1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2b46e756fSKirill A. Shutemov #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 3b46e756fSKirill A. Shutemov 4b46e756fSKirill A. Shutemov #include <linux/mm.h> 5b46e756fSKirill A. Shutemov #include <linux/sched.h> 66e84f315SIngo Molnar #include <linux/sched/mm.h> 7f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 8b46e756fSKirill A. Shutemov #include <linux/mmu_notifier.h> 9b46e756fSKirill A. Shutemov #include <linux/rmap.h> 10b46e756fSKirill A. Shutemov #include <linux/swap.h> 11b46e756fSKirill A. Shutemov #include <linux/mm_inline.h> 12b46e756fSKirill A. Shutemov #include <linux/kthread.h> 13b46e756fSKirill A. Shutemov #include <linux/khugepaged.h> 14b46e756fSKirill A. Shutemov #include <linux/freezer.h> 15b46e756fSKirill A. Shutemov #include <linux/mman.h> 16b46e756fSKirill A. Shutemov #include <linux/hashtable.h> 17b46e756fSKirill A. Shutemov #include <linux/userfaultfd_k.h> 18b46e756fSKirill A. Shutemov #include <linux/page_idle.h> 1980110bbfSPasha Tatashin #include <linux/page_table_check.h> 20b46e756fSKirill A. Shutemov #include <linux/swapops.h> 21f3f0e1d2SKirill A. Shutemov #include <linux/shmem_fs.h> 22b46e756fSKirill A. Shutemov 23b46e756fSKirill A. Shutemov #include <asm/tlb.h> 24b46e756fSKirill A. Shutemov #include <asm/pgalloc.h> 25b46e756fSKirill A. Shutemov #include "internal.h" 26b46e756fSKirill A. Shutemov 27b46e756fSKirill A. Shutemov enum scan_result { 28b46e756fSKirill A. Shutemov SCAN_FAIL, 29b46e756fSKirill A. Shutemov SCAN_SUCCEED, 30b46e756fSKirill A. Shutemov SCAN_PMD_NULL, 31b46e756fSKirill A. Shutemov SCAN_EXCEED_NONE_PTE, 3271a2c112SKirill A. Shutemov SCAN_EXCEED_SWAP_PTE, 3371a2c112SKirill A. Shutemov SCAN_EXCEED_SHARED_PTE, 34b46e756fSKirill A. Shutemov SCAN_PTE_NON_PRESENT, 35e1e267c7SPeter Xu SCAN_PTE_UFFD_WP, 36b46e756fSKirill A. Shutemov SCAN_PAGE_RO, 370db501f7SEbru Akagunduz SCAN_LACK_REFERENCED_PAGE, 38b46e756fSKirill A. Shutemov SCAN_PAGE_NULL, 39b46e756fSKirill A. Shutemov SCAN_SCAN_ABORT, 40b46e756fSKirill A. Shutemov SCAN_PAGE_COUNT, 41b46e756fSKirill A. Shutemov SCAN_PAGE_LRU, 42b46e756fSKirill A. Shutemov SCAN_PAGE_LOCK, 43b46e756fSKirill A. Shutemov SCAN_PAGE_ANON, 44b46e756fSKirill A. Shutemov SCAN_PAGE_COMPOUND, 45b46e756fSKirill A. Shutemov SCAN_ANY_PROCESS, 46b46e756fSKirill A. Shutemov SCAN_VMA_NULL, 47b46e756fSKirill A. Shutemov SCAN_VMA_CHECK, 48b46e756fSKirill A. Shutemov SCAN_ADDRESS_RANGE, 49b46e756fSKirill A. Shutemov SCAN_DEL_PAGE_LRU, 50b46e756fSKirill A. Shutemov SCAN_ALLOC_HUGE_PAGE_FAIL, 51b46e756fSKirill A. Shutemov SCAN_CGROUP_CHARGE_FAIL, 52f3f0e1d2SKirill A. Shutemov SCAN_TRUNCATED, 5399cb0dbdSSong Liu SCAN_PAGE_HAS_PRIVATE, 54b46e756fSKirill A. Shutemov }; 55b46e756fSKirill A. Shutemov 56b46e756fSKirill A. Shutemov #define CREATE_TRACE_POINTS 57b46e756fSKirill A. Shutemov #include <trace/events/huge_memory.h> 58b46e756fSKirill A. Shutemov 594aab2be0SVijay Balakrishna static struct task_struct *khugepaged_thread __read_mostly; 604aab2be0SVijay Balakrishna static DEFINE_MUTEX(khugepaged_mutex); 614aab2be0SVijay Balakrishna 62b46e756fSKirill A. Shutemov /* default scan 8*512 pte (or vmas) every 30 second */ 63b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_to_scan __read_mostly; 64b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_collapsed; 65b46e756fSKirill A. Shutemov static unsigned int khugepaged_full_scans; 66b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 67b46e756fSKirill A. Shutemov /* during fragmentation poll the hugepage allocator once every minute */ 68b46e756fSKirill A. Shutemov static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 69b46e756fSKirill A. Shutemov static unsigned long khugepaged_sleep_expire; 70b46e756fSKirill A. Shutemov static DEFINE_SPINLOCK(khugepaged_mm_lock); 71b46e756fSKirill A. Shutemov static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 72b46e756fSKirill A. Shutemov /* 73b46e756fSKirill A. Shutemov * default collapse hugepages if there is at least one pte mapped like 74b46e756fSKirill A. Shutemov * it would have happened if the vma was large enough during page 75b46e756fSKirill A. Shutemov * fault. 76b46e756fSKirill A. Shutemov */ 77b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_none __read_mostly; 78b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_swap __read_mostly; 7971a2c112SKirill A. Shutemov static unsigned int khugepaged_max_ptes_shared __read_mostly; 80b46e756fSKirill A. Shutemov 81b46e756fSKirill A. Shutemov #define MM_SLOTS_HASH_BITS 10 82b46e756fSKirill A. Shutemov static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 83b46e756fSKirill A. Shutemov 84b46e756fSKirill A. Shutemov static struct kmem_cache *mm_slot_cache __read_mostly; 85b46e756fSKirill A. Shutemov 8627e1f827SSong Liu #define MAX_PTE_MAPPED_THP 8 8727e1f827SSong Liu 88b46e756fSKirill A. Shutemov /** 89b46e756fSKirill A. Shutemov * struct mm_slot - hash lookup from mm to mm_slot 90b46e756fSKirill A. Shutemov * @hash: hash collision list 91b46e756fSKirill A. Shutemov * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 92b46e756fSKirill A. Shutemov * @mm: the mm that this information is valid for 93336e6b53SAlex Shi * @nr_pte_mapped_thp: number of pte mapped THP 94336e6b53SAlex Shi * @pte_mapped_thp: address array corresponding pte mapped THP 95b46e756fSKirill A. Shutemov */ 96b46e756fSKirill A. Shutemov struct mm_slot { 97b46e756fSKirill A. Shutemov struct hlist_node hash; 98b46e756fSKirill A. Shutemov struct list_head mm_node; 99b46e756fSKirill A. Shutemov struct mm_struct *mm; 10027e1f827SSong Liu 10127e1f827SSong Liu /* pte-mapped THP in this mm */ 10227e1f827SSong Liu int nr_pte_mapped_thp; 10327e1f827SSong Liu unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; 104b46e756fSKirill A. Shutemov }; 105b46e756fSKirill A. Shutemov 106b46e756fSKirill A. Shutemov /** 107b46e756fSKirill A. Shutemov * struct khugepaged_scan - cursor for scanning 108b46e756fSKirill A. Shutemov * @mm_head: the head of the mm list to scan 109b46e756fSKirill A. Shutemov * @mm_slot: the current mm_slot we are scanning 110b46e756fSKirill A. Shutemov * @address: the next address inside that to be scanned 111b46e756fSKirill A. Shutemov * 112b46e756fSKirill A. Shutemov * There is only the one khugepaged_scan instance of this cursor structure. 113b46e756fSKirill A. Shutemov */ 114b46e756fSKirill A. Shutemov struct khugepaged_scan { 115b46e756fSKirill A. Shutemov struct list_head mm_head; 116b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 117b46e756fSKirill A. Shutemov unsigned long address; 118b46e756fSKirill A. Shutemov }; 119b46e756fSKirill A. Shutemov 120b46e756fSKirill A. Shutemov static struct khugepaged_scan khugepaged_scan = { 121b46e756fSKirill A. Shutemov .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 122b46e756fSKirill A. Shutemov }; 123b46e756fSKirill A. Shutemov 124e1465d12SJérémy Lefaure #ifdef CONFIG_SYSFS 125b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 126b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 127b46e756fSKirill A. Shutemov char *buf) 128b46e756fSKirill A. Shutemov { 129ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); 130b46e756fSKirill A. Shutemov } 131b46e756fSKirill A. Shutemov 132b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 133b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 134b46e756fSKirill A. Shutemov const char *buf, size_t count) 135b46e756fSKirill A. Shutemov { 136dfefd226SAlexey Dobriyan unsigned int msecs; 137b46e756fSKirill A. Shutemov int err; 138b46e756fSKirill A. Shutemov 139dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 140dfefd226SAlexey Dobriyan if (err) 141b46e756fSKirill A. Shutemov return -EINVAL; 142b46e756fSKirill A. Shutemov 143b46e756fSKirill A. Shutemov khugepaged_scan_sleep_millisecs = msecs; 144b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 145b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 146b46e756fSKirill A. Shutemov 147b46e756fSKirill A. Shutemov return count; 148b46e756fSKirill A. Shutemov } 149b46e756fSKirill A. Shutemov static struct kobj_attribute scan_sleep_millisecs_attr = 150b46e756fSKirill A. Shutemov __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 151b46e756fSKirill A. Shutemov scan_sleep_millisecs_store); 152b46e756fSKirill A. Shutemov 153b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 154b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 155b46e756fSKirill A. Shutemov char *buf) 156b46e756fSKirill A. Shutemov { 157ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 158b46e756fSKirill A. Shutemov } 159b46e756fSKirill A. Shutemov 160b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 161b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 162b46e756fSKirill A. Shutemov const char *buf, size_t count) 163b46e756fSKirill A. Shutemov { 164dfefd226SAlexey Dobriyan unsigned int msecs; 165b46e756fSKirill A. Shutemov int err; 166b46e756fSKirill A. Shutemov 167dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 168dfefd226SAlexey Dobriyan if (err) 169b46e756fSKirill A. Shutemov return -EINVAL; 170b46e756fSKirill A. Shutemov 171b46e756fSKirill A. Shutemov khugepaged_alloc_sleep_millisecs = msecs; 172b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 173b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 174b46e756fSKirill A. Shutemov 175b46e756fSKirill A. Shutemov return count; 176b46e756fSKirill A. Shutemov } 177b46e756fSKirill A. Shutemov static struct kobj_attribute alloc_sleep_millisecs_attr = 178b46e756fSKirill A. Shutemov __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 179b46e756fSKirill A. Shutemov alloc_sleep_millisecs_store); 180b46e756fSKirill A. Shutemov 181b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_show(struct kobject *kobj, 182b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 183b46e756fSKirill A. Shutemov char *buf) 184b46e756fSKirill A. Shutemov { 185ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); 186b46e756fSKirill A. Shutemov } 187b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_store(struct kobject *kobj, 188b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 189b46e756fSKirill A. Shutemov const char *buf, size_t count) 190b46e756fSKirill A. Shutemov { 191dfefd226SAlexey Dobriyan unsigned int pages; 192b46e756fSKirill A. Shutemov int err; 193b46e756fSKirill A. Shutemov 194dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &pages); 195dfefd226SAlexey Dobriyan if (err || !pages) 196b46e756fSKirill A. Shutemov return -EINVAL; 197b46e756fSKirill A. Shutemov 198b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = pages; 199b46e756fSKirill A. Shutemov 200b46e756fSKirill A. Shutemov return count; 201b46e756fSKirill A. Shutemov } 202b46e756fSKirill A. Shutemov static struct kobj_attribute pages_to_scan_attr = 203b46e756fSKirill A. Shutemov __ATTR(pages_to_scan, 0644, pages_to_scan_show, 204b46e756fSKirill A. Shutemov pages_to_scan_store); 205b46e756fSKirill A. Shutemov 206b46e756fSKirill A. Shutemov static ssize_t pages_collapsed_show(struct kobject *kobj, 207b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 208b46e756fSKirill A. Shutemov char *buf) 209b46e756fSKirill A. Shutemov { 210ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); 211b46e756fSKirill A. Shutemov } 212b46e756fSKirill A. Shutemov static struct kobj_attribute pages_collapsed_attr = 213b46e756fSKirill A. Shutemov __ATTR_RO(pages_collapsed); 214b46e756fSKirill A. Shutemov 215b46e756fSKirill A. Shutemov static ssize_t full_scans_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_full_scans); 220b46e756fSKirill A. Shutemov } 221b46e756fSKirill A. Shutemov static struct kobj_attribute full_scans_attr = 222b46e756fSKirill A. Shutemov __ATTR_RO(full_scans); 223b46e756fSKirill A. Shutemov 224b46e756fSKirill A. Shutemov static ssize_t khugepaged_defrag_show(struct kobject *kobj, 225b46e756fSKirill A. Shutemov struct kobj_attribute *attr, char *buf) 226b46e756fSKirill A. Shutemov { 227b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 228b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 229b46e756fSKirill A. Shutemov } 230b46e756fSKirill A. Shutemov static ssize_t khugepaged_defrag_store(struct kobject *kobj, 231b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 232b46e756fSKirill A. Shutemov const char *buf, size_t count) 233b46e756fSKirill A. Shutemov { 234b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 235b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 236b46e756fSKirill A. Shutemov } 237b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_defrag_attr = 238b46e756fSKirill A. Shutemov __ATTR(defrag, 0644, khugepaged_defrag_show, 239b46e756fSKirill A. Shutemov khugepaged_defrag_store); 240b46e756fSKirill A. Shutemov 241b46e756fSKirill A. Shutemov /* 242b46e756fSKirill A. Shutemov * max_ptes_none controls if khugepaged should collapse hugepages over 243b46e756fSKirill A. Shutemov * any unmapped ptes in turn potentially increasing the memory 244b46e756fSKirill A. Shutemov * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 245b46e756fSKirill A. Shutemov * reduce the available free memory in the system as it 246b46e756fSKirill A. Shutemov * runs. Increasing max_ptes_none will instead potentially reduce the 247b46e756fSKirill A. Shutemov * free memory in the system during the khugepaged scan. 248b46e756fSKirill A. Shutemov */ 249b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 250b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 251b46e756fSKirill A. Shutemov char *buf) 252b46e756fSKirill A. Shutemov { 253ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); 254b46e756fSKirill A. Shutemov } 255b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 256b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 257b46e756fSKirill A. Shutemov const char *buf, size_t count) 258b46e756fSKirill A. Shutemov { 259b46e756fSKirill A. Shutemov int err; 260b46e756fSKirill A. Shutemov unsigned long max_ptes_none; 261b46e756fSKirill A. Shutemov 262b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_none); 263b46e756fSKirill A. Shutemov if (err || max_ptes_none > HPAGE_PMD_NR-1) 264b46e756fSKirill A. Shutemov return -EINVAL; 265b46e756fSKirill A. Shutemov 266b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = max_ptes_none; 267b46e756fSKirill A. Shutemov 268b46e756fSKirill A. Shutemov return count; 269b46e756fSKirill A. Shutemov } 270b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_none_attr = 271b46e756fSKirill A. Shutemov __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 272b46e756fSKirill A. Shutemov khugepaged_max_ptes_none_store); 273b46e756fSKirill A. Shutemov 274b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj, 275b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 276b46e756fSKirill A. Shutemov char *buf) 277b46e756fSKirill A. Shutemov { 278ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); 279b46e756fSKirill A. Shutemov } 280b46e756fSKirill A. Shutemov 281b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj, 282b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 283b46e756fSKirill A. Shutemov const char *buf, size_t count) 284b46e756fSKirill A. Shutemov { 285b46e756fSKirill A. Shutemov int err; 286b46e756fSKirill A. Shutemov unsigned long max_ptes_swap; 287b46e756fSKirill A. Shutemov 288b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_swap); 289b46e756fSKirill A. Shutemov if (err || max_ptes_swap > HPAGE_PMD_NR-1) 