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> 201e2f2d31SKent Overstreet #include <linux/rcupdate_wait.h> 21b46e756fSKirill A. Shutemov #include <linux/swapops.h> 22f3f0e1d2SKirill A. Shutemov #include <linux/shmem_fs.h> 23e2942062Sxu xin #include <linux/ksm.h> 24b46e756fSKirill A. Shutemov 25b46e756fSKirill A. Shutemov #include <asm/tlb.h> 26b46e756fSKirill A. Shutemov #include <asm/pgalloc.h> 27b46e756fSKirill A. Shutemov #include "internal.h" 28b26e2701SQi Zheng #include "mm_slot.h" 29b46e756fSKirill A. Shutemov 30b46e756fSKirill A. Shutemov enum scan_result { 31b46e756fSKirill A. Shutemov SCAN_FAIL, 32b46e756fSKirill A. Shutemov SCAN_SUCCEED, 33b46e756fSKirill A. Shutemov SCAN_PMD_NULL, 3434488399SZach O'Keefe SCAN_PMD_NONE, 3550722804SZach O'Keefe SCAN_PMD_MAPPED, 36b46e756fSKirill A. Shutemov SCAN_EXCEED_NONE_PTE, 3771a2c112SKirill A. Shutemov SCAN_EXCEED_SWAP_PTE, 3871a2c112SKirill A. Shutemov SCAN_EXCEED_SHARED_PTE, 39b46e756fSKirill A. Shutemov SCAN_PTE_NON_PRESENT, 40e1e267c7SPeter Xu SCAN_PTE_UFFD_WP, 4158ac9a89SZach O'Keefe SCAN_PTE_MAPPED_HUGEPAGE, 42b46e756fSKirill A. Shutemov SCAN_PAGE_RO, 430db501f7SEbru Akagunduz SCAN_LACK_REFERENCED_PAGE, 44b46e756fSKirill A. Shutemov SCAN_PAGE_NULL, 45b46e756fSKirill A. Shutemov SCAN_SCAN_ABORT, 46b46e756fSKirill A. Shutemov SCAN_PAGE_COUNT, 47b46e756fSKirill A. Shutemov SCAN_PAGE_LRU, 48b46e756fSKirill A. Shutemov SCAN_PAGE_LOCK, 49b46e756fSKirill A. Shutemov SCAN_PAGE_ANON, 50b46e756fSKirill A. Shutemov SCAN_PAGE_COMPOUND, 51b46e756fSKirill A. Shutemov SCAN_ANY_PROCESS, 52b46e756fSKirill A. Shutemov SCAN_VMA_NULL, 53b46e756fSKirill