(mapping->host, 1)) { 1599f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1600042a3082SHugh Dickins goto xa_locked; 1601f3f0e1d2SKirill A. Shutemov } 16024101196bSMatthew Wilcox (Oracle) xas_store(&xas, new_page); 160377da9389SMatthew Wilcox nr_none++; 160477da9389SMatthew Wilcox continue; 1605f3f0e1d2SKirill A. Shutemov } 1606f3f0e1d2SKirill A. Shutemov 16073159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 160877da9389SMatthew Wilcox xas_unlock_irq(&xas); 1609f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 1610f3f0e1d2SKirill A. Shutemov if (shmem_getpage(mapping->host, index, &page, 1611acdd9f8eSHugh Dickins SGP_NOALLOC)) { 1612f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 161377da9389SMatthew Wilcox goto xa_unlocked; 1614f3f0e1d2SKirill A. Shutemov } 1615f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1616f3f0e1d2SKirill A. Shutemov get_page(page); 1617042a3082SHugh Dickins xas_unlock_irq(&xas); 1618f3f0e1d2SKirill A. Shutemov } else { 1619f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1620042a3082SHugh Dickins goto xa_locked; 1621f3f0e1d2SKirill A. Shutemov } 162299cb0dbdSSong Liu } else { /* !is_shmem */ 162399cb0dbdSSong Liu if (!page || xa_is_value(page)) { 162499cb0dbdSSong Liu xas_unlock_irq(&xas); 162599cb0dbdSSong Liu page_cache_sync_readahead(mapping, &file->f_ra, 162699cb0dbdSSong Liu file, index, 1627e5a59d30SDavid Howells end - index); 162899cb0dbdSSong Liu /* drain pagevecs to help isolate_lru_page() */ 162999cb0dbdSSong Liu lru_add_drain(); 163099cb0dbdSSong Liu page = find_lock_page(mapping, index); 163199cb0dbdSSong Liu if (unlikely(page == NULL)) { 163299cb0dbdSSong Liu result = SCAN_FAIL; 163399cb0dbdSSong Liu goto xa_unlocked; 163499cb0dbdSSong Liu } 163575f36069SSong Liu } else if (PageDirty(page)) { 163675f36069SSong Liu /* 163775f36069SSong Liu * khugepaged only works on read-only fd, 163875f36069SSong Liu * so this page is dirty because it hasn't 163975f36069SSong Liu * been flushed since first write. There 164075f36069SSong Liu * won't be new dirty pages. 164175f36069SSong Liu * 164275f36069SSong Liu * Trigger async flush here and hope the 164375f36069SSong Liu * writeback is done when khugepaged 164475f36069SSong Liu * revisits this page. 164575f36069SSong Liu * 164675f36069SSong Liu * This is a one-off situation. We are not 164775f36069SSong Liu * forcing writeback in loop. 164875f36069SSong Liu */ 164975f36069SSong Liu xas_unlock_irq(&xas); 165075f36069SSong Liu filemap_flush(mapping); 165175f36069SSong Liu result = SCAN_FAIL; 165275f36069SSong Liu goto xa_unlocked; 165374c42e1bSRongwei Wang } else if (PageWriteback(page)) { 165474c42e1bSRongwei Wang xas_unlock_irq(&xas); 165574c42e1bSRongwei Wang result = SCAN_FAIL; 165674c42e1bSRongwei Wang goto xa_unlocked; 165799cb0dbdSSong Liu } else if (trylock_page(page)) { 165899cb0dbdSSong Liu get_page(page); 165999cb0dbdSSong Liu xas_unlock_irq(&xas); 166099cb0dbdSSong Liu } else { 166199cb0dbdSSong Liu result = SCAN_PAGE_LOCK; 166299cb0dbdSSong Liu goto xa_locked; 166399cb0dbdSSong Liu } 166499cb0dbdSSong Liu } 1665f3f0e1d2SKirill A. Shutemov 1666f3f0e1d2SKirill A. Shutemov /* 1667b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1668f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1669f3f0e1d2SKirill A. Shutemov */ 1670f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 16714655e5e5SSong Liu 16724655e5e5SSong Liu /* make sure the page is up to date */ 16734655e5e5SSong Liu if (unlikely(!PageUptodate(page))) { 16744655e5e5SSong Liu result = SCAN_FAIL; 16754655e5e5SSong Liu goto out_unlock; 16764655e5e5SSong Liu } 167706a5e126SHugh Dickins 167806a5e126SHugh Dickins /* 167906a5e126SHugh Dickins * If file was truncated then extended, or hole-punched, before 168006a5e126SHugh Dickins * we locked the first page, then a THP might be there already. 168106a5e126SHugh Dickins */ 168206a5e126SHugh Dickins if (PageTransCompound(page)) { 168306a5e126SHugh Dickins result = SCAN_PAGE_COMPOUND; 168406a5e126SHugh Dickins goto out_unlock; 168506a5e126SHugh Dickins } 1686f3f0e1d2SKirill A. Shutemov 1687f3f0e1d2SKirill A. Shutemov if (page_mapping(page) != mapping) { 1688f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1689f3f0e1d2SKirill A. Shutemov goto out_unlock; 1690f3f0e1d2SKirill A. Shutemov } 1691f3f0e1d2SKirill A. Shutemov 169274c42e1bSRongwei Wang if (!is_shmem && (PageDirty(page) || 169374c42e1bSRongwei Wang PageWriteback(page))) { 16944655e5e5SSong Liu /* 16954655e5e5SSong Liu * khugepaged only works on read-only fd, so this 16964655e5e5SSong Liu * page is dirty because it hasn't been flushed 16974655e5e5SSong Liu * since first write. 