290b46e756fSKirill A. Shutemov return -EINVAL; 291b46e756fSKirill A. Shutemov 292b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = max_ptes_swap; 293b46e756fSKirill A. Shutemov 294b46e756fSKirill A. Shutemov return count; 295b46e756fSKirill A. Shutemov } 296b46e756fSKirill A. Shutemov 297b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_swap_attr = 298b46e756fSKirill A. Shutemov __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show, 299b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap_store); 300b46e756fSKirill A. Shutemov 30171a2c112SKirill A. Shutemov static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj, 30271a2c112SKirill A. Shutemov struct kobj_attribute *attr, 30371a2c112SKirill A. Shutemov char *buf) 30471a2c112SKirill A. Shutemov { 305ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); 30671a2c112SKirill A. Shutemov } 30771a2c112SKirill A. Shutemov 30871a2c112SKirill A. Shutemov static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj, 30971a2c112SKirill A. Shutemov struct kobj_attribute *attr, 31071a2c112SKirill A. Shutemov const char *buf, size_t count) 31171a2c112SKirill A. Shutemov { 31271a2c112SKirill A. Shutemov int err; 31371a2c112SKirill A. Shutemov unsigned long max_ptes_shared; 31471a2c112SKirill A. Shutemov 31571a2c112SKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_shared); 31671a2c112SKirill A. Shutemov if (err || max_ptes_shared > HPAGE_PMD_NR-1) 31771a2c112SKirill A. Shutemov return -EINVAL; 31871a2c112SKirill A. Shutemov 31971a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = max_ptes_shared; 32071a2c112SKirill A. Shutemov 32171a2c112SKirill A. Shutemov return count; 32271a2c112SKirill A. Shutemov } 32371a2c112SKirill A. Shutemov 32471a2c112SKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_shared_attr = 32571a2c112SKirill A. Shutemov __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show, 32671a2c112SKirill A. Shutemov khugepaged_max_ptes_shared_store); 32771a2c112SKirill A. Shutemov 328b46e756fSKirill A. Shutemov static struct attribute *khugepaged_attr[] = { 329b46e756fSKirill A. Shutemov &khugepaged_defrag_attr.attr, 330b46e756fSKirill A. Shutemov &khugepaged_max_ptes_none_attr.attr, 33171a2c112SKirill A. Shutemov &khugepaged_max_ptes_swap_attr.attr, 33271a2c112SKirill A. Shutemov &khugepaged_max_ptes_shared_attr.attr, 333b46e756fSKirill A. Shutemov &pages_to_scan_attr.attr, 334b46e756fSKirill A. Shutemov &pages_collapsed_attr.attr, 335b46e756fSKirill A. Shutemov &full_scans_attr.attr, 336b46e756fSKirill A. Shutemov &scan_sleep_millisecs_attr.attr, 337b46e756fSKirill A. Shutemov &alloc_sleep_millisecs_attr.attr, 338b46e756fSKirill A. Shutemov NULL, 339b46e756fSKirill A. Shutemov }; 340b46e756fSKirill A. Shutemov 341b46e756fSKirill A. Shutemov struct attribute_group khugepaged_attr_group = { 342b46e756fSKirill A. Shutemov .attrs = khugepaged_attr, 343b46e756fSKirill A. Shutemov .name = "khugepaged", 344b46e756fSKirill A. Shutemov }; 345e1465d12SJérémy Lefaure #endif /* CONFIG_SYSFS */ 346b46e756fSKirill A. Shutemov 347b46e756fSKirill A. Shutemov int hugepage_madvise(struct vm_area_struct *vma, 348b46e756fSKirill A. Shutemov unsigned long *vm_flags, int advice) 349b46e756fSKirill A. Shutemov { 350b46e756fSKirill A. Shutemov switch (advice) { 351b46e756fSKirill A. Shutemov case MADV_HUGEPAGE: 352b46e756fSKirill A. Shutemov #ifdef CONFIG_S390 353b46e756fSKirill A. Shutemov /* 354b46e756fSKirill A. Shutemov * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 355b46e756fSKirill A. Shutemov * can't handle this properly after s390_enable_sie, so we simply 356b46e756fSKirill A. Shutemov * ignore the madvise to prevent qemu from causing a SIGSEGV. 357b46e756fSKirill A. Shutemov */ 358b46e756fSKirill A. Shutemov if (mm_has_pgste(vma->vm_mm)) 359b46e756fSKirill A. Shutemov return 0; 360b46e756fSKirill A. Shutemov #endif 361b46e756fSKirill A. Shutemov *vm_flags &= ~VM_NOHUGEPAGE; 362b46e756fSKirill A. Shutemov *vm_flags |= VM_HUGEPAGE; 363b46e756fSKirill A. Shutemov /* 364b46e756fSKirill A. Shutemov * If the vma become good for khugepaged to scan, 365b46e756fSKirill A. Shutemov * register it here without waiting a page fault that 366b46e756fSKirill A. Shutemov * may not happen any time soon. 367b46e756fSKirill A. Shutemov */ 368*d2081b2bSYang Shi khugepaged_enter_vma_merge(vma, *vm_flags); 369b46e756fSKirill A. Shutemov break; 370b46e756fSKirill A. Shutemov case MADV_NOHUGEPAGE: 371b46e756fSKirill A. Shutemov *vm_flags &= ~VM_HUGEPAGE; 372b46e756fSKirill A. Shutemov *vm_flags |= VM_NOHUGEPAGE; 373b46e756fSKirill A. Shutemov /* 374b46e756fSKirill A. Shutemov * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 375b46e756fSKirill A. Shutemov * this vma even if we leave the mm registered in khugepaged if 376b46e756fSKirill A. Shutemov * it got registered before VM_NOHUGEPAGE was set. 377b46e756fSKirill A. Shutemov */ 378b46e756fSKirill A. Shutemov break; 379b46e756fSKirill A. Shutemov } 380b46e756fSKirill A. Shutemov 381b46e756fSKirill A. Shutemov return 0; 382b46e756fSKirill A. Shutemov } 383b46e756fSKirill A. Shutemov 384b46e756fSKirill A. Shutemov int __init khugepaged_init(void) 385b46e756fSKirill A. Shutemov { 386b46e756fSKirill A. Shutemov mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 387b46e756fSKirill A. Shutemov sizeof(struct mm_slot), 388b46e756fSKirill A. Shutemov __alignof__(struct mm_slot), 0, NULL); 389b46e756fSKirill A. Shutemov if (!mm_slot_cache) 390b46e756fSKirill A. Shutemov return -ENOMEM; 391b46e756fSKirill A. Shutemov 392b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 393b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 394b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 39571a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; 396b46e756fSKirill A. Shutemov 397b46e756fSKirill A. Shutemov return 0; 398b46e756fSKirill A. Shutemov } 399b46e756fSKirill A. Shutemov 400b46e756fSKirill A. Shutemov void __init khugepaged_destroy(void) 401b46e756fSKirill A. Shutemov { 402b46e756fSKirill A. Shutemov kmem_cache_destroy(mm_slot_cache); 403b46e756fSKirill A. Shutemov } 404b46e756fSKirill A. Shutemov 405b46e756fSKirill A. Shutemov static inline struct mm_slot *alloc_mm_slot(void) 406b46e756fSKirill A. Shutemov { 407b46e756fSKirill A. Shutemov if (!mm_slot_cache) /* initialization failed */ 408b46e756fSKirill A. Shutemov return NULL; 409b46e756fSKirill A. Shutemov return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 410b46e756fSKirill A. Shutemov } 411b46e756fSKirill A. Shutemov 412b46e756fSKirill A. Shutemov static inline void free_mm_slot(struct mm_slot *mm_slot) 413b46e756fSKirill A. Shutemov { 414b46e756fSKirill A. Shutemov kmem_cache_free(mm_slot_cache, mm_slot); 415b46e756fSKirill A. Shutemov } 416b46e756fSKirill A. Shutemov 417b46e756fSKirill A. Shutemov static struct mm_slot *get_mm_slot(struct mm_struct *mm) 418b46e756fSKirill A. Shutemov { 419b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 420b46e756fSKirill A. Shutemov 421b46e756fSKirill A. Shutemov hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 422b46e756fSKirill A. Shutemov if (mm == mm_slot->mm) 423b46e756fSKirill A. Shutemov return mm_slot; 424b46e756fSKirill A. Shutemov 425b46e756fSKirill A. Shutemov return NULL; 426b46e756fSKirill A. Shutemov } 427b46e756fSKirill A. Shutemov 428b46e756fSKirill A. Shutemov static void insert_to_mm_slots_hash(struct mm_struct *mm, 429b46e756fSKirill A. Shutemov struct mm_slot *mm_slot) 430b46e756fSKirill A. Shutemov { 431b46e756fSKirill A. Shutemov mm_slot->mm = mm; 432b46e756fSKirill A. Shutemov hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 433b46e756fSKirill A. Shutemov } 434b46e756fSKirill A. Shutemov 435b46e756fSKirill A. Shutemov static inline int khugepaged_test_exit(struct mm_struct *mm) 436b46e756fSKirill A. Shutemov { 4374d45e75aSJann Horn return atomic_read(&mm->mm_users) == 0; 438b46e756fSKirill A. Shutemov } 439b46e756fSKirill A. Shutemov 44050f8b92fSSong Liu static bool hugepage_vma_check(struct vm_area_struct *vma, 44150f8b92fSSong Liu unsigned long vm_flags) 442c2231020SYang Shi { 443e6be37b2SMiaohe Lin if (!transhuge_vma_enabled(vma, vm_flags)) 444c2231020SYang Shi return false; 44599cb0dbdSSong Liu 446cb648754SYang Shi if (vm_flags & VM_NO_KHUGEPAGED) 447cb648754SYang Shi return false; 448cb648754SYang Shi 44952b52bf1SYang Shi /* Don't run khugepaged against DAX vma */ 45052b52bf1SYang Shi if (vma_is_dax(vma)) 45152b52bf1SYang Shi return false; 45252b52bf1SYang Shi 453a4aeaa06SYang Shi if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - 454a4aeaa06SYang Shi vma->vm_pgoff, HPAGE_PMD_NR)) 455a4aeaa06SYang Shi return false; 456a4aeaa06SYang Shi 457cd89fb06SRik van Riel /* Enabled via shmem mount options or sysfs settings. */ 458a4aeaa06SYang Shi if (shmem_file(vma->vm_file)) 459a4aeaa06SYang Shi return shmem_huge_enabled(vma); 460cd89fb06SRik van Riel 461cd89fb06SRik van Riel /* THP settings require madvise. */ 462cd89fb06SRik van Riel if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) 463cd89fb06SRik van Riel return false; 464cd89fb06SRik van Riel 465a4aeaa06SYang Shi /* Only regular file is valid */ 46678d12c19SYang Shi if (file_thp_enabled(vma)) 46778d12c19SYang Shi return true; 468cd89fb06SRik van Riel 46925fa414aSxu xin if (!vma->anon_vma || !vma_is_anonymous(vma)) 470c2231020SYang Shi return false; 471222100eeSAnshuman Khandual if (vma_is_temporary_stack(vma)) 472c2231020SYang Shi return false; 473cb648754SYang Shi 474cb648754SYang Shi return true; 475c2231020SYang Shi } 476c2231020SYang Shi 477*d2081b2bSYang Shi void __khugepaged_enter(struct mm_struct *mm) 478b46e756fSKirill A. Shutemov { 479b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 480b46e756fSKirill A. Shutemov int wakeup; 481b46e756fSKirill A. Shutemov 482b46e756fSKirill A. Shutemov mm_slot = alloc_mm_slot(); 483b46e756fSKirill A. Shutemov if (!mm_slot) 484*d2081b2bSYang Shi return; 485b46e756fSKirill A. Shutemov 486b46e756fSKirill A. Shutemov /* __khugepaged_exit() must not run from under us */ 48728ff0a3cSMiaohe Lin VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); 488b46e756fSKirill A. Shutemov if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 489b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 490*d2081b2bSYang Shi return; 491b46e756fSKirill A. Shutemov } 492b46e756fSKirill A. Shutemov 493b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 494b46e756fSKirill A. Shutemov insert_to_mm_slots_hash(mm, mm_slot); 495b46e756fSKirill A. Shutemov /* 496b46e756fSKirill A. Shutemov * Insert just behind the scanning cursor, to let the area settle 497b46e756fSKirill A. Shutemov * down a little. 498b46e756fSKirill A. Shutemov */ 499b46e756fSKirill A. Shutemov wakeup = list_empty(&khugepaged_scan.mm_head); 500b46e756fSKirill A. Shutemov list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 501b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 502b46e756fSKirill A. Shutemov 503f1f10076SVegard Nossum mmgrab(mm); 504b46e756fSKirill A. Shutemov if (wakeup) 505b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 506b46e756fSKirill A. Shutemov } 507b46e756fSKirill A. Shutemov 508*d2081b2bSYang Shi void khugepaged_enter_vma_merge(struct vm_area_struct *vma, 509b46e756fSKirill A. Shutemov unsigned long vm_flags) 510b46e756fSKirill A. Shutemov { 511b46e756fSKirill A. Shutemov unsigned long hstart, hend; 512c2231020SYang Shi 513b46e756fSKirill A. Shutemov /* 51499cb0dbdSSong Liu * khugepaged only supports read-only files for non-shmem files. 51599cb0dbdSSong Liu * khugepaged does not yet work on special mappings. And 51699cb0dbdSSong Liu * file-private shmem THP is not supported. 517b46e756fSKirill A. Shutemov */ 51850f8b92fSSong Liu if (!hugepage_vma_check(vma, vm_flags)) 519*d2081b2bSYang Shi return; 520c2231020SYang Shi 521b46e756fSKirill A. Shutemov hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 522b46e756fSKirill A. Shutemov hend = vma->vm_end & HPAGE_PMD_MASK; 523b46e756fSKirill A. Shutemov if (hstart < hend) 524*d2081b2bSYang Shi khugepaged_enter(vma, vm_flags); 525b46e756fSKirill A. Shutemov } 526b46e756fSKirill A. Shutemov 527b46e756fSKirill A. Shutemov void __khugepaged_exit(struct mm_struct *mm) 528b46e756fSKirill A. Shutemov { 529b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 530b46e756fSKirill A. Shutemov int free = 0; 531b46e756fSKirill A. Shutemov 532b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 533b46e756fSKirill A. Shutemov mm_slot = get_mm_slot(mm); 534b46e756fSKirill A. Shutemov if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 535b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 536b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 537b46e756fSKirill A. Shutemov free = 1; 538b46e756fSKirill A. Shutemov } 539b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 540b46e756fSKirill A. Shutemov 541b46e756fSKirill A. Shutemov if (free) { 542b46e756fSKirill A. Shutemov clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 543b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 544b46e756fSKirill A. Shutemov mmdrop(mm); 545b46e756fSKirill A. Shutemov } else if (mm_slot) { 546b46e756fSKirill A. Shutemov /* 547b46e756fSKirill A. Shutemov * This is required to serialize against 548b46e756fSKirill A. Shutemov * khugepaged_test_exit() (which is guaranteed to run 549b46e756fSKirill A. Shutemov * under mmap sem read mode). Stop here (after we 550b46e756fSKirill A. Shutemov * return all pagetables will be destroyed) until 551b46e756fSKirill A. Shutemov * khugepaged has finished working on the pagetables 552c1e8d7c6SMichel Lespinasse * under the mmap_lock. 553b46e756fSKirill A. Shutemov */ 554d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 555d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 556b46e756fSKirill A. Shutemov } 557b46e756fSKirill A. Shutemov } 558b46e756fSKirill A. Shutemov 559b46e756fSKirill A. Shutemov static void release_pte_page(struct page *page) 560b46e756fSKirill A. Shutemov { 5615503fbf2SKirill A. Shutemov mod_node_page_state(page_pgdat(page), 5625503fbf2SKirill A. Shutemov NR_ISOLATED_ANON + page_is_file_lru(page), 5635503fbf2SKirill A. Shutemov -compound_nr(page)); 564b46e756fSKirill A. Shutemov unlock_page(page); 565b46e756fSKirill A. Shutemov putback_lru_page(page); 566b46e756fSKirill A. Shutemov } 567b46e756fSKirill A. Shutemov 5685503fbf2SKirill A. Shutemov static void release_pte_pages(pte_t *pte, pte_t *_pte, 5695503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 570b46e756fSKirill A. Shutemov { 5715503fbf2SKirill A. Shutemov struct page *page, *tmp; 5725503fbf2SKirill A. Shutemov 573b46e756fSKirill A. Shutemov while (--_pte >= pte) { 574b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 5755503fbf2SKirill A. Shutemov 5765503fbf2SKirill A. Shutemov page = pte_page(pteval); 5775503fbf2SKirill A. Shutemov if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) && 5785503fbf2SKirill A. Shutemov !PageCompound(page)) 5795503fbf2SKirill A. Shutemov release_pte_page(page); 5805503fbf2SKirill A. Shutemov } 5815503fbf2SKirill A. Shutemov 5825503fbf2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, compound_pagelist, lru) { 5835503fbf2SKirill A. Shutemov list_del(&page->lru); 5845503fbf2SKirill A. Shutemov release_pte_page(page); 585b46e756fSKirill A. Shutemov } 586b46e756fSKirill A. Shutemov } 587b46e756fSKirill A. Shutemov 5889445689fSKirill A. Shutemov static bool is_refcount_suitable(struct