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" 26b26e2701SQi Zheng #include "mm_slot.h" 27b46e756fSKirill A. Shutemov 28b46e756fSKirill A. Shutemov enum scan_result { 29b46e756fSKirill A. Shutemov SCAN_FAIL, 30b46e756fSKirill A. Shutemov SCAN_SUCCEED, 31b46e756fSKirill A. Shutemov SCAN_PMD_NULL, 3234488399SZach O'Keefe SCAN_PMD_NONE, 3350722804SZach O'Keefe SCAN_PMD_MAPPED, 34b46e756fSKirill A. Shutemov SCAN_EXCEED_NONE_PTE, 3571a2c112SKirill A. Shutemov SCAN_EXCEED_SWAP_PTE, 3671a2c112SKirill A. Shutemov SCAN_EXCEED_SHARED_PTE, 37b46e756fSKirill A. Shutemov SCAN_PTE_NON_PRESENT, 38e1e267c7SPeter Xu SCAN_PTE_UFFD_WP, 3958ac9a89SZach O'Keefe SCAN_PTE_MAPPED_HUGEPAGE, 40b46e756fSKirill A. Shutemov SCAN_PAGE_RO, 410db501f7SEbru Akagunduz SCAN_LACK_REFERENCED_PAGE, 42b46e756fSKirill A. Shutemov SCAN_PAGE_NULL, 43b46e756fSKirill A. Shutemov SCAN_SCAN_ABORT, 44b46e756fSKirill A. Shutemov SCAN_PAGE_COUNT, 45b46e756fSKirill A. Shutemov SCAN_PAGE_LRU, 46b46e756fSKirill A. Shutemov SCAN_PAGE_LOCK, 47b46e756fSKirill A. Shutemov SCAN_PAGE_ANON, 48b46e756fSKirill A. Shutemov SCAN_PAGE_COMPOUND, 49b46e756fSKirill A. Shutemov SCAN_ANY_PROCESS, 50b46e756fSKirill A. Shutemov SCAN_VMA_NULL, 51b46e756fSKirill A. Shutemov SCAN_VMA_CHECK, 52b46e756fSKirill A. Shutemov SCAN_ADDRESS_RANGE, 53b46e756fSKirill A. Shutemov SCAN_DEL_PAGE_LRU, 54b46e756fSKirill A. Shutemov SCAN_ALLOC_HUGE_PAGE_FAIL, 55b46e756fSKirill A. Shutemov SCAN_CGROUP_CHARGE_FAIL, 56f3f0e1d2SKirill A. Shutemov SCAN_TRUNCATED, 5799cb0dbdSSong Liu SCAN_PAGE_HAS_PRIVATE, 58b46e756fSKirill A. Shutemov }; 59b46e756fSKirill A. Shutemov 60b46e756fSKirill A. Shutemov #define CREATE_TRACE_POINTS 61b46e756fSKirill A. Shutemov #include <trace/events/huge_memory.h> 62b46e756fSKirill A. Shutemov 634aab2be0SVijay Balakrishna static struct task_struct *khugepaged_thread __read_mostly; 644aab2be0SVijay Balakrishna static DEFINE_MUTEX(khugepaged_mutex); 654aab2be0SVijay Balakrishna 66b46e756fSKirill A. Shutemov /* default scan 8*512 pte (or vmas) every 30 second */ 67b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_to_scan __read_mostly; 68b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_collapsed; 69b46e756fSKirill A. Shutemov static unsigned int khugepaged_full_scans; 70b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 71b46e756fSKirill A. Shutemov /* during fragmentation poll the hugepage allocator once every minute */ 72b46e756fSKirill A. Shutemov static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 73b46e756fSKirill A. Shutemov static unsigned long khugepaged_sleep_expire; 74b46e756fSKirill A. Shutemov static DEFINE_SPINLOCK(khugepaged_mm_lock); 75b46e756fSKirill A. Shutemov static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 76b46e756fSKirill A. Shutemov /* 77b46e756fSKirill A. Shutemov * default collapse hugepages if there is at least one pte mapped like 78b46e756fSKirill A. Shutemov * it would have happened if the vma was large enough during page 79b46e756fSKirill A. Shutemov * fault. 80d8ea7cc8SZach O'Keefe * 81d8ea7cc8SZach O'Keefe * Note that these are only respected if collapse was initiated by khugepaged. 82b46e756fSKirill A. Shutemov */ 83b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_none __read_mostly; 84b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_swap __read_mostly; 8571a2c112SKirill A. Shutemov static unsigned int khugepaged_max_ptes_shared __read_mostly; 86b46e756fSKirill A. Shutemov 87b46e756fSKirill A. Shutemov #define MM_SLOTS_HASH_BITS 10 88b46e756fSKirill A. Shutemov static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 89b46e756fSKirill A. Shutemov 90b46e756fSKirill A. Shutemov static struct kmem_cache *mm_slot_cache __read_mostly; 91b46e756fSKirill A. Shutemov 9227e1f827SSong Liu #define MAX_PTE_MAPPED_THP 8 9327e1f827SSong Liu 9434d6b470SZach O'Keefe struct collapse_control { 95d8ea7cc8SZach O'Keefe bool is_khugepaged; 96d8ea7cc8SZach O'Keefe 9734d6b470SZach O'Keefe /* Num pages scanned per node */ 9834d6b470SZach O'Keefe u32 node_load[MAX_NUMNODES]; 9934d6b470SZach O'Keefe 100e031ff96SYang Shi /* nodemask for allocation fallback */ 101e031ff96SYang Shi nodemask_t alloc_nmask; 10234d6b470SZach O'Keefe }; 10334d6b470SZach O'Keefe 104b46e756fSKirill A. Shutemov /** 105b26e2701SQi Zheng * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned 106b26e2701SQi Zheng * @slot: hash lookup from mm to mm_slot 107336e6b53SAlex Shi * @nr_pte_mapped_thp: number of pte mapped THP 108336e6b53SAlex Shi * @pte_mapped_thp: address array corresponding pte mapped THP 109b46e756fSKirill A. Shutemov */ 110b26e2701SQi Zheng struct khugepaged_mm_slot { 111b26e2701SQi Zheng struct mm_slot slot; 11227e1f827SSong Liu 11327e1f827SSong Liu /* pte-mapped THP in this mm */ 11427e1f827SSong Liu int nr_pte_mapped_thp; 11527e1f827SSong Liu unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; 116b46e756fSKirill A. Shutemov }; 117b46e756fSKirill A. Shutemov 118b46e756fSKirill A. Shutemov /** 119b46e756fSKirill A. Shutemov * struct khugepaged_scan - cursor for scanning 120b46e756fSKirill A. Shutemov * @mm_head: the head of the mm list to scan 121b46e756fSKirill A. Shutemov * @mm_slot: the current mm_slot we are scanning 122b46e756fSKirill A. Shutemov * @address: the next address inside that to be scanned 123b46e756fSKirill A. Shutemov * 124b46e756fSKirill A. Shutemov * There is only the one khugepaged_scan instance of this cursor structure. 125b46e756fSKirill A. Shutemov */ 126b46e756fSKirill A. Shutemov struct khugepaged_scan { 127b46e756fSKirill A. Shutemov struct list_head mm_head; 128b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 129b46e756fSKirill A. Shutemov unsigned long address; 130b46e756fSKirill A. Shutemov }; 131b46e756fSKirill A. Shutemov 132b46e756fSKirill A. Shutemov static struct khugepaged_scan khugepaged_scan = { 133b46e756fSKirill A. Shutemov .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 134b46e756fSKirill A. Shutemov }; 135b46e756fSKirill A. Shutemov 136e1465d12SJérémy Lefaure #ifdef CONFIG_SYSFS 137b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 138b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 139b46e756fSKirill A. Shutemov char *buf) 140b46e756fSKirill A. Shutemov { 141ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); 142b46e756fSKirill A. Shutemov } 143b46e756fSKirill A. Shutemov 144b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 145b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 146b46e756fSKirill A. Shutemov const char *buf, size_t count) 147b46e756fSKirill A. Shutemov { 148dfefd226SAlexey Dobriyan unsigned int msecs; 149b46e756fSKirill A. Shutemov int err; 150b46e756fSKirill A. Shutemov 151dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 152dfefd226SAlexey Dobriyan if (err) 153b46e756fSKirill A. Shutemov return -EINVAL; 154b46e756fSKirill A. Shutemov 155b46e756fSKirill A. Shutemov khugepaged_scan_sleep_millisecs = msecs; 156b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 157b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 158b46e756fSKirill A. Shutemov 159b46e756fSKirill A. Shutemov return count; 160b46e756fSKirill A. Shutemov } 161b46e756fSKirill A. Shutemov static struct kobj_attribute scan_sleep_millisecs_attr = 1626dcdc94dSMiaohe Lin __ATTR_RW(scan_sleep_millisecs); 163b46e756fSKirill A. Shutemov 164b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 165b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 166b46e756fSKirill A. Shutemov char *buf) 167b46e756fSKirill A. Shutemov { 168ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 169b46e756fSKirill A. Shutemov } 170b46e756fSKirill A. Shutemov 171b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 172b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 173b46e756fSKirill A. Shutemov const char *buf, size_t count) 174b46e756fSKirill A. Shutemov { 175dfefd226SAlexey Dobriyan unsigned int msecs; 176b46e756fSKirill A. Shutemov int err; 177b46e756fSKirill A. Shutemov 178dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 179dfefd226SAlexey Dobriyan if (err) 180b46e756fSKirill A. Shutemov return -EINVAL; 181b46e756fSKirill A. Shutemov 182b46e756fSKirill A. Shutemov khugepaged_alloc_sleep_millisecs = msecs; 183b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 184b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 185b46e756fSKirill A. Shutemov 186b46e756fSKirill A. Shutemov return count; 187b46e756fSKirill A. Shutemov } 188b46e756fSKirill A. Shutemov static struct kobj_attribute alloc_sleep_millisecs_attr = 1896dcdc94dSMiaohe Lin __ATTR_RW(alloc_sleep_millisecs); 190b46e756fSKirill A. Shutemov 191b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_show(struct kobject *kobj, 192b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 193b46e756fSKirill A. Shutemov char *buf) 194b46e756fSKirill A. Shutemov { 195ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); 196b46e756fSKirill A. Shutemov } 197b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_store(struct kobject *kobj, 198b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 199b46e756fSKirill A. Shutemov const char *buf, size_t count) 200b46e756fSKirill A. Shutemov { 201dfefd226SAlexey Dobriyan unsigned int pages; 202b46e756fSKirill A. Shutemov int err; 203b46e756fSKirill A. Shutemov 204dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &pages); 205dfefd226SAlexey Dobriyan if (err || !pages) 206b46e756fSKirill A. Shutemov return -EINVAL; 207b46e756fSKirill A. Shutemov 208b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = pages; 209b46e756fSKirill A. Shutemov 210b46e756fSKirill A. Shutemov return count; 211b46e756fSKirill A. Shutemov } 212b46e756fSKirill A. Shutemov static struct kobj_attribute pages_to_scan_attr = 2136dcdc94dSMiaohe Lin __ATTR_RW(pages_to_scan); 214b46e756fSKirill A. Shutemov 215b46e756fSKirill A. Shutemov static ssize_t pages_collapsed_show(struct kobject *kobj, 216b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 217b46e756fSKirill A. Shutemov char *buf) 218b46e756fSKirill A. Shutemov { 219ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); 220b46e756fSKirill A. Shutemov } 221b46e756fSKirill A. Shutemov static struct kobj_attribute pages_collapsed_attr = 222b46e756fSKirill A. Shutemov __ATTR_RO(pages_collapsed); 223b46e756fSKirill A. Shutemov 224b46e756fSKirill A. Shutemov static ssize_t full_scans_show(struct kobject *kobj, 225b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 226b46e756fSKirill A. Shutemov char *buf) 227b46e756fSKirill A. Shutemov { 228ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_full_scans); 229b46e756fSKirill A. Shutemov } 230b46e756fSKirill A. Shutemov static struct kobj_attribute full_scans_attr = 231b46e756fSKirill A. Shutemov __ATTR_RO(full_scans); 232b46e756fSKirill A. Shutemov 2336dcdc94dSMiaohe Lin static ssize_t defrag_show(struct kobject *kobj, 234b46e756fSKirill A. Shutemov struct kobj_attribute *attr, char *buf) 235b46e756fSKirill A. Shutemov { 236b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 237b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 238b46e756fSKirill A. Shutemov } 2396dcdc94dSMiaohe Lin static ssize_t defrag_store(struct kobject *kobj, 240b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 241b46e756fSKirill A. Shutemov const char *buf, size_t count) 242b46e756fSKirill A. Shutemov { 243b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 244b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 245b46e756fSKirill A. Shutemov } 246b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_defrag_attr = 2476dcdc94dSMiaohe Lin __ATTR_RW(defrag); 248b46e756fSKirill A. Shutemov 249b46e756fSKirill A. Shutemov /* 250b46e756fSKirill A. Shutemov * max_ptes_none controls if khugepaged should collapse hugepages over 251b46e756fSKirill A. Shutemov * any unmapped ptes in turn potentially increasing the memory 252b46e756fSKirill A. Shutemov * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 253b46e756fSKirill A. Shutemov * reduce the available free memory in the system as it 254b46e756fSKirill A. Shutemov * runs. Increasing max_ptes_none will instead potentially reduce the 255b46e756fSKirill A. Shutemov * free memory in the system during the khugepaged scan. 256b46e756fSKirill A. Shutemov */ 2576dcdc94dSMiaohe Lin static ssize_t max_ptes_none_show(struct kobject *kobj, 258b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 259b46e756fSKirill A. Shutemov char *buf) 260b46e756fSKirill A. Shutemov { 261ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); 262b46e756fSKirill A. Shutemov } 2636dcdc94dSMiaohe Lin static ssize_t max_ptes_none_store(struct kobject *kobj, 264b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 265b46e756fSKirill A. Shutemov const char *buf, size_t count) 266b46e756fSKirill A. Shutemov { 267b46e756fSKirill A. Shutemov int err; 268b46e756fSKirill A. Shutemov unsigned long max_ptes_none; 269b46e756fSKirill A. Shutemov 270b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_none); 271b46e756fSKirill A. Shutemov if (err || max_ptes_none > HPAGE_PMD_NR - 1) 272b46e756fSKirill A. Shutemov return -EINVAL; 273b46e756fSKirill A. Shutemov 274b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = max_ptes_none; 275b46e756fSKirill A. Shutemov 276b46e756fSKirill A. Shutemov return count; 277b46e756fSKirill A. Shutemov } 278b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_none_attr = 2796dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_none); 280b46e756fSKirill A. Shutemov 2816dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_show(struct kobject *kobj, 282b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 283b46e756fSKirill A. Shutemov char *buf) 284b46e756fSKirill A. Shutemov { 285ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); 286b46e756fSKirill A. Shutemov } 287b46e756fSKirill A. Shutemov 2886dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_store(struct kobject *kobj, 289b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 290b46e756fSKirill A. Shutemov const char *buf, size_t count) 291b46e756fSKirill A. Shutemov { 292b46e756fSKirill A. Shutemov int err; 293b46e756fSKirill A. Shutemov unsigned long max_ptes_swap; 294b46e756fSKirill A. Shutemov 295b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_swap); 296b46e756fSKirill A. Shutemov if (err || max_ptes_swap > HPAGE_PMD_NR - 1) 297b46e756fSKirill A. Shutemov return -EINVAL; 298b46e756fSKirill A. Shutemov 299b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = max_ptes_swap; 300b46e756fSKirill A. Shutemov 301b46e756fSKirill A. Shutemov return count; 302b46e756fSKirill A. Shutemov } 303b46e756fSKirill A. Shutemov 304b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_swap_attr = 3056dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_swap); 306b46e756fSKirill A. Shutemov 3076dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_show(struct kobject *kobj, 30871a2c112SKirill A. Shutemov struct kobj_attribute *attr, 30971a2c112SKirill A. Shutemov char *buf) 31071a2c112SKirill A. Shutemov { 311ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); 31271a2c112SKirill A. Shutemov } 31371a2c112SKirill A. Shutemov 3146dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_store(struct kobject *kobj, 31571a2c112SKirill A. Shutemov struct kobj_attribute *attr, 31671a2c112SKirill A. Shutemov const char *buf, size_t count) 31771a2c112SKirill A. Shutemov { 31871a2c112SKirill A. Shutemov int err; 31971a2c112SKirill A. Shutemov unsigned long max_ptes_shared; 32071a2c112SKirill A. Shutemov 32171a2c112SKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_shared); 32271a2c112SKirill A. Shutemov if (err || max_ptes_shared > HPAGE_PMD_NR - 1) 32371a2c112SKirill A. Shutemov return -EINVAL; 32471a2c112SKirill A. Shutemov 32571a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = max_ptes_shared; 32671a2c112SKirill A. Shutemov 32771a2c112SKirill A. Shutemov return count; 32871a2c112SKirill A. Shutemov } 32971a2c112SKirill A. Shutemov 33071a2c112SKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_shared_attr = 3316dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_shared); 33271a2c112SKirill A. Shutemov 333b46e756fSKirill A. Shutemov static struct attribute *khugepaged_attr[] = { 334b46e756fSKirill A. Shutemov &khugepaged_defrag_attr.attr, 335b46e756fSKirill A. Shutemov &khugepaged_max_ptes_none_attr.attr, 33671a2c112SKirill A. Shutemov &khugepaged_max_ptes_swap_attr.attr, 33771a2c112SKirill A. Shutemov &khugepaged_max_ptes_shared_attr.attr, 338b46e756fSKirill A. Shutemov &pages_to_scan_attr.attr, 339b46e756fSKirill A. Shutemov &pages_collapsed_attr.attr, 340b46e756fSKirill A. Shutemov &full_scans_attr.attr, 341b46e756fSKirill A. Shutemov &scan_sleep_millisecs_attr.attr, 342b46e756fSKirill A. Shutemov &alloc_sleep_millisecs_attr.attr, 343b46e756fSKirill A. Shutemov NULL, 344b46e756fSKirill A. Shutemov }; 345b46e756fSKirill A. Shutemov 346b46e756fSKirill A. Shutemov struct attribute_group khugepaged_attr_group = { 347b46e756fSKirill A. Shutemov .attrs = khugepaged_attr, 348b46e756fSKirill A. Shutemov .name = "khugepaged", 349b46e756fSKirill A. Shutemov }; 350e1465d12SJérémy Lefaure #endif /* CONFIG_SYSFS */ 351b46e756fSKirill A. Shutemov 352b46e756fSKirill A. Shutemov int hugepage_madvise(struct vm_area_struct *vma, 353b46e756fSKirill A. Shutemov unsigned long *vm_flags, int advice) 354b46e756fSKirill A. Shutemov { 355b46e756fSKirill A. Shutemov switch (advice) { 356b46e756fSKirill A. Shutemov case MADV_HUGEPAGE: 357b46e756fSKirill A. Shutemov #ifdef CONFIG_S390 358b46e756fSKirill A. Shutemov /* 359b46e756fSKirill A. Shutemov * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 360b46e756fSKirill A. Shutemov * can't handle this properly after s390_enable_sie, so we simply 361b46e756fSKirill A. Shutemov * ignore the madvise to prevent qemu from causing a SIGSEGV. 