16984655e5e5SSong Liu */ 16994655e5e5SSong Liu result = SCAN_FAIL; 17004655e5e5SSong Liu goto out_unlock; 17014655e5e5SSong Liu } 17024655e5e5SSong Liu 1703f3f0e1d2SKirill A. Shutemov if (isolate_lru_page(page)) { 1704f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1705042a3082SHugh Dickins goto out_unlock; 1706f3f0e1d2SKirill A. Shutemov } 1707f3f0e1d2SKirill A. Shutemov 170899cb0dbdSSong Liu if (page_has_private(page) && 170999cb0dbdSSong Liu !try_to_release_page(page, GFP_KERNEL)) { 171099cb0dbdSSong Liu result = SCAN_PAGE_HAS_PRIVATE; 17112f33a706SHugh Dickins putback_lru_page(page); 171299cb0dbdSSong Liu goto out_unlock; 171399cb0dbdSSong Liu } 171499cb0dbdSSong Liu 1715f3f0e1d2SKirill A. Shutemov if (page_mapped(page)) 1716869f7ee6SMatthew Wilcox (Oracle) try_to_unmap(page_folio(page), 1717869f7ee6SMatthew Wilcox (Oracle) TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); 1718f3f0e1d2SKirill A. Shutemov 171977da9389SMatthew Wilcox xas_lock_irq(&xas); 172077da9389SMatthew Wilcox xas_set(&xas, index); 1721f3f0e1d2SKirill A. Shutemov 172277da9389SMatthew Wilcox VM_BUG_ON_PAGE(page != xas_load(&xas), page); 1723f3f0e1d2SKirill A. Shutemov 1724f3f0e1d2SKirill A. Shutemov /* 1725f3f0e1d2SKirill A. Shutemov * The page is expected to have page_count() == 3: 1726f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 172777da9389SMatthew Wilcox * - one reference from page cache; 1728f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1729f3f0e1d2SKirill A. Shutemov */ 1730f3f0e1d2SKirill A. Shutemov if (!page_ref_freeze(page, 3)) { 1731f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1732042a3082SHugh Dickins xas_unlock_irq(&xas); 1733042a3082SHugh Dickins putback_lru_page(page); 1734042a3082SHugh Dickins goto out_unlock; 1735f3f0e1d2SKirill A. Shutemov } 1736f3f0e1d2SKirill A. Shutemov 1737f3f0e1d2SKirill A. Shutemov /* 1738f3f0e1d2SKirill A. Shutemov * Add the page to the list to be able to undo the collapse if 1739f3f0e1d2SKirill A. Shutemov * something go wrong. 1740f3f0e1d2SKirill A. Shutemov */ 1741f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 1742f3f0e1d2SKirill A. Shutemov 1743f3f0e1d2SKirill A. Shutemov /* Finally, replace with the new page. */ 17444101196bSMatthew Wilcox (Oracle) xas_store(&xas, new_page); 1745f3f0e1d2SKirill A. Shutemov continue; 1746f3f0e1d2SKirill A. Shutemov out_unlock: 1747f3f0e1d2SKirill A. Shutemov unlock_page(page); 1748f3f0e1d2SKirill A. Shutemov put_page(page); 1749042a3082SHugh Dickins goto xa_unlocked; 1750f3f0e1d2SKirill A. Shutemov } 1751bf9eceadSMuchun Song nr = thp_nr_pages(new_page); 1752f3f0e1d2SKirill A. Shutemov 175399cb0dbdSSong Liu if (is_shmem) 175457b2847dSMuchun Song __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr); 175509d91cdaSSong Liu else { 1756bf9eceadSMuchun Song __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr); 175709d91cdaSSong Liu filemap_nr_thps_inc(mapping); 1758eb6ecbedSCollin Fijalkovich /* 1759eb6ecbedSCollin Fijalkovich * Paired with smp_mb() in do_dentry_open() to ensure 1760eb6ecbedSCollin Fijalkovich * i_writecount is up to date and the update to nr_thps is 1761eb6ecbedSCollin Fijalkovich * visible. Ensures the page cache will be truncated if the 1762eb6ecbedSCollin Fijalkovich * file is opened writable. 