page *page) 5899445689fSKirill A. Shutemov { 5909445689fSKirill A. Shutemov int expected_refcount; 5919445689fSKirill A. Shutemov 5929445689fSKirill A. Shutemov expected_refcount = total_mapcount(page); 5939445689fSKirill A. Shutemov if (PageSwapCache(page)) 5949445689fSKirill A. Shutemov expected_refcount += compound_nr(page); 5959445689fSKirill A. Shutemov 5969445689fSKirill A. Shutemov return page_count(page) == expected_refcount; 5979445689fSKirill A. Shutemov } 5989445689fSKirill A. Shutemov 599b46e756fSKirill A. Shutemov static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 600b46e756fSKirill A. Shutemov unsigned long address, 6015503fbf2SKirill A. Shutemov pte_t *pte, 6025503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 603b46e756fSKirill A. Shutemov { 604b46e756fSKirill A. Shutemov struct page *page = NULL; 605b46e756fSKirill A. Shutemov pte_t *_pte; 60671a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0, result = 0, referenced = 0; 6070db501f7SEbru Akagunduz bool writable = false; 608b46e756fSKirill A. Shutemov 609b46e756fSKirill A. Shutemov for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 610b46e756fSKirill A. Shutemov _pte++, address += PAGE_SIZE) { 611b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 612b46e756fSKirill A. Shutemov if (pte_none(pteval) || (pte_present(pteval) && 613b46e756fSKirill A. Shutemov is_zero_pfn(pte_pfn(pteval)))) { 614b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 615b46e756fSKirill A. Shutemov ++none_or_zero <= khugepaged_max_ptes_none) { 616b46e756fSKirill A. Shutemov continue; 617b46e756fSKirill A. Shutemov } else { 618b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 619e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 620b46e756fSKirill A. Shutemov goto out; 621b46e756fSKirill A. Shutemov } 622b46e756fSKirill A. Shutemov } 623b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 624b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 625b46e756fSKirill A. Shutemov goto out; 626b46e756fSKirill A. Shutemov } 627b46e756fSKirill A. Shutemov page = vm_normal_page(vma, address, pteval); 628b46e756fSKirill A. Shutemov if (unlikely(!page)) { 629b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 630b46e756fSKirill A. Shutemov goto out; 631b46e756fSKirill A. Shutemov } 632b46e756fSKirill A. Shutemov 633b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageAnon(page), page); 634b46e756fSKirill A. Shutemov 63571a2c112SKirill A. Shutemov if (page_mapcount(page) > 1 && 63671a2c112SKirill A. Shutemov ++shared > khugepaged_max_ptes_shared) { 63771a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 638e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 63971a2c112SKirill A. Shutemov goto out; 64071a2c112SKirill A. Shutemov } 64171a2c112SKirill A. Shutemov 6425503fbf2SKirill A. Shutemov if (PageCompound(page)) { 6435503fbf2SKirill A. Shutemov struct page *p; 6445503fbf2SKirill A. Shutemov page = compound_head(page); 6455503fbf2SKirill A. Shutemov 6465503fbf2SKirill A. Shutemov /* 6475503fbf2SKirill A. Shutemov * Check if we have dealt with the compound page 6485503fbf2SKirill A. Shutemov * already 6495503fbf2SKirill A. Shutemov */ 6505503fbf2SKirill A. Shutemov list_for_each_entry(p, compound_pagelist, lru) { 6515503fbf2SKirill A. Shutemov if (page == p) 6525503fbf2SKirill A. Shutemov goto next; 6535503fbf2SKirill A. Shutemov } 6545503fbf2SKirill A. Shutemov } 6555503fbf2SKirill A. Shutemov 656b46e756fSKirill A. Shutemov /* 657b46e756fSKirill A. Shutemov * We can do it before isolate_lru_page because the 658b46e756fSKirill A. Shutemov * page can't be freed from under us. NOTE: PG_lock 659b46e756fSKirill A. Shutemov * is needed to serialize against split_huge_page 660b46e756fSKirill A. Shutemov * when invoked from the VM. 661b46e756fSKirill A. Shutemov */ 662b46e756fSKirill A. Shutemov if (!trylock_page(page)) { 663b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 664b46e756fSKirill A. Shutemov goto out; 665b46e756fSKirill A. Shutemov } 666b46e756fSKirill A. Shutemov 667b46e756fSKirill A. Shutemov /* 6689445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 6699445689fSKirill A. Shutemov * 6709445689fSKirill A. Shutemov * The page table that maps the page has been already unlinked 6719445689fSKirill A. Shutemov * from the page table tree and this process cannot get 672f0953a1bSIngo Molnar * an additional pin on the page. 6739445689fSKirill A. Shutemov * 6749445689fSKirill A. Shutemov * New pins can come later if the page is shared across fork, 6759445689fSKirill A. Shutemov * but not from this process. The other process cannot write to 6769445689fSKirill A. Shutemov * the page, only trigger CoW. 677b46e756fSKirill A. Shutemov */ 6789445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 679b46e756fSKirill A. Shutemov unlock_page(page); 680b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 681b46e756fSKirill A. Shutemov goto out; 682b46e756fSKirill A. Shutemov } 683b46e756fSKirill A. Shutemov 684b46e756fSKirill A. Shutemov /* 685b46e756fSKirill A. Shutemov * Isolate the page to avoid collapsing an hugepage 686b46e756fSKirill A. Shutemov * currently in use by the VM. 687b46e756fSKirill A. Shutemov */ 688b46e756fSKirill A. Shutemov if (isolate_lru_page(page)) { 689b46e756fSKirill A. Shutemov unlock_page(page); 690b46e756fSKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 691b46e756fSKirill A. Shutemov goto out; 692b46e756fSKirill A. Shutemov } 6935503fbf2SKirill A. Shutemov mod_node_page_state(page_pgdat(page), 6945503fbf2SKirill A. Shutemov NR_ISOLATED_ANON + page_is_file_lru(page), 6955503fbf2SKirill A. Shutemov compound_nr(page)); 696b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 697b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(PageLRU(page), page); 698b46e756fSKirill A. Shutemov 6995503fbf2SKirill A. Shutemov if (PageCompound(page)) 7005503fbf2SKirill A. Shutemov list_add_tail(&page->lru, compound_pagelist); 7015503fbf2SKirill A. Shutemov next: 7020db501f7SEbru Akagunduz /* There should be enough young pte to collapse the page */ 703b46e756fSKirill A. Shutemov if (pte_young(pteval) || 704b46e756fSKirill A. Shutemov page_is_young(page) || PageReferenced(page) || 705b46e756fSKirill A. Shutemov mmu_notifier_test_young(vma->vm_mm, address)) 7060db501f7SEbru Akagunduz referenced++; 7075503fbf2SKirill A. Shutemov 7085503fbf2SKirill A. Shutemov if (pte_write(pteval)) 7095503fbf2SKirill A. Shutemov writable = true; 710b46e756fSKirill A. Shutemov } 71174e579bfSMiaohe Lin 71274e579bfSMiaohe Lin if (unlikely(!writable)) { 71374e579bfSMiaohe Lin result = SCAN_PAGE_RO; 71474e579bfSMiaohe Lin } else if (unlikely(!referenced)) { 71574e579bfSMiaohe Lin result = SCAN_LACK_REFERENCED_PAGE; 71674e579bfSMiaohe Lin } else { 717b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 718b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 719b46e756fSKirill A. Shutemov referenced, writable, result); 720b46e756fSKirill A. Shutemov return 1; 721b46e756fSKirill A. Shutemov } 722b46e756fSKirill A. Shutemov out: 7235503fbf2SKirill A. Shutemov release_pte_pages(pte, _pte, compound_pagelist); 724b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 725b46e756fSKirill A. Shutemov referenced, writable, result); 726b46e756fSKirill A. Shutemov return 0; 727b46e756fSKirill A. Shutemov } 728b46e756fSKirill A. Shutemov 729b46e756fSKirill A. Shutemov static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 730b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 731b46e756fSKirill A. Shutemov unsigned long address, 7325503fbf2SKirill A. Shutemov spinlock_t *ptl, 7335503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 734b46e756fSKirill A. Shutemov { 7355503fbf2SKirill A. Shutemov struct page *src_page, *tmp; 736b46e756fSKirill A. Shutemov pte_t *_pte; 737338a16baSDavid Rientjes for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 738338a16baSDavid Rientjes _pte++, page++, address += PAGE_SIZE) { 739b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 740b46e756fSKirill A. Shutemov 741b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 742b46e756fSKirill A. Shutemov clear_user_highpage(page, address); 743b46e756fSKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 744b46e756fSKirill A. Shutemov if (is_zero_pfn(pte_pfn(pteval))) { 745b46e756fSKirill A. Shutemov /* 746b46e756fSKirill A. Shutemov * ptl mostly unnecessary. 747b46e756fSKirill A. Shutemov */ 748b46e756fSKirill A. Shutemov spin_lock(ptl); 74908d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 750b46e756fSKirill A. Shutemov spin_unlock(ptl); 751b46e756fSKirill A. Shutemov } 752b46e756fSKirill A. Shutemov } else { 753b46e756fSKirill A. Shutemov src_page = pte_page(pteval); 754b46e756fSKirill A. Shutemov copy_user_highpage(page, src_page, address, vma); 7555503fbf2SKirill A. Shutemov if (!PageCompound(src_page)) 756b46e756fSKirill A. Shutemov release_pte_page(src_page); 757b46e756fSKirill A. Shutemov /* 758b46e756fSKirill A. Shutemov * ptl mostly unnecessary, but preempt has to 759b46e756fSKirill A. Shutemov * be disabled to update the per-cpu stats 760b46e756fSKirill A. Shutemov * inside page_remove_rmap(). 761b46e756fSKirill A. Shutemov */ 762b46e756fSKirill A. Shutemov spin_lock(ptl); 76308d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 764cea86fe2SHugh Dickins page_remove_rmap(src_page, vma, false); 765b46e756fSKirill A. Shutemov spin_unlock(ptl); 766b46e756fSKirill A. Shutemov free_page_and_swap_cache(src_page); 767b46e756fSKirill A. Shutemov } 768b46e756fSKirill A. Shutemov } 7695503fbf2SKirill A. Shutemov 7705503fbf2SKirill A. Shutemov list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { 7715503fbf2SKirill A. Shutemov list_del(&src_page->lru); 7725503fbf2SKirill A. Shutemov release_pte_page(src_page); 7735503fbf2SKirill A. Shutemov } 774b46e756fSKirill A. Shutemov } 775b46e756fSKirill A. Shutemov 776b46e756fSKirill A. Shutemov static void khugepaged_alloc_sleep(void) 777b46e756fSKirill A. Shutemov { 778b46e756fSKirill A. Shutemov DEFINE_WAIT(wait); 779b46e756fSKirill A. Shutemov 780b46e756fSKirill A. Shutemov add_wait_queue(&khugepaged_wait, &wait); 781b46e756fSKirill A. Shutemov freezable_schedule_timeout_interruptible( 782b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 783b46e756fSKirill A. Shutemov remove_wait_queue(&khugepaged_wait, &wait); 784b46e756fSKirill A. Shutemov } 785b46e756fSKirill A. Shutemov 786b46e756fSKirill A. Shutemov static int khugepaged_node_load[MAX_NUMNODES]; 787b46e756fSKirill A. Shutemov 788b46e756fSKirill A. Shutemov static bool khugepaged_scan_abort(int nid) 789b46e756fSKirill A. Shutemov { 790b46e756fSKirill A. Shutemov int i; 791b46e756fSKirill A. Shutemov 792b46e756fSKirill A. Shutemov /* 793a5f5f91dSMel Gorman * If node_reclaim_mode is disabled, then no extra effort is made to 794b46e756fSKirill A. Shutemov * allocate memory locally. 795b46e756fSKirill A. Shutemov */ 796202e35dbSDave Hansen if (!node_reclaim_enabled()) 797b46e756fSKirill A. Shutemov return false; 798b46e756fSKirill A. Shutemov 799b46e756fSKirill A. Shutemov /* If there is a count for this node already, it must be acceptable */ 800b46e756fSKirill A. Shutemov if (khugepaged_node_load[nid]) 801b46e756fSKirill A. Shutemov return false; 802b46e756fSKirill A. Shutemov 803b46e756fSKirill A. Shutemov for (i = 0; i < MAX_NUMNODES; i++) { 804b46e756fSKirill A. Shutemov if (!khugepaged_node_load[i]) 805b46e756fSKirill A. Shutemov continue; 806a55c7454SMatt Fleming if (node_distance(nid, i) > node_reclaim_distance) 807b46e756fSKirill A. Shutemov return true; 808b46e756fSKirill A. Shutemov } 809b46e756fSKirill A. Shutemov return false; 810b46e756fSKirill A. Shutemov } 811b46e756fSKirill A. Shutemov 812b46e756fSKirill A. Shutemov /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 813b46e756fSKirill A. Shutemov static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 814b46e756fSKirill A. Shutemov { 81525160354SVlastimil Babka return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 816b46e756fSKirill A. Shutemov } 817b46e756fSKirill A. Shutemov 818b46e756fSKirill A. Shutemov #ifdef CONFIG_NUMA 819b46e756fSKirill A. Shutemov static int khugepaged_find_target_node(void) 820b46e756fSKirill A. Shutemov { 821b46e756fSKirill A. Shutemov static int last_khugepaged_target_node = NUMA_NO_NODE; 822b46e756fSKirill A. Shutemov int nid, target_node = 0, max_value = 0; 823b46e756fSKirill A. Shutemov 824b46e756fSKirill A. Shutemov /* find first node with max normal pages hit */ 825b46e756fSKirill A. Shutemov for (nid = 0; nid < MAX_NUMNODES; nid++) 826b46e756fSKirill A. Shutemov if (khugepaged_node_load[nid] > max_value) { 827b46e756fSKirill A. Shutemov max_value = khugepaged_node_load[nid]; 828b46e756fSKirill A. Shutemov target_node = nid; 829b46e756fSKirill A. Shutemov } 830b46e756fSKirill A. Shutemov 831b46e756fSKirill A. Shutemov /* do some balance if several nodes have the same hit record */ 832b46e756fSKirill A. Shutemov if (target_node <= last_khugepaged_target_node) 833b46e756fSKirill A. Shutemov for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; 834b46e756fSKirill A. Shutemov nid++) 835b46e756fSKirill A. Shutemov if (max_value == khugepaged_node_load[nid]) { 836b46e756fSKirill A. Shutemov target_node = nid; 837b46e756fSKirill A. Shutemov break; 838b46e756fSKirill A. Shutemov } 839b46e756fSKirill A. Shutemov 840b46e756fSKirill A. Shutemov last_khugepaged_target_node = target_node; 841b46e756fSKirill A. Shutemov return target_node; 842b46e756fSKirill A. Shutemov } 843b46e756fSKirill A. Shutemov 844b46e756fSKirill A. Shutemov static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 845b46e756fSKirill A. Shutemov { 846b46e756fSKirill A. Shutemov if (IS_ERR(*hpage)) { 847b46e756fSKirill A. Shutemov if (!*wait) 848b46e756fSKirill A. Shutemov return false; 849b46e756fSKirill A. Shutemov 850b46e756fSKirill A. Shutemov *wait = false; 851b46e756fSKirill A. Shutemov *hpage = NULL; 852b46e756fSKirill A. Shutemov khugepaged_alloc_sleep(); 853b46e756fSKirill A. Shutemov } else if (*hpage) { 854b46e756fSKirill A. Shutemov put_page(*hpage); 855b46e756fSKirill A. Shutemov *hpage = NULL; 856b46e756fSKirill A. Shutemov } 857b46e756fSKirill A. Shutemov 858b46e756fSKirill A. Shutemov return true; 859b46e756fSKirill A. Shutemov } 860b46e756fSKirill A. Shutemov 861b46e756fSKirill A. Shutemov static struct page * 862988ddb71SKirill A. Shutemov khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) 863b46e756fSKirill A. Shutemov { 864b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(*hpage, *hpage); 865b46e756fSKirill A. Shutemov 866b46e756fSKirill A. Shutemov *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); 867b46e756fSKirill A. Shutemov if (unlikely(!