362b46e756fSKirill A. Shutemov */ 363b46e756fSKirill A. Shutemov if (mm_has_pgste(vma->vm_mm)) 364b46e756fSKirill A. Shutemov return 0; 365b46e756fSKirill A. Shutemov #endif 366b46e756fSKirill A. Shutemov *vm_flags &= ~VM_NOHUGEPAGE; 367b46e756fSKirill A. Shutemov *vm_flags |= VM_HUGEPAGE; 368b46e756fSKirill A. Shutemov /* 369b46e756fSKirill A. Shutemov * If the vma become good for khugepaged to scan, 370b46e756fSKirill A. Shutemov * register it here without waiting a page fault that 371b46e756fSKirill A. Shutemov * may not happen any time soon. 372b46e756fSKirill A. Shutemov */ 373c791576cSYang Shi khugepaged_enter_vma(vma, *vm_flags); 374b46e756fSKirill A. Shutemov break; 375b46e756fSKirill A. Shutemov case MADV_NOHUGEPAGE: 376b46e756fSKirill A. Shutemov *vm_flags &= ~VM_HUGEPAGE; 377b46e756fSKirill A. Shutemov *vm_flags |= VM_NOHUGEPAGE; 378b46e756fSKirill A. Shutemov /* 379b46e756fSKirill A. Shutemov * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 380b46e756fSKirill A. Shutemov * this vma even if we leave the mm registered in khugepaged if 381b46e756fSKirill A. Shutemov * it got registered before VM_NOHUGEPAGE was set. 382b46e756fSKirill A. Shutemov */ 383b46e756fSKirill A. Shutemov break; 384b46e756fSKirill A. Shutemov } 385b46e756fSKirill A. Shutemov 386b46e756fSKirill A. Shutemov return 0; 387b46e756fSKirill A. Shutemov } 388b46e756fSKirill A. Shutemov 389b46e756fSKirill A. Shutemov int __init khugepaged_init(void) 390b46e756fSKirill A. Shutemov { 391b46e756fSKirill A. Shutemov mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 392b26e2701SQi Zheng sizeof(struct khugepaged_mm_slot), 393b26e2701SQi Zheng __alignof__(struct khugepaged_mm_slot), 394b26e2701SQi Zheng 0, NULL); 395b46e756fSKirill A. Shutemov if (!mm_slot_cache) 396b46e756fSKirill A. Shutemov return -ENOMEM; 397b46e756fSKirill A. Shutemov 398b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 399b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 400b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 40171a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; 402b46e756fSKirill A. Shutemov 403b46e756fSKirill A. Shutemov return 0; 404b46e756fSKirill A. Shutemov } 405b46e756fSKirill A. Shutemov 406b46e756fSKirill A. Shutemov void __init khugepaged_destroy(void) 407b46e756fSKirill A. Shutemov { 408b46e756fSKirill A. Shutemov kmem_cache_destroy(mm_slot_cache); 409b46e756fSKirill A. Shutemov } 410b46e756fSKirill A. Shutemov 4117d2c4385SZach O'Keefe static inline int hpage_collapse_test_exit(struct mm_struct *mm) 412b46e756fSKirill A. Shutemov { 4134d45e75aSJann Horn return atomic_read(&mm->mm_users) == 0; 414b46e756fSKirill A. Shutemov } 415b46e756fSKirill A. Shutemov 416d2081b2bSYang Shi void __khugepaged_enter(struct mm_struct *mm) 417b46e756fSKirill A. Shutemov { 418b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 419b26e2701SQi Zheng struct mm_slot *slot; 420b46e756fSKirill A. Shutemov int wakeup; 421b46e756fSKirill A. Shutemov 422b26e2701SQi Zheng mm_slot = mm_slot_alloc(mm_slot_cache); 423b46e756fSKirill A. Shutemov if (!mm_slot) 424d2081b2bSYang Shi return; 425b46e756fSKirill A. Shutemov 426b26e2701SQi Zheng slot = &mm_slot->slot; 427b26e2701SQi Zheng 428b46e756fSKirill A. Shutemov /* __khugepaged_exit() must not run from under us */ 4297d2c4385SZach O'Keefe VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm); 430b46e756fSKirill A. Shutemov if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 431b26e2701SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 432d2081b2bSYang Shi return; 433b46e756fSKirill A. Shutemov } 434b46e756fSKirill A. Shutemov 435b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 436b26e2701SQi Zheng mm_slot_insert(mm_slots_hash, mm, slot); 437b46e756fSKirill A. Shutemov /* 438b46e756fSKirill A. Shutemov * Insert just behind the scanning cursor, to let the area settle 439b46e756fSKirill A. Shutemov * down a little. 440b46e756fSKirill A. Shutemov */ 441b46e756fSKirill A. Shutemov wakeup = list_empty(&khugepaged_scan.mm_head); 442b26e2701SQi Zheng list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head); 443b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 444b46e756fSKirill A. Shutemov 445f1f10076SVegard Nossum mmgrab(mm); 446b46e756fSKirill A. Shutemov if (wakeup) 447b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 448b46e756fSKirill A. Shutemov } 449b46e756fSKirill A. Shutemov 450c791576cSYang Shi void khugepaged_enter_vma(struct vm_area_struct *vma, 451b46e756fSKirill A. Shutemov unsigned long vm_flags) 452b46e756fSKirill A. Shutemov { 4532647d11bSYang Shi if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && 4541064026bSYang Shi hugepage_flags_enabled()) { 455a7f4e6e4SZach O'Keefe if (hugepage_vma_check(vma, vm_flags, false, false, true)) 4562647d11bSYang Shi __khugepaged_enter(vma->vm_mm); 4572647d11bSYang Shi } 458b46e756fSKirill A. Shutemov } 459b46e756fSKirill A. Shutemov 460b46e756fSKirill A. Shutemov void __khugepaged_exit(struct mm_struct *mm) 461b46e756fSKirill A. Shutemov { 462b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 463b26e2701SQi Zheng struct mm_slot *slot; 464b46e756fSKirill A. Shutemov int free = 0; 465b46e756fSKirill A. Shutemov 466b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 467b26e2701SQi Zheng slot = mm_slot_lookup(mm_slots_hash, mm); 468b26e2701SQi Zheng mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 469b46e756fSKirill A. Shutemov if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 470b26e2701SQi Zheng hash_del(&slot->hash); 471b26e2701SQi Zheng list_del(&slot->mm_node); 472b46e756fSKirill A. Shutemov free = 1; 473b46e756fSKirill A. Shutemov } 474b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 475b46e756fSKirill A. Shutemov 476b46e756fSKirill A. Shutemov if (free) { 477b46e756fSKirill A. Shutemov clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 478b26e2701SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 479b46e756fSKirill A. Shutemov mmdrop(mm); 480b46e756fSKirill A. Shutemov } else if (mm_slot) { 481b46e756fSKirill A. Shutemov /* 482b46e756fSKirill A. Shutemov * This is required to serialize against 4837d2c4385SZach O'Keefe * hpage_collapse_test_exit() (which is guaranteed to run 4847d2c4385SZach O'Keefe * under mmap sem read mode). Stop here (after we return all 4857d2c4385SZach O'Keefe * pagetables will be destroyed) until khugepaged has finished 4867d2c4385SZach O'Keefe * working on the pagetables under the mmap_lock. 487b46e756fSKirill A. Shutemov */ 488d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 489d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 490b46e756fSKirill A. Shutemov } 491b46e756fSKirill A. Shutemov } 492b46e756fSKirill A. Shutemov 493*92644f58SVishal Moola (Oracle) static void release_pte_folio(struct folio *folio) 494*92644f58SVishal Moola (Oracle) { 495*92644f58SVishal Moola (Oracle) node_stat_mod_folio(folio, 496*92644f58SVishal Moola (Oracle) NR_ISOLATED_ANON + folio_is_file_lru(folio), 497*92644f58SVishal Moola (Oracle) -folio_nr_pages(folio)); 498*92644f58SVishal Moola (Oracle) folio_unlock(folio); 499*92644f58SVishal Moola (Oracle) folio_putback_lru(folio); 500*92644f58SVishal Moola (Oracle) } 501*92644f58SVishal Moola (Oracle) 502b46e756fSKirill A. Shutemov static void release_pte_page(struct page *page) 503b46e756fSKirill A. Shutemov { 504*92644f58SVishal Moola (Oracle) release_pte_folio(page_folio(page)); 505b46e756fSKirill A. Shutemov } 506b46e756fSKirill A. Shutemov 5075503fbf2SKirill A. Shutemov static void release_pte_pages(pte_t *pte, pte_t *_pte, 5085503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 509b46e756fSKirill A. Shutemov { 5105503fbf2SKirill A. Shutemov struct page *page, *tmp; 5115503fbf2SKirill A. Shutemov 512b46e756fSKirill A. Shutemov while (--_pte >= pte) { 513b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 5145503fbf2SKirill A. Shutemov 5155503fbf2SKirill A. Shutemov page = pte_page(pteval); 5165503fbf2SKirill A. Shutemov if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) && 5175503fbf2SKirill A. Shutemov !PageCompound(page)) 5185503fbf2SKirill A. Shutemov release_pte_page(page); 5195503fbf2SKirill A. Shutemov } 5205503fbf2SKirill A. Shutemov 5215503fbf2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, compound_pagelist, lru) { 5225503fbf2SKirill A. Shutemov list_del(&page->lru); 5235503fbf2SKirill A. Shutemov release_pte_page(page); 524b46e756fSKirill A. Shutemov } 525b46e756fSKirill A. Shutemov } 526b46e756fSKirill A. Shutemov 5279445689fSKirill A. Shutemov static bool is_refcount_suitable(struct page *page) 5289445689fSKirill A. Shutemov { 5299445689fSKirill A. Shutemov int expected_refcount; 5309445689fSKirill A. Shutemov 5319445689fSKirill A. Shutemov expected_refcount = total_mapcount(page); 5329445689fSKirill A. Shutemov if (PageSwapCache(page)) 5339445689fSKirill A. Shutemov expected_refcount += compound_nr(page); 5349445689fSKirill A. Shutemov 5359445689fSKirill A. Shutemov return page_count(page) == expected_refcount; 5369445689fSKirill A. Shutemov } 5379445689fSKirill A. Shutemov 538b46e756fSKirill A. Shutemov static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 539b46e756fSKirill A. Shutemov unsigned long address, 5405503fbf2SKirill A. Shutemov pte_t *pte, 541d8ea7cc8SZach O'Keefe struct collapse_control *cc, 5425503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 543b46e756fSKirill A. Shutemov { 544b46e756fSKirill A. Shutemov struct page *page = NULL; 545b46e756fSKirill A. Shutemov pte_t *_pte; 54650ad2f24SZach O'Keefe int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0; 5470db501f7SEbru Akagunduz bool writable = false; 548b46e756fSKirill A. Shutemov 549b46e756fSKirill A. Shutemov for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 550b46e756fSKirill A. Shutemov _pte++, address += PAGE_SIZE) { 551b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 552b46e756fSKirill A. Shutemov if (pte_none(pteval) || (pte_present(pteval) && 553b46e756fSKirill A. Shutemov is_zero_pfn(pte_pfn(pteval)))) { 554d8ea7cc8SZach O'Keefe ++none_or_zero; 555b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 556d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 557d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 558b46e756fSKirill A. Shutemov continue; 559b46e756fSKirill A. Shutemov } else { 560b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 561e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 562b46e756fSKirill A. Shutemov goto out; 563b46e756fSKirill A. Shutemov } 564b46e756fSKirill A. Shutemov } 565b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 566b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 567b46e756fSKirill A. Shutemov goto out; 568b46e756fSKirill A. Shutemov } 569b46e756fSKirill A. Shutemov page = vm_normal_page(vma, address, pteval); 5703218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 571b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 572b46e756fSKirill A. Shutemov goto out; 573b46e756fSKirill A. Shutemov } 574b46e756fSKirill A. Shutemov 575b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageAnon(page), page); 576b46e756fSKirill A. Shutemov 577d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 578d8ea7cc8SZach O'Keefe ++shared; 579d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 580d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 58171a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 582e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 58371a2c112SKirill A. Shutemov goto out; 58471a2c112SKirill A. Shutemov } 585d8ea7cc8SZach O'Keefe } 58671a2c112SKirill A. Shutemov 5875503fbf2SKirill A. Shutemov if (PageCompound(page)) { 5885503fbf2SKirill A. Shutemov struct page *p; 5895503fbf2SKirill A. Shutemov page = compound_head(page); 5905503fbf2SKirill A. Shutemov 5915503fbf2SKirill A. Shutemov /* 5925503fbf2SKirill A. Shutemov * Check if we have dealt with the compound page 5935503fbf2SKirill A. Shutemov * already 5945503fbf2SKirill A. Shutemov */ 5955503fbf2SKirill A. Shutemov list_for_each_entry(p, compound_pagelist, lru) { 5965503fbf2SKirill A. Shutemov if (page == p) 5975503fbf2SKirill A. Shutemov goto next; 5985503fbf2SKirill A. Shutemov } 5995503fbf2SKirill A. Shutemov } 6005503fbf2SKirill A. Shutemov 601b46e756fSKirill A. Shutemov /* 602b46e756fSKirill A. Shutemov * We can do it before isolate_lru_page because the 603b46e756fSKirill A. Shutemov * page can't be freed from under us. NOTE: PG_lock 604b46e756fSKirill A. Shutemov * is needed to serialize against split_huge_page 605b46e756fSKirill A. Shutemov * when invoked from the VM. 606b46e756fSKirill A. Shutemov */ 607b46e756fSKirill A. Shutemov if (!trylock_page(page)) { 608b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 609b46e756fSKirill A. Shutemov goto out; 610b46e756fSKirill A. Shutemov } 611b46e756fSKirill A. Shutemov 612b46e756fSKirill A. Shutemov /* 6139445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 6149445689fSKirill A. Shutemov * 6159445689fSKirill A. Shutemov * The page table that maps the page has been already unlinked 6169445689fSKirill A. Shutemov * from the page table tree and this process cannot get 617f0953a1bSIngo Molnar * an additional pin on the page. 6189445689fSKirill A. Shutemov * 6199445689fSKirill A. Shutemov * New pins can come later if the page is shared across fork, 6209445689fSKirill A. Shutemov * but not from this process. The other process cannot write to 6219445689fSKirill A. Shutemov * the page, only trigger CoW. 622b46e756fSKirill A. Shutemov */ 6239445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 624b46e756fSKirill A. Shutemov unlock_page(page); 625b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 626b46e756fSKirill A. Shutemov goto out; 627b46e756fSKirill A. Shutemov } 628b46e756fSKirill A. Shutemov 629b46e756fSKirill A. Shutemov /* 630b46e756fSKirill A. Shutemov * Isolate the page to avoid collapsing an hugepage 631b46e756fSKirill A. Shutemov * currently in use by the VM. 632b46e756fSKirill A. Shutemov */ 633b46e756fSKirill A. Shutemov if (isolate_lru_page(page)) { 634b46e756fSKirill A. Shutemov unlock_page(page); 635b46e756fSKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 636b46e756fSKirill A. Shutemov goto out; 637b46e756fSKirill A. Shutemov } 6385503fbf2SKirill A. Shutemov mod_node_page_state(page_pgdat(page), 6395503fbf2SKirill A. Shutemov NR_ISOLATED_ANON + page_is_file_lru(page), 6405503fbf2SKirill A. Shutemov compound_nr(page)); 641b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 642b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(PageLRU(page), page); 643b46e756fSKirill A. Shutemov 6445503fbf2SKirill A. Shutemov if (PageCompound(page)) 6455503fbf2SKirill A. Shutemov list_add_tail(&page->lru, compound_pagelist); 6465503fbf2SKirill A. Shutemov next: 647d8ea7cc8SZach O'Keefe /* 648d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 649d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 650d8ea7cc8SZach O'Keefe */ 651d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 652d8ea7cc8SZach O'Keefe (pte_young(pteval) || page_is_young(page) || 653d8ea7cc8SZach O'Keefe PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, 654d8ea7cc8SZach O'Keefe address))) 6550db501f7SEbru Akagunduz referenced++; 6565503fbf2SKirill A. Shutemov 6575503fbf2SKirill A. Shutemov if (pte_write(pteval)) 6585503fbf2SKirill A. Shutemov writable = true; 659b46e756fSKirill A. Shutemov } 66074e579bfSMiaohe Lin 66174e579bfSMiaohe Lin if (unlikely(!writable)) { 66274e579bfSMiaohe Lin result = SCAN_PAGE_RO; 663d8ea7cc8SZach O'Keefe } else if (unlikely(cc->is_khugepaged && !referenced)) { 66474e579bfSMiaohe Lin result = SCAN_LACK_REFERENCED_PAGE; 66574e579bfSMiaohe Lin } else { 666b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 667b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 668b46e756fSKirill A. Shutemov referenced, writable, result); 66950ad2f24SZach O'Keefe return result; 670b46e756fSKirill A. Shutemov } 671b46e756fSKirill A. Shutemov out: 6725503fbf2SKirill A. Shutemov release_pte_pages(pte, _pte, compound_pagelist); 673b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 674b46e756fSKirill A. Shutemov referenced, writable, result); 67550ad2f24SZach O'Keefe return result; 676b46e756fSKirill A. Shutemov } 677b46e756fSKirill A. Shutemov 678b46e756fSKirill A. Shutemov static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 679b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 680b46e756fSKirill A. Shutemov unsigned long address, 6815503fbf2SKirill A. Shutemov spinlock_t *ptl, 6825503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 683b46e756fSKirill A. Shutemov { 6845503fbf2SKirill A. Shutemov struct page *src_page, *tmp; 685b46e756fSKirill A. Shutemov pte_t *_pte; 686338a16baSDavid Rientjes for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 687338a16baSDavid Rientjes _pte++, page++, address += PAGE_SIZE) { 688b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 689b46e756fSKirill A. Shutemov 690b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 691b46e756fSKirill A. Shutemov clear_user_highpage(page, address); 692b46e756fSKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 693b46e756fSKirill A. Shutemov if (is_zero_pfn(pte_pfn(pteval))) { 694b46e756fSKirill A. Shutemov /* 695b46e756fSKirill A. Shutemov * ptl mostly unnecessary. 