1763eb6ecbedSCollin Fijalkovich */ 1764eb6ecbedSCollin Fijalkovich smp_mb(); 1765eb6ecbedSCollin Fijalkovich if (inode_is_open_for_write(mapping->host)) { 1766eb6ecbedSCollin Fijalkovich result = SCAN_FAIL; 1767eb6ecbedSCollin Fijalkovich __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr); 1768eb6ecbedSCollin Fijalkovich filemap_nr_thps_dec(mapping); 1769eb6ecbedSCollin Fijalkovich goto xa_locked; 1770eb6ecbedSCollin Fijalkovich } 177109d91cdaSSong Liu } 177299cb0dbdSSong Liu 1773042a3082SHugh Dickins if (nr_none) { 17749d82c694SJohannes Weiner __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none); 17752f55f070SMiaohe Lin /* nr_none is always 0 for non-shmem. */ 17769d82c694SJohannes Weiner __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none); 1777042a3082SHugh Dickins } 1778042a3082SHugh Dickins 17796b24ca4aSMatthew Wilcox (Oracle) /* Join all the small entries into a single multi-index entry */ 17806b24ca4aSMatthew Wilcox (Oracle) xas_set_order(&xas, start, HPAGE_PMD_ORDER); 17816b24ca4aSMatthew Wilcox (Oracle) xas_store(&xas, new_page); 1782042a3082SHugh Dickins xa_locked: 1783042a3082SHugh Dickins xas_unlock_irq(&xas); 178477da9389SMatthew Wilcox xa_unlocked: 1785042a3082SHugh Dickins 17866d9df8a5SHugh Dickins /* 17876d9df8a5SHugh Dickins * If collapse is successful, flush must be done now before copying. 17886d9df8a5SHugh Dickins * If collapse is unsuccessful, does flush actually need to be done? 17896d9df8a5SHugh Dickins * Do it anyway, to clear the state. 17906d9df8a5SHugh Dickins */ 17916d9df8a5SHugh Dickins try_to_unmap_flush(); 17926d9df8a5SHugh Dickins 1793f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 179477da9389SMatthew Wilcox struct page *page, *tmp; 1795f3f0e1d2SKirill A. Shutemov 1796f3f0e1d2SKirill A. Shutemov /* 179777da9389SMatthew Wilcox * Replacing old pages with new one has succeeded, now we 179877da9389SMatthew Wilcox * need to copy the content and free the old pages. 1799f3f0e1d2SKirill A. Shutemov */ 18002af8ff29SHugh Dickins index = start; 1801f3f0e1d2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, &pagelist, lru) { 18022af8ff29SHugh Dickins while (index < page->index) { 18032af8ff29SHugh Dickins clear_highpage(new_page + (index % HPAGE_PMD_NR)); 18042af8ff29SHugh Dickins index++; 18052af8ff29SHugh Dickins } 1806f3f0e1d2SKirill A. Shutemov copy_highpage(new_page + (page->index % HPAGE_PMD_NR), 1807f3f0e1d2SKirill A. Shutemov page); 1808f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1809f3f0e1d2SKirill A. Shutemov page->mapping = NULL; 1810042a3082SHugh Dickins page_ref_unfreeze(page, 1); 1811f3f0e1d2SKirill A. Shutemov ClearPageActive(page); 1812f3f0e1d2SKirill A. Shutemov ClearPageUnevictable(page); 1813042a3082SHugh Dickins unlock_page(page); 1814f3f0e1d2SKirill A. Shutemov put_page(page); 18152af8ff29SHugh Dickins index++; 18162af8ff29SHugh Dickins } 18172af8ff29SHugh Dickins while (index < end) { 18182af8ff29SHugh Dickins clear_highpage(new_page + (index % HPAGE_PMD_NR)); 18192af8ff29SHugh Dickins index++; 1820f3f0e1d2SKirill A. Shutemov } 1821f3f0e1d2SKirill A. Shutemov 1822f3f0e1d2SKirill A. Shutemov SetPageUptodate(new_page); 182387c460a0SHugh Dickins page_ref_add(new_page, HPAGE_PMD_NR - 1); 18246058eaecSJohannes Weiner if (is_shmem) 182599cb0dbdSSong Liu set_page_dirty(new_page); 18266058eaecSJohannes Weiner lru_cache_add(new_page); 1827f3f0e1d2SKirill A. Shutemov 1828042a3082SHugh Dickins /* 1829042a3082SHugh Dickins * Remove pte page tables, so we can re-fault the page as huge. 1830042a3082SHugh Dickins */ 1831042a3082SHugh Dickins retract_page_tables(mapping, start); 1832f3f0e1d2SKirill A. Shutemov *hpage = NULL; 183387aa7529SYang Shi 183487aa7529SYang Shi khugepaged_pages_collapsed++; 1835f3f0e1d2SKirill A. Shutemov } else { 183677da9389SMatthew Wilcox struct page *page; 1837aaa52e34SHugh Dickins 183877da9389SMatthew Wilcox /* Something went wrong: roll back page cache changes */ 183977da9389SMatthew Wilcox xas_lock_irq(&xas); 18402f55f070SMiaohe Lin if (nr_none) { 1841aaa52e34SHugh Dickins mapping->nrpages -= nr_none; 1842aaa52e34SHugh Dickins shmem_uncharge(mapping->host, nr_none); 18432f55f070SMiaohe Lin } 1844aaa52e34SHugh Dickins 184577da9389SMatthew Wilcox xas_set(&xas, start); 184677da9389SMatthew Wilcox xas_for_each(&xas, page, end - 1) { 1847f3f0e1d2SKirill A. Shutemov page = list_first_entry_or_null(&pagelist, 1848f3f0e1d2SKirill A. Shutemov struct page, lru); 184977da9389SMatthew Wilcox if (!page || xas.xa_index < page->index) { 1850f3f0e1d2SKirill A. Shutemov if (!nr_none) 1851f3f0e1d2SKirill A. Shutemov break; 1852f3f0e1d2SKirill A. Shutemov nr_none--; 185359749e6cSJohannes Weiner /* Put holes back where they were */ 185477da9389SMatthew Wilcox xas_store(&xas, NULL); 1855f3f0e1d2SKirill A. Shutemov continue; 1856f3f0e1d2SKirill A. Shutemov } 1857f3f0e1d2SKirill