*hpage)) { 868b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 869b46e756fSKirill A. Shutemov *hpage = ERR_PTR(-ENOMEM); 870b46e756fSKirill A. Shutemov return NULL; 871b46e756fSKirill A. Shutemov } 872b46e756fSKirill A. Shutemov 873b46e756fSKirill A. Shutemov prep_transhuge_page(*hpage); 874b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 875b46e756fSKirill A. Shutemov return *hpage; 876b46e756fSKirill A. Shutemov } 877b46e756fSKirill A. Shutemov #else 878b46e756fSKirill A. Shutemov static int khugepaged_find_target_node(void) 879b46e756fSKirill A. Shutemov { 880b46e756fSKirill A. Shutemov return 0; 881b46e756fSKirill A. Shutemov } 882b46e756fSKirill A. Shutemov 883b46e756fSKirill A. Shutemov static inline struct page *alloc_khugepaged_hugepage(void) 884b46e756fSKirill A. Shutemov { 885b46e756fSKirill A. Shutemov struct page *page; 886b46e756fSKirill A. Shutemov 887b46e756fSKirill A. Shutemov page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(), 888b46e756fSKirill A. Shutemov HPAGE_PMD_ORDER); 889b46e756fSKirill A. Shutemov if (page) 890b46e756fSKirill A. Shutemov prep_transhuge_page(page); 891b46e756fSKirill A. Shutemov return page; 892b46e756fSKirill A. Shutemov } 893b46e756fSKirill A. Shutemov 894b46e756fSKirill A. Shutemov static struct page *khugepaged_alloc_hugepage(bool *wait) 895b46e756fSKirill A. Shutemov { 896b46e756fSKirill A. Shutemov struct page *hpage; 897b46e756fSKirill A. Shutemov 898b46e756fSKirill A. Shutemov do { 899b46e756fSKirill A. Shutemov hpage = alloc_khugepaged_hugepage(); 900b46e756fSKirill A. Shutemov if (!hpage) { 901b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 902b46e756fSKirill A. Shutemov if (!*wait) 903b46e756fSKirill A. Shutemov return NULL; 904b46e756fSKirill A. Shutemov 905b46e756fSKirill A. Shutemov *wait = false; 906b46e756fSKirill A. Shutemov khugepaged_alloc_sleep(); 907b46e756fSKirill A. Shutemov } else 908b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 909b46e756fSKirill A. Shutemov } while (unlikely(!hpage) && likely(khugepaged_enabled())); 910b46e756fSKirill A. Shutemov 911b46e756fSKirill A. Shutemov return hpage; 912b46e756fSKirill A. Shutemov } 913b46e756fSKirill A. Shutemov 914b46e756fSKirill A. Shutemov static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 915b46e756fSKirill A. Shutemov { 916033b5d77SHugh Dickins /* 917033b5d77SHugh Dickins * If the hpage allocated earlier was briefly exposed in page cache 918033b5d77SHugh Dickins * before collapse_file() failed, it is possible that racing lookups 919033b5d77SHugh Dickins * have not yet completed, and would then be unpleasantly surprised by 920033b5d77SHugh Dickins * finding the hpage reused for the same mapping at a different offset. 921033b5d77SHugh Dickins * Just release the previous allocation if there is any danger of that. 922033b5d77SHugh Dickins */ 923033b5d77SHugh Dickins if (*hpage && page_count(*hpage) > 1) { 924033b5d77SHugh Dickins put_page(*hpage); 925033b5d77SHugh Dickins *hpage = NULL; 926033b5d77SHugh Dickins } 927033b5d77SHugh Dickins 928b46e756fSKirill A. Shutemov if (!*hpage) 929b46e756fSKirill A. Shutemov *hpage = khugepaged_alloc_hugepage(wait); 930b46e756fSKirill A. Shutemov 931b46e756fSKirill A. Shutemov if (unlikely(!*hpage)) 932b46e756fSKirill A. Shutemov return false; 933b46e756fSKirill A. Shutemov 934b46e756fSKirill A. Shutemov return true; 935b46e756fSKirill A. Shutemov } 936b46e756fSKirill A. Shutemov 937b46e756fSKirill A. Shutemov static struct page * 938988ddb71SKirill A. Shutemov khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) 939b46e756fSKirill A. Shutemov { 940b46e756fSKirill A. Shutemov VM_BUG_ON(!*hpage); 941b46e756fSKirill A. Shutemov 942b46e756fSKirill A. Shutemov return *hpage; 943b46e756fSKirill A. Shutemov } 944b46e756fSKirill A. Shutemov #endif 945b46e756fSKirill A. Shutemov 946b46e756fSKirill A. Shutemov /* 947c1e8d7c6SMichel Lespinasse * If mmap_lock temporarily dropped, revalidate vma 948c1e8d7c6SMichel Lespinasse * before taking mmap_lock. 949b46e756fSKirill A. Shutemov * Return 0 if succeeds, otherwise return none-zero 950b46e756fSKirill A. Shutemov * value (scan code). 951b46e756fSKirill A. Shutemov */ 952b46e756fSKirill A. Shutemov 953c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 954c131f751SKirill A. Shutemov struct vm_area_struct **vmap) 955b46e756fSKirill A. Shutemov { 956b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 957b46e756fSKirill A. Shutemov unsigned long hstart, hend; 958b46e756fSKirill A. Shutemov 959b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 960b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 961b46e756fSKirill A. Shutemov 962c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 963b46e756fSKirill A. Shutemov if (!vma) 964b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 965b46e756fSKirill A. Shutemov 966b46e756fSKirill A. Shutemov hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 967b46e756fSKirill A. Shutemov hend = vma->vm_end & HPAGE_PMD_MASK; 968b46e756fSKirill A. Shutemov if (address < hstart || address + HPAGE_PMD_SIZE > hend) 969b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 97050f8b92fSSong Liu if (!hugepage_vma_check(vma, vma->vm_flags)) 971b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 972594cced1SKirill A. Shutemov /* Anon VMA expected */ 97325fa414aSxu xin if (!vma->anon_vma || !vma_is_anonymous(vma)) 974594cced1SKirill A. Shutemov return SCAN_VMA_CHECK; 975b46e756fSKirill A. Shutemov return 0; 976b46e756fSKirill A. Shutemov } 977b46e756fSKirill A. Shutemov 978b46e756fSKirill A. Shutemov /* 979b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 980b46e756fSKirill A. Shutemov * Only done if khugepaged_scan_pmd believes it is worthwhile. 981b46e756fSKirill A. Shutemov * 982b46e756fSKirill A. Shutemov * Called and returns without pte mapped or spinlocks held, 983c1e8d7c6SMichel Lespinasse * but with mmap_lock held to protect against vma changes. 984b46e756fSKirill A. Shutemov */ 985b46e756fSKirill A. Shutemov 986b46e756fSKirill A. Shutemov static bool __collapse_huge_page_swapin(struct mm_struct *mm, 987b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 9882b635dd3SWill Deacon unsigned long haddr, pmd_t *pmd, 9890db501f7SEbru Akagunduz int referenced) 990b46e756fSKirill A. Shutemov { 9912b740303SSouptick Joarder int swapped_in = 0; 9922b740303SSouptick Joarder vm_fault_t ret = 0; 9932b635dd3SWill Deacon unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); 9942b635dd3SWill Deacon 9952b635dd3SWill Deacon for (address = haddr; address < end; address += PAGE_SIZE) { 99682b0f8c3SJan Kara struct vm_fault vmf = { 997b46e756fSKirill A. Shutemov .vma = vma, 998b46e756fSKirill A. Shutemov .address = address, 9992b635dd3SWill Deacon .pgoff = linear_page_index(vma, haddr), 1000b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 1001b46e756fSKirill A. Shutemov .pmd = pmd, 1002b46e756fSKirill A. Shutemov }; 1003b46e756fSKirill A. Shutemov 100482b0f8c3SJan Kara vmf.pte = pte_offset_map(pmd, address); 10052994302bSJan Kara vmf.orig_pte = *vmf.pte; 10062b635dd3SWill Deacon if (!is_swap_pte(vmf.orig_pte)) { 10072b635dd3SWill Deacon pte_unmap(vmf.pte); 1008b46e756fSKirill A. Shutemov continue; 10092b635dd3SWill Deacon } 1010b46e756fSKirill A. Shutemov swapped_in++; 10112994302bSJan Kara ret = do_swap_page(&vmf); 10120db501f7SEbru Akagunduz 1013c1e8d7c6SMichel Lespinasse /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */ 1014b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 1015d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 10162b635dd3SWill Deacon if (hugepage_vma_revalidate(mm, haddr, &vma)) { 1017b46e756fSKirill A. Shutemov /* vma is no longer available, don't continue to swapin */ 10180db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 1019b46e756fSKirill A. Shutemov return false; 102047f863eaSEbru Akagunduz } 1021b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 10222b635dd3SWill Deacon if (mm_find_pmd(mm, haddr) != pmd) { 1023835152a2SSeongJae Park trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 1024b46e756fSKirill A. Shutemov return false; 1025b46e756fSKirill A. Shutemov } 1026835152a2SSeongJae Park } 1027b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 10280db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 1029b46e756fSKirill A. Shutemov return false; 1030b46e756fSKirill A. Shutemov } 1031b46e756fSKirill A. Shutemov } 1032ae2c5d80SKirill A. Shutemov 1033ae2c5d80SKirill A. Shutemov /* Drain LRU add pagevec to remove extra pin on the swapped in pages */ 1034ae2c5d80SKirill A. Shutemov if (swapped_in) 1035ae2c5d80SKirill A. Shutemov lru_add_drain(); 1036ae2c5d80SKirill A. Shutemov 10370db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); 1038b46e756fSKirill A. Shutemov return true; 1039b46e756fSKirill A. Shutemov } 1040b46e756fSKirill A. Shutemov 1041b46e756fSKirill A. Shutemov static void collapse_huge_page(struct mm_struct *mm, 1042b46e756fSKirill A. Shutemov unsigned long address, 1043b46e756fSKirill A. Shutemov struct page **hpage, 1044ffe945e6SKirill A. Shutemov int node, int referenced, int unmapped) 1045b46e756fSKirill A. Shutemov { 10465503fbf2SKirill A. Shutemov LIST_HEAD(compound_pagelist); 1047b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 1048b46e756fSKirill A. Shutemov pte_t *pte; 1049b46e756fSKirill A. Shutemov pgtable_t pgtable; 1050b46e756fSKirill A. Shutemov struct page *new_page; 1051b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 1052b46e756fSKirill A. Shutemov int isolated = 0, result = 0; 1053c131f751SKirill A. Shutemov struct vm_area_struct *vma; 1054ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1055b46e756fSKirill A. Shutemov gfp_t gfp; 1056b46e756fSKirill A. Shutemov 1057b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1058b46e756fSKirill A. Shutemov 1059b46e756fSKirill A. Shutemov /* Only allocate from the target node */ 106041b6167eSMichal Hocko gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 1061b46e756fSKirill A. Shutemov 1062988ddb71SKirill A. Shutemov /* 1063c1e8d7c6SMichel Lespinasse * Before allocating the hugepage, release the mmap_lock read lock. 1064988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 1065c1e8d7c6SMichel Lespinasse * sync compaction, and we do not need to hold the mmap_lock during 1066988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 1067988ddb71SKirill A. Shutemov */ 1068d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1069988ddb71SKirill A. Shutemov new_page = khugepaged_alloc_page(hpage, gfp, node); 1070b46e756fSKirill A. Shutemov if (!new_page) { 1071b46e756fSKirill A. Shutemov result = SCAN_ALLOC_HUGE_PAGE_FAIL; 1072b46e756fSKirill A. Shutemov goto out_nolock; 1073b46e756fSKirill A. Shutemov } 1074b46e756fSKirill A. Shutemov 10758f425e4eSMatthew Wilcox (Oracle) if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) { 1076b46e756fSKirill A. Shutemov result = SCAN_CGROUP_CHARGE_FAIL; 1077b46e756fSKirill A. Shutemov goto out_nolock; 1078b46e756fSKirill A. Shutemov } 10799d82c694SJohannes Weiner count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC); 1080b46e756fSKirill A. Shutemov 1081d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 1082c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 1083b46e756fSKirill A. Shutemov if (result) { 1084d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1085b46e756fSKirill A. Shutemov goto out_nolock; 1086b46e756fSKirill A. Shutemov } 1087b46e756fSKirill A. Shutemov 1088b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 1089b46e756fSKirill A. Shutemov if (!pmd) { 1090b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 1091d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1092b46e756fSKirill A. Shutemov goto out_nolock; 1093b46e756fSKirill A. Shutemov } 1094b46e756fSKirill A. Shutemov 1095b46e756fSKirill A. Shutemov /* 1096c1e8d7c6SMichel Lespinasse * __collapse_huge_page_swapin always returns with mmap_lock locked. 1097c1e8d7c6SMichel Lespinasse * If it fails, we release mmap_lock and jump out_nolock. 1098b46e756fSKirill A. Shutemov * Continuing to collapse causes inconsistency. 1099b46e756fSKirill A. Shutemov */ 1100ffe945e6SKirill A. Shutemov if (unmapped && !__collapse_huge_page_swapin(mm, vma, address, 1101ffe945e6SKirill A. Shutemov pmd, referenced)) { 1102d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1103b46e756fSKirill A. Shutemov goto out_nolock; 1104b46e756fSKirill A. Shutemov } 1105b46e756fSKirill A. Shutemov 1106d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1107b46e756fSKirill A. Shutemov /* 1108b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 1109b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 1110b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 1111b46e756fSKirill A. Shutemov */ 1112d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 1113c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 1114b46e756fSKirill A. Shutemov if (result) 111518d24a7cSMiaohe Lin goto out_up_write; 1116b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 1117b46e756fSKirill A. Shutemov if (mm_find_pmd(mm, address) != pmd) 111818d24a7cSMiaohe Lin goto out_up_write; 1119b46e756fSKirill A. Shutemov 1120b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1121b46e756fSKirill A. Shutemov 11227269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 11236f4f13e8SJérôme Glisse address, address + HPAGE_PMD_SIZE); 1124ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1125ec649c9dSVille Syrjälä 1126ec649c9dSVille Syrjälä pte = pte_offset_map(pmd, address); 1127ec649c9dSVille Syrjälä pte_ptl = pte_lockptr(mm, pmd); 1128ec649c9dSVille Syrjälä 1129b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1130b46e756fSKirill A. Shutemov /* 1131b46e756fSKirill A. Shutemov * After this gup_fast can't run anymore. This also removes 1132b46e756fSKirill A. Shutemov * any huge TLB entry from the CPU so we won't allow 1133b46e756fSKirill A. Shutemov * huge and small TLB entries for the same virtual address 1134b46e756fSKirill A. Shutemov * to avoid the risk of CPU bugs in that area. 1135b46e756fSKirill A. Shutemov */ 1136b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1137b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1138ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1139b46e756fSKirill A. Shutemov 1140b46e756fSKirill A. Shutemov spin_lock(pte_ptl); 11415503fbf2SKirill A. Shutemov isolated = __collapse_huge_page_isolate(vma, address, pte, 11425503fbf2SKirill A. Shutemov &compound_pagelist); 1143b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1144b46e756fSKirill A. Shutemov 1145b46e756fSKirill A. Shutemov if (unlikely(!isolated)) { 1146b46e756fSKirill A. Shutemov pte_unmap(pte); 1147b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1148b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1149b46e756fSKirill A. Shutemov /* 1150b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1151b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1152b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1153b46e756fSKirill A. Shutemov */ 1154b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1155b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1156b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1157b46e756fSKirill A. Shutemov result = SCAN_FAIL; 115818d24a7cSMiaohe Lin goto out_up_write; 1159b46e756fSKirill A. Shutemov } 1160b46e756fSKirill A. Shutemov 1161b46e756fSKirill A. Shutemov /* 1162b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1163b46e756fSKirill A. Shutemov * can't run anymore. 1164b46e756fSKirill A. Shutemov */ 1165b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1166b46e756fSKirill A. Shutemov 11675503fbf2SKirill A. Shutemov __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl, 11685503fbf2SKirill A. Shutemov &compound_pagelist); 1169b46e756fSKirill A. Shutemov pte_unmap(pte); 1170588d01f9SMiaohe Lin /* 1171588d01f9SMiaohe Lin * spin_lock() below is not the equivalent of smp_wmb(), but 1172588d01f9SMiaohe Lin * the smp_wmb() inside __SetPageUptodate() can be reused to 1173588d01f9SMiaohe Lin * avoid the copy_huge_page writes to become visible after 1174588d01f9SMiaohe Lin * the set_pmd_at() write. 