696b46e756fSKirill A. Shutemov */ 697b46e756fSKirill A. Shutemov spin_lock(ptl); 69808d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 699b46e756fSKirill A. Shutemov spin_unlock(ptl); 700b46e756fSKirill A. Shutemov } 701b46e756fSKirill A. Shutemov } else { 702b46e756fSKirill A. Shutemov src_page = pte_page(pteval); 703b46e756fSKirill A. Shutemov copy_user_highpage(page, src_page, address, vma); 7045503fbf2SKirill A. Shutemov if (!PageCompound(src_page)) 705b46e756fSKirill A. Shutemov release_pte_page(src_page); 706b46e756fSKirill A. Shutemov /* 707b46e756fSKirill A. Shutemov * ptl mostly unnecessary, but preempt has to 708b46e756fSKirill A. Shutemov * be disabled to update the per-cpu stats 709b46e756fSKirill A. Shutemov * inside page_remove_rmap(). 710b46e756fSKirill A. Shutemov */ 711b46e756fSKirill A. Shutemov spin_lock(ptl); 71208d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 713cea86fe2SHugh Dickins page_remove_rmap(src_page, vma, false); 714b46e756fSKirill A. Shutemov spin_unlock(ptl); 715b46e756fSKirill A. Shutemov free_page_and_swap_cache(src_page); 716b46e756fSKirill A. Shutemov } 717b46e756fSKirill A. Shutemov } 7185503fbf2SKirill A. Shutemov 7195503fbf2SKirill A. Shutemov list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { 7205503fbf2SKirill A. Shutemov list_del(&src_page->lru); 7211baec203SMiaohe Lin mod_node_page_state(page_pgdat(src_page), 7221baec203SMiaohe Lin NR_ISOLATED_ANON + page_is_file_lru(src_page), 7231baec203SMiaohe Lin -compound_nr(src_page)); 7241baec203SMiaohe Lin unlock_page(src_page); 7251baec203SMiaohe Lin free_swap_cache(src_page); 7261baec203SMiaohe Lin putback_lru_page(src_page); 7275503fbf2SKirill A. Shutemov } 728b46e756fSKirill A. Shutemov } 729b46e756fSKirill A. Shutemov 730b46e756fSKirill A. Shutemov static void khugepaged_alloc_sleep(void) 731b46e756fSKirill A. Shutemov { 732b46e756fSKirill A. Shutemov DEFINE_WAIT(wait); 733b46e756fSKirill A. Shutemov 734b46e756fSKirill A. Shutemov add_wait_queue(&khugepaged_wait, &wait); 735f5d39b02SPeter Zijlstra __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE); 736f5d39b02SPeter Zijlstra schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 737b46e756fSKirill A. Shutemov remove_wait_queue(&khugepaged_wait, &wait); 738b46e756fSKirill A. Shutemov } 739b46e756fSKirill A. Shutemov 74034d6b470SZach O'Keefe struct collapse_control khugepaged_collapse_control = { 741d8ea7cc8SZach O'Keefe .is_khugepaged = true, 74234d6b470SZach O'Keefe }; 743b46e756fSKirill A. Shutemov 7447d2c4385SZach O'Keefe static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc) 745b46e756fSKirill A. Shutemov { 746b46e756fSKirill A. Shutemov int i; 747b46e756fSKirill A. Shutemov 748b46e756fSKirill A. Shutemov /* 749a5f5f91dSMel Gorman * If node_reclaim_mode is disabled, then no extra effort is made to 750b46e756fSKirill A. Shutemov * allocate memory locally. 751b46e756fSKirill A. Shutemov */ 752202e35dbSDave Hansen if (!node_reclaim_enabled()) 753b46e756fSKirill A. Shutemov return false; 754b46e756fSKirill A. Shutemov 755b46e756fSKirill A. Shutemov /* If there is a count for this node already, it must be acceptable */ 75634d6b470SZach O'Keefe if (cc->node_load[nid]) 757b46e756fSKirill A. Shutemov return false; 758b46e756fSKirill A. Shutemov 759b46e756fSKirill A. Shutemov for (i = 0; i < MAX_NUMNODES; i++) { 76034d6b470SZach O'Keefe if (!cc->node_load[i]) 761b46e756fSKirill A. Shutemov continue; 762a55c7454SMatt Fleming if (node_distance(nid, i) > node_reclaim_distance) 763b46e756fSKirill A. Shutemov return true; 764b46e756fSKirill A. Shutemov } 765b46e756fSKirill A. Shutemov return false; 766b46e756fSKirill A. Shutemov } 767b46e756fSKirill A. Shutemov 7681064026bSYang Shi #define khugepaged_defrag() \ 7691064026bSYang Shi (transparent_hugepage_flags & \ 7701064026bSYang Shi (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)) 7711064026bSYang Shi 772b46e756fSKirill A. Shutemov /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 773b46e756fSKirill A. Shutemov static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 774b46e756fSKirill A. Shutemov { 77525160354SVlastimil Babka return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 776b46e756fSKirill A. Shutemov } 777b46e756fSKirill A. Shutemov 778b46e756fSKirill A. Shutemov #ifdef CONFIG_NUMA 7797d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 780b46e756fSKirill A. Shutemov { 781b46e756fSKirill A. Shutemov int nid, target_node = 0, max_value = 0; 782b46e756fSKirill A. Shutemov 783b46e756fSKirill A. Shutemov /* find first node with max normal pages hit */ 784b46e756fSKirill A. Shutemov for (nid = 0; nid < MAX_NUMNODES; nid++) 78534d6b470SZach O'Keefe if (cc->node_load[nid] > max_value) { 78634d6b470SZach O'Keefe max_value = cc->node_load[nid]; 787b46e756fSKirill A. Shutemov target_node = nid; 788b46e756fSKirill A. Shutemov } 789b46e756fSKirill A. Shutemov 790e031ff96SYang Shi for_each_online_node(nid) { 791e031ff96SYang Shi if (max_value == cc->node_load[nid]) 792e031ff96SYang Shi node_set(nid, cc->alloc_nmask); 793b46e756fSKirill A. Shutemov } 794b46e756fSKirill A. Shutemov 795b46e756fSKirill A. Shutemov return target_node; 796b46e756fSKirill A. Shutemov } 797c6a7f445SYang Shi #else 7987d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 799b46e756fSKirill A. Shutemov { 800c6a7f445SYang Shi return 0; 801b46e756fSKirill A. Shutemov } 802c6a7f445SYang Shi #endif 803b46e756fSKirill A. Shutemov 804e031ff96SYang Shi static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node, 805e031ff96SYang Shi nodemask_t *nmask) 806b46e756fSKirill A. Shutemov { 807e031ff96SYang Shi *hpage = __alloc_pages(gfp, HPAGE_PMD_ORDER, node, nmask); 808b46e756fSKirill A. Shutemov if (unlikely(!*hpage)) { 809b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 8109710a78aSZach O'Keefe return false; 811b46e756fSKirill A. Shutemov } 812b46e756fSKirill A. Shutemov 813b46e756fSKirill A. Shutemov prep_transhuge_page(*hpage); 814b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 815b46e756fSKirill A. Shutemov return true; 816b46e756fSKirill A. Shutemov } 817b46e756fSKirill A. Shutemov 818b46e756fSKirill A. Shutemov /* 819c1e8d7c6SMichel Lespinasse * If mmap_lock temporarily dropped, revalidate vma 820c1e8d7c6SMichel Lespinasse * before taking mmap_lock. 82150ad2f24SZach O'Keefe * Returns enum scan_result value. 822b46e756fSKirill A. Shutemov */ 823b46e756fSKirill A. Shutemov 824c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 82534488399SZach O'Keefe bool expect_anon, 826a7f4e6e4SZach O'Keefe struct vm_area_struct **vmap, 827a7f4e6e4SZach O'Keefe struct collapse_control *cc) 828b46e756fSKirill A. Shutemov { 829b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 830b46e756fSKirill A. Shutemov 8317d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 832b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 833b46e756fSKirill A. Shutemov 834c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 835b46e756fSKirill A. Shutemov if (!vma) 836b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 837b46e756fSKirill A. Shutemov 8384fa6893fSYang Shi if (!transhuge_vma_suitable(vma, address)) 839b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 840a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, 841a7f4e6e4SZach O'Keefe cc->is_khugepaged)) 842b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 843f707fa49SYang Shi /* 844f707fa49SYang Shi * Anon VMA expected, the address may be unmapped then 845f707fa49SYang Shi * remapped to file after khugepaged reaquired the mmap_lock. 846f707fa49SYang Shi * 847f707fa49SYang Shi * hugepage_vma_check may return true for qualified file 848f707fa49SYang Shi * vmas. 849f707fa49SYang Shi */ 85034488399SZach O'Keefe if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) 85134488399SZach O'Keefe return SCAN_PAGE_ANON; 85250ad2f24SZach O'Keefe return SCAN_SUCCEED; 853b46e756fSKirill A. Shutemov } 854b46e756fSKirill A. Shutemov 855edb5d0cfSZach O'Keefe /* 856edb5d0cfSZach O'Keefe * See pmd_trans_unstable() for how the result may change out from 857edb5d0cfSZach O'Keefe * underneath us, even if we hold mmap_lock in read. 858edb5d0cfSZach O'Keefe */ 85950722804SZach O'Keefe static int find_pmd_or_thp_or_none(struct mm_struct *mm, 86050722804SZach O'Keefe unsigned long address, 86150722804SZach O'Keefe pmd_t **pmd) 86250722804SZach O'Keefe { 86350722804SZach O'Keefe pmd_t pmde; 86450722804SZach O'Keefe 86550722804SZach O'Keefe *pmd = mm_find_pmd(mm, address); 86650722804SZach O'Keefe if (!*pmd) 86750722804SZach O'Keefe return SCAN_PMD_NULL; 86850722804SZach O'Keefe 869dab6e717SPeter Zijlstra pmde = pmdp_get_lockless(*pmd); 87050722804SZach O'Keefe 87150722804SZach O'Keefe #ifdef CONFIG_TRANSPARENT_HUGEPAGE 87250722804SZach O'Keefe /* See comments in pmd_none_or_trans_huge_or_clear_bad() */ 87350722804SZach O'Keefe barrier(); 87450722804SZach O'Keefe #endif 87534488399SZach O'Keefe if (pmd_none(pmde)) 87634488399SZach O'Keefe return SCAN_PMD_NONE; 877edb5d0cfSZach O'Keefe if (!pmd_present(pmde)) 878edb5d0cfSZach O'Keefe return SCAN_PMD_NULL; 87950722804SZach O'Keefe if (pmd_trans_huge(pmde)) 88050722804SZach O'Keefe return SCAN_PMD_MAPPED; 881edb5d0cfSZach O'Keefe if (pmd_devmap(pmde)) 882edb5d0cfSZach O'Keefe return SCAN_PMD_NULL; 88350722804SZach O'Keefe if (pmd_bad(pmde)) 88450722804SZach O'Keefe return SCAN_PMD_NULL; 88550722804SZach O'Keefe return SCAN_SUCCEED; 88650722804SZach O'Keefe } 88750722804SZach O'Keefe 88850722804SZach O'Keefe static int check_pmd_still_valid(struct mm_struct *mm, 88950722804SZach O'Keefe unsigned long address, 89050722804SZach O'Keefe pmd_t *pmd) 89150722804SZach O'Keefe { 89250722804SZach O'Keefe pmd_t *new_pmd; 89350722804SZach O'Keefe int result = find_pmd_or_thp_or_none(mm, address, &new_pmd); 89450722804SZach O'Keefe 89550722804SZach O'Keefe if (result != SCAN_SUCCEED) 89650722804SZach O'Keefe return result; 89750722804SZach O'Keefe if (new_pmd != pmd) 89850722804SZach O'Keefe return SCAN_FAIL; 89950722804SZach O'Keefe return SCAN_SUCCEED; 900b46e756fSKirill A. Shutemov } 901b46e756fSKirill A. Shutemov 902b46e756fSKirill A. Shutemov /* 903b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 9047d2c4385SZach O'Keefe * Only done if hpage_collapse_scan_pmd believes it is worthwhile. 905b46e756fSKirill A. Shutemov * 9064d928e20SMiaohe Lin * Called and returns without pte mapped or spinlocks held. 9074d928e20SMiaohe Lin * Note that if false is returned, mmap_lock will be released. 908b46e756fSKirill A. Shutemov */ 909b46e756fSKirill A. Shutemov 91050ad2f24SZach O'Keefe static int __collapse_huge_page_swapin(struct mm_struct *mm, 911b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 9122b635dd3SWill Deacon unsigned long haddr, pmd_t *pmd, 9130db501f7SEbru Akagunduz int referenced) 914b46e756fSKirill A. Shutemov { 9152b740303SSouptick Joarder int swapped_in = 0; 9162b740303SSouptick Joarder vm_fault_t ret = 0; 9172b635dd3SWill Deacon unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); 9182b635dd3SWill Deacon 9192b635dd3SWill Deacon for (address = haddr; address < end; address += PAGE_SIZE) { 92082b0f8c3SJan Kara struct vm_fault vmf = { 921b46e756fSKirill A. Shutemov .vma = vma, 922b46e756fSKirill A. Shutemov .address = address, 9232b635dd3SWill Deacon .pgoff = linear_page_index(vma, haddr), 924b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 925b46e756fSKirill A. Shutemov .pmd = pmd, 926b46e756fSKirill A. Shutemov }; 927b46e756fSKirill A. Shutemov 92882b0f8c3SJan Kara vmf.pte = pte_offset_map(pmd, address); 9292994302bSJan Kara vmf.orig_pte = *vmf.pte; 9302b635dd3SWill Deacon if (!is_swap_pte(vmf.orig_pte)) { 9312b635dd3SWill Deacon pte_unmap(vmf.pte); 932b46e756fSKirill A. Shutemov continue; 9332b635dd3SWill Deacon } 9342994302bSJan Kara ret = do_swap_page(&vmf); 9350db501f7SEbru Akagunduz 9364d928e20SMiaohe Lin /* 9374d928e20SMiaohe Lin * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. 9384d928e20SMiaohe Lin * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because 9394d928e20SMiaohe Lin * we do not retry here and swap entry will remain in pagetable 9404d928e20SMiaohe Lin * resulting in later failure. 9414d928e20SMiaohe Lin */ 942b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 9430db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 94450ad2f24SZach O'Keefe /* Likely, but not guaranteed, that page lock failed */ 94550ad2f24SZach O'Keefe return SCAN_PAGE_LOCK; 94647f863eaSEbru Akagunduz } 947b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 9484d928e20SMiaohe Lin mmap_read_unlock(mm); 9490db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 95050ad2f24SZach O'Keefe return SCAN_FAIL; 951b46e756fSKirill A. Shutemov } 9524d928e20SMiaohe Lin swapped_in++; 953b46e756fSKirill A. Shutemov } 954ae2c5d80SKirill A. Shutemov 955ae2c5d80SKirill A. Shutemov /* Drain LRU add pagevec to remove extra pin on the swapped in pages */ 956ae2c5d80SKirill A. Shutemov if (swapped_in) 957ae2c5d80SKirill A. Shutemov lru_add_drain(); 958ae2c5d80SKirill A. Shutemov 9590db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); 96050ad2f24SZach O'Keefe return SCAN_SUCCEED; 961b46e756fSKirill A. Shutemov } 962b46e756fSKirill A. Shutemov 9639710a78aSZach O'Keefe static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, 9649710a78aSZach O'Keefe struct collapse_control *cc) 9659710a78aSZach O'Keefe { 9667d8faaf1SZach O'Keefe gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() : 967e031ff96SYang Shi GFP_TRANSHUGE); 9687d2c4385SZach O'Keefe int node = hpage_collapse_find_target_node(cc); 9699710a78aSZach O'Keefe 970e031ff96SYang Shi if (!hpage_collapse_alloc_page(hpage, gfp, node, &cc->alloc_nmask)) 9719710a78aSZach O'Keefe return SCAN_ALLOC_HUGE_PAGE_FAIL; 9729710a78aSZach O'Keefe if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp))) 9739710a78aSZach O'Keefe return SCAN_CGROUP_CHARGE_FAIL; 9749710a78aSZach O'Keefe count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC); 9759710a78aSZach O'Keefe return SCAN_SUCCEED; 9769710a78aSZach O'Keefe } 9779710a78aSZach O'Keefe 97850ad2f24SZach O'Keefe static int collapse_huge_page(struct mm_struct *mm, unsigned long address, 97950ad2f24SZach O'Keefe int referenced, int unmapped, 98050ad2f24SZach O'Keefe struct collapse_control *cc) 981b46e756fSKirill A. Shutemov { 9825503fbf2SKirill A. Shutemov LIST_HEAD(compound_pagelist); 983b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 984b46e756fSKirill A. Shutemov pte_t *pte; 985b46e756fSKirill A. Shutemov pgtable_t pgtable; 98650ad2f24SZach O'Keefe struct page *hpage; 987b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 98850ad2f24SZach O'Keefe int result = SCAN_FAIL; 989c131f751SKirill A. Shutemov struct vm_area_struct *vma; 990ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 991b46e756fSKirill A. Shutemov 992b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 993b46e756fSKirill A. Shutemov 994988ddb71SKirill A. Shutemov /* 995c1e8d7c6SMichel Lespinasse * Before allocating the hugepage, release the mmap_lock read lock. 996988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 997c1e8d7c6SMichel Lespinasse * sync compaction, and we do not need to hold the mmap_lock during 998988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 999988ddb71SKirill A. Shutemov */ 1000d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1001b46e756fSKirill A. Shutemov 100250ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 10039710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1004b46e756fSKirill A. Shutemov goto out_nolock; 1005b46e756fSKirill A. Shutemov 1006d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 100734488399SZach O'Keefe result = hugepage_vma_revalidate(mm, address, true, &vma, cc); 100850ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) { 1009d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1010b46e756fSKirill A. Shutemov goto out_nolock; 1011b46e756fSKirill A. Shutemov } 1012b46e756fSKirill A. Shutemov 101350722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 101450722804SZach O'Keefe if (result != SCAN_SUCCEED) { 1015d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1016b46e756fSKirill A. Shutemov goto out_nolock; 1017b46e756fSKirill A. Shutemov } 1018b46e756fSKirill A. Shutemov 101950ad2f24SZach O'Keefe if (unmapped) { 1020b46e756fSKirill A. Shutemov /* 102150ad2f24SZach O'Keefe * __collapse_huge_page_swapin will return with mmap_lock 102250ad2f24SZach O'Keefe * released when it fails. So we jump out_nolock directly in 102350ad2f24SZach O'Keefe * that case. Continuing to collapse causes inconsistency. 