A. Shutemov 185877da9389SMatthew Wilcox VM_BUG_ON_PAGE(page->index != xas.xa_index, page); 1859f3f0e1d2SKirill A. Shutemov 1860f3f0e1d2SKirill A. Shutemov /* Unfreeze the page. */ 1861f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1862f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(page, 2); 186377da9389SMatthew Wilcox xas_store(&xas, page); 186477da9389SMatthew Wilcox xas_pause(&xas); 186577da9389SMatthew Wilcox xas_unlock_irq(&xas); 1866f3f0e1d2SKirill A. Shutemov unlock_page(page); 1867042a3082SHugh Dickins putback_lru_page(page); 186877da9389SMatthew Wilcox xas_lock_irq(&xas); 1869f3f0e1d2SKirill A. Shutemov } 1870f3f0e1d2SKirill A. Shutemov VM_BUG_ON(nr_none); 187177da9389SMatthew Wilcox xas_unlock_irq(&xas); 1872f3f0e1d2SKirill A. Shutemov 1873f3f0e1d2SKirill A. Shutemov new_page->mapping = NULL; 1874f3f0e1d2SKirill A. Shutemov } 1875042a3082SHugh Dickins 1876042a3082SHugh Dickins unlock_page(new_page); 1877f3f0e1d2SKirill A. Shutemov out: 1878f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 1879c6a7f445SYang Shi if (!IS_ERR_OR_NULL(*hpage)) { 1880bbc6b703SMatthew Wilcox (Oracle) mem_cgroup_uncharge(page_folio(*hpage)); 1881c6a7f445SYang Shi put_page(*hpage); 1882c6a7f445SYang Shi } 1883f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 1884f3f0e1d2SKirill A. Shutemov } 1885f3f0e1d2SKirill A. Shutemov 188634d6b470SZach O'Keefe static void khugepaged_scan_file(struct mm_struct *mm, struct file *file, 188734d6b470SZach O'Keefe pgoff_t start, struct page **hpage, 188834d6b470SZach O'Keefe struct collapse_control *cc) 1889f3f0e1d2SKirill A. Shutemov { 1890f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 1891579c571eSSong Liu struct address_space *mapping = file->f_mapping; 189285b392dbSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, start); 1893f3f0e1d2SKirill A. Shutemov int present, swap; 1894f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 1895f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 1896f3f0e1d2SKirill A. Shutemov 1897f3f0e1d2SKirill A. Shutemov present = 0; 1898f3f0e1d2SKirill A. Shutemov swap = 0; 189934d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1900f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 190185b392dbSMatthew Wilcox xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 190285b392dbSMatthew Wilcox if (xas_retry(&xas, page)) 1903f3f0e1d2SKirill A. Shutemov continue; 1904f3f0e1d2SKirill A. Shutemov 190585b392dbSMatthew Wilcox if (xa_is_value(page)) { 1906f3f0e1d2SKirill A. Shutemov if (++swap > khugepaged_max_ptes_swap) { 1907f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1908e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1909f3f0e1d2SKirill A. Shutemov break; 1910f3f0e1d2SKirill A. Shutemov } 1911f3f0e1d2SKirill A. Shutemov continue; 1912f3f0e1d2SKirill A. Shutemov } 1913f3f0e1d2SKirill A. Shutemov 19146b24ca4aSMatthew Wilcox (Oracle) /* 19156b24ca4aSMatthew Wilcox (Oracle) * XXX: khugepaged should compact smaller compound pages 19166b24ca4aSMatthew Wilcox (Oracle) * into a PMD sized page 19176b24ca4aSMatthew Wilcox (Oracle) */ 1918f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 1919f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COMPOUND; 1920f3f0e1d2SKirill A. Shutemov break; 1921f3f0e1d2SKirill A. Shutemov } 1922f3f0e1d2SKirill A. Shutemov 1923f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 192434d6b470SZach O'Keefe if (khugepaged_scan_abort(node, cc)) { 1925f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 1926f3f0e1d2SKirill A. Shutemov break; 1927f3f0e1d2SKirill A. Shutemov } 192834d6b470SZach O'Keefe cc->node_load[node]++; 1929f3f0e1d2SKirill A. Shutemov 1930f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 1931f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 1932f3f0e1d2SKirill A. Shutemov break; 1933f3f0e1d2SKirill A. Shutemov } 1934f3f0e1d2SKirill A. Shutemov 193599cb0dbdSSong Liu if (page_count(page) != 193699cb0dbdSSong Liu 1 + page_mapcount(page) + page_has_private(page)) { 1937f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1938f3f0e1d2SKirill A. Shutemov break; 1939f3f0e1d2SKirill A. Shutemov } 1940f3f0e1d2SKirill A. Shutemov 1941f3f0e1d2SKirill A. Shutemov /* 1942f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 