1175588d01f9SMiaohe Lin */ 1176b46e756fSKirill A. Shutemov __SetPageUptodate(new_page); 1177b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1178b46e756fSKirill A. Shutemov 1179b46e756fSKirill A. Shutemov _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); 1180f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1181b46e756fSKirill A. Shutemov 1182b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1183b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 118440f2bbf7SDavid Hildenbrand page_add_new_anon_rmap(new_page, vma, address); 1185b518154eSJoonsoo Kim lru_cache_add_inactive_or_unevictable(new_page, vma); 1186b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1187b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1188b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1189b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1190b46e756fSKirill A. Shutemov 1191b46e756fSKirill A. Shutemov *hpage = NULL; 1192b46e756fSKirill A. Shutemov 1193b46e756fSKirill A. Shutemov khugepaged_pages_collapsed++; 1194b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1195b46e756fSKirill A. Shutemov out_up_write: 1196d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 1197b46e756fSKirill A. Shutemov out_nolock: 11989d82c694SJohannes Weiner if (!IS_ERR_OR_NULL(*hpage)) 1199bbc6b703SMatthew Wilcox (Oracle) mem_cgroup_uncharge(page_folio(*hpage)); 1200b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page(mm, isolated, result); 1201b46e756fSKirill A. Shutemov return; 1202b46e756fSKirill A. Shutemov } 1203b46e756fSKirill A. Shutemov 1204b46e756fSKirill A. Shutemov static int khugepaged_scan_pmd(struct mm_struct *mm, 1205b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 1206b46e756fSKirill A. Shutemov unsigned long address, 1207b46e756fSKirill A. Shutemov struct page **hpage) 1208b46e756fSKirill A. Shutemov { 1209b46e756fSKirill A. Shutemov pmd_t *pmd; 1210b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 121171a2c112SKirill A. Shutemov int ret = 0, result = 0, referenced = 0; 121271a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0; 1213b46e756fSKirill A. Shutemov struct page *page = NULL; 1214b46e756fSKirill A. Shutemov unsigned long _address; 1215b46e756fSKirill A. Shutemov spinlock_t *ptl; 1216b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 12170db501f7SEbru Akagunduz bool writable = false; 1218b46e756fSKirill A. Shutemov 1219b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1220b46e756fSKirill A. Shutemov 1221b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 1222b46e756fSKirill A. Shutemov if (!pmd) { 1223b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 1224b46e756fSKirill A. Shutemov goto out; 1225b46e756fSKirill A. Shutemov } 1226b46e756fSKirill A. Shutemov 1227b46e756fSKirill A. Shutemov memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 1228b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1229b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 1230b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1231b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 1232b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1233b46e756fSKirill A. Shutemov if (++unmapped <= khugepaged_max_ptes_swap) { 1234e1e267c7SPeter Xu /* 1235e1e267c7SPeter Xu * Always be strict with uffd-wp 1236e1e267c7SPeter Xu * enabled swap entries. Please see 1237e1e267c7SPeter Xu * comment below for pte_uffd_wp(). 1238e1e267c7SPeter Xu */ 1239e1e267c7SPeter Xu if (pte_swp_uffd_wp(pteval)) { 1240e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1241e1e267c7SPeter Xu goto out_unmap; 1242e1e267c7SPeter Xu } 1243b46e756fSKirill A. Shutemov continue; 1244b46e756fSKirill A. Shutemov } else { 1245b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1246e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1247b46e756fSKirill A. Shutemov goto out_unmap; 1248b46e756fSKirill A. Shutemov } 1249b46e756fSKirill A. Shutemov } 1250b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1251b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1252b46e756fSKirill A. Shutemov ++none_or_zero <= khugepaged_max_ptes_none) { 1253b46e756fSKirill A. Shutemov continue; 1254b46e756fSKirill A. Shutemov } else { 1255b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1256e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1257b46e756fSKirill A. Shutemov goto out_unmap; 1258b46e756fSKirill A. Shutemov } 1259b46e756fSKirill A. Shutemov } 1260e1e267c7SPeter Xu if (pte_uffd_wp(pteval)) { 1261e1e267c7SPeter Xu /* 1262e1e267c7SPeter Xu * Don't collapse the page if any of the small 1263e1e267c7SPeter Xu * PTEs are armed with uffd write protection. 1264e1e267c7SPeter Xu * Here we can also mark the new huge pmd as 1265e1e267c7SPeter Xu * write protected if any of the small ones is 12668958b249SHaitao Shi * marked but that could bring unknown 1267e1e267c7SPeter Xu * userfault messages that falls outside of 1268e1e267c7SPeter Xu * the registered range. So, just be simple. 1269e1e267c7SPeter Xu */ 1270e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1271e1e267c7SPeter Xu goto out_unmap; 1272e1e267c7SPeter Xu } 1273b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1274b46e756fSKirill A. Shutemov writable = true; 1275b46e756fSKirill A. Shutemov 1276b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 1277b46e756fSKirill A. Shutemov if (unlikely(!page)) { 1278b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1279b46e756fSKirill A. Shutemov goto out_unmap; 1280b46e756fSKirill A. Shutemov } 1281b46e756fSKirill A. Shutemov 128271a2c112SKirill A. Shutemov if (page_mapcount(page) > 1 && 128371a2c112SKirill A. Shutemov ++shared > khugepaged_max_ptes_shared) { 128471a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 1285e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 128671a2c112SKirill A. Shutemov goto out_unmap; 128771a2c112SKirill A. Shutemov } 128871a2c112SKirill A. Shutemov 12895503fbf2SKirill A. Shutemov page = compound_head(page); 1290b46e756fSKirill A. Shutemov 1291b46e756fSKirill A. Shutemov /* 1292b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 1293b46e756fSKirill A. Shutemov * information to khugepaged_node_load[]. 12940b8f0d87SQuanfa Fu * Khugepaged will allocate hugepage from the node has the max 1295b46e756fSKirill A. Shutemov * hit record. 1296b46e756fSKirill A. Shutemov */ 1297b46e756fSKirill A. Shutemov node = page_to_nid(page); 1298b46e756fSKirill A. Shutemov if (khugepaged_scan_abort(node)) { 1299b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1300b46e756fSKirill A. Shutemov goto out_unmap; 1301b46e756fSKirill A. Shutemov } 1302b46e756fSKirill A. Shutemov khugepaged_node_load[node]++; 1303b46e756fSKirill A. Shutemov if (!PageLRU(page)) { 1304b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1305b46e756fSKirill A. Shutemov goto out_unmap; 1306b46e756fSKirill A. Shutemov } 1307b46e756fSKirill A. Shutemov if (PageLocked(page)) { 1308b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1309b46e756fSKirill A. Shutemov goto out_unmap; 1310b46e756fSKirill A. Shutemov } 1311b46e756fSKirill A. Shutemov if (!PageAnon(page)) { 1312b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1313b46e756fSKirill A. Shutemov goto out_unmap; 1314b46e756fSKirill A. Shutemov } 1315b46e756fSKirill A. Shutemov 1316b46e756fSKirill A. Shutemov /* 13179445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 13189445689fSKirill A. Shutemov * 13199445689fSKirill A. Shutemov * Here the check is racy it may see totmal_mapcount > refcount 13209445689fSKirill A. Shutemov * in some cases. 13219445689fSKirill A. Shutemov * For example, one process with one forked child process. 13229445689fSKirill A. Shutemov * The parent has the PMD split due to MADV_DONTNEED, then 13239445689fSKirill A. Shutemov * the child is trying unmap the whole PMD, but khugepaged 13249445689fSKirill A. Shutemov * may be scanning the parent between the child has 13259445689fSKirill A. Shutemov * PageDoubleMap flag cleared and dec the mapcount. So 13269445689fSKirill A. Shutemov * khugepaged may see total_mapcount > refcount. 13279445689fSKirill A. Shutemov * 13289445689fSKirill A. Shutemov * But such case is ephemeral we could always retry collapse 13299445689fSKirill A. Shutemov * later. However it may report false positive if the page 13309445689fSKirill A. Shutemov * has excessive GUP pins (i.e. 512). Anyway the same check 13319445689fSKirill A. Shutemov * will be done again later the risk seems low. 1332b46e756fSKirill A. Shutemov */ 13339445689fSKirill A. Shutemov if (!is_refcount_suitable(page)) { 1334b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1335b46e756fSKirill A. Shutemov goto out_unmap; 1336b46e756fSKirill A. Shutemov } 1337b46e756fSKirill A. Shutemov if (pte_young(pteval) || 1338b46e756fSKirill A. Shutemov page_is_young(page) || PageReferenced(page) || 1339b46e756fSKirill A. Shutemov mmu_notifier_test_young(vma->vm_mm, address)) 13400db501f7SEbru Akagunduz referenced++; 1341b46e756fSKirill A. Shutemov } 1342ffe945e6SKirill A. Shutemov if (!writable) { 1343ffe945e6SKirill A. Shutemov result = SCAN_PAGE_RO; 1344ffe945e6SKirill A. Shutemov } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) { 1345ffe945e6SKirill A. Shutemov result = SCAN_LACK_REFERENCED_PAGE; 1346ffe945e6SKirill A. Shutemov } else { 1347b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1348b46e756fSKirill A. Shutemov ret = 1; 1349b46e756fSKirill A. Shutemov } 1350b46e756fSKirill A. Shutemov out_unmap: 1351b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 1352b46e756fSKirill A. Shutemov if (ret) { 1353b46e756fSKirill A. Shutemov node = khugepaged_find_target_node(); 1354c1e8d7c6SMichel Lespinasse /* collapse_huge_page will return with the mmap_lock released */ 1355ffe945e6SKirill A. Shutemov collapse_huge_page(mm, address, hpage, node, 1356ffe945e6SKirill A. Shutemov referenced, unmapped); 1357b46e756fSKirill A. Shutemov } 1358b46e756fSKirill A. Shutemov out: 1359b46e756fSKirill A. Shutemov trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, 1360b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 1361b46e756fSKirill A. Shutemov return ret; 1362b46e756fSKirill A. Shutemov } 1363b46e756fSKirill A. Shutemov 1364b46e756fSKirill A. Shutemov static void collect_mm_slot(struct mm_slot *mm_slot) 1365b46e756fSKirill A. Shutemov { 1366b46e756fSKirill A. Shutemov struct mm_struct *mm = mm_slot->mm; 1367b46e756fSKirill A. Shutemov 136835f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1369b46e756fSKirill A. Shutemov 1370b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm)) { 1371b46e756fSKirill A. Shutemov /* free mm_slot */ 1372b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 1373b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 1374b46e756fSKirill A. Shutemov 1375b46e756fSKirill A. Shutemov /* 1376b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1377b46e756fSKirill A. Shutemov * 1378b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1379b46e756fSKirill A. Shutemov */ 1380b46e756fSKirill A. Shutemov 1381b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1382b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 1383b46e756fSKirill A. Shutemov mmdrop(mm); 1384b46e756fSKirill A. Shutemov } 1385b46e756fSKirill A. Shutemov } 1386b46e756fSKirill A. Shutemov 1387396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 138827e1f827SSong Liu /* 138927e1f827SSong Liu * Notify khugepaged that given addr of the mm is pte-mapped THP. Then 139027e1f827SSong Liu * khugepaged should try to collapse the page table. 139127e1f827SSong Liu */ 139227e1f827SSong Liu static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm, 139327e1f827SSong Liu unsigned long addr) 139427e1f827SSong Liu { 139527e1f827SSong Liu struct mm_slot *mm_slot; 139627e1f827SSong Liu 139727e1f827SSong Liu VM_BUG_ON(addr & ~HPAGE_PMD_MASK); 139827e1f827SSong Liu 139927e1f827SSong Liu spin_lock(&khugepaged_mm_lock); 140027e1f827SSong Liu mm_slot = get_mm_slot(mm); 140127e1f827SSong Liu if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) 140227e1f827SSong Liu mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; 140327e1f827SSong Liu spin_unlock(&khugepaged_mm_lock); 140427e1f827SSong Liu return 0; 140527e1f827SSong Liu } 140627e1f827SSong Liu 1407e59a47b8SPasha Tatashin static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 1408e59a47b8SPasha Tatashin unsigned long addr, pmd_t *pmdp) 1409e59a47b8SPasha Tatashin { 1410e59a47b8SPasha Tatashin spinlock_t *ptl; 1411e59a47b8SPasha Tatashin pmd_t pmd; 1412e59a47b8SPasha Tatashin 141380110bbfSPasha Tatashin mmap_assert_write_locked(mm); 1414e59a47b8SPasha Tatashin ptl = pmd_lock(vma->vm_mm, pmdp); 1415e59a47b8SPasha Tatashin pmd = pmdp_collapse_flush(vma, addr, pmdp); 1416e59a47b8SPasha Tatashin spin_unlock(ptl); 1417e59a47b8SPasha Tatashin mm_dec_nr_ptes(mm); 141880110bbfSPasha Tatashin page_table_check_pte_clear_range(mm, addr, pmd); 1419e59a47b8SPasha Tatashin pte_free(mm, pmd_pgtable(pmd)); 1420e59a47b8SPasha Tatashin } 1421e59a47b8SPasha Tatashin 142227e1f827SSong Liu /** 1423336e6b53SAlex Shi * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at 1424336e6b53SAlex Shi * address haddr. 1425336e6b53SAlex Shi * 1426336e6b53SAlex Shi * @mm: process address space where collapse happens 1427336e6b53SAlex Shi * @addr: THP collapse address 142827e1f827SSong Liu * 142927e1f827SSong Liu * This function checks whether all the PTEs in the PMD are pointing to the 143027e1f827SSong Liu * right THP. If so, retract the page table so the THP can refault in with 143127e1f827SSong Liu * as pmd-mapped. 143227e1f827SSong Liu */ 143327e1f827SSong Liu void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr) 143427e1f827SSong Liu { 143527e1f827SSong Liu unsigned long haddr = addr & HPAGE_PMD_MASK; 143627e1f827SSong Liu struct vm_area_struct *vma = find_vma(mm, haddr); 1437119a5fc1SHugh Dickins struct page *hpage; 143827e1f827SSong Liu pte_t *start_pte, *pte; 1439e59a47b8SPasha Tatashin pmd_t *pmd; 144027e1f827SSong Liu spinlock_t *ptl; 144127e1f827SSong Liu int count = 0; 144227e1f827SSong Liu int i; 144327e1f827SSong Liu 144427e1f827SSong Liu if (!vma || !vma->vm_file || 1445fef792a4SMiaohe Lin !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) 144627e1f827SSong Liu return; 144727e1f827SSong Liu 144827e1f827SSong Liu /* 144927e1f827SSong Liu * This vm_flags may not have VM_HUGEPAGE if the page was not 145027e1f827SSong Liu * collapsed by this mm. But we can still collapse if the page is 145127e1f827SSong Liu * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check() 145227e1f827SSong Liu * will not fail the vma for missing VM_HUGEPAGE 145327e1f827SSong Liu */ 145427e1f827SSong Liu if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE)) 145527e1f827SSong Liu return; 145627e1f827SSong Liu 1457deb4c93aSPeter Xu /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ 1458deb4c93aSPeter Xu if (userfaultfd_wp(vma)) 1459deb4c93aSPeter Xu return; 1460deb4c93aSPeter Xu 1461119a5fc1SHugh Dickins hpage = find_lock_page(vma->vm_file->f_mapping, 1462119a5fc1SHugh Dickins linear_page_index(vma, haddr)); 1463119a5fc1SHugh Dickins if (!hpage) 1464119a5fc1SHugh Dickins return; 1465119a5fc1SHugh Dickins 1466119a5fc1SHugh Dickins if (!PageHead(hpage)) 1467119a5fc1SHugh Dickins goto drop_hpage; 1468119a5fc1SHugh Dickins 146927e1f827SSong Liu pmd = mm_find_pmd(mm, haddr); 147027e1f827SSong Liu if (!pmd) 1471119a5fc1SHugh Dickins goto drop_hpage; 147227e1f827SSong Liu 147327e1f827SSong Liu start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 147427e1f827SSong Liu 147527e1f827SSong Liu /* step 1: check all mapped PTEs are to the right huge page */ 147627e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 147727e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 147827e1f827SSong Liu struct page *page; 147927e1f827SSong Liu 148027e1f827SSong Liu /* empty pte, skip */ 148127e1f827SSong Liu if (pte_none(*pte)) 148227e1f827SSong Liu continue; 148327e1f827SSong Liu 148427e1f827SSong Liu /* page swapped out, abort */ 148527e1f827SSong Liu if (!pte_present(*pte)) 148627e1f827SSong Liu goto abort; 148727e1f827SSong Liu 148827e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 148927e1f827SSong Liu 149027e1f827SSong Liu /* 1491119a5fc1SHugh Dickins * Note that uprobe, debugger, or MAP_PRIVATE may change the 1492119a5fc1SHugh Dickins * page table, but the new page will not be a subpage of hpage. 