1024b46e756fSKirill A. Shutemov */ 102550ad2f24SZach O'Keefe result = __collapse_huge_page_swapin(mm, vma, address, pmd, 102650ad2f24SZach O'Keefe referenced); 102750ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 1028b46e756fSKirill A. Shutemov goto out_nolock; 1029b46e756fSKirill A. Shutemov } 1030b46e756fSKirill A. Shutemov 1031d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1032b46e756fSKirill A. Shutemov /* 1033b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 1034b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 1035b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 1036b46e756fSKirill A. Shutemov */ 1037d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 103834488399SZach O'Keefe result = hugepage_vma_revalidate(mm, address, true, &vma, cc); 103950ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 104018d24a7cSMiaohe Lin goto out_up_write; 1041b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 104250722804SZach O'Keefe result = check_pmd_still_valid(mm, address, pmd); 104350722804SZach O'Keefe if (result != SCAN_SUCCEED) 104418d24a7cSMiaohe Lin goto out_up_write; 1045b46e756fSKirill A. Shutemov 1046b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1047b46e756fSKirill A. Shutemov 10487d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address, 10497d4a8be0SAlistair Popple address + HPAGE_PMD_SIZE); 1050ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1051ec649c9dSVille Syrjälä 1052ec649c9dSVille Syrjälä pte = pte_offset_map(pmd, address); 1053ec649c9dSVille Syrjälä pte_ptl = pte_lockptr(mm, pmd); 1054ec649c9dSVille Syrjälä 1055b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1056b46e756fSKirill A. Shutemov /* 105770cbc3ccSYang Shi * This removes any huge TLB entry from the CPU so we won't allow 105870cbc3ccSYang Shi * huge and small TLB entries for the same virtual address to 105970cbc3ccSYang Shi * avoid the risk of CPU bugs in that area. 106070cbc3ccSYang Shi * 106170cbc3ccSYang Shi * Parallel fast GUP is fine since fast GUP will back off when 106270cbc3ccSYang Shi * it detects PMD is changed. 1063b46e756fSKirill A. Shutemov */ 1064b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1065b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1066ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 10672ba99c5eSJann Horn tlb_remove_table_sync_one(); 1068b46e756fSKirill A. Shutemov 1069b46e756fSKirill A. Shutemov spin_lock(pte_ptl); 1070d8ea7cc8SZach O'Keefe result = __collapse_huge_page_isolate(vma, address, pte, cc, 10715503fbf2SKirill A. Shutemov &compound_pagelist); 1072b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1073b46e756fSKirill A. Shutemov 107450ad2f24SZach O'Keefe if (unlikely(result != SCAN_SUCCEED)) { 1075b46e756fSKirill A. Shutemov pte_unmap(pte); 1076b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1077b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1078b46e756fSKirill A. Shutemov /* 1079b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1080b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1081b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1082b46e756fSKirill A. Shutemov */ 1083b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1084b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1085b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 108618d24a7cSMiaohe Lin goto out_up_write; 1087b46e756fSKirill A. Shutemov } 1088b46e756fSKirill A. Shutemov 1089b46e756fSKirill A. Shutemov /* 1090b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1091b46e756fSKirill A. Shutemov * can't run anymore. 1092b46e756fSKirill A. Shutemov */ 1093b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1094b46e756fSKirill A. Shutemov 109550ad2f24SZach O'Keefe __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl, 10965503fbf2SKirill A. Shutemov &compound_pagelist); 1097b46e756fSKirill A. Shutemov pte_unmap(pte); 1098588d01f9SMiaohe Lin /* 1099588d01f9SMiaohe Lin * spin_lock() below is not the equivalent of smp_wmb(), but 1100588d01f9SMiaohe Lin * the smp_wmb() inside __SetPageUptodate() can be reused to 1101588d01f9SMiaohe Lin * avoid the copy_huge_page writes to become visible after 1102588d01f9SMiaohe Lin * the set_pmd_at() write. 1103588d01f9SMiaohe Lin */ 110450ad2f24SZach O'Keefe __SetPageUptodate(hpage); 1105b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1106b46e756fSKirill A. Shutemov 110750ad2f24SZach O'Keefe _pmd = mk_huge_pmd(hpage, vma->vm_page_prot); 1108f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1109b46e756fSKirill A. Shutemov 1110b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1111b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 111250ad2f24SZach O'Keefe page_add_new_anon_rmap(hpage, vma, address); 111350ad2f24SZach O'Keefe lru_cache_add_inactive_or_unevictable(hpage, vma); 1114b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1115b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1116b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1117b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1118b46e756fSKirill A. Shutemov 111950ad2f24SZach O'Keefe hpage = NULL; 1120b46e756fSKirill A. Shutemov 1121b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1122b46e756fSKirill A. Shutemov out_up_write: 1123d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 1124b46e756fSKirill A. Shutemov out_nolock: 112550ad2f24SZach O'Keefe if (hpage) { 112650ad2f24SZach O'Keefe mem_cgroup_uncharge(page_folio(hpage)); 112750ad2f24SZach O'Keefe put_page(hpage); 1128c6a7f445SYang Shi } 112950ad2f24SZach O'Keefe trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); 113050ad2f24SZach O'Keefe return result; 1131b46e756fSKirill A. Shutemov } 1132b46e756fSKirill A. Shutemov 11337d2c4385SZach O'Keefe static int hpage_collapse_scan_pmd(struct mm_struct *mm, 1134b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 113550ad2f24SZach O'Keefe unsigned long address, bool *mmap_locked, 113634d6b470SZach O'Keefe struct collapse_control *cc) 1137b46e756fSKirill A. Shutemov { 1138b46e756fSKirill A. Shutemov pmd_t *pmd; 1139b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 114050ad2f24SZach O'Keefe int result = SCAN_FAIL, referenced = 0; 114171a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0; 1142b46e756fSKirill A. Shutemov struct page *page = NULL; 1143b46e756fSKirill A. Shutemov unsigned long _address; 1144b46e756fSKirill A. Shutemov spinlock_t *ptl; 1145b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 11460db501f7SEbru Akagunduz bool writable = false; 1147b46e756fSKirill A. Shutemov 1148b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1149b46e756fSKirill A. Shutemov 115050722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 115150722804SZach O'Keefe if (result != SCAN_SUCCEED) 1152b46e756fSKirill A. Shutemov goto out; 1153b46e756fSKirill A. Shutemov 115434d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1155e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 1156b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1157b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; 1158b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1159b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 1160b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1161d8ea7cc8SZach O'Keefe ++unmapped; 1162d8ea7cc8SZach O'Keefe if (!cc->is_khugepaged || 1163d8ea7cc8SZach O'Keefe unmapped <= khugepaged_max_ptes_swap) { 1164e1e267c7SPeter Xu /* 1165e1e267c7SPeter Xu * Always be strict with uffd-wp 1166e1e267c7SPeter Xu * enabled swap entries. Please see 1167e1e267c7SPeter Xu * comment below for pte_uffd_wp(). 1168e1e267c7SPeter Xu */ 1169e1e267c7SPeter Xu if (pte_swp_uffd_wp(pteval)) { 1170e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1171e1e267c7SPeter Xu goto out_unmap; 1172e1e267c7SPeter Xu } 1173b46e756fSKirill A. Shutemov continue; 1174b46e756fSKirill A. Shutemov } else { 1175b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1176e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1177b46e756fSKirill A. Shutemov goto out_unmap; 1178b46e756fSKirill A. Shutemov } 1179b46e756fSKirill A. Shutemov } 1180b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1181d8ea7cc8SZach O'Keefe ++none_or_zero; 1182b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1183d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 1184d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 1185b46e756fSKirill A. Shutemov continue; 1186b46e756fSKirill A. Shutemov } else { 1187b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1188e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1189b46e756fSKirill A. Shutemov goto out_unmap; 1190b46e756fSKirill A. Shutemov } 1191b46e756fSKirill A. Shutemov } 1192e1e267c7SPeter Xu if (pte_uffd_wp(pteval)) { 1193e1e267c7SPeter Xu /* 1194e1e267c7SPeter Xu * Don't collapse the page if any of the small 1195e1e267c7SPeter Xu * PTEs are armed with uffd write protection. 1196e1e267c7SPeter Xu * Here we can also mark the new huge pmd as 1197e1e267c7SPeter Xu * write protected if any of the small ones is 11988958b249SHaitao Shi * marked but that could bring unknown 1199e1e267c7SPeter Xu * userfault messages that falls outside of 1200e1e267c7SPeter Xu * the registered range. So, just be simple. 1201e1e267c7SPeter Xu */ 1202e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1203e1e267c7SPeter Xu goto out_unmap; 1204e1e267c7SPeter Xu } 1205b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1206b46e756fSKirill A. Shutemov writable = true; 1207b46e756fSKirill A. Shutemov 1208b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 12093218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 1210b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1211b46e756fSKirill A. Shutemov goto out_unmap; 1212b46e756fSKirill A. Shutemov } 1213b46e756fSKirill A. Shutemov 1214d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 1215d8ea7cc8SZach O'Keefe ++shared; 1216d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 1217d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 121871a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 1219e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 122071a2c112SKirill A. Shutemov goto out_unmap; 122171a2c112SKirill A. Shutemov } 1222d8ea7cc8SZach O'Keefe } 122371a2c112SKirill A. Shutemov 12245503fbf2SKirill A. Shutemov page = compound_head(page); 1225b46e756fSKirill A. Shutemov 1226b46e756fSKirill A. Shutemov /* 1227b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 122834d6b470SZach O'Keefe * information to cc->node_load[]. 12290b8f0d87SQuanfa Fu * Khugepaged will allocate hugepage from the node has the max 1230b46e756fSKirill A. Shutemov * hit record. 1231b46e756fSKirill A. Shutemov */ 1232b46e756fSKirill A. Shutemov node = page_to_nid(page); 12337d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 1234b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1235b46e756fSKirill A. Shutemov goto out_unmap; 1236b46e756fSKirill A. Shutemov } 123734d6b470SZach O'Keefe cc->node_load[node]++; 1238b46e756fSKirill A. Shutemov if (!PageLRU(page)) { 1239b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1240b46e756fSKirill A. Shutemov goto out_unmap; 1241b46e756fSKirill A. Shutemov } 1242b46e756fSKirill A. Shutemov if (PageLocked(page)) { 1243b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1244b46e756fSKirill A. Shutemov goto out_unmap; 1245b46e756fSKirill A. Shutemov } 1246b46e756fSKirill A. Shutemov if (!PageAnon(page)) { 1247b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1248b46e756fSKirill A. Shutemov goto out_unmap; 1249b46e756fSKirill A. Shutemov } 1250b46e756fSKirill A. Shutemov 1251b46e756fSKirill A. Shutemov /* 12529445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 12539445689fSKirill A. Shutemov * 1254cb67f428SHugh Dickins * Here the check may be racy: 1255cb67f428SHugh Dickins * it may see total_mapcount > refcount in some cases? 12569445689fSKirill A. Shutemov * But such case is ephemeral we could always retry collapse 12579445689fSKirill A. Shutemov * later. However it may report false positive if the page 12589445689fSKirill A. Shutemov * has excessive GUP pins (i.e. 512). Anyway the same check 12599445689fSKirill A. Shutemov * will be done again later the risk seems low. 1260b46e756fSKirill A. Shutemov */ 12619445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 1262b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1263b46e756fSKirill A. Shutemov goto out_unmap; 1264b46e756fSKirill A. Shutemov } 1265d8ea7cc8SZach O'Keefe 1266d8ea7cc8SZach O'Keefe /* 1267d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 1268d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 1269d8ea7cc8SZach O'Keefe */ 1270d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 1271d8ea7cc8SZach O'Keefe (pte_young(pteval) || page_is_young(page) || 1272d8ea7cc8SZach O'Keefe PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, 1273d8ea7cc8SZach O'Keefe address))) 12740db501f7SEbru Akagunduz referenced++; 1275b46e756fSKirill A. Shutemov } 1276ffe945e6SKirill A. Shutemov if (!writable) { 1277ffe945e6SKirill A. Shutemov result = SCAN_PAGE_RO; 1278d8ea7cc8SZach O'Keefe } else if (cc->is_khugepaged && 1279d8ea7cc8SZach O'Keefe (!referenced || 1280d8ea7cc8SZach O'Keefe (unmapped && referenced < HPAGE_PMD_NR / 2))) { 1281ffe945e6SKirill A. Shutemov result = SCAN_LACK_REFERENCED_PAGE; 1282ffe945e6SKirill A. Shutemov } else { 1283b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1284b46e756fSKirill A. Shutemov } 1285b46e756fSKirill A. Shutemov out_unmap: 1286b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 128750ad2f24SZach O'Keefe if (result == SCAN_SUCCEED) { 128850ad2f24SZach O'Keefe result = collapse_huge_page(mm, address, referenced, 128950ad2f24SZach O'Keefe unmapped, cc); 1290c1e8d7c6SMichel Lespinasse /* collapse_huge_page will return with the mmap_lock released */ 129150ad2f24SZach O'Keefe *mmap_locked = false; 1292b46e756fSKirill A. Shutemov } 1293b46e756fSKirill A. Shutemov out: 1294b46e756fSKirill A. Shutemov trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, 1295b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 129650ad2f24SZach O'Keefe return result; 1297b46e756fSKirill A. Shutemov } 1298b46e756fSKirill A. Shutemov 1299b26e2701SQi Zheng static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot) 1300b46e756fSKirill A. Shutemov { 1301b26e2701SQi Zheng struct mm_slot *slot = &mm_slot->slot; 1302b26e2701SQi Zheng struct mm_struct *mm = slot->mm; 1303b46e756fSKirill A. Shutemov 130435f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1305b46e756fSKirill A. Shutemov 13067d2c4385SZach O'Keefe if (hpage_collapse_test_exit(mm)) { 1307b46e756fSKirill A. Shutemov /* free mm_slot */ 1308b26e2701SQi Zheng hash_del(&slot->hash); 1309b26e2701SQi Zheng list_del(&slot->mm_node); 1310b46e756fSKirill A. Shutemov 1311b46e756fSKirill A. Shutemov /* 1312b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1313b46e756fSKirill A. Shutemov * 1314b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1315b46e756fSKirill A. Shutemov */ 1316b46e756fSKirill A. Shutemov 1317b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1318b26e2701SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 1319b46e756fSKirill A. Shutemov mmdrop(mm); 1320b46e756fSKirill A. Shutemov } 1321b46e756fSKirill A. Shutemov } 1322b46e756fSKirill A. Shutemov 1323396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 132427e1f827SSong Liu /* 132527e1f827SSong Liu * Notify khugepaged that given addr of the mm is pte-mapped THP. Then 132627e1f827SSong Liu * khugepaged should try to collapse the page table. 132734488399SZach O'Keefe * 132834488399SZach O'Keefe * Note that following race exists: 132934488399SZach O'Keefe * (1) khugepaged calls khugepaged_collapse_pte_mapped_thps() for mm_struct A, 133034488399SZach O'Keefe * emptying the A's ->pte_mapped_thp[] array. 133134488399SZach O'Keefe * (2) MADV_COLLAPSE collapses some file extent with target mm_struct B, and 133234488399SZach O'Keefe * retract_page_tables() finds a VMA in mm_struct A mapping the same extent 133334488399SZach O'Keefe * (at virtual address X) and adds an entry (for X) into mm_struct A's 133434488399SZach O'Keefe * ->pte-mapped_thp[] array. 133534488399SZach O'Keefe * (3) khugepaged calls khugepaged_collapse_scan_file() for mm_struct A at X, 133634488399SZach O'Keefe * sees a pte-mapped THP (SCAN_PTE_MAPPED_HUGEPAGE) and adds an entry 133734488399SZach O'Keefe * (for X) into mm_struct A's ->pte-mapped_thp[] array. 133834488399SZach O'Keefe * Thus, it's possible the same address is added multiple times for the same 133934488399SZach O'Keefe * mm_struct. Should this happen, we'll simply attempt 134034488399SZach O'Keefe * collapse_pte_mapped_thp() multiple times for the same address, under the same 134134488399SZach O'Keefe * exclusive mmap_lock, and assuming the first call is successful, subsequent 134234488399SZach O'Keefe * attempts will return quickly (without grabbing any additional locks) when 134334488399SZach O'Keefe * a huge pmd is found in find_pmd_or_thp_or_none(). Since this is a cheap 134434488399SZach O'Keefe * check, and since this is a rare occurrence, the cost of preventing this 134534488399SZach O'Keefe * "multiple-add" is thought to be more expensive than just handling it, should 134634488399SZach O'Keefe * it occur. 