1943f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 1944f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 1945f3f0e1d2SKirill A. Shutemov */ 1946f3f0e1d2SKirill A. Shutemov 1947f3f0e1d2SKirill A. Shutemov present++; 1948f3f0e1d2SKirill A. Shutemov 1949f3f0e1d2SKirill A. Shutemov if (need_resched()) { 195085b392dbSMatthew Wilcox xas_pause(&xas); 1951f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 1952f3f0e1d2SKirill A. Shutemov } 1953f3f0e1d2SKirill A. Shutemov } 1954f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 1955f3f0e1d2SKirill A. Shutemov 1956f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 1957f3f0e1d2SKirill A. Shutemov if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 1958f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1959e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1960f3f0e1d2SKirill A. Shutemov } else { 1961*9710a78aSZach O'Keefe collapse_file(mm, file, start, hpage, cc); 1962f3f0e1d2SKirill A. Shutemov } 1963f3f0e1d2SKirill A. Shutemov } 1964f3f0e1d2SKirill A. Shutemov 1965f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 1966f3f0e1d2SKirill A. Shutemov } 1967f3f0e1d2SKirill A. Shutemov #else 196834d6b470SZach O'Keefe static void khugepaged_scan_file(struct mm_struct *mm, struct file *file, 196934d6b470SZach O'Keefe pgoff_t start, struct page **hpage, 197034d6b470SZach O'Keefe struct collapse_control *cc) 1971f3f0e1d2SKirill A. Shutemov { 1972f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 1973f3f0e1d2SKirill A. Shutemov } 197427e1f827SSong Liu 19750edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 197627e1f827SSong Liu { 197727e1f827SSong Liu } 1978f3f0e1d2SKirill A. Shutemov #endif 1979f3f0e1d2SKirill A. Shutemov 1980b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 198134d6b470SZach O'Keefe struct page **hpage, 198234d6b470SZach O'Keefe struct collapse_control *cc) 1983b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 1984b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 1985b46e756fSKirill A. Shutemov { 1986b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 1987b46e756fSKirill A. Shutemov struct mm_struct *mm; 1988b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 1989b46e756fSKirill A. Shutemov int progress = 0; 1990b46e756fSKirill A. Shutemov 1991b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 199235f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1993b46e756fSKirill A. Shutemov 1994b46e756fSKirill A. Shutemov if (khugepaged_scan.mm_slot) 1995b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 1996b46e756fSKirill A. Shutemov else { 1997b46e756fSKirill A. Shutemov mm_slot = list_entry(khugepaged_scan.mm_head.next, 1998b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 1999b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2000b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 2001b46e756fSKirill A. Shutemov } 2002b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 200327e1f827SSong Liu khugepaged_collapse_pte_mapped_thps(mm_slot); 2004b46e756fSKirill A. Shutemov 2005b46e756fSKirill A. Shutemov mm = mm_slot->mm; 20063b454ad3SYang Shi /* 20073b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 20083b454ad3SYang Shi * the next mm on the list. 20093b454ad3SYang Shi */ 2010b46e756fSKirill A. Shutemov vma = NULL; 2011d8ed45c5SMichel Lespinasse if (unlikely(!mmap_read_trylock(mm))) 2012c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 20133b454ad3SYang Shi if (likely(!khugepaged_test_exit(mm))) 2014b46e756fSKirill A. Shutemov vma = find_vma(mm, khugepaged_scan.address); 2015b46e756fSKirill A. Shutemov 2016b46e756fSKirill A. Shutemov progress++; 2017b46e756fSKirill A. Shutemov for (; vma; vma = vma->vm_next) { 2018b46e756fSKirill A. Shutemov unsigned long hstart, hend; 2019b46e756fSKirill A. Shutemov 2020b46e756fSKirill A. Shutemov cond_resched(); 2021b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) { 2022b46e756fSKirill A. Shutemov progress++; 2023b46e756fSKirill A. Shutemov break; 2024b46e756fSKirill A. Shutemov } 20257da4e2cbSYang Shi if (!hugepage_vma_check(vma, vma->vm_flags, false, false)) { 2026b46e756fSKirill A. Shutemov skip: 2027b46e756fSKirill A. Shutemov progress++; 2028b46e756fSKirill A. Shutemov continue; 2029b46e756fSKirill