149327e1f827SSong Liu */ 1494119a5fc1SHugh Dickins if (hpage + i != page) 149527e1f827SSong Liu goto abort; 149627e1f827SSong Liu count++; 149727e1f827SSong Liu } 149827e1f827SSong Liu 149927e1f827SSong Liu /* step 2: adjust rmap */ 150027e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 150127e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 150227e1f827SSong Liu struct page *page; 150327e1f827SSong Liu 150427e1f827SSong Liu if (pte_none(*pte)) 150527e1f827SSong Liu continue; 150627e1f827SSong Liu page = vm_normal_page(vma, addr, *pte); 1507cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 150827e1f827SSong Liu } 150927e1f827SSong Liu 151027e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 151127e1f827SSong Liu 151227e1f827SSong Liu /* step 3: set proper refcount and mm_counters. */ 1513119a5fc1SHugh Dickins if (count) { 151427e1f827SSong Liu page_ref_sub(hpage, count); 151527e1f827SSong Liu add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); 151627e1f827SSong Liu } 151727e1f827SSong Liu 151827e1f827SSong Liu /* step 4: collapse pmd */ 1519e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, haddr, pmd); 1520119a5fc1SHugh Dickins drop_hpage: 1521119a5fc1SHugh Dickins unlock_page(hpage); 1522119a5fc1SHugh Dickins put_page(hpage); 152327e1f827SSong Liu return; 152427e1f827SSong Liu 152527e1f827SSong Liu abort: 152627e1f827SSong Liu pte_unmap_unlock(start_pte, ptl); 1527119a5fc1SHugh Dickins goto drop_hpage; 152827e1f827SSong Liu } 152927e1f827SSong Liu 15300edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 153127e1f827SSong Liu { 153227e1f827SSong Liu struct mm_struct *mm = mm_slot->mm; 153327e1f827SSong Liu int i; 153427e1f827SSong Liu 153527e1f827SSong Liu if (likely(mm_slot->nr_pte_mapped_thp == 0)) 15360edf61e5SMiaohe Lin return; 153727e1f827SSong Liu 1538d8ed45c5SMichel Lespinasse if (!mmap_write_trylock(mm)) 15390edf61e5SMiaohe Lin return; 154027e1f827SSong Liu 154127e1f827SSong Liu if (unlikely(khugepaged_test_exit(mm))) 154227e1f827SSong Liu goto out; 154327e1f827SSong Liu 154427e1f827SSong Liu for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) 154527e1f827SSong Liu collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]); 154627e1f827SSong Liu 154727e1f827SSong Liu out: 154827e1f827SSong Liu mm_slot->nr_pte_mapped_thp = 0; 1549d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 155027e1f827SSong Liu } 155127e1f827SSong Liu 1552f3f0e1d2SKirill A. Shutemov static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1553f3f0e1d2SKirill A. Shutemov { 1554f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 155518e77600SHugh Dickins struct mm_struct *mm; 1556f3f0e1d2SKirill A. Shutemov unsigned long addr; 1557e59a47b8SPasha Tatashin pmd_t *pmd; 1558f3f0e1d2SKirill A. Shutemov 1559f3f0e1d2SKirill A. Shutemov i_mmap_lock_write(mapping); 1560f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 156127e1f827SSong Liu /* 156227e1f827SSong Liu * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 156327e1f827SSong Liu * got written to. These VMAs are likely not worth investing 15643e4e28c5SMichel Lespinasse * mmap_write_lock(mm) as PMD-mapping is likely to be split 156527e1f827SSong Liu * later. 156627e1f827SSong Liu * 156727e1f827SSong Liu * Not that vma->anon_vma check is racy: it can be set up after 1568c1e8d7c6SMichel Lespinasse * the check but before we took mmap_lock by the fault path. 156927e1f827SSong Liu * But page lock would prevent establishing any new ptes of the 157027e1f827SSong Liu * page, so we are safe. 157127e1f827SSong Liu * 157227e1f827SSong Liu * An alternative would be drop the check, but check that page 157327e1f827SSong Liu * table is clear before calling pmdp_collapse_flush() under 157427e1f827SSong Liu * ptl. It has higher chance to recover THP for the VMA, but 157527e1f827SSong Liu * has higher cost too. 157627e1f827SSong Liu */ 1577f3f0e1d2SKirill A. Shutemov if (vma->anon_vma) 1578f3f0e1d2SKirill A. Shutemov continue; 1579f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 1580f3f0e1d2SKirill A. Shutemov if (addr & ~HPAGE_PMD_MASK) 1581f3f0e1d2SKirill A. Shutemov continue; 1582f3f0e1d2SKirill A. Shutemov if (vma->vm_end < addr + HPAGE_PMD_SIZE) 1583f3f0e1d2SKirill A. Shutemov continue; 158418e77600SHugh Dickins mm = vma->vm_mm; 158518e77600SHugh Dickins pmd = mm_find_pmd(mm, addr); 1586f3f0e1d2SKirill A. Shutemov if (!pmd) 1587f3f0e1d2SKirill A. Shutemov continue; 1588f3f0e1d2SKirill A. Shutemov /* 1589c1e8d7c6SMichel Lespinasse * We need exclusive mmap_lock to retract page table. 159027e1f827SSong Liu * 159127e1f827SSong Liu * We use trylock due to lock inversion: we need to acquire 1592c1e8d7c6SMichel Lespinasse * mmap_lock while holding page lock. Fault path does it in 159327e1f827SSong Liu * reverse order. Trylock is a way to avoid deadlock. 1594f3f0e1d2SKirill A. Shutemov */ 159518e77600SHugh Dickins if (mmap_write_trylock(mm)) { 1596deb4c93aSPeter Xu /* 1597deb4c93aSPeter Xu * When a vma is registered with uffd-wp, we can't 1598deb4c93aSPeter Xu * recycle the pmd pgtable because there can be pte 1599deb4c93aSPeter Xu * markers installed. Skip it only, so the rest mm/vma 1600deb4c93aSPeter Xu * can still have the same file mapped hugely, however 1601deb4c93aSPeter Xu * it'll always mapped in small page size for uffd-wp 1602deb4c93aSPeter Xu * registered ranges. 1603deb4c93aSPeter Xu */ 1604deb4c93aSPeter Xu if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma)) 1605e59a47b8SPasha Tatashin collapse_and_free_pmd(mm, vma, addr, pmd); 160618e77600SHugh Dickins mmap_write_unlock(mm); 160727e1f827SSong Liu } else { 160827e1f827SSong Liu /* Try again later */ 160918e77600SHugh Dickins khugepaged_add_pte_mapped_thp(mm, addr); 1610f3f0e1d2SKirill A. Shutemov } 1611f3f0e1d2SKirill A. Shutemov } 1612f3f0e1d2SKirill A. Shutemov i_mmap_unlock_write(mapping); 1613f3f0e1d2SKirill A. Shutemov } 1614f3f0e1d2SKirill A. Shutemov 1615f3f0e1d2SKirill A. Shutemov /** 161699cb0dbdSSong Liu * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1617f3f0e1d2SKirill A. Shutemov * 1618336e6b53SAlex Shi * @mm: process address space where collapse happens 1619336e6b53SAlex Shi * @file: file that collapse on 1620336e6b53SAlex Shi * @start: collapse start address 1621336e6b53SAlex Shi * @hpage: new allocated huge page for collapse 1622336e6b53SAlex Shi * @node: appointed node the new huge page allocate from 1623336e6b53SAlex Shi * 1624f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 162587c460a0SHugh Dickins * - allocate and lock a new huge page; 162677da9389SMatthew Wilcox * - scan page cache replacing old pages with the new one 162799cb0dbdSSong Liu * + swap/gup in pages if necessary; 1628f3f0e1d2SKirill A. Shutemov * + fill in gaps; 162977da9389SMatthew Wilcox * + keep old pages around in case rollback is required; 163077da9389SMatthew Wilcox * - if replacing succeeds: 1631f3f0e1d2SKirill A. Shutemov * + copy data over; 1632f3f0e1d2SKirill A. Shutemov * + free old pages; 163387c460a0SHugh Dickins * + unlock huge page; 1634f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1635f3f0e1d2SKirill A. Shutemov * + put all pages back and unfreeze them; 163677da9389SMatthew Wilcox * + restore gaps in the page cache; 163787c460a0SHugh Dickins * + unlock and free huge page; 1638f3f0e1d2SKirill A. Shutemov */ 1639579c571eSSong Liu static void collapse_file(struct mm_struct *mm, 1640579c571eSSong Liu struct file *file, pgoff_t start, 1641f3f0e1d2SKirill A. Shutemov struct page **hpage, int node) 1642f3f0e1d2SKirill A. Shutemov { 1643579c571eSSong Liu struct address_space *mapping = file->f_mapping; 1644f3f0e1d2SKirill A. Shutemov gfp_t gfp; 164577da9389SMatthew Wilcox struct page *new_page; 1646f3f0e1d2SKirill A. Shutemov pgoff_t index, end = start + HPAGE_PMD_NR; 1647f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 164877da9389SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1649f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 165099cb0dbdSSong Liu bool is_shmem = shmem_file(file); 1651bf9eceadSMuchun Song int nr; 1652f3f0e1d2SKirill A. Shutemov 165399cb0dbdSSong Liu VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1654f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1655f3f0e1d2SKirill A. Shutemov 1656f3f0e1d2SKirill A. Shutemov /* Only allocate from the target node */ 165741b6167eSMichal Hocko gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 1658f3f0e1d2SKirill A. Shutemov 1659f3f0e1d2SKirill A. Shutemov new_page = khugepaged_alloc_page(hpage, gfp, node); 1660f3f0e1d2SKirill A. Shutemov if (!new_page) { 1661f3f0e1d2SKirill A. Shutemov result = SCAN_ALLOC_HUGE_PAGE_FAIL; 1662f3f0e1d2SKirill A. Shutemov goto out; 1663f3f0e1d2SKirill A. Shutemov } 1664f3f0e1d2SKirill A. Shutemov 16658f425e4eSMatthew Wilcox (Oracle) if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) { 1666f3f0e1d2SKirill A. Shutemov result = SCAN_CGROUP_CHARGE_FAIL; 1667f3f0e1d2SKirill A. Shutemov goto out; 1668f3f0e1d2SKirill A. Shutemov } 16699d82c694SJohannes Weiner count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC); 1670f3f0e1d2SKirill A. Shutemov 16716b24ca4aSMatthew Wilcox (Oracle) /* 16726b24ca4aSMatthew Wilcox (Oracle) * Ensure we have slots for all the pages in the range. This is 16736b24ca4aSMatthew Wilcox (Oracle) * almost certainly a no-op because most of the pages must be present 16746b24ca4aSMatthew Wilcox (Oracle) */ 167595feeabbSHugh Dickins do { 167695feeabbSHugh Dickins xas_lock_irq(&xas); 167795feeabbSHugh Dickins xas_create_range(&xas); 167895feeabbSHugh Dickins if (!xas_error(&xas)) 167995feeabbSHugh Dickins break; 168095feeabbSHugh Dickins xas_unlock_irq(&xas); 168195feeabbSHugh Dickins if (!xas_nomem(&xas, GFP_KERNEL)) { 168295feeabbSHugh Dickins result = SCAN_FAIL; 168395feeabbSHugh Dickins goto out; 168495feeabbSHugh Dickins } 168595feeabbSHugh Dickins } while (1); 168695feeabbSHugh Dickins 1687042a3082SHugh Dickins __SetPageLocked(new_page); 168899cb0dbdSSong Liu if (is_shmem) 1689042a3082SHugh Dickins __SetPageSwapBacked(new_page); 1690f3f0e1d2SKirill A. Shutemov new_page->index = start; 1691f3f0e1d2SKirill A. Shutemov new_page->mapping = mapping; 1692f3f0e1d2SKirill A. Shutemov 1693f3f0e1d2SKirill A. Shutemov /* 169487c460a0SHugh Dickins * At this point the new_page is locked and not up-to-date. 169587c460a0SHugh Dickins * It's safe to insert it into the page cache, because nobody would 169687c460a0SHugh Dickins * be able to map it or use it in another way until we unlock it. 1697f3f0e1d2SKirill A. Shutemov */ 1698f3f0e1d2SKirill A. Shutemov 169977da9389SMatthew Wilcox xas_set(&xas, start); 170077da9389SMatthew Wilcox for (index = start; index < end; index++) { 170177da9389SMatthew Wilcox struct page *page = xas_next(&xas); 170277da9389SMatthew Wilcox 170377da9389SMatthew Wilcox VM_BUG_ON(index != xas.xa_index); 170499cb0dbdSSong Liu if (is_shmem) { 170577da9389SMatthew Wilcox if (!page) { 1706701270faSHugh Dickins /* 170799cb0dbdSSong Liu * Stop if extent has been truncated or 170899cb0dbdSSong Liu * hole-punched, and is now completely 170999cb0dbdSSong Liu * empty. 1710701270faSHugh Dickins */ 1711701270faSHugh Dickins if (index == start) { 1712701270faSHugh Dickins if (!xas_next_entry(&xas, end - 1)) { 1713701270faSHugh Dickins result = SCAN_TRUNCATED; 1714042a3082SHugh Dickins goto xa_locked; 1715701270faSHugh Dickins } 1716701270faSHugh Dickins xas_set(&xas, index); 1717701270faSHugh Dickins } 171877da9389SMatthew Wilcox if (!shmem_charge(mapping->host, 1)) { 1719f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1720042a3082SHugh Dickins goto xa_locked; 1721f3f0e1d2SKirill A. Shutemov } 17224101196bSMatthew Wilcox (Oracle) xas_store(&xas, new_page); 172377da9389SMatthew Wilcox nr_none++; 172477da9389SMatthew Wilcox continue; 1725f3f0e1d2SKirill A. Shutemov } 1726f3f0e1d2SKirill A. Shutemov 17273159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 172877da9389SMatthew Wilcox xas_unlock_irq(&xas); 1729f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 1730f3f0e1d2SKirill A. Shutemov if (shmem_getpage(mapping->host, index, &page, 1731acdd9f8eSHugh Dickins SGP_NOALLOC)) { 1732f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 173377da9389SMatthew Wilcox goto xa_unlocked; 1734f3f0e1d2SKirill A. Shutemov } 1735f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1736f3f0e1d2SKirill A. Shutemov get_page(page); 1737042a3082SHugh Dickins xas_unlock_irq(&xas); 1738f3f0e1d2SKirill A. Shutemov } else { 1739f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1740042a3082SHugh Dickins goto xa_locked; 1741f3f0e1d2SKirill A. Shutemov } 174299cb0dbdSSong Liu } else { /* !is_shmem */ 174399cb0dbdSSong Liu if (!page || xa_is_value(page)) { 174499cb0dbdSSong Liu xas_unlock_irq(&xas); 174599cb0dbdSSong Liu page_cache_sync_readahead(mapping, &file->f_ra, 174699cb0dbdSSong Liu file, index, 1747e5a59d30SDavid Howells end - index); 174899cb0dbdSSong Liu /* drain pagevecs to help isolate_lru_page() */ 174999cb0dbdSSong Liu lru_add_drain(); 175099cb0dbdSSong Liu page = find_lock_page(mapping, index); 175199cb0dbdSSong Liu if (unlikely(page == NULL)) { 175299cb0dbdSSong Liu result = SCAN_FAIL; 175399cb0dbdSSong Liu goto xa_unlocked; 175499cb0dbdSSong Liu } 175575f36069SSong Liu } else if (PageDirty(page)) { 175675f36069SSong Liu /* 175775f36069SSong Liu * khugepaged only works on read-only fd, 175875f36069SSong Liu * so this page is dirty because it hasn't 175975f36069SSong Liu * been flushed since first write. There 176075f36069SSong Liu * won't be new dirty pages. 176175f36069SSong Liu * 176275f36069SSong Liu * Trigger async flush here and hope the 176375f36069SSong Liu * writeback is done when khugepaged 176475f36069SSong Liu * revisits this page. 176575f36069SSong Liu * 176675f36069SSong Liu * This is a one-off situation. We are not 176775f36069SSong Liu * forcing writeback in loop. 176875f36069SSong Liu */ 176975f36069SSong Liu xas_unlock_irq(&xas); 177075f36069SSong Liu filemap_flush(mapping); 177175f36069SSong Liu result = SCAN_FAIL; 177275f36069SSong Liu goto xa_unlocked; 177374c42e1bSRongwei Wang } else if (PageWriteback(page)) { 177474c42e1bSRongwei Wang xas_unlock_irq(&xas); 177574c42e1bSRongwei Wang result = SCAN_FAIL; 177674c42e1bSRongwei Wang goto xa_unlocked; 177799cb0dbdSSong Liu } else if (trylock_page(page)) { 177899cb0dbdSSong Liu get_page(page); 177999cb0dbdSSong Liu xas_unlock_irq(&xas); 178099cb0dbdSSong Liu } else { 178199cb0dbdSSong Liu result = SCAN_PAGE_LOCK; 178299cb0dbdSSong Liu goto xa_locked; 178399cb0dbdSSong Liu } 178499cb0dbdSSong Liu } 1785f3f0e1d2SKirill A. Shutemov 1786f3f0e1d2SKirill A. Shutemov /* 1787b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1788f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1789f3f0e1d2SKirill A. Shutemov */ 1790f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 17914655e5e5SSong Liu 17924655e5e5SSong Liu /* make sure the page is up to date */ 17934655e5e5SSong Liu if (unlikely(!PageUptodate(page))) { 17944655e5e5SSong Liu result = SCAN_FAIL; 17954655e5e5SSong Liu goto out_unlock; 17964655e5e5SSong Liu } 179706a5e126SHugh Dickins 179806a5e126SHugh Dickins /* 179906a5e126SHugh Dickins * If file was truncated then extended, or hole-punched, before 180006a5e126SHugh Dickins * we locked the first page, then a THP might be there already. 