134727e1f827SSong Liu */ 134858ac9a89SZach O'Keefe static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm, 134927e1f827SSong Liu unsigned long addr) 135027e1f827SSong Liu { 1351b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 1352b26e2701SQi Zheng struct mm_slot *slot; 135358ac9a89SZach O'Keefe bool ret = false; 135427e1f827SSong Liu 135527e1f827SSong Liu VM_BUG_ON(addr & ~HPAGE_PMD_MASK); 135627e1f827SSong Liu 135727e1f827SSong Liu spin_lock(&khugepaged_mm_lock); 1358b26e2701SQi Zheng slot = mm_slot_lookup(mm_slots_hash, mm); 1359b26e2701SQi Zheng mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 136058ac9a89SZach O'Keefe if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) { 136127e1f827SSong Liu mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; 136258ac9a89SZach O'Keefe ret = true; 136358ac9a89SZach O'Keefe } 136427e1f827SSong Liu spin_unlock(&khugepaged_mm_lock); 136558ac9a89SZach O'Keefe return ret; 136627e1f827SSong Liu } 136727e1f827SSong Liu 136834488399SZach O'Keefe /* hpage must be locked, and mmap_lock must be held in write */ 136934488399SZach O'Keefe static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, 137034488399SZach O'Keefe pmd_t *pmdp, struct page *hpage) 137134488399SZach O'Keefe { 137234488399SZach O'Keefe struct vm_fault vmf = { 137334488399SZach O'Keefe .vma = vma, 137434488399SZach O'Keefe .address = addr, 137534488399SZach O'Keefe .flags = 0, 137634488399SZach O'Keefe .pmd = pmdp, 137734488399SZach O'Keefe }; 137834488399SZach O'Keefe 137934488399SZach O'Keefe VM_BUG_ON(!PageTransHuge(hpage)); 138034488399SZach O'Keefe mmap_assert_write_locked(vma->vm_mm); 138134488399SZach O'Keefe 138234488399SZach O'Keefe if (do_set_pmd(&vmf, hpage)) 138334488399SZach O'Keefe return SCAN_FAIL; 138434488399SZach O'Keefe 138534488399SZach O'Keefe get_page(hpage); 138634488399SZach O'Keefe return SCAN_SUCCEED; 138727e1f827SSong Liu } 138827e1f827SSong Liu 13898d3c106eSJann Horn /* 13908d3c106eSJann Horn * A note about locking: 13918d3c106eSJann Horn * Trying to take the page table spinlocks would be useless here because those 13928d3c106eSJann Horn * are only used to synchronize: 13938d3c106eSJann Horn * 13948d3c106eSJann Horn * - modifying terminal entries (ones that point to a data page, not to another 13958d3c106eSJann Horn * page table) 13968d3c106eSJann Horn * - installing *new* non-terminal entries 13978d3c106eSJann Horn * 13988d3c106eSJann Horn * Instead, we need roughly the same kind of protection as free_pgtables() or 13998d3c106eSJann Horn * mm_take_all_locks() (but only for a single VMA): 14008d3c106eSJann Horn * The mmap lock together with this VMA's rmap locks covers all paths towards 14018d3c106eSJann Horn * the page table entries we're messing with here, except for hardware page 14028d3c106eSJann Horn * table walks and lockless_pages_from_mm(). 14038d3c106eSJann Horn */ 1404e59a47b8SPasha Tatashin static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 1405e59a47b8SPasha Tatashin unsigned long addr, pmd_t *pmdp) 1406e59a47b8SPasha Tatashin { 1407e59a47b8SPasha Tatashin pmd_t pmd; 1408f268f6cfSJann Horn struct mmu_notifier_range range; 1409e59a47b8SPasha Tatashin 141080110bbfSPasha Tatashin mmap_assert_write_locked(mm); 14118d3c106eSJann Horn if (vma->vm_file) 14128d3c106eSJann Horn lockdep_assert_held_write(&vma->vm_file->f_mapping->i_mmap_rwsem); 14138d3c106eSJann Horn /* 14148d3c106eSJann Horn * All anon_vmas attached to the VMA have the same root and are 14158d3c106eSJann Horn * therefore locked by the same lock. 14168d3c106eSJann Horn */ 14178d3c106eSJann Horn if (vma->anon_vma) 14188d3c106eSJann Horn lockdep_assert_held_write(&vma->anon_vma->root->rwsem); 14198d3c106eSJann Horn 14207d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr, 1421f268f6cfSJann Horn addr + HPAGE_PMD_SIZE); 1422f268f6cfSJann Horn mmu_notifier_invalidate_range_start(&range); 1423e59a47b8SPasha Tatashin pmd = pmdp_collapse_flush(vma, addr, pmdp); 14242ba99c5eSJann Horn tlb_remove_table_sync_one(); 1425f268f6cfSJann Horn mmu_notifier_invalidate_range_end(&range); 1426e59a47b8SPasha Tatashin mm_dec_nr_ptes(mm); 142780110bbfSPasha Tatashin page_table_check_pte_clear_range(mm, addr, pmd); 1428e59a47b8SPasha Tatashin pte_free(mm, pmd_pgtable(pmd)); 1429e59a47b8SPasha Tatashin } 1430e59a47b8SPasha Tatashin 143127e1f827SSong Liu /** 1432336e6b53SAlex Shi * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at 1433336e6b53SAlex Shi * address haddr. 1434336e6b53SAlex Shi * 1435336e6b53SAlex Shi * @mm: process address space where collapse happens 1436336e6b53SAlex Shi * @addr: THP collapse address 143734488399SZach O'Keefe * @install_pmd: If a huge PMD should be installed 143827e1f827SSong Liu * 143927e1f827SSong Liu * This function checks whether all the PTEs in the PMD are pointing to the 144027e1f827SSong Liu * right THP. If so, retract the page table so the THP can refault in with 144134488399SZach O'Keefe * as pmd-mapped. Possibly install a huge PMD mapping the THP. 144227e1f827SSong Liu */ 144334488399SZach O'Keefe int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, 144434488399SZach O'Keefe bool install_pmd) 144527e1f827SSong Liu { 144627e1f827SSong Liu unsigned long haddr = addr & HPAGE_PMD_MASK; 144794d815b2SLiam R. Howlett struct vm_area_struct *vma = vma_lookup(mm, haddr); 1448119a5fc1SHugh Dickins struct page *hpage; 144927e1f827SSong Liu pte_t *start_pte, *pte; 1450e59a47b8SPasha Tatashin pmd_t *pmd; 145127e1f827SSong Liu spinlock_t *ptl; 145258ac9a89SZach O'Keefe int count = 0, result = SCAN_FAIL; 145327e1f827SSong Liu int i; 145427e1f827SSong Liu 145558ac9a89SZach O'Keefe mmap_assert_write_locked(mm); 145658ac9a89SZach O'Keefe 145734488399SZach O'Keefe /* Fast check before locking page if already PMD-mapped */ 145858ac9a89SZach O'Keefe result = find_pmd_or_thp_or_none(mm, haddr, &pmd); 145934488399SZach O'Keefe if (result == SCAN_PMD_MAPPED) 146034488399SZach O'Keefe return result; 146158ac9a89SZach O'Keefe 146227e1f827SSong Liu if (!vma || !vma->vm_file || 1463fef792a4SMiaohe Lin !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) 146434488399SZach O'Keefe return SCAN_VMA_CHECK; 146527e1f827SSong Liu 146627e1f827SSong Liu /* 1467a7f4e6e4SZach O'Keefe * If we are here, we've succeeded in replacing all the native pages 1468a7f4e6e4SZach O'Keefe * in the page cache with a single hugepage. If a mm were to fault-in 1469a7f4e6e4SZach O'Keefe * this memory (mapped by a suitably aligned VMA), we'd get the hugepage 1470a7f4e6e4SZach O'Keefe * and map it by a PMD, regardless of sysfs THP settings. As such, let's 1471a7f4e6e4SZach O'Keefe * analogously elide sysfs THP settings here. 147227e1f827SSong Liu */ 1473a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) 147434488399SZach O'Keefe return SCAN_VMA_CHECK; 147527e1f827SSong Liu 1476deb4c93aSPeter Xu /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ 1477deb4c93aSPeter Xu if (userfaultfd_wp(vma)) 147834488399SZach O'Keefe return SCAN_PTE_UFFD_WP; 1479deb4c93aSPeter Xu 1480119a5fc1SHugh Dickins hpage = find_lock_page(vma->vm_file->f_mapping, 1481119a5fc1SHugh Dickins linear_page_index(vma, haddr)); 1482119a5fc1SHugh Dickins if (!hpage) 148334488399SZach O'Keefe return SCAN_PAGE_NULL; 1484119a5fc1SHugh Dickins 148534488399SZach O'Keefe if (!PageHead(hpage)) { 148634488399SZach O'Keefe result = SCAN_FAIL; 1487119a5fc1SHugh Dickins goto drop_hpage; 148834488399SZach O'Keefe } 1489119a5fc1SHugh Dickins 149034488399SZach O'Keefe if (compound_order(hpage) != HPAGE_PMD_ORDER) { 149134488399SZach O'Keefe result = SCAN_PAGE_COMPOUND; 1492119a5fc1SHugh Dickins goto drop_hpage; 149334488399SZach O'Keefe } 1494780a4b6fSZach O'Keefe 149534488399SZach O'Keefe switch (result) { 149634488399SZach O'Keefe case SCAN_SUCCEED: 149734488399SZach O'Keefe break; 149834488399SZach O'Keefe case SCAN_PMD_NONE: 149934488399SZach O'Keefe /* 150034488399SZach O'Keefe * In MADV_COLLAPSE path, possible race with khugepaged where 150134488399SZach O'Keefe * all pte entries have been removed and pmd cleared. If so, 150234488399SZach O'Keefe * skip all the pte checks and just update the pmd mapping. 150334488399SZach O'Keefe */ 150434488399SZach O'Keefe goto maybe_install_pmd; 150534488399SZach O'Keefe default: 150627e1f827SSong Liu goto drop_hpage; 150734488399SZach O'Keefe } 150827e1f827SSong Liu 15098d3c106eSJann Horn /* 15108d3c106eSJann Horn * We need to lock the mapping so that from here on, only GUP-fast and 15118d3c106eSJann Horn * hardware page walks can access the parts of the page tables that 15128d3c106eSJann Horn * we're operating on. 15138d3c106eSJann Horn * See collapse_and_free_pmd(). 15148d3c106eSJann Horn */ 15158d3c106eSJann Horn i_mmap_lock_write(vma->vm_file->f_mapping); 15168d3c106eSJann Horn 15178d3c106eSJann Horn /* 15188d3c106eSJann Horn * This spinlock should be unnecessary: Nobody else should be accessing 15198d3c106eSJann Horn * the page tables under spinlock protection here, only 15208d3c106eSJann Horn * lockless_pages_from_mm() and the hardware page walker can access page 15218d3c106eSJann Horn * tables while all the high-level locks are held in write mode. 15228d3c106eSJann Horn */ 152327e1f827SSong Liu start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 152434488399SZach O'Keefe result = SCAN_FAIL; 152527e1f827SSong Liu 152627e1f827SSong Liu /* step 1: check all mapped PTEs are to the right huge page */ 152727e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 152827e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 152927e1f827SSong Liu struct page *page; 153027e1f827SSong Liu 153127e1f827SSong Liu /* empty pte, skip */ 153227e1f827SSong Liu if (pte_none(*pte)) 153327e1f827SSong Liu continue; 153427e1f827SSong Liu 153527e1f827SSong Liu /* page swapped out, abort */ 153634488399SZach O'Keefe if (!pte_present(*pte)) { 153734488399SZach O'Keefe result = SCAN_PTE_NON_PRESENT; 153827e1f827SSong Liu goto abort; 153934488399SZach O'Keefe } 154027e1f827SSong Liu 154127e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 15423218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 15433218f871SAlex Sierra page = NULL; 154427e1f827SSong Liu /* 1545119a5fc1SHugh Dickins * Note that uprobe, debugger, or MAP_PRIVATE may change the 1546119a5fc1SHugh Dickins * page table, but the new page will not be a subpage of hpage. 154727e1f827SSong Liu */ 1548119a5fc1SHugh Dickins if (hpage + i != page) 154927e1f827SSong Liu goto abort; 155027e1f827SSong Liu count++; 155127e1f827SSong Liu } 155227e1f827SSong Liu 155327e1f827SSong Liu /* step 2: adjust rmap */ 155427e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 155527e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 155627e1f827SSong Liu struct page *page; 155727e1f827SSong Liu 155827e1f827SSong Liu if (pte_none(*pte)) 155927e1f827SSong Liu continue; 156027e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 15613218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 15623218f871SAlex Sierra goto abort; 1563cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 156427e1f827SSong Liu } 156527e1f827SSong Liu 156627e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 156727e1f827SSong Liu 156827e1f827SSong Liu /* step 3: set proper refcount and mm_counters. */ 1569119a5fc1SHugh Dickins if (count) { 157027e1f827SSong Liu page_ref_sub(hpage, count); 157127e1f827SSong Liu add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); 157227e1f827SSong Liu } 157327e1f827SSong Liu 157434488399SZach O'Keefe /* step 4: remove pte entries */ 1575ab0c3f12SHugh Dickins /* we make no change to anon, but protect concurrent anon page lookup */ 1576ab0c3f12SHugh Dickins if (vma->anon_vma) 1577ab0c3f12SHugh Dickins anon_vma_lock_write(vma->anon_vma); 1578ab0c3f12SHugh Dickins 1579e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, haddr, pmd); 158034488399SZach O'Keefe 1581ab0c3f12SHugh Dickins if (vma->anon_vma) 1582ab0c3f12SHugh Dickins anon_vma_unlock_write(vma->anon_vma); 15838d3c106eSJann Horn i_mmap_unlock_write(vma->vm_file->f_mapping); 15848d3c106eSJann Horn 158534488399SZach O'Keefe maybe_install_pmd: 158634488399SZach O'Keefe /* step 5: install pmd entry */ 158734488399SZach O'Keefe result = install_pmd 158834488399SZach O'Keefe ? set_huge_pmd(vma, haddr, pmd, hpage) 158934488399SZach O'Keefe : SCAN_SUCCEED; 159034488399SZach O'Keefe 1591119a5fc1SHugh Dickins drop_hpage: 1592119a5fc1SHugh Dickins unlock_page(hpage); 1593119a5fc1SHugh Dickins put_page(hpage); 159434488399SZach O'Keefe return result; 159527e1f827SSong Liu 159627e1f827SSong Liu abort: 159727e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 15988d3c106eSJann Horn i_mmap_unlock_write(vma->vm_file->f_mapping); 1599119a5fc1SHugh Dickins goto drop_hpage; 160027e1f827SSong Liu } 160127e1f827SSong Liu 1602b26e2701SQi Zheng static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot) 160327e1f827SSong Liu { 1604b26e2701SQi Zheng struct mm_slot *slot = &mm_slot->slot; 1605b26e2701SQi Zheng struct mm_struct *mm = slot->mm; 160627e1f827SSong Liu int i; 160727e1f827SSong Liu 160827e1f827SSong Liu if (likely(mm_slot->nr_pte_mapped_thp == 0)) 16090edf61e5SMiaohe Lin return; 161027e1f827SSong Liu 1611d8ed45c5SMichel Lespinasse if (!mmap_write_trylock(mm)) 16120edf61e5SMiaohe Lin return; 161327e1f827SSong Liu 16147d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 161527e1f827SSong Liu goto out; 161627e1f827SSong Liu 161727e1f827SSong Liu for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) 161834488399SZach O'Keefe collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i], false); 161927e1f827SSong Liu 162027e1f827SSong Liu out: 162127e1f827SSong Liu mm_slot->nr_pte_mapped_thp = 0; 1622d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 162327e1f827SSong Liu } 162427e1f827SSong Liu 162534488399SZach O'Keefe static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff, 162634488399SZach O'Keefe struct mm_struct *target_mm, 162734488399SZach O'Keefe unsigned long target_addr, struct page *hpage, 162834488399SZach O'Keefe struct collapse_control *cc) 1629f3f0e1d2SKirill A. Shutemov { 1630f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 163134488399SZach O'Keefe int target_result = SCAN_FAIL; 1632f3f0e1d2SKirill A. Shutemov 1633f3f0e1d2SKirill A. Shutemov i_mmap_lock_write(mapping); 1634f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 163534488399SZach O'Keefe int result = SCAN_FAIL; 163634488399SZach O'Keefe struct mm_struct *mm = NULL; 163734488399SZach O'Keefe unsigned long addr = 0; 163834488399SZach O'Keefe pmd_t *pmd; 163934488399SZach O'Keefe bool is_target = false; 164034488399SZach O'Keefe 164127e1f827SSong Liu /* 164227e1f827SSong Liu * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 164327e1f827SSong Liu * got written to. These VMAs are likely not worth investing 16443e4e28c5SMichel Lespinasse * mmap_write_lock(mm) as PMD-mapping is likely to be split 164527e1f827SSong Liu * later. 164627e1f827SSong Liu * 164736ee2c78SMiaohe Lin * Note that vma->anon_vma check is racy: it can be set up after 1648c1e8d7c6SMichel Lespinasse * the check but before we took mmap_lock by the fault path. 164927e1f827SSong Liu * But page lock would prevent establishing any new ptes of the 165027e1f827SSong Liu * page, so we are safe. 165127e1f827SSong Liu * 165227e1f827SSong Liu * An alternative would be drop the check, but check that page 165327e1f827SSong Liu * table is clear before calling pmdp_collapse_flush() under 165427e1f827SSong Liu * ptl. It has higher chance to recover THP for the VMA, but 16558d3c106eSJann Horn * has higher cost too. It would also probably require locking 16568d3c106eSJann Horn * the anon_vma. 165727e1f827SSong Liu */ 1658023f47a8SJann Horn if (READ_ONCE(vma->anon_vma)) { 165934488399SZach O'Keefe result = SCAN_PAGE_ANON; 166034488399SZach O'Keefe goto next; 166134488399SZach O'Keefe } 1662f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 166334488399SZach O'Keefe if (addr & ~HPAGE_PMD_MASK || 166434488399SZach O'Keefe vma->vm_end < addr + HPAGE_PMD_SIZE) { 166534488399SZach O'Keefe result = SCAN_VMA_CHECK; 166634488399SZach O'Keefe goto next; 166734488399SZach O'Keefe } 166818e77600SHugh Dickins mm = vma->vm_mm; 166934488399SZach O'Keefe is_target = mm == target_mm && addr == target_addr; 167034488399SZach O'Keefe result = find_pmd_or_thp_or_none(mm, addr, &pmd); 167134488399SZach O'Keefe if (result != SCAN_SUCCEED) 167234488399SZach O'Keefe goto next; 1673f3f0e1d2SKirill A. Shutemov /* 1674c1e8d7c6SMichel Lespinasse * We need exclusive mmap_lock to retract page table. 167527e1f827SSong Liu * 167627e1f827SSong Liu * We use trylock due to lock inversion: we need to acquire 1677c1e8d7c6SMichel Lespinasse * mmap_lock while holding page lock. Fault path does it in 167827e1f827SSong Liu * reverse order. Trylock is a way to avoid deadlock. 167934488399SZach O'Keefe * 168034488399SZach O'Keefe * Also, it's not MADV_COLLAPSE's job to collapse other 168134488399SZach O'Keefe * mappings - let khugepaged take care of them later. 1682f3f0e1d2SKirill A. Shutemov */ 168334488399SZach O'Keefe result = SCAN_PTE_MAPPED_HUGEPAGE; 168434488399SZach O'Keefe if ((cc->is_khugepaged || is_target) && 168534488399SZach O'Keefe mmap_write_trylock(mm)) { 1686deb4c93aSPeter Xu /* 1687023f47a8SJann Horn * Re-check whether we have an ->anon_vma, because 1688023f47a8SJann Horn * collapse_and_free_pmd() requires that either no 1689023f47a8SJann Horn * ->anon_vma exists or the anon_vma is locked. 