A. Shutemov } 20304fa6893fSYang Shi hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); 20314fa6893fSYang Shi hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); 2032b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 2033b46e756fSKirill A. Shutemov goto skip; 2034b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 2035b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 2036b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2037b46e756fSKirill A. Shutemov 2038b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 2039b46e756fSKirill A. Shutemov int ret; 2040b46e756fSKirill A. Shutemov cond_resched(); 2041b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 2042b46e756fSKirill A. Shutemov goto breakouterloop; 2043b46e756fSKirill A. Shutemov 2044b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 2045b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 2046b46e756fSKirill A. Shutemov hend); 204799cb0dbdSSong Liu if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 2048396bcc52SMatthew Wilcox (Oracle) struct file *file = get_file(vma->vm_file); 2049f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 2050f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 205199cb0dbdSSong Liu 2052d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 2053f3f0e1d2SKirill A. Shutemov ret = 1; 205434d6b470SZach O'Keefe khugepaged_scan_file(mm, file, pgoff, hpage, 205534d6b470SZach O'Keefe cc); 2056f3f0e1d2SKirill A. Shutemov fput(file); 2057f3f0e1d2SKirill A. Shutemov } else { 2058b46e756fSKirill A. Shutemov ret = khugepaged_scan_pmd(mm, vma, 2059b46e756fSKirill A. Shutemov khugepaged_scan.address, 206034d6b470SZach O'Keefe hpage, cc); 2061f3f0e1d2SKirill A. Shutemov } 2062b46e756fSKirill A. Shutemov /* move to next address */ 2063b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 2064b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 2065b46e756fSKirill A. Shutemov if (ret) 2066c1e8d7c6SMichel Lespinasse /* we released mmap_lock so break loop */ 2067c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2068b46e756fSKirill A. Shutemov if (progress >= pages) 2069b46e756fSKirill A. Shutemov goto breakouterloop; 2070b46e756fSKirill A. Shutemov } 2071b46e756fSKirill A. Shutemov } 2072b46e756fSKirill A. Shutemov breakouterloop: 2073d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ 2074c1e8d7c6SMichel Lespinasse breakouterloop_mmap_lock: 2075b46e756fSKirill A. Shutemov 2076b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2077b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2078b46e756fSKirill A. Shutemov /* 2079b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 2080b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 2081b46e756fSKirill A. Shutemov */ 2082b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm) || !vma) { 2083b46e756fSKirill A. Shutemov /* 2084b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 2085b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 2086b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 2087b46e756fSKirill A. Shutemov */ 2088b46e756fSKirill A. Shutemov if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2089b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = list_entry( 2090b46e756fSKirill A. Shutemov mm_slot->mm_node.next, 2091b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2092b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2093b46e756fSKirill A. Shutemov } else { 2094b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2095b46e756fSKirill A. Shutemov khugepaged_full_scans++; 2096b46e756fSKirill A. Shutemov } 2097b46e756fSKirill A. Shutemov 2098b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2099b46e756fSKirill A. Shutemov } 2100b46e756fSKirill A. Shutemov 2101b46e756fSKirill A. Shutemov return progress; 2102b46e756fSKirill A. Shutemov } 2103b46e756fSKirill A. Shutemov 2104b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 2105b46e756fSKirill A. Shutemov { 2106b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 21071064026bSYang Shi hugepage_flags_enabled(); 2108b46e756fSKirill A. Shutemov } 2109b46e756fSKirill A. Shutemov 2110b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 2111b46e756fSKirill A. Shutemov { 2112b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 