180106a5e126SHugh Dickins */ 180206a5e126SHugh Dickins if (PageTransCompound(page)) { 180306a5e126SHugh Dickins result = SCAN_PAGE_COMPOUND; 180406a5e126SHugh Dickins goto out_unlock; 180506a5e126SHugh Dickins } 1806f3f0e1d2SKirill A. Shutemov 1807f3f0e1d2SKirill A. Shutemov if (page_mapping(page) != mapping) { 1808f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1809f3f0e1d2SKirill A. Shutemov goto out_unlock; 1810f3f0e1d2SKirill A. Shutemov } 1811f3f0e1d2SKirill A. Shutemov 181274c42e1bSRongwei Wang if (!is_shmem && (PageDirty(page) || 181374c42e1bSRongwei Wang PageWriteback(page))) { 18144655e5e5SSong Liu /* 18154655e5e5SSong Liu * khugepaged only works on read-only fd, so this 18164655e5e5SSong Liu * page is dirty because it hasn't been flushed 18174655e5e5SSong Liu * since first write. 18184655e5e5SSong Liu */ 18194655e5e5SSong Liu result = SCAN_FAIL; 18204655e5e5SSong Liu goto out_unlock; 18214655e5e5SSong Liu } 18224655e5e5SSong Liu 1823f3f0e1d2SKirill A. Shutemov if (isolate_lru_page(page)) { 1824f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1825042a3082SHugh Dickins goto out_unlock; 1826f3f0e1d2SKirill A. Shutemov } 1827f3f0e1d2SKirill A. Shutemov 182899cb0dbdSSong Liu if (page_has_private(page) && 182999cb0dbdSSong Liu !try_to_release_page(page, GFP_KERNEL)) { 183099cb0dbdSSong Liu result = SCAN_PAGE_HAS_PRIVATE; 18312f33a706SHugh Dickins putback_lru_page(page); 183299cb0dbdSSong Liu goto out_unlock; 183399cb0dbdSSong Liu } 183499cb0dbdSSong Liu 1835f3f0e1d2SKirill A. Shutemov if (page_mapped(page)) 1836869f7ee6SMatthew Wilcox (Oracle) try_to_unmap(page_folio(page), 1837869f7ee6SMatthew Wilcox (Oracle) TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); 1838f3f0e1d2SKirill A. Shutemov 183977da9389SMatthew Wilcox xas_lock_irq(&xas); 184077da9389SMatthew Wilcox xas_set(&xas, index); 1841f3f0e1d2SKirill A. Shutemov 184277da9389SMatthew Wilcox VM_BUG_ON_PAGE(page != xas_load(&xas), page); 1843f3f0e1d2SKirill A. Shutemov 1844f3f0e1d2SKirill A. Shutemov /* 1845f3f0e1d2SKirill A. Shutemov * The page is expected to have page_count() == 3: 1846f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 184777da9389SMatthew Wilcox * - one reference from page cache; 1848f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1849f3f0e1d2SKirill A. Shutemov */ 1850f3f0e1d2SKirill A. Shutemov if (!page_ref_freeze(page, 3)) { 1851f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1852042a3082SHugh Dickins xas_unlock_irq(&xas); 1853042a3082SHugh Dickins putback_lru_page(page); 1854042a3082SHugh Dickins goto out_unlock; 1855f3f0e1d2SKirill A. Shutemov } 1856f3f0e1d2SKirill A. Shutemov 1857f3f0e1d2SKirill A. Shutemov /* 1858f3f0e1d2SKirill A. Shutemov * Add the page to the list to be able to undo the collapse if 1859f3f0e1d2SKirill A. Shutemov * something go wrong. 1860f3f0e1d2SKirill A. Shutemov */ 1861f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 1862f3f0e1d2SKirill A. Shutemov 1863f3f0e1d2SKirill A. Shutemov /* Finally, replace with the new page. */ 18644101196bSMatthew Wilcox (Oracle) xas_store(&xas, new_page); 1865f3f0e1d2SKirill A. Shutemov continue; 1866f3f0e1d2SKirill A. Shutemov out_unlock: 1867f3f0e1d2SKirill A. Shutemov unlock_page(page); 1868f3f0e1d2SKirill A. Shutemov put_page(page); 1869042a3082SHugh Dickins goto xa_unlocked; 1870f3f0e1d2SKirill A. Shutemov } 1871bf9eceadSMuchun Song nr = thp_nr_pages(new_page); 1872f3f0e1d2SKirill A. Shutemov 187399cb0dbdSSong Liu if (is_shmem) 187457b2847dSMuchun Song __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr); 187509d91cdaSSong Liu else { 1876bf9eceadSMuchun Song __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr); 187709d91cdaSSong Liu filemap_nr_thps_inc(mapping); 1878eb6ecbedSCollin Fijalkovich /* 1879eb6ecbedSCollin Fijalkovich * Paired with smp_mb() in do_dentry_open() to ensure 1880eb6ecbedSCollin Fijalkovich * i_writecount is up to date and the update to nr_thps is 1881eb6ecbedSCollin Fijalkovich * visible. Ensures the page cache will be truncated if the 1882eb6ecbedSCollin Fijalkovich * file is opened writable. 1883eb6ecbedSCollin Fijalkovich */ 1884eb6ecbedSCollin Fijalkovich smp_mb(); 1885eb6ecbedSCollin Fijalkovich if (inode_is_open_for_write(mapping->host)) { 1886eb6ecbedSCollin Fijalkovich result = SCAN_FAIL; 1887eb6ecbedSCollin Fijalkovich __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr); 1888eb6ecbedSCollin Fijalkovich filemap_nr_thps_dec(mapping); 1889eb6ecbedSCollin Fijalkovich goto xa_locked; 1890eb6ecbedSCollin Fijalkovich } 189109d91cdaSSong Liu } 189299cb0dbdSSong Liu 1893042a3082SHugh Dickins if (nr_none) { 18949d82c694SJohannes Weiner __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none); 189599cb0dbdSSong Liu if (is_shmem) 18969d82c694SJohannes Weiner __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none); 1897042a3082SHugh Dickins } 1898042a3082SHugh Dickins 18996b24ca4aSMatthew Wilcox (Oracle) /* Join all the small entries into a single multi-index entry */ 19006b24ca4aSMatthew Wilcox (Oracle) xas_set_order(&xas, start, HPAGE_PMD_ORDER); 19016b24ca4aSMatthew Wilcox (Oracle) xas_store(&xas, new_page); 1902042a3082SHugh Dickins xa_locked: 1903042a3082SHugh Dickins xas_unlock_irq(&xas); 190477da9389SMatthew Wilcox xa_unlocked: 1905042a3082SHugh Dickins 19066d9df8a5SHugh Dickins /* 19076d9df8a5SHugh Dickins * If collapse is successful, flush must be done now before copying. 19086d9df8a5SHugh Dickins * If collapse is unsuccessful, does flush actually need to be done? 19096d9df8a5SHugh Dickins * Do it anyway, to clear the state. 19106d9df8a5SHugh Dickins */ 19116d9df8a5SHugh Dickins try_to_unmap_flush(); 19126d9df8a5SHugh Dickins 1913f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 191477da9389SMatthew Wilcox struct page *page, *tmp; 1915f3f0e1d2SKirill A. Shutemov 1916f3f0e1d2SKirill A. Shutemov /* 191777da9389SMatthew Wilcox * Replacing old pages with new one has succeeded, now we 191877da9389SMatthew Wilcox * need to copy the content and free the old pages. 1919f3f0e1d2SKirill A. Shutemov */ 19202af8ff29SHugh Dickins index = start; 1921f3f0e1d2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, &pagelist, lru) { 19222af8ff29SHugh Dickins while (index < page->index) { 19232af8ff29SHugh Dickins clear_highpage(new_page + (index % HPAGE_PMD_NR)); 19242af8ff29SHugh Dickins index++; 19252af8ff29SHugh Dickins } 1926f3f0e1d2SKirill A. Shutemov copy_highpage(new_page + (page->index % HPAGE_PMD_NR), 1927f3f0e1d2SKirill A. Shutemov page); 1928f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1929f3f0e1d2SKirill A. Shutemov page->mapping = NULL; 1930042a3082SHugh Dickins page_ref_unfreeze(page, 1); 1931f3f0e1d2SKirill A. Shutemov ClearPageActive(page); 1932f3f0e1d2SKirill A. Shutemov ClearPageUnevictable(page); 1933042a3082SHugh Dickins unlock_page(page); 1934f3f0e1d2SKirill A. Shutemov put_page(page); 19352af8ff29SHugh Dickins index++; 19362af8ff29SHugh Dickins } 19372af8ff29SHugh Dickins while (index < end) { 19382af8ff29SHugh Dickins clear_highpage(new_page + (index % HPAGE_PMD_NR)); 19392af8ff29SHugh Dickins index++; 1940f3f0e1d2SKirill A. Shutemov } 1941f3f0e1d2SKirill A. Shutemov 1942f3f0e1d2SKirill A. Shutemov SetPageUptodate(new_page); 194387c460a0SHugh Dickins page_ref_add(new_page, HPAGE_PMD_NR - 1); 19446058eaecSJohannes Weiner if (is_shmem) 194599cb0dbdSSong Liu set_page_dirty(new_page); 19466058eaecSJohannes Weiner lru_cache_add(new_page); 1947f3f0e1d2SKirill A. Shutemov 1948042a3082SHugh Dickins /* 1949042a3082SHugh Dickins * Remove pte page tables, so we can re-fault the page as huge. 1950042a3082SHugh Dickins */ 1951042a3082SHugh Dickins retract_page_tables(mapping, start); 1952f3f0e1d2SKirill A. Shutemov *hpage = NULL; 195387aa7529SYang Shi 195487aa7529SYang Shi khugepaged_pages_collapsed++; 1955f3f0e1d2SKirill A. Shutemov } else { 195677da9389SMatthew Wilcox struct page *page; 1957aaa52e34SHugh Dickins 195877da9389SMatthew Wilcox /* Something went wrong: roll back page cache changes */ 195977da9389SMatthew Wilcox xas_lock_irq(&xas); 1960aaa52e34SHugh Dickins mapping->nrpages -= nr_none; 196199cb0dbdSSong Liu 196299cb0dbdSSong Liu if (is_shmem) 1963aaa52e34SHugh Dickins shmem_uncharge(mapping->host, nr_none); 1964aaa52e34SHugh Dickins 196577da9389SMatthew Wilcox xas_set(&xas, start); 196677da9389SMatthew Wilcox xas_for_each(&xas, page, end - 1) { 1967f3f0e1d2SKirill A. Shutemov page = list_first_entry_or_null(&pagelist, 1968f3f0e1d2SKirill A. Shutemov struct page, lru); 196977da9389SMatthew Wilcox if (!page || xas.xa_index < page->index) { 1970f3f0e1d2SKirill A. Shutemov if (!nr_none) 1971f3f0e1d2SKirill A. Shutemov break; 1972f3f0e1d2SKirill A. Shutemov nr_none--; 197359749e6cSJohannes Weiner /* Put holes back where they were */ 197477da9389SMatthew Wilcox xas_store(&xas, NULL); 1975f3f0e1d2SKirill A. Shutemov continue; 1976f3f0e1d2SKirill A. Shutemov } 1977f3f0e1d2SKirill A. Shutemov 197877da9389SMatthew Wilcox VM_BUG_ON_PAGE(page->index != xas.xa_index, page); 1979f3f0e1d2SKirill A. Shutemov 1980f3f0e1d2SKirill A. Shutemov /* Unfreeze the page. */ 1981f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1982f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(page, 2); 198377da9389SMatthew Wilcox xas_store(&xas, page); 198477da9389SMatthew Wilcox xas_pause(&xas); 198577da9389SMatthew Wilcox xas_unlock_irq(&xas); 1986f3f0e1d2SKirill A. Shutemov unlock_page(page); 1987042a3082SHugh Dickins putback_lru_page(page); 198877da9389SMatthew Wilcox xas_lock_irq(&xas); 1989f3f0e1d2SKirill A. Shutemov } 1990f3f0e1d2SKirill A. Shutemov VM_BUG_ON(nr_none); 199177da9389SMatthew Wilcox xas_unlock_irq(&xas); 1992f3f0e1d2SKirill A. Shutemov 1993f3f0e1d2SKirill A. Shutemov new_page->mapping = NULL; 1994f3f0e1d2SKirill A. Shutemov } 1995042a3082SHugh Dickins 1996042a3082SHugh Dickins unlock_page(new_page); 1997f3f0e1d2SKirill A. Shutemov out: 1998f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 19999d82c694SJohannes Weiner if (!IS_ERR_OR_NULL(*hpage)) 2000bbc6b703SMatthew Wilcox (Oracle) mem_cgroup_uncharge(page_folio(*hpage)); 2001f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 2002f3f0e1d2SKirill A. Shutemov } 2003f3f0e1d2SKirill A. Shutemov 2004579c571eSSong Liu static void khugepaged_scan_file(struct mm_struct *mm, 2005579c571eSSong Liu struct file *file, pgoff_t start, struct page **hpage) 2006f3f0e1d2SKirill A. Shutemov { 2007f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 2008579c571eSSong Liu struct address_space *mapping = file->f_mapping; 200985b392dbSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, start); 2010f3f0e1d2SKirill A. Shutemov int present, swap; 2011f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 2012f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 2013f3f0e1d2SKirill A. Shutemov 2014f3f0e1d2SKirill A. Shutemov present = 0; 2015f3f0e1d2SKirill A. Shutemov swap = 0; 2016f3f0e1d2SKirill A. Shutemov memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 2017f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 201885b392dbSMatthew Wilcox xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 201985b392dbSMatthew Wilcox if (xas_retry(&xas, page)) 2020f3f0e1d2SKirill A. Shutemov continue; 2021f3f0e1d2SKirill A. Shutemov 202285b392dbSMatthew Wilcox if (xa_is_value(page)) { 2023f3f0e1d2SKirill A. Shutemov if (++swap > khugepaged_max_ptes_swap) { 2024f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 2025e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 2026f3f0e1d2SKirill A. Shutemov break; 2027f3f0e1d2SKirill A. Shutemov } 2028f3f0e1d2SKirill A. Shutemov continue; 2029f3f0e1d2SKirill A. Shutemov } 2030f3f0e1d2SKirill A. Shutemov 20316b24ca4aSMatthew Wilcox (Oracle) /* 20326b24ca4aSMatthew Wilcox (Oracle) * XXX: khugepaged should compact smaller compound pages 20336b24ca4aSMatthew Wilcox (Oracle) * into a PMD sized page 20346b24ca4aSMatthew Wilcox (Oracle) */ 2035f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 2036f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COMPOUND; 2037f3f0e1d2SKirill A. Shutemov break; 2038f3f0e1d2SKirill A. Shutemov } 2039f3f0e1d2SKirill A. Shutemov 2040f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 2041f3f0e1d2SKirill A. Shutemov if (khugepaged_scan_abort(node)) { 2042f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 2043f3f0e1d2SKirill A. Shutemov break; 2044f3f0e1d2SKirill A. Shutemov } 2045f3f0e1d2SKirill A. Shutemov khugepaged_node_load[node]++; 2046f3f0e1d2SKirill A. Shutemov 2047f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 2048f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 2049f3f0e1d2SKirill A. Shutemov break; 2050f3f0e1d2SKirill A. Shutemov } 2051f3f0e1d2SKirill A. Shutemov 205299cb0dbdSSong Liu if (page_count(page) != 205399cb0dbdSSong Liu 1 + page_mapcount(page) + page_has_private(page)) { 2054f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 2055f3f0e1d2SKirill A. Shutemov break; 2056f3f0e1d2SKirill A. Shutemov } 2057f3f0e1d2SKirill A. Shutemov 2058f3f0e1d2SKirill A. Shutemov /* 2059f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 2060f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 2061f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 2062f3f0e1d2SKirill A. Shutemov */ 2063f3f0e1d2SKirill A. Shutemov 2064f3f0e1d2SKirill A. Shutemov present++; 2065f3f0e1d2SKirill A. Shutemov 2066f3f0e1d2SKirill A. Shutemov if (need_resched()) { 206785b392dbSMatthew Wilcox xas_pause(&xas); 2068f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 2069f3f0e1d2SKirill A. Shutemov } 2070f3f0e1d2SKirill A. Shutemov } 2071f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 2072f3f0e1d2SKirill A. Shutemov 2073f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 2074f3f0e1d2SKirill A. Shutemov if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 2075f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 2076e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 2077f3f0e1d2SKirill A. Shutemov } else { 2078f3f0e1d2SKirill A. Shutemov node = khugepaged_find_target_node(); 2079579c571eSSong Liu collapse_file(mm, file, start, hpage, node); 2080f3f0e1d2SKirill A. Shutemov } 2081f3f0e1d2SKirill A. Shutemov } 2082f3f0e1d2SKirill A. Shutemov 2083f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 2084f3f0e1d2SKirill A. Shutemov } 2085f3f0e1d2SKirill A. Shutemov #else 2086579c571eSSong Liu static void khugepaged_scan_file(struct mm_struct *mm, 2087579c571eSSong Liu struct file *file, pgoff_t start, struct page **hpage) 2088f3f0e1d2SKirill A. Shutemov { 2089f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 