1690023f47a8SJann Horn * We already checked ->anon_vma above, but that check 1691023f47a8SJann Horn * is racy because ->anon_vma can be populated under the 1692023f47a8SJann Horn * mmap lock in read mode. 1693023f47a8SJann Horn */ 1694023f47a8SJann Horn if (vma->anon_vma) { 1695023f47a8SJann Horn result = SCAN_PAGE_ANON; 1696023f47a8SJann Horn goto unlock_next; 1697023f47a8SJann Horn } 1698023f47a8SJann Horn /* 1699deb4c93aSPeter Xu * When a vma is registered with uffd-wp, we can't 1700deb4c93aSPeter Xu * recycle the pmd pgtable because there can be pte 1701deb4c93aSPeter Xu * markers installed. Skip it only, so the rest mm/vma 1702deb4c93aSPeter Xu * can still have the same file mapped hugely, however 1703deb4c93aSPeter Xu * it'll always mapped in small page size for uffd-wp 1704deb4c93aSPeter Xu * registered ranges. 1705deb4c93aSPeter Xu */ 170634488399SZach O'Keefe if (hpage_collapse_test_exit(mm)) { 170734488399SZach O'Keefe result = SCAN_ANY_PROCESS; 170834488399SZach O'Keefe goto unlock_next; 1709f3f0e1d2SKirill A. Shutemov } 171034488399SZach O'Keefe if (userfaultfd_wp(vma)) { 171134488399SZach O'Keefe result = SCAN_PTE_UFFD_WP; 171234488399SZach O'Keefe goto unlock_next; 171334488399SZach O'Keefe } 171434488399SZach O'Keefe collapse_and_free_pmd(mm, vma, addr, pmd); 171534488399SZach O'Keefe if (!cc->is_khugepaged && is_target) 171634488399SZach O'Keefe result = set_huge_pmd(vma, addr, pmd, hpage); 171734488399SZach O'Keefe else 171834488399SZach O'Keefe result = SCAN_SUCCEED; 171934488399SZach O'Keefe 172034488399SZach O'Keefe unlock_next: 172134488399SZach O'Keefe mmap_write_unlock(mm); 172234488399SZach O'Keefe goto next; 172334488399SZach O'Keefe } 172434488399SZach O'Keefe /* 172534488399SZach O'Keefe * Calling context will handle target mm/addr. Otherwise, let 172634488399SZach O'Keefe * khugepaged try again later. 172734488399SZach O'Keefe */ 172834488399SZach O'Keefe if (!is_target) { 172934488399SZach O'Keefe khugepaged_add_pte_mapped_thp(mm, addr); 173034488399SZach O'Keefe continue; 173134488399SZach O'Keefe } 173234488399SZach O'Keefe next: 173334488399SZach O'Keefe if (is_target) 173434488399SZach O'Keefe target_result = result; 1735f3f0e1d2SKirill A. Shutemov } 1736f3f0e1d2SKirill A. Shutemov i_mmap_unlock_write(mapping); 173734488399SZach O'Keefe return target_result; 1738f3f0e1d2SKirill A. Shutemov } 1739f3f0e1d2SKirill A. Shutemov 1740f3f0e1d2SKirill A. Shutemov /** 174199cb0dbdSSong Liu * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1742f3f0e1d2SKirill A. Shutemov * 1743336e6b53SAlex Shi * @mm: process address space where collapse happens 174434488399SZach O'Keefe * @addr: virtual collapse start address 1745336e6b53SAlex Shi * @file: file that collapse on 1746336e6b53SAlex Shi * @start: collapse start address 17479710a78aSZach O'Keefe * @cc: collapse context and scratchpad 1748336e6b53SAlex Shi * 1749f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 175087c460a0SHugh Dickins * - allocate and lock a new huge page; 175177da9389SMatthew Wilcox * - scan page cache replacing old pages with the new one 175299cb0dbdSSong Liu * + swap/gup in pages if necessary; 1753f3f0e1d2SKirill A. Shutemov * + fill in gaps; 175477da9389SMatthew Wilcox * + keep old pages around in case rollback is required; 175577da9389SMatthew Wilcox * - if replacing succeeds: 1756f3f0e1d2SKirill A. Shutemov * + copy data over; 1757f3f0e1d2SKirill A. Shutemov * + free old pages; 175887c460a0SHugh Dickins * + unlock huge page; 1759f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1760f3f0e1d2SKirill A. Shutemov * + put all pages back and unfreeze them; 176177da9389SMatthew Wilcox * + restore gaps in the page cache; 176287c460a0SHugh Dickins * + unlock and free huge page; 1763f3f0e1d2SKirill A. Shutemov */ 176434488399SZach O'Keefe static int collapse_file(struct mm_struct *mm, unsigned long addr, 1765579c571eSSong Liu struct file *file, pgoff_t start, 176634488399SZach O'Keefe struct collapse_control *cc) 1767f3f0e1d2SKirill A. Shutemov { 1768579c571eSSong Liu struct address_space *mapping = file->f_mapping; 176950ad2f24SZach O'Keefe struct page *hpage; 17704c9473e8SGautam Menghani pgoff_t index = 0, end = start + HPAGE_PMD_NR; 1771f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 177277da9389SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1773f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 177499cb0dbdSSong Liu bool is_shmem = shmem_file(file); 17754c9473e8SGautam Menghani int nr = 0; 1776f3f0e1d2SKirill A. Shutemov 177799cb0dbdSSong Liu VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1778f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1779f3f0e1d2SKirill A. Shutemov 178050ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 17819710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1782f3f0e1d2SKirill A. Shutemov goto out; 1783f3f0e1d2SKirill A. Shutemov 17846b24ca4aSMatthew Wilcox (Oracle) /* 17856b24ca4aSMatthew Wilcox (Oracle) * Ensure we have slots for all the pages in the range. This is 17866b24ca4aSMatthew Wilcox (Oracle) * almost certainly a no-op because most of the pages must be present 17876b24ca4aSMatthew Wilcox (Oracle) */ 178895feeabbSHugh Dickins do { 178995feeabbSHugh Dickins xas_lock_irq(&xas); 179095feeabbSHugh Dickins xas_create_range(&xas); 179195feeabbSHugh Dickins if (!xas_error(&xas)) 179295feeabbSHugh Dickins break; 179395feeabbSHugh Dickins xas_unlock_irq(&xas); 179495feeabbSHugh Dickins if (!xas_nomem(&xas, GFP_KERNEL)) { 179595feeabbSHugh Dickins result = SCAN_FAIL; 179695feeabbSHugh Dickins goto out; 179795feeabbSHugh Dickins } 179895feeabbSHugh Dickins } while (1); 179995feeabbSHugh Dickins 180050ad2f24SZach O'Keefe __SetPageLocked(hpage); 180199cb0dbdSSong Liu if (is_shmem) 180250ad2f24SZach O'Keefe __SetPageSwapBacked(hpage); 180350ad2f24SZach O'Keefe hpage->index = start; 180450ad2f24SZach O'Keefe hpage->mapping = mapping; 1805f3f0e1d2SKirill A. Shutemov 1806f3f0e1d2SKirill A. Shutemov /* 180750ad2f24SZach O'Keefe * At this point the hpage is locked and not up-to-date. 180887c460a0SHugh Dickins * It's safe to insert it into the page cache, because nobody would 180987c460a0SHugh Dickins * be able to map it or use it in another way until we unlock it. 1810f3f0e1d2SKirill A. Shutemov */ 1811f3f0e1d2SKirill A. Shutemov 181277da9389SMatthew Wilcox xas_set(&xas, start); 181377da9389SMatthew Wilcox for (index = start; index < end; index++) { 181477da9389SMatthew Wilcox struct page *page = xas_next(&xas); 181564ab3195SVishal Moola (Oracle) struct folio *folio; 181677da9389SMatthew Wilcox 181777da9389SMatthew Wilcox VM_BUG_ON(index != xas.xa_index); 181899cb0dbdSSong Liu if (is_shmem) { 181977da9389SMatthew Wilcox if (!page) { 1820701270faSHugh Dickins /* 182199cb0dbdSSong Liu * Stop if extent has been truncated or 182299cb0dbdSSong Liu * hole-punched, and is now completely 182399cb0dbdSSong Liu * empty. 1824701270faSHugh Dickins */ 1825701270faSHugh Dickins if (index == start) { 1826701270faSHugh Dickins if (!xas_next_entry(&xas, end - 1)) { 1827701270faSHugh Dickins result = SCAN_TRUNCATED; 1828042a3082SHugh Dickins goto xa_locked; 1829701270faSHugh Dickins } 1830701270faSHugh Dickins xas_set(&xas, index); 1831701270faSHugh Dickins } 183277da9389SMatthew Wilcox if (!shmem_charge(mapping->host, 1)) { 1833f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1834042a3082SHugh Dickins goto xa_locked; 1835f3f0e1d2SKirill A. Shutemov } 183650ad2f24SZach O'Keefe xas_store(&xas, hpage); 183777da9389SMatthew Wilcox nr_none++; 183877da9389SMatthew Wilcox continue; 1839f3f0e1d2SKirill A. Shutemov } 1840f3f0e1d2SKirill A. Shutemov 18413159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 184277da9389SMatthew Wilcox xas_unlock_irq(&xas); 1843f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 18447459c149SMatthew Wilcox (Oracle) if (shmem_get_folio(mapping->host, index, 18457459c149SMatthew Wilcox (Oracle) &folio, SGP_NOALLOC)) { 1846f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 184777da9389SMatthew Wilcox goto xa_unlocked; 1848f3f0e1d2SKirill A. Shutemov } 18497459c149SMatthew Wilcox (Oracle) page = folio_file_page(folio, index); 1850f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1851f3f0e1d2SKirill A. Shutemov get_page(page); 1852042a3082SHugh Dickins xas_unlock_irq(&xas); 1853f3f0e1d2SKirill A. Shutemov } else { 1854f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1855042a3082SHugh Dickins goto xa_locked; 1856f3f0e1d2SKirill A. Shutemov } 185799cb0dbdSSong Liu } else { /* !is_shmem */ 185899cb0dbdSSong Liu if (!page || xa_is_value(page)) { 185999cb0dbdSSong Liu xas_unlock_irq(&xas); 186099cb0dbdSSong Liu page_cache_sync_readahead(mapping, &file->f_ra, 186199cb0dbdSSong Liu file, index, 1862e5a59d30SDavid Howells end - index); 186399cb0dbdSSong Liu /* drain pagevecs to help isolate_lru_page() */ 186499cb0dbdSSong Liu lru_add_drain(); 186599cb0dbdSSong Liu page = find_lock_page(mapping, index); 186699cb0dbdSSong Liu if (unlikely(page == NULL)) { 186799cb0dbdSSong Liu result = SCAN_FAIL; 186899cb0dbdSSong Liu goto xa_unlocked; 186999cb0dbdSSong Liu } 187075f36069SSong Liu } else if (PageDirty(page)) { 187175f36069SSong Liu /* 187275f36069SSong Liu * khugepaged only works on read-only fd, 187375f36069SSong Liu * so this page is dirty because it hasn't 187475f36069SSong Liu * been flushed since first write. There 187575f36069SSong Liu * won't be new dirty pages. 187675f36069SSong Liu * 187775f36069SSong Liu * Trigger async flush here and hope the 187875f36069SSong Liu * writeback is done when khugepaged 187975f36069SSong Liu * revisits this page. 188075f36069SSong Liu * 188175f36069SSong Liu * This is a one-off situation. We are not 188275f36069SSong Liu * forcing writeback in loop. 188375f36069SSong Liu */ 188475f36069SSong Liu xas_unlock_irq(&xas); 188575f36069SSong Liu filemap_flush(mapping); 188675f36069SSong Liu result = SCAN_FAIL; 188775f36069SSong Liu goto xa_unlocked; 188874c42e1bSRongwei Wang } else if (PageWriteback(page)) { 188974c42e1bSRongwei Wang xas_unlock_irq(&xas); 189074c42e1bSRongwei Wang result = SCAN_FAIL; 189174c42e1bSRongwei Wang goto xa_unlocked; 189299cb0dbdSSong Liu } else if (trylock_page(page)) { 189399cb0dbdSSong Liu get_page(page); 189499cb0dbdSSong Liu xas_unlock_irq(&xas); 189599cb0dbdSSong Liu } else { 189699cb0dbdSSong Liu result = SCAN_PAGE_LOCK; 189799cb0dbdSSong Liu goto xa_locked; 189899cb0dbdSSong Liu } 189999cb0dbdSSong Liu } 1900f3f0e1d2SKirill A. Shutemov 1901f3f0e1d2SKirill A. Shutemov /* 1902b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1903f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1904f3f0e1d2SKirill A. Shutemov */ 1905f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 19064655e5e5SSong Liu 19074655e5e5SSong Liu /* make sure the page is up to date */ 19084655e5e5SSong Liu if (unlikely(!PageUptodate(page))) { 19094655e5e5SSong Liu result = SCAN_FAIL; 19104655e5e5SSong Liu goto out_unlock; 19114655e5e5SSong Liu } 191206a5e126SHugh Dickins 191306a5e126SHugh Dickins /* 191406a5e126SHugh Dickins * If file was truncated then extended, or hole-punched, before 191506a5e126SHugh Dickins * we locked the first page, then a THP might be there already. 191658ac9a89SZach O'Keefe * This will be discovered on the first iteration. 191706a5e126SHugh Dickins */ 191806a5e126SHugh Dickins if (PageTransCompound(page)) { 191958ac9a89SZach O'Keefe struct page *head = compound_head(page); 192058ac9a89SZach O'Keefe 192158ac9a89SZach O'Keefe result = compound_order(head) == HPAGE_PMD_ORDER && 192258ac9a89SZach O'Keefe head->index == start 192358ac9a89SZach O'Keefe /* Maybe PMD-mapped */ 192458ac9a89SZach O'Keefe ? SCAN_PTE_MAPPED_HUGEPAGE 192558ac9a89SZach O'Keefe : SCAN_PAGE_COMPOUND; 192606a5e126SHugh Dickins goto out_unlock; 192706a5e126SHugh Dickins } 1928f3f0e1d2SKirill A. Shutemov 192964ab3195SVishal Moola (Oracle) folio = page_folio(page); 193064ab3195SVishal Moola (Oracle) 193164ab3195SVishal Moola (Oracle) if (folio_mapping(folio) != mapping) { 1932f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1933f3f0e1d2SKirill A. Shutemov goto out_unlock; 1934f3f0e1d2SKirill A. Shutemov } 1935f3f0e1d2SKirill A. Shutemov 193664ab3195SVishal Moola (Oracle) if (!is_shmem && (folio_test_dirty(folio) || 193764ab3195SVishal Moola (Oracle) folio_test_writeback(folio))) { 19384655e5e5SSong Liu /* 19394655e5e5SSong Liu * khugepaged only works on read-only fd, so this 19404655e5e5SSong Liu * page is dirty because it hasn't been flushed 19414655e5e5SSong Liu * since first write. 19424655e5e5SSong Liu */ 19434655e5e5SSong Liu result = SCAN_FAIL; 19444655e5e5SSong Liu goto out_unlock; 19454655e5e5SSong Liu } 19464655e5e5SSong Liu 194764ab3195SVishal Moola (Oracle) if (folio_isolate_lru(folio)) { 1948f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1949042a3082SHugh Dickins goto out_unlock; 1950f3f0e1d2SKirill A. Shutemov } 1951f3f0e1d2SKirill A. Shutemov 195264ab3195SVishal Moola (Oracle) if (folio_has_private(folio) && 195364ab3195SVishal Moola (Oracle) !filemap_release_folio(folio, GFP_KERNEL)) { 195499cb0dbdSSong Liu result = SCAN_PAGE_HAS_PRIVATE; 195564ab3195SVishal Moola (Oracle) folio_putback_lru(folio); 195699cb0dbdSSong Liu goto out_unlock; 195799cb0dbdSSong Liu } 195899cb0dbdSSong Liu 195964ab3195SVishal Moola (Oracle) if (folio_mapped(folio)) 196064ab3195SVishal Moola (Oracle) try_to_unmap(folio, 1961869f7ee6SMatthew Wilcox (Oracle) TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); 1962f3f0e1d2SKirill A. Shutemov 196377da9389SMatthew Wilcox xas_lock_irq(&xas); 196477da9389SMatthew Wilcox xas_set(&xas, index); 1965f3f0e1d2SKirill A. Shutemov 196677da9389SMatthew Wilcox VM_BUG_ON_PAGE(page != xas_load(&xas), page); 1967f3f0e1d2SKirill A. Shutemov 1968f3f0e1d2SKirill A. Shutemov /* 1969f3f0e1d2SKirill A. Shutemov * The page is expected to have page_count() == 3: 1970f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 197177da9389SMatthew Wilcox * - one reference from page cache; 1972f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1973f3f0e1d2SKirill A. Shutemov */ 1974f3f0e1d2SKirill A. Shutemov if (!page_ref_freeze(page, 3)) { 1975f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1976042a3082SHugh Dickins xas_unlock_irq(&xas); 1977042a3082SHugh Dickins putback_lru_page(page); 1978042a3082SHugh Dickins goto out_unlock; 1979f3f0e1d2SKirill A. Shutemov } 1980f3f0e1d2SKirill A. Shutemov 1981f3f0e1d2SKirill A. Shutemov /* 1982f3f0e1d2SKirill A. Shutemov * Add the page to the list to be able to undo the collapse if 1983f3f0e1d2SKirill A. Shutemov * something go wrong. 1984f3f0e1d2SKirill A. Shutemov */ 1985f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 1986f3f0e1d2SKirill A. Shutemov 1987f3f0e1d2SKirill A. Shutemov /* Finally, replace with the new page. */ 198850ad2f24SZach O'Keefe xas_store(&xas, hpage); 1989f3f0e1d2SKirill A. Shutemov continue; 1990f3f0e1d2SKirill A. Shutemov out_unlock: 1991f3f0e1d2SKirill A. Shutemov unlock_page(page); 1992f3f0e1d2SKirill A. Shutemov put_page(page); 1993042a3082SHugh Dickins goto xa_unlocked; 1994f3f0e1d2SKirill A. Shutemov } 199550ad2f24SZach O'Keefe nr = thp_nr_pages(hpage); 1996f3f0e1d2SKirill A. Shutemov 199799cb0dbdSSong Liu if (is_shmem) 199850ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr); 199909d91cdaSSong Liu else { 200050ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr); 200109d91cdaSSong Liu filemap_nr_thps_inc(mapping); 2002eb6ecbedSCollin Fijalkovich /* 2003eb6ecbedSCollin Fijalkovich * Paired with smp_mb() in do_dentry_open() to ensure 2004eb6ecbedSCollin Fijalkovich * i_writecount is up to date and the update to nr_thps is 2005eb6ecbedSCollin Fijalkovich * visible. Ensures the page cache will be truncated if the 2006eb6ecbedSCollin Fijalkovich * file is opened writable. 2007eb6ecbedSCollin Fijalkovich */ 2008eb6ecbedSCollin Fijalkovich smp_mb(); 2009eb6ecbedSCollin Fijalkovich if (inode_is_open_for_write(mapping->host)) { 2010eb6ecbedSCollin Fijalkovich result = SCAN_FAIL; 201150ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr); 2012eb6ecbedSCollin Fijalkovich filemap_nr_thps_dec(mapping); 2013eb6ecbedSCollin Fijalkovich goto xa_locked; 2014eb6ecbedSCollin Fijalkovich } 201509d91cdaSSong Liu } 201699cb0dbdSSong Liu 2017042a3082SHugh Dickins if (nr_none) { 201850ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none); 20192f55f070SMiaohe Lin /* nr_none is always 0 for non-shmem. */ 202050ad2f24SZach O'Keefe __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none); 2021042a3082SHugh Dickins } 2022042a3082SHugh Dickins 20236b24ca4aSMatthew Wilcox (Oracle) /* Join all the small entries into a single multi-index entry */ 20246b24ca4aSMatthew Wilcox (Oracle) xas_set_order(&xas, start, HPAGE_PMD_ORDER); 202550ad2f24SZach O'Keefe xas_store(&xas, hpage); 2026042a3082SHugh Dickins xa_locked: 2027042a3082SHugh Dickins xas_unlock_irq(&xas); 202877da9389SMatthew Wilcox xa_unlocked: 2029042a3082SHugh Dickins 20306d9df8a5SHugh Dickins /* 20316d9df8a5SHugh Dickins * If collapse is successful, flush must be done now before copying. 20326d9df8a5SHugh Dickins * If collapse is unsuccessful, does flush actually need to be done? 20336d9df8a5SHugh Dickins * Do it anyway, to clear the state. 