2113b46e756fSKirill A. Shutemov kthread_should_stop(); 2114b46e756fSKirill A. Shutemov } 2115b46e756fSKirill A. Shutemov 211634d6b470SZach O'Keefe static void khugepaged_do_scan(struct collapse_control *cc) 2117b46e756fSKirill A. Shutemov { 2118b46e756fSKirill A. Shutemov struct page *hpage = NULL; 2119b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 212089dc6a96SYanfei Xu unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); 2121b46e756fSKirill A. Shutemov bool wait = true; 2122b46e756fSKirill A. Shutemov 2123a980df33SKirill A. Shutemov lru_add_drain_all(); 2124a980df33SKirill A. Shutemov 2125c6a7f445SYang Shi while (true) { 2126b46e756fSKirill A. Shutemov cond_resched(); 2127b46e756fSKirill A. Shutemov 2128b46e756fSKirill A. Shutemov if (unlikely(kthread_should_stop() || try_to_freeze())) 2129b46e756fSKirill A. Shutemov break; 2130b46e756fSKirill A. Shutemov 2131b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2132b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 2133b46e756fSKirill A. Shutemov pass_through_head++; 2134b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 2135b46e756fSKirill A. Shutemov pass_through_head < 2) 2136b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 213734d6b470SZach O'Keefe &hpage, cc); 2138b46e756fSKirill A. Shutemov else 2139b46e756fSKirill A. Shutemov progress = pages; 2140b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2141b46e756fSKirill A. Shutemov 2142c6a7f445SYang Shi if (progress >= pages) 2143c6a7f445SYang Shi break; 2144c6a7f445SYang Shi 2145c6a7f445SYang Shi if (IS_ERR(hpage)) { 2146c6a7f445SYang Shi /* 2147c6a7f445SYang Shi * If fail to allocate the first time, try to sleep for 2148c6a7f445SYang Shi * a while. When hit again, cancel the scan. 2149c6a7f445SYang Shi */ 2150c6a7f445SYang Shi if (!wait) 2151c6a7f445SYang Shi break; 2152c6a7f445SYang Shi wait = false; 2153c6a7f445SYang Shi hpage = NULL; 2154c6a7f445SYang Shi khugepaged_alloc_sleep(); 2155c6a7f445SYang Shi } 2156c6a7f445SYang Shi } 2157b46e756fSKirill A. Shutemov } 2158b46e756fSKirill A. Shutemov 2159b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 2160b46e756fSKirill A. Shutemov { 2161b46e756fSKirill A. Shutemov return kthread_should_stop() || 2162b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 2163b46e756fSKirill A. Shutemov } 2164b46e756fSKirill A. Shutemov 2165b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 2166b46e756fSKirill A. Shutemov { 2167b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 2168b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 2169b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2170b46e756fSKirill A. Shutemov 2171b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 2172b46e756fSKirill A. Shutemov return; 2173b46e756fSKirill A. Shutemov 2174b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2175b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 2176b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 2177b46e756fSKirill A. Shutemov scan_sleep_jiffies); 2178b46e756fSKirill A. Shutemov return; 2179b46e756fSKirill A. Shutemov } 2180b46e756fSKirill A. Shutemov 21811064026bSYang Shi if (hugepage_flags_enabled()) 2182b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2183b46e756fSKirill A. Shutemov } 2184b46e756fSKirill A. Shutemov 2185b46e756fSKirill A. Shutemov static int khugepaged(void *none) 2186b46e756fSKirill A. Shutemov { 2187b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 2188b46e756fSKirill A. Shutemov 2189b46e756fSKirill A. Shutemov set_freezable(); 2190b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 2191b46e756fSKirill A. Shutemov 2192b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 219334d6b470SZach O'Keefe khugepaged_do_scan(&khugepaged_collapse_control); 2194b46e756fSKirill A. Shutemov khugepaged_wait_work(); 2195b46e756fSKirill A. Shutemov } 2196b46e756fSKirill A. Shutemov 2197b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2198b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2199b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2200b46e756fSKirill A. Shutemov if (mm_slot) 2201b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2202b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2203b46e756fSKirill