2090f3f0e1d2SKirill A. Shutemov } 209127e1f827SSong Liu 20920edf61e5SMiaohe Lin static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 209327e1f827SSong Liu { 209427e1f827SSong Liu } 2095f3f0e1d2SKirill A. Shutemov #endif 2096f3f0e1d2SKirill A. Shutemov 2097b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 2098b46e756fSKirill A. Shutemov struct page **hpage) 2099b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 2100b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 2101b46e756fSKirill A. Shutemov { 2102b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 2103b46e756fSKirill A. Shutemov struct mm_struct *mm; 2104b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 2105b46e756fSKirill A. Shutemov int progress = 0; 2106b46e756fSKirill A. Shutemov 2107b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 210835f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 2109b46e756fSKirill A. Shutemov 2110b46e756fSKirill A. Shutemov if (khugepaged_scan.mm_slot) 2111b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2112b46e756fSKirill A. Shutemov else { 2113b46e756fSKirill A. Shutemov mm_slot = list_entry(khugepaged_scan.mm_head.next, 2114b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2115b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2116b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 2117b46e756fSKirill A. Shutemov } 2118b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 211927e1f827SSong Liu khugepaged_collapse_pte_mapped_thps(mm_slot); 2120b46e756fSKirill A. Shutemov 2121b46e756fSKirill A. Shutemov mm = mm_slot->mm; 21223b454ad3SYang Shi /* 21233b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 21243b454ad3SYang Shi * the next mm on the list. 21253b454ad3SYang Shi */ 2126b46e756fSKirill A. Shutemov vma = NULL; 2127d8ed45c5SMichel Lespinasse if (unlikely(!mmap_read_trylock(mm))) 2128c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 21293b454ad3SYang Shi if (likely(!khugepaged_test_exit(mm))) 2130b46e756fSKirill A. Shutemov vma = find_vma(mm, khugepaged_scan.address); 2131b46e756fSKirill A. Shutemov 2132b46e756fSKirill A. Shutemov progress++; 2133b46e756fSKirill A. Shutemov for (; vma; vma = vma->vm_next) { 2134b46e756fSKirill A. Shutemov unsigned long hstart, hend; 2135b46e756fSKirill A. Shutemov 2136b46e756fSKirill A. Shutemov cond_resched(); 2137b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) { 2138b46e756fSKirill A. Shutemov progress++; 2139b46e756fSKirill A. Shutemov break; 2140b46e756fSKirill A. Shutemov } 214150f8b92fSSong Liu if (!hugepage_vma_check(vma, vma->vm_flags)) { 2142b46e756fSKirill A. Shutemov skip: 2143b46e756fSKirill A. Shutemov progress++; 2144b46e756fSKirill A. Shutemov continue; 2145b46e756fSKirill A. Shutemov } 2146b46e756fSKirill A. Shutemov hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2147b46e756fSKirill A. Shutemov hend = vma->vm_end & HPAGE_PMD_MASK; 2148b46e756fSKirill A. Shutemov if (hstart >= hend) 2149b46e756fSKirill A. Shutemov goto skip; 2150b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 2151b46e756fSKirill A. Shutemov goto skip; 2152b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 2153b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 2154b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2155396bcc52SMatthew Wilcox (Oracle) if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma)) 2156396bcc52SMatthew Wilcox (Oracle) goto skip; 2157b46e756fSKirill A. Shutemov 2158b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 2159b46e756fSKirill A. Shutemov int ret; 2160b46e756fSKirill A. Shutemov cond_resched(); 2161b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 2162b46e756fSKirill A. Shutemov goto breakouterloop; 2163b46e756fSKirill A. Shutemov 2164b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 2165b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 2166b46e756fSKirill A. Shutemov hend); 216799cb0dbdSSong Liu if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 2168396bcc52SMatthew Wilcox (Oracle) struct file *file = get_file(vma->vm_file); 2169f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 2170f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 217199cb0dbdSSong Liu 2172d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 2173f3f0e1d2SKirill A. Shutemov ret = 1; 2174579c571eSSong Liu khugepaged_scan_file(mm, file, pgoff, hpage); 2175f3f0e1d2SKirill A. Shutemov fput(file); 2176f3f0e1d2SKirill A. Shutemov } else { 2177b46e756fSKirill A. Shutemov ret = khugepaged_scan_pmd(mm, vma, 2178b46e756fSKirill A. Shutemov khugepaged_scan.address, 2179b46e756fSKirill A. Shutemov hpage); 2180f3f0e1d2SKirill A. Shutemov } 2181b46e756fSKirill A. Shutemov /* move to next address */ 2182b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 2183b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 2184b46e756fSKirill A. Shutemov if (ret) 2185c1e8d7c6SMichel Lespinasse /* we released mmap_lock so break loop */ 2186c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2187b46e756fSKirill A. Shutemov if (progress >= pages) 2188b46e756fSKirill A. Shutemov goto breakouterloop; 2189b46e756fSKirill A. Shutemov } 2190b46e756fSKirill A. Shutemov } 2191b46e756fSKirill A. Shutemov breakouterloop: 2192d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ 2193c1e8d7c6SMichel Lespinasse breakouterloop_mmap_lock: 2194b46e756fSKirill A. Shutemov 2195b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2196b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2197b46e756fSKirill A. Shutemov /* 2198b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 2199b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 2200b46e756fSKirill A. Shutemov */ 2201b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm) || !vma) { 2202b46e756fSKirill A. Shutemov /* 2203b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 2204b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 2205b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 2206b46e756fSKirill A. Shutemov */ 2207b46e756fSKirill A. Shutemov if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2208b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = list_entry( 2209b46e756fSKirill A. Shutemov mm_slot->mm_node.next, 2210b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2211b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2212b46e756fSKirill A. Shutemov } else { 2213b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2214b46e756fSKirill A. Shutemov khugepaged_full_scans++; 2215b46e756fSKirill A. Shutemov } 2216b46e756fSKirill A. Shutemov 2217b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2218b46e756fSKirill A. Shutemov } 2219b46e756fSKirill A. Shutemov 2220b46e756fSKirill A. Shutemov return progress; 2221b46e756fSKirill A. Shutemov } 2222b46e756fSKirill A. Shutemov 2223b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 2224b46e756fSKirill A. Shutemov { 2225b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 2226b46e756fSKirill A. Shutemov khugepaged_enabled(); 2227b46e756fSKirill A. Shutemov } 2228b46e756fSKirill A. Shutemov 2229b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 2230b46e756fSKirill A. Shutemov { 2231b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 2232b46e756fSKirill A. Shutemov kthread_should_stop(); 2233b46e756fSKirill A. Shutemov } 2234b46e756fSKirill A. Shutemov 2235b46e756fSKirill A. Shutemov static void khugepaged_do_scan(void) 2236b46e756fSKirill A. Shutemov { 2237b46e756fSKirill A. Shutemov struct page *hpage = NULL; 2238b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 223989dc6a96SYanfei Xu unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); 2240b46e756fSKirill A. Shutemov bool wait = true; 2241b46e756fSKirill A. Shutemov 2242a980df33SKirill A. Shutemov lru_add_drain_all(); 2243a980df33SKirill A. Shutemov 2244b46e756fSKirill A. Shutemov while (progress < pages) { 2245b46e756fSKirill A. Shutemov if (!khugepaged_prealloc_page(&hpage, &wait)) 2246b46e756fSKirill A. Shutemov break; 2247b46e756fSKirill A. Shutemov 2248b46e756fSKirill A. Shutemov cond_resched(); 2249b46e756fSKirill A. Shutemov 2250b46e756fSKirill A. Shutemov if (unlikely(kthread_should_stop() || try_to_freeze())) 2251b46e756fSKirill A. Shutemov break; 2252b46e756fSKirill A. Shutemov 2253b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2254b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 2255b46e756fSKirill A. Shutemov pass_through_head++; 2256b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 2257b46e756fSKirill A. Shutemov pass_through_head < 2) 2258b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 2259b46e756fSKirill A. Shutemov &hpage); 2260b46e756fSKirill A. Shutemov else 2261b46e756fSKirill A. Shutemov progress = pages; 2262b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2263b46e756fSKirill A. Shutemov } 2264b46e756fSKirill A. Shutemov 2265b46e756fSKirill A. Shutemov if (!IS_ERR_OR_NULL(hpage)) 2266b46e756fSKirill A. Shutemov put_page(hpage); 2267b46e756fSKirill A. Shutemov } 2268b46e756fSKirill A. Shutemov 2269b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 2270b46e756fSKirill A. Shutemov { 2271b46e756fSKirill A. Shutemov return kthread_should_stop() || 2272b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 2273b46e756fSKirill A. Shutemov } 2274b46e756fSKirill A. Shutemov 2275b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 2276b46e756fSKirill A. Shutemov { 2277b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 2278b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 2279b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2280b46e756fSKirill A. Shutemov 2281b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 2282b46e756fSKirill A. Shutemov return; 2283b46e756fSKirill A. Shutemov 2284b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2285b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 2286b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 2287b46e756fSKirill A. Shutemov scan_sleep_jiffies); 2288b46e756fSKirill A. Shutemov return; 2289b46e756fSKirill A. Shutemov } 2290b46e756fSKirill A. Shutemov 2291b46e756fSKirill A. Shutemov if (khugepaged_enabled()) 2292b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2293b46e756fSKirill A. Shutemov } 2294b46e756fSKirill A. Shutemov 2295b46e756fSKirill A. Shutemov static int khugepaged(void *none) 2296b46e756fSKirill A. Shutemov { 2297b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 2298b46e756fSKirill A. Shutemov 2299b46e756fSKirill A. Shutemov set_freezable(); 2300b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 2301b46e756fSKirill A. Shutemov 2302b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 2303b46e756fSKirill A. Shutemov khugepaged_do_scan(); 2304b46e756fSKirill A. Shutemov khugepaged_wait_work(); 2305b46e756fSKirill A. Shutemov } 2306b46e756fSKirill A. Shutemov 2307b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2308b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2309b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2310b46e756fSKirill A. Shutemov if (mm_slot) 2311b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2312b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2313b46e756fSKirill A. Shutemov return 0; 2314b46e756fSKirill A. Shutemov } 2315b46e756fSKirill A. Shutemov 2316b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 2317b46e756fSKirill A. Shutemov { 2318b46e756fSKirill A. Shutemov struct zone *zone; 2319b46e756fSKirill A. Shutemov int nr_zones = 0; 2320b46e756fSKirill A. Shutemov unsigned long recommended_min; 2321b46e756fSKirill A. Shutemov 2322bd3400eaSLiangcai Fan if (!khugepaged_enabled()) { 2323bd3400eaSLiangcai Fan calculate_min_free_kbytes(); 2324bd3400eaSLiangcai Fan goto update_wmarks; 2325bd3400eaSLiangcai Fan } 2326bd3400eaSLiangcai Fan 2327b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 2328b7d349c7SJoonsoo Kim /* 2329b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 2330b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 2331b7d349c7SJoonsoo Kim */ 2332b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 2333b7d349c7SJoonsoo Kim continue; 2334b7d349c7SJoonsoo Kim 2335b46e756fSKirill A. Shutemov nr_zones++; 2336b7d349c7SJoonsoo Kim } 2337b46e756fSKirill A. Shutemov 2338b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2339b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 2340b46e756fSKirill A. Shutemov 2341b46e756fSKirill A. Shutemov /* 2342b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 2343b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 2344b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 2345b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 2346b46e756fSKirill A. Shutemov */ 2347b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 2348b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2349b46e756fSKirill A. Shutemov 2350b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 2351b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 2352b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 2353b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 2354b46e756fSKirill A. Shutemov 2355b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 2356b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 2357b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2358b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 2359b46e756fSKirill A. Shutemov 2360b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 2361b46e756fSKirill A. Shutemov } 2362bd3400eaSLiangcai Fan 2363bd3400eaSLiangcai Fan update_wmarks: 2364b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 2365b46e756fSKirill A. Shutemov } 2366b46e756fSKirill A. Shutemov 2367b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 2368b46e756fSKirill A. Shutemov { 2369b46e756fSKirill A. Shutemov int err = 0; 2370b46e756fSKirill A. Shutemov 2371b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 2372b46e756fSKirill A. Shutemov if (khugepaged_enabled()) { 2373b46e756fSKirill A. Shutemov if (!khugepaged_thread) 2374b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 2375b46e756fSKirill A. Shutemov "khugepaged"); 2376b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 2377b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2378b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 2379b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2380b46e756fSKirill A. Shutemov goto fail; 2381b46e756fSKirill A. Shutemov } 2382b46e756fSKirill A. Shutemov 2383b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 2384b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 2385b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 2386b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 2387b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2388b46e756fSKirill A. Shutemov } 2389bd3400eaSLiangcai Fan set_recommended_min_free_kbytes(); 2390b46e756fSKirill A. Shutemov fail: 2391b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 2392b46e756fSKirill A. Shutemov return err; 2393b46e756fSKirill A. Shutemov } 23944aab2be0SVijay Balakrishna 23954aab2be0SVijay Balakrishna void khugepaged_min_free_kbytes_update(void) 23964aab2be0SVijay Balakrishna { 23974aab2be0SVijay Balakrishna mutex_lock(&khugepaged_mutex); 23984aab2be0SVijay Balakrishna if (khugepaged_enabled() && khugepaged_thread) 23994aab2be0SVijay Balakrishna set_recommended_min_free_kbytes(); 24004aab2be0SVijay Balakrishna mutex_unlock(&khugepaged_mutex); 24014aab2be0SVijay Balakrishna } 2402