20346d9df8a5SHugh Dickins */ 20356d9df8a5SHugh Dickins try_to_unmap_flush(); 20366d9df8a5SHugh Dickins 2037f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 203877da9389SMatthew Wilcox struct page *page, *tmp; 2039284a344eSVishal Moola (Oracle) struct folio *folio; 2040f3f0e1d2SKirill A. Shutemov 2041f3f0e1d2SKirill A. Shutemov /* 204277da9389SMatthew Wilcox * Replacing old pages with new one has succeeded, now we 204377da9389SMatthew Wilcox * need to copy the content and free the old pages. 2044f3f0e1d2SKirill A. Shutemov */ 20452af8ff29SHugh Dickins index = start; 2046f3f0e1d2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, &pagelist, lru) { 20472af8ff29SHugh Dickins while (index < page->index) { 204850ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 20492af8ff29SHugh Dickins index++; 20502af8ff29SHugh Dickins } 205150ad2f24SZach O'Keefe copy_highpage(hpage + (page->index % HPAGE_PMD_NR), 2052f3f0e1d2SKirill A. Shutemov page); 2053f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 2054f3f0e1d2SKirill A. Shutemov page->mapping = NULL; 2055042a3082SHugh Dickins page_ref_unfreeze(page, 1); 2056f3f0e1d2SKirill A. Shutemov ClearPageActive(page); 2057f3f0e1d2SKirill A. Shutemov ClearPageUnevictable(page); 2058042a3082SHugh Dickins unlock_page(page); 2059f3f0e1d2SKirill A. Shutemov put_page(page); 20602af8ff29SHugh Dickins index++; 20612af8ff29SHugh Dickins } 20622af8ff29SHugh Dickins while (index < end) { 206350ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 20642af8ff29SHugh Dickins index++; 2065f3f0e1d2SKirill A. Shutemov } 2066f3f0e1d2SKirill A. Shutemov 2067284a344eSVishal Moola (Oracle) folio = page_folio(hpage); 2068284a344eSVishal Moola (Oracle) folio_mark_uptodate(folio); 2069284a344eSVishal Moola (Oracle) folio_ref_add(folio, HPAGE_PMD_NR - 1); 2070284a344eSVishal Moola (Oracle) 20716058eaecSJohannes Weiner if (is_shmem) 2072284a344eSVishal Moola (Oracle) folio_mark_dirty(folio); 2073284a344eSVishal Moola (Oracle) folio_add_lru(folio); 2074f3f0e1d2SKirill A. Shutemov 2075042a3082SHugh Dickins /* 2076042a3082SHugh Dickins * Remove pte page tables, so we can re-fault the page as huge. 2077042a3082SHugh Dickins */ 207834488399SZach O'Keefe result = retract_page_tables(mapping, start, mm, addr, hpage, 207934488399SZach O'Keefe cc); 208050ad2f24SZach O'Keefe unlock_page(hpage); 208150ad2f24SZach O'Keefe hpage = NULL; 2082f3f0e1d2SKirill A. Shutemov } else { 208377da9389SMatthew Wilcox struct page *page; 2084aaa52e34SHugh Dickins 208577da9389SMatthew Wilcox /* Something went wrong: roll back page cache changes */ 208677da9389SMatthew Wilcox xas_lock_irq(&xas); 20872f55f070SMiaohe Lin if (nr_none) { 2088aaa52e34SHugh Dickins mapping->nrpages -= nr_none; 2089aaa52e34SHugh Dickins shmem_uncharge(mapping->host, nr_none); 20902f55f070SMiaohe Lin } 2091aaa52e34SHugh Dickins 209277da9389SMatthew Wilcox xas_set(&xas, start); 209377da9389SMatthew Wilcox xas_for_each(&xas, page, end - 1) { 2094f3f0e1d2SKirill A. Shutemov page = list_first_entry_or_null(&pagelist, 2095f3f0e1d2SKirill A. Shutemov struct page, lru); 209677da9389SMatthew Wilcox if (!page || xas.xa_index < page->index) { 2097f3f0e1d2SKirill A. Shutemov if (!nr_none) 2098f3f0e1d2SKirill A. Shutemov break; 2099f3f0e1d2SKirill A. Shutemov nr_none--; 210059749e6cSJohannes Weiner /* Put holes back where they were */ 210177da9389SMatthew Wilcox xas_store(&xas, NULL); 2102f3f0e1d2SKirill A. Shutemov continue; 2103f3f0e1d2SKirill A. Shutemov } 2104f3f0e1d2SKirill A. Shutemov 210577da9389SMatthew Wilcox VM_BUG_ON_PAGE(page->index != xas.xa_index, page); 2106f3f0e1d2SKirill A. Shutemov 2107f3f0e1d2SKirill A. Shutemov /* Unfreeze the page. */ 2108f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 2109f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(page, 2); 211077da9389SMatthew Wilcox xas_store(&xas, page); 211177da9389SMatthew Wilcox xas_pause(&xas); 211277da9389SMatthew Wilcox xas_unlock_irq(&xas); 2113f3f0e1d2SKirill A. Shutemov unlock_page(page); 2114042a3082SHugh Dickins putback_lru_page(page); 211577da9389SMatthew Wilcox xas_lock_irq(&xas); 2116f3f0e1d2SKirill A. Shutemov } 2117f3f0e1d2SKirill A. Shutemov VM_BUG_ON(nr_none); 211877da9389SMatthew Wilcox xas_unlock_irq(&xas); 2119f3f0e1d2SKirill A. Shutemov 212050ad2f24SZach O'Keefe hpage->mapping = NULL; 2121f3f0e1d2SKirill A. Shutemov } 2122042a3082SHugh Dickins 212350ad2f24SZach O'Keefe if (hpage) 212450ad2f24SZach O'Keefe unlock_page(hpage); 2125f3f0e1d2SKirill A. Shutemov out: 2126f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 212750ad2f24SZach O'Keefe if (hpage) { 212850ad2f24SZach O'Keefe mem_cgroup_uncharge(page_folio(hpage)); 212950ad2f24SZach O'Keefe put_page(hpage); 2130c6a7f445SYang Shi } 21314c9473e8SGautam Menghani 21324c9473e8SGautam Menghani trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result); 213350ad2f24SZach O'Keefe return result; 2134f3f0e1d2SKirill A. Shutemov } 2135f3f0e1d2SKirill A. Shutemov 213634488399SZach O'Keefe static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, 213734488399SZach O'Keefe struct file *file, pgoff_t start, 213834488399SZach O'Keefe struct collapse_control *cc) 2139f3f0e1d2SKirill A. Shutemov { 2140f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 2141579c571eSSong Liu struct address_space *mapping = file->f_mapping; 214285b392dbSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, start); 2143f3f0e1d2SKirill A. Shutemov int present, swap; 2144f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 2145f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 2146f3f0e1d2SKirill A. Shutemov 2147f3f0e1d2SKirill A. Shutemov present = 0; 2148f3f0e1d2SKirill A. Shutemov swap = 0; 214934d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 2150e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 2151f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 215285b392dbSMatthew Wilcox xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 215385b392dbSMatthew Wilcox if (xas_retry(&xas, page)) 2154f3f0e1d2SKirill A. Shutemov continue; 2155f3f0e1d2SKirill A. Shutemov 215685b392dbSMatthew Wilcox if (xa_is_value(page)) { 2157d8ea7cc8SZach O'Keefe ++swap; 2158d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 2159d8ea7cc8SZach O'Keefe swap > khugepaged_max_ptes_swap) { 2160f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 2161e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 2162f3f0e1d2SKirill A. Shutemov break; 2163f3f0e1d2SKirill A. Shutemov } 2164f3f0e1d2SKirill A. Shutemov continue; 2165f3f0e1d2SKirill A. Shutemov } 2166f3f0e1d2SKirill A. Shutemov 21676b24ca4aSMatthew Wilcox (Oracle) /* 216858ac9a89SZach O'Keefe * TODO: khugepaged should compact smaller compound pages 21696b24ca4aSMatthew Wilcox (Oracle) * into a PMD sized page 21706b24ca4aSMatthew Wilcox (Oracle) */ 2171f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 217258ac9a89SZach O'Keefe struct page *head = compound_head(page); 217358ac9a89SZach O'Keefe 217458ac9a89SZach O'Keefe result = compound_order(head) == HPAGE_PMD_ORDER && 217558ac9a89SZach O'Keefe head->index == start 217658ac9a89SZach O'Keefe /* Maybe PMD-mapped */ 217758ac9a89SZach O'Keefe ? SCAN_PTE_MAPPED_HUGEPAGE 217858ac9a89SZach O'Keefe : SCAN_PAGE_COMPOUND; 217958ac9a89SZach O'Keefe /* 218058ac9a89SZach O'Keefe * For SCAN_PTE_MAPPED_HUGEPAGE, further processing 218158ac9a89SZach O'Keefe * by the caller won't touch the page cache, and so 218258ac9a89SZach O'Keefe * it's safe to skip LRU and refcount checks before 218358ac9a89SZach O'Keefe * returning. 218458ac9a89SZach O'Keefe */ 2185f3f0e1d2SKirill A. Shutemov break; 2186f3f0e1d2SKirill A. Shutemov } 2187f3f0e1d2SKirill A. Shutemov 2188f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 21897d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 2190f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 2191f3f0e1d2SKirill A. Shutemov break; 2192f3f0e1d2SKirill A. Shutemov } 219334d6b470SZach O'Keefe cc->node_load[node]++; 2194f3f0e1d2SKirill A. Shutemov 2195f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 2196f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 2197f3f0e1d2SKirill A. Shutemov break; 2198f3f0e1d2SKirill A. Shutemov } 2199f3f0e1d2SKirill A. Shutemov 220099cb0dbdSSong Liu if (page_count(page) != 220199cb0dbdSSong Liu 1 + page_mapcount(page) + page_has_private(page)) { 2202f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 2203f3f0e1d2SKirill A. Shutemov break; 2204f3f0e1d2SKirill A. Shutemov } 2205f3f0e1d2SKirill A. Shutemov 2206f3f0e1d2SKirill A. Shutemov /* 2207f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 2208f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 2209f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 2210f3f0e1d2SKirill A. Shutemov */ 2211f3f0e1d2SKirill A. Shutemov 2212f3f0e1d2SKirill A. Shutemov present++; 2213f3f0e1d2SKirill A. Shutemov 2214f3f0e1d2SKirill A. Shutemov if (need_resched()) { 221585b392dbSMatthew Wilcox xas_pause(&xas); 2216f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 2217f3f0e1d2SKirill A. Shutemov } 2218f3f0e1d2SKirill A. Shutemov } 2219f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 2220f3f0e1d2SKirill A. Shutemov 2221f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 2222d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 2223d8ea7cc8SZach O'Keefe present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 2224f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 2225e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 2226f3f0e1d2SKirill A. Shutemov } else { 222734488399SZach O'Keefe result = collapse_file(mm, addr, file, start, cc); 2228f3f0e1d2SKirill A. Shutemov } 2229f3f0e1d2SKirill A. Shutemov } 2230f3f0e1d2SKirill A. Shutemov 2231045634ffSGautam Menghani trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result); 223250ad2f24SZach O'Keefe return result; 2233f3f0e1d2SKirill A. Shutemov } 2234f3f0e1d2SKirill A. Shutemov #else 223534488399SZach O'Keefe static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, 223634488399SZach O'Keefe struct file *file, pgoff_t start, 223734488399SZach O'Keefe struct collapse_control *cc) 2238f3f0e1d2SKirill A. Shutemov { 2239f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 2240f3f0e1d2SKirill A. Shutemov } 224127e1f827SSong Liu 2242b26e2701SQi Zheng static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot) 224327e1f827SSong Liu { 224427e1f827SSong Liu } 224558ac9a89SZach O'Keefe 224658ac9a89SZach O'Keefe static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm, 224758ac9a89SZach O'Keefe unsigned long addr) 224858ac9a89SZach O'Keefe { 224958ac9a89SZach O'Keefe return false; 225058ac9a89SZach O'Keefe } 2251f3f0e1d2SKirill A. Shutemov #endif 2252f3f0e1d2SKirill A. Shutemov 225350ad2f24SZach O'Keefe static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, 225434d6b470SZach O'Keefe struct collapse_control *cc) 2255b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 2256b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 2257b46e756fSKirill A. Shutemov { 225868540502SMatthew Wilcox (Oracle) struct vma_iterator vmi; 2259b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 2260b26e2701SQi Zheng struct mm_slot *slot; 2261b46e756fSKirill A. Shutemov struct mm_struct *mm; 2262b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 2263b46e756fSKirill A. Shutemov int progress = 0; 2264b46e756fSKirill A. Shutemov 2265b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 226635f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 226750ad2f24SZach O'Keefe *result = SCAN_FAIL; 2268b46e756fSKirill A. Shutemov 2269b26e2701SQi Zheng if (khugepaged_scan.mm_slot) { 2270b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2271b26e2701SQi Zheng slot = &mm_slot->slot; 2272b26e2701SQi Zheng } else { 2273b26e2701SQi Zheng slot = list_entry(khugepaged_scan.mm_head.next, 2274b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2275b26e2701SQi Zheng mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 2276b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2277b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 2278b46e756fSKirill A. Shutemov } 2279b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 228027e1f827SSong Liu khugepaged_collapse_pte_mapped_thps(mm_slot); 2281b46e756fSKirill A. Shutemov 2282b26e2701SQi Zheng mm = slot->mm; 22833b454ad3SYang Shi /* 22843b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 22853b454ad3SYang Shi * the next mm on the list. 22863b454ad3SYang Shi */ 2287b46e756fSKirill A. Shutemov vma = NULL; 2288d8ed45c5SMichel Lespinasse if (unlikely(!mmap_read_trylock(mm))) 2289c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2290b46e756fSKirill A. Shutemov 2291b46e756fSKirill A. Shutemov progress++; 229268540502SMatthew Wilcox (Oracle) if (unlikely(hpage_collapse_test_exit(mm))) 229368540502SMatthew Wilcox (Oracle) goto breakouterloop; 229468540502SMatthew Wilcox (Oracle) 229568540502SMatthew Wilcox (Oracle) vma_iter_init(&vmi, mm, khugepaged_scan.address); 229668540502SMatthew Wilcox (Oracle) for_each_vma(vmi, vma) { 2297b46e756fSKirill A. Shutemov unsigned long hstart, hend; 2298b46e756fSKirill A. Shutemov 2299b46e756fSKirill A. Shutemov cond_resched(); 23007d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) { 2301b46e756fSKirill A. Shutemov progress++; 2302b46e756fSKirill A. Shutemov break; 2303b46e756fSKirill A. Shutemov } 2304a7f4e6e4SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) { 2305b46e756fSKirill A. Shutemov skip: 2306b46e756fSKirill A. Shutemov progress++; 2307b46e756fSKirill A. Shutemov continue; 2308b46e756fSKirill A. Shutemov } 23094fa6893fSYang Shi hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); 23104fa6893fSYang Shi hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); 2311b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 2312b46e756fSKirill A. Shutemov goto skip; 2313b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 2314b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 2315b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2316b46e756fSKirill A. Shutemov 2317b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 231850ad2f24SZach O'Keefe bool mmap_locked = true; 231950ad2f24SZach O'Keefe 2320b46e756fSKirill A. Shutemov cond_resched(); 23217d2c4385SZach O'Keefe if (unlikely(hpage_collapse_test_exit(mm))) 2322b46e756fSKirill A. Shutemov goto breakouterloop; 2323b46e756fSKirill A. Shutemov 2324b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 2325b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 2326b46e756fSKirill A. Shutemov hend); 232799cb0dbdSSong Liu if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 2328396bcc52SMatthew Wilcox (Oracle) struct file *file = get_file(vma->vm_file); 2329f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 2330f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 233199cb0dbdSSong Liu 2332d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 233334488399SZach O'Keefe *result = hpage_collapse_scan_file(mm, 233434488399SZach O'Keefe khugepaged_scan.address, 233534488399SZach O'Keefe file, pgoff, cc); 233650ad2f24SZach O'Keefe mmap_locked = false; 2337f3f0e1d2SKirill A. Shutemov fput(file); 2338f3f0e1d2SKirill A. Shutemov } else { 23397d2c4385SZach O'Keefe *result = hpage_collapse_scan_pmd(mm, vma, 2340b46e756fSKirill A. Shutemov khugepaged_scan.address, 23417d2c4385SZach O'Keefe &mmap_locked, 23427d2c4385SZach O'Keefe cc); 2343f3f0e1d2SKirill A. Shutemov } 234458ac9a89SZach O'Keefe switch (*result) { 234558ac9a89SZach O'Keefe case SCAN_PTE_MAPPED_HUGEPAGE: { 234658ac9a89SZach O'Keefe pmd_t *pmd; 234758ac9a89SZach O'Keefe 234858ac9a89SZach O'Keefe *result = find_pmd_or_thp_or_none(mm, 234958ac9a89SZach O'Keefe khugepaged_scan.address, 235058ac9a89SZach O'Keefe &pmd); 235158ac9a89SZach O'Keefe if (*result != SCAN_SUCCEED) 235258ac9a89SZach O'Keefe break; 235358ac9a89SZach O'Keefe if (!khugepaged_add_pte_mapped_thp(mm, 235458ac9a89SZach O'Keefe khugepaged_scan.address)) 235558ac9a89SZach O'Keefe break; 235658ac9a89SZach O'Keefe } fallthrough; 235758ac9a89SZach O'Keefe case SCAN_SUCCEED: 235850ad2f24SZach O'Keefe ++khugepaged_pages_collapsed; 235958ac9a89SZach O'Keefe break; 236058ac9a89SZach O'Keefe default: 236158ac9a89SZach O'Keefe break; 236258ac9a89SZach O'Keefe } 236358ac9a89SZach O'Keefe 2364b46e756fSKirill A. Shutemov /* move to next address */ 2365b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 2366b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 236750ad2f24SZach O'Keefe if (!mmap_locked) 236850ad2f24SZach O'Keefe /* 236950ad2f24SZach O'Keefe * We released mmap_lock so break loop. Note 237050ad2f24SZach O'Keefe * that we drop mmap_lock before all hugepage 237150ad2f24SZach O'Keefe * allocations, so if allocation fails, we are 237250ad2f24SZach O'Keefe * guaranteed to break here and report the 237350ad2f24SZach O'Keefe * correct result back to caller. 