A. Shutemov return 0; 2204b46e756fSKirill A. Shutemov } 2205b46e756fSKirill A. Shutemov 2206b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 2207b46e756fSKirill A. Shutemov { 2208b46e756fSKirill A. Shutemov struct zone *zone; 2209b46e756fSKirill A. Shutemov int nr_zones = 0; 2210b46e756fSKirill A. Shutemov unsigned long recommended_min; 2211b46e756fSKirill A. Shutemov 22121064026bSYang Shi if (!hugepage_flags_enabled()) { 2213bd3400eaSLiangcai Fan calculate_min_free_kbytes(); 2214bd3400eaSLiangcai Fan goto update_wmarks; 2215bd3400eaSLiangcai Fan } 2216bd3400eaSLiangcai Fan 2217b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 2218b7d349c7SJoonsoo Kim /* 2219b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 2220b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 2221b7d349c7SJoonsoo Kim */ 2222b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 2223b7d349c7SJoonsoo Kim continue; 2224b7d349c7SJoonsoo Kim 2225b46e756fSKirill A. Shutemov nr_zones++; 2226b7d349c7SJoonsoo Kim } 2227b46e756fSKirill A. Shutemov 2228b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2229b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 2230b46e756fSKirill A. Shutemov 2231b46e756fSKirill A. Shutemov /* 2232b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 2233b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 2234b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 2235b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 2236b46e756fSKirill A. Shutemov */ 2237b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 2238b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2239b46e756fSKirill A. Shutemov 2240b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 2241b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 2242b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 2243b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 2244b46e756fSKirill A. Shutemov 2245b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 2246b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 2247b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2248b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 2249b46e756fSKirill A. Shutemov 2250b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 2251b46e756fSKirill A. Shutemov } 2252bd3400eaSLiangcai Fan 2253bd3400eaSLiangcai Fan update_wmarks: 2254b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 2255b46e756fSKirill A. Shutemov } 2256b46e756fSKirill A. Shutemov 2257b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 2258b46e756fSKirill A. Shutemov { 2259b46e756fSKirill A. Shutemov int err = 0; 2260b46e756fSKirill A. Shutemov 2261b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 22621064026bSYang Shi if (hugepage_flags_enabled()) { 2263b46e756fSKirill A. Shutemov if (!khugepaged_thread) 2264b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 2265b46e756fSKirill A. Shutemov "khugepaged"); 2266b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 2267b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2268b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 2269b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2270b46e756fSKirill A. Shutemov goto fail; 2271b46e756fSKirill A. Shutemov } 2272b46e756fSKirill A. Shutemov 2273b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 2274b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 2275b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 2276b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 2277b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2278b46e756fSKirill A. Shutemov } 2279bd3400eaSLiangcai Fan set_recommended_min_free_kbytes(); 2280b46e756fSKirill A. Shutemov fail: 2281b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 2282b46e756fSKirill A. Shutemov return err; 2283b46e756fSKirill A. Shutemov } 22844aab2be0SVijay Balakrishna 22854aab2be0SVijay Balakrishna void khugepaged_min_free_kbytes_update(void) 22864aab2be0SVijay Balakrishna { 22874aab2be0SVijay Balakrishna mutex_lock(&khugepaged_mutex); 22881064026bSYang Shi if (hugepage_flags_enabled() && khugepaged_thread) 22894aab2be0SVijay Balakrishna set_recommended_min_free_kbytes(); 22904aab2be0SVijay Balakrishna mutex_unlock(&khugepaged_mutex); 22914aab2be0SVijay Balakrishna } 2292