237450ad2f24SZach O'Keefe */ 2375c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2376b46e756fSKirill A. Shutemov if (progress >= pages) 2377b46e756fSKirill A. Shutemov goto breakouterloop; 2378b46e756fSKirill A. Shutemov } 2379b46e756fSKirill A. Shutemov } 2380b46e756fSKirill A. Shutemov breakouterloop: 2381d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ 2382c1e8d7c6SMichel Lespinasse breakouterloop_mmap_lock: 2383b46e756fSKirill A. Shutemov 2384b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2385b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2386b46e756fSKirill A. Shutemov /* 2387b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 2388b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 2389b46e756fSKirill A. Shutemov */ 23907d2c4385SZach O'Keefe if (hpage_collapse_test_exit(mm) || !vma) { 2391b46e756fSKirill A. Shutemov /* 2392b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 2393b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 2394b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 2395b46e756fSKirill A. Shutemov */ 2396b26e2701SQi Zheng if (slot->mm_node.next != &khugepaged_scan.mm_head) { 2397b26e2701SQi Zheng slot = list_entry(slot->mm_node.next, 2398b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2399b26e2701SQi Zheng khugepaged_scan.mm_slot = 2400b26e2701SQi Zheng mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 2401b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2402b46e756fSKirill A. Shutemov } else { 2403b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2404b46e756fSKirill A. Shutemov khugepaged_full_scans++; 2405b46e756fSKirill A. Shutemov } 2406b46e756fSKirill A. Shutemov 2407b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2408b46e756fSKirill A. Shutemov } 2409b46e756fSKirill A. Shutemov 2410b46e756fSKirill A. Shutemov return progress; 2411b46e756fSKirill A. Shutemov } 2412b46e756fSKirill A. Shutemov 2413b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 2414b46e756fSKirill A. Shutemov { 2415b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 24161064026bSYang Shi hugepage_flags_enabled(); 2417b46e756fSKirill A. Shutemov } 2418b46e756fSKirill A. Shutemov 2419b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 2420b46e756fSKirill A. Shutemov { 2421b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 2422b46e756fSKirill A. Shutemov kthread_should_stop(); 2423b46e756fSKirill A. Shutemov } 2424b46e756fSKirill A. Shutemov 242534d6b470SZach O'Keefe static void khugepaged_do_scan(struct collapse_control *cc) 2426b46e756fSKirill A. Shutemov { 2427b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 242889dc6a96SYanfei Xu unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); 2429b46e756fSKirill A. Shutemov bool wait = true; 243050ad2f24SZach O'Keefe int result = SCAN_SUCCEED; 2431b46e756fSKirill A. Shutemov 2432a980df33SKirill A. Shutemov lru_add_drain_all(); 2433a980df33SKirill A. Shutemov 2434c6a7f445SYang Shi while (true) { 2435b46e756fSKirill A. Shutemov cond_resched(); 2436b46e756fSKirill A. Shutemov 2437b46e756fSKirill A. Shutemov if (unlikely(kthread_should_stop() || try_to_freeze())) 2438b46e756fSKirill A. Shutemov break; 2439b46e756fSKirill A. Shutemov 2440b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2441b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 2442b46e756fSKirill A. Shutemov pass_through_head++; 2443b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 2444b46e756fSKirill A. Shutemov pass_through_head < 2) 2445b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 244650ad2f24SZach O'Keefe &result, cc); 2447b46e756fSKirill A. Shutemov else 2448b46e756fSKirill A. Shutemov progress = pages; 2449b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2450b46e756fSKirill A. Shutemov 2451c6a7f445SYang Shi if (progress >= pages) 2452c6a7f445SYang Shi break; 2453c6a7f445SYang Shi 245450ad2f24SZach O'Keefe if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) { 2455c6a7f445SYang Shi /* 2456c6a7f445SYang Shi * If fail to allocate the first time, try to sleep for 2457c6a7f445SYang Shi * a while. When hit again, cancel the scan. 2458c6a7f445SYang Shi */ 2459c6a7f445SYang Shi if (!wait) 2460c6a7f445SYang Shi break; 2461c6a7f445SYang Shi wait = false; 2462c6a7f445SYang Shi khugepaged_alloc_sleep(); 2463c6a7f445SYang Shi } 2464c6a7f445SYang Shi } 2465b46e756fSKirill A. Shutemov } 2466b46e756fSKirill A. Shutemov 2467b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 2468b46e756fSKirill A. Shutemov { 2469b46e756fSKirill A. Shutemov return kthread_should_stop() || 2470b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 2471b46e756fSKirill A. Shutemov } 2472b46e756fSKirill A. Shutemov 2473b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 2474b46e756fSKirill A. Shutemov { 2475b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 2476b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 2477b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2478b46e756fSKirill A. Shutemov 2479b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 2480b46e756fSKirill A. Shutemov return; 2481b46e756fSKirill A. Shutemov 2482b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2483b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 2484b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 2485b46e756fSKirill A. Shutemov scan_sleep_jiffies); 2486b46e756fSKirill A. Shutemov return; 2487b46e756fSKirill A. Shutemov } 2488b46e756fSKirill A. Shutemov 24891064026bSYang Shi if (hugepage_flags_enabled()) 2490b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2491b46e756fSKirill A. Shutemov } 2492b46e756fSKirill A. Shutemov 2493b46e756fSKirill A. Shutemov static int khugepaged(void *none) 2494b46e756fSKirill A. Shutemov { 2495b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 2496b46e756fSKirill A. Shutemov 2497b46e756fSKirill A. Shutemov set_freezable(); 2498b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 2499b46e756fSKirill A. Shutemov 2500b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 250134d6b470SZach O'Keefe khugepaged_do_scan(&khugepaged_collapse_control); 2502b46e756fSKirill A. Shutemov khugepaged_wait_work(); 2503b46e756fSKirill A. Shutemov } 2504b46e756fSKirill A. Shutemov 2505b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2506b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2507b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2508b46e756fSKirill A. Shutemov if (mm_slot) 2509b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2510b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2511b46e756fSKirill A. Shutemov return 0; 2512b46e756fSKirill A. Shutemov } 2513b46e756fSKirill A. Shutemov 2514b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 2515b46e756fSKirill A. Shutemov { 2516b46e756fSKirill A. Shutemov struct zone *zone; 2517b46e756fSKirill A. Shutemov int nr_zones = 0; 2518b46e756fSKirill A. Shutemov unsigned long recommended_min; 2519b46e756fSKirill A. Shutemov 25201064026bSYang Shi if (!hugepage_flags_enabled()) { 2521bd3400eaSLiangcai Fan calculate_min_free_kbytes(); 2522bd3400eaSLiangcai Fan goto update_wmarks; 2523bd3400eaSLiangcai Fan } 2524bd3400eaSLiangcai Fan 2525b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 2526b7d349c7SJoonsoo Kim /* 2527b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 2528b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 2529b7d349c7SJoonsoo Kim */ 2530b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 2531b7d349c7SJoonsoo Kim continue; 2532b7d349c7SJoonsoo Kim 2533b46e756fSKirill A. Shutemov nr_zones++; 2534b7d349c7SJoonsoo Kim } 2535b46e756fSKirill A. Shutemov 2536b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2537b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 2538b46e756fSKirill A. Shutemov 2539b46e756fSKirill A. Shutemov /* 2540b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 2541b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 2542b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 2543b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 2544b46e756fSKirill A. Shutemov */ 2545b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 2546b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2547b46e756fSKirill A. Shutemov 2548b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 2549b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 2550b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 2551b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 2552b46e756fSKirill A. Shutemov 2553b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 2554b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 2555b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2556b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 2557b46e756fSKirill A. Shutemov 2558b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 2559b46e756fSKirill A. Shutemov } 2560bd3400eaSLiangcai Fan 2561bd3400eaSLiangcai Fan update_wmarks: 2562b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 2563b46e756fSKirill A. Shutemov } 2564b46e756fSKirill A. Shutemov 2565b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 2566b46e756fSKirill A. Shutemov { 2567b46e756fSKirill A. Shutemov int err = 0; 2568b46e756fSKirill A. Shutemov 2569b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 25701064026bSYang Shi if (hugepage_flags_enabled()) { 2571b46e756fSKirill A. Shutemov if (!khugepaged_thread) 2572b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 2573b46e756fSKirill A. Shutemov "khugepaged"); 2574b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 2575b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2576b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 2577b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2578b46e756fSKirill A. Shutemov goto fail; 2579b46e756fSKirill A. Shutemov } 2580b46e756fSKirill A. Shutemov 2581b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 2582b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 2583b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 2584b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 2585b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2586b46e756fSKirill A. Shutemov } 2587bd3400eaSLiangcai Fan set_recommended_min_free_kbytes(); 2588b46e756fSKirill A. Shutemov fail: 2589b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 2590b46e756fSKirill A. Shutemov return err; 2591b46e756fSKirill A. Shutemov } 25924aab2be0SVijay Balakrishna 25934aab2be0SVijay Balakrishna void khugepaged_min_free_kbytes_update(void) 25944aab2be0SVijay Balakrishna { 25954aab2be0SVijay Balakrishna mutex_lock(&khugepaged_mutex); 25961064026bSYang Shi if (hugepage_flags_enabled() && khugepaged_thread) 25974aab2be0SVijay Balakrishna set_recommended_min_free_kbytes(); 25984aab2be0SVijay Balakrishna mutex_unlock(&khugepaged_mutex); 25994aab2be0SVijay Balakrishna } 26007d8faaf1SZach O'Keefe 260157e9cc50SJohannes Weiner bool current_is_khugepaged(void) 260257e9cc50SJohannes Weiner { 260357e9cc50SJohannes Weiner return kthread_func(current) == khugepaged; 260457e9cc50SJohannes Weiner } 260557e9cc50SJohannes Weiner 26067d8faaf1SZach O'Keefe static int madvise_collapse_errno(enum scan_result r) 26077d8faaf1SZach O'Keefe { 26087d8faaf1SZach O'Keefe /* 26097d8faaf1SZach O'Keefe * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide 26107d8faaf1SZach O'Keefe * actionable feedback to caller, so they may take an appropriate 26117d8faaf1SZach O'Keefe * fallback measure depending on the nature of the failure. 26127d8faaf1SZach O'Keefe */ 26137d8faaf1SZach O'Keefe switch (r) { 26147d8faaf1SZach O'Keefe case SCAN_ALLOC_HUGE_PAGE_FAIL: 26157d8faaf1SZach O'Keefe return -ENOMEM; 26167d8faaf1SZach O'Keefe case SCAN_CGROUP_CHARGE_FAIL: 26177d8faaf1SZach O'Keefe return -EBUSY; 26187d8faaf1SZach O'Keefe /* Resource temporary unavailable - trying again might succeed */ 26197d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 26207d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 26210f3e2a2cSZach O'Keefe case SCAN_DEL_PAGE_LRU: 26227d8faaf1SZach O'Keefe return -EAGAIN; 26237d8faaf1SZach O'Keefe /* 26247d8faaf1SZach O'Keefe * Other: Trying again likely not to succeed / error intrinsic to 26257d8faaf1SZach O'Keefe * specified memory range. khugepaged likely won't be able to collapse 26267d8faaf1SZach O'Keefe * either. 26277d8faaf1SZach O'Keefe */ 26287d8faaf1SZach O'Keefe default: 26297d8faaf1SZach O'Keefe return -EINVAL; 26307d8faaf1SZach O'Keefe } 26317d8faaf1SZach O'Keefe } 26327d8faaf1SZach O'Keefe 26337d8faaf1SZach O'Keefe int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev, 26347d8faaf1SZach O'Keefe unsigned long start, unsigned long end) 26357d8faaf1SZach O'Keefe { 26367d8faaf1SZach O'Keefe struct collapse_control *cc; 26377d8faaf1SZach O'Keefe struct mm_struct *mm = vma->vm_mm; 26387d8faaf1SZach O'Keefe unsigned long hstart, hend, addr; 26397d8faaf1SZach O'Keefe int thps = 0, last_fail = SCAN_FAIL; 26407d8faaf1SZach O'Keefe bool mmap_locked = true; 26417d8faaf1SZach O'Keefe 26427d8faaf1SZach O'Keefe BUG_ON(vma->vm_start > start); 26437d8faaf1SZach O'Keefe BUG_ON(vma->vm_end < end); 26447d8faaf1SZach O'Keefe 26457d8faaf1SZach O'Keefe *prev = vma; 26467d8faaf1SZach O'Keefe 26477d8faaf1SZach O'Keefe if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) 26487d8faaf1SZach O'Keefe return -EINVAL; 26497d8faaf1SZach O'Keefe 26507d8faaf1SZach O'Keefe cc = kmalloc(sizeof(*cc), GFP_KERNEL); 26517d8faaf1SZach O'Keefe if (!cc) 26527d8faaf1SZach O'Keefe return -ENOMEM; 26537d8faaf1SZach O'Keefe cc->is_khugepaged = false; 26547d8faaf1SZach O'Keefe 26557d8faaf1SZach O'Keefe mmgrab(mm); 26567d8faaf1SZach O'Keefe lru_add_drain_all(); 26577d8faaf1SZach O'Keefe 26587d8faaf1SZach O'Keefe hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 26597d8faaf1SZach O'Keefe hend = end & HPAGE_PMD_MASK; 26607d8faaf1SZach O'Keefe 26617d8faaf1SZach O'Keefe for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) { 26627d8faaf1SZach O'Keefe int result = SCAN_FAIL; 26637d8faaf1SZach O'Keefe 26647d8faaf1SZach O'Keefe if (!mmap_locked) { 26657d8faaf1SZach O'Keefe cond_resched(); 26667d8faaf1SZach O'Keefe mmap_read_lock(mm); 26677d8faaf1SZach O'Keefe mmap_locked = true; 266834488399SZach O'Keefe result = hugepage_vma_revalidate(mm, addr, false, &vma, 266934488399SZach O'Keefe cc); 26707d8faaf1SZach O'Keefe if (result != SCAN_SUCCEED) { 26717d8faaf1SZach O'Keefe last_fail = result; 26727d8faaf1SZach O'Keefe goto out_nolock; 26737d8faaf1SZach O'Keefe } 26744d24de94SYang Shi 267552dc0310SZach O'Keefe hend = min(hend, vma->vm_end & HPAGE_PMD_MASK); 26767d8faaf1SZach O'Keefe } 26777d8faaf1SZach O'Keefe mmap_assert_locked(mm); 26787d8faaf1SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 2679e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 268034488399SZach O'Keefe if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 268134488399SZach O'Keefe struct file *file = get_file(vma->vm_file); 268234488399SZach O'Keefe pgoff_t pgoff = linear_page_index(vma, addr); 268334488399SZach O'Keefe 268434488399SZach O'Keefe mmap_read_unlock(mm); 268534488399SZach O'Keefe mmap_locked = false; 268634488399SZach O'Keefe result = hpage_collapse_scan_file(mm, addr, file, pgoff, 26877d2c4385SZach O'Keefe cc); 268834488399SZach O'Keefe fput(file); 268934488399SZach O'Keefe } else { 269034488399SZach O'Keefe result = hpage_collapse_scan_pmd(mm, vma, addr, 269134488399SZach O'Keefe &mmap_locked, cc); 269234488399SZach O'Keefe } 26937d8faaf1SZach O'Keefe if (!mmap_locked) 26947d8faaf1SZach O'Keefe *prev = NULL; /* Tell caller we dropped mmap_lock */ 26957d8faaf1SZach O'Keefe 269634488399SZach O'Keefe handle_result: 26977d8faaf1SZach O'Keefe switch (result) { 26987d8faaf1SZach O'Keefe case SCAN_SUCCEED: 26997d8faaf1SZach O'Keefe case SCAN_PMD_MAPPED: 27007d8faaf1SZach O'Keefe ++thps; 27017d8faaf1SZach O'Keefe break; 270234488399SZach O'Keefe case SCAN_PTE_MAPPED_HUGEPAGE: 270334488399SZach O'Keefe BUG_ON(mmap_locked); 270434488399SZach O'Keefe BUG_ON(*prev); 270534488399SZach O'Keefe mmap_write_lock(mm); 270634488399SZach O'Keefe result = collapse_pte_mapped_thp(mm, addr, true); 270734488399SZach O'Keefe mmap_write_unlock(mm); 270834488399SZach O'Keefe goto handle_result; 27097d8faaf1SZach O'Keefe /* Whitelisted set of results where continuing OK */ 27107d8faaf1SZach O'Keefe case SCAN_PMD_NULL: 27117d8faaf1SZach O'Keefe case SCAN_PTE_NON_PRESENT: 27127d8faaf1SZach O'Keefe case SCAN_PTE_UFFD_WP: 27137d8faaf1SZach O'Keefe case SCAN_PAGE_RO: 27147d8faaf1SZach O'Keefe case SCAN_LACK_REFERENCED_PAGE: 27157d8faaf1SZach O'Keefe case SCAN_PAGE_NULL: 27167d8faaf1SZach O'Keefe case SCAN_PAGE_COUNT: 27177d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 27187d8faaf1SZach O'Keefe case SCAN_PAGE_COMPOUND: 27197d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 27200f3e2a2cSZach O'Keefe case SCAN_DEL_PAGE_LRU: 27217d8faaf1SZach O'Keefe last_fail = result; 27227d8faaf1SZach O'Keefe break; 27237d8faaf1SZach O'Keefe default: 27247d8faaf1SZach O'Keefe last_fail = result; 27257d8faaf1SZach O'Keefe /* Other error, exit */ 27267d8faaf1SZach O'Keefe goto out_maybelock; 27277d8faaf1SZach O'Keefe } 27287d8faaf1SZach O'Keefe } 27297d8faaf1SZach O'Keefe 27307d8faaf1SZach O'Keefe out_maybelock: 27317d8faaf1SZach O'Keefe /* Caller expects us to hold mmap_lock on return */ 27327d8faaf1SZach O'Keefe if (!mmap_locked) 27337d8faaf1SZach O'Keefe mmap_read_lock(mm); 27347d8faaf1SZach O'Keefe out_nolock: 27357d8faaf1SZach O'Keefe mmap_assert_locked(mm); 27367d8faaf1SZach O'Keefe mmdrop(mm); 27377d8faaf1SZach O'Keefe kfree(cc); 27387d8faaf1SZach O'Keefe 27397d8faaf1SZach O'Keefe return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0 27407d8faaf1SZach O'Keefe : madvise_collapse_errno(last_fail); 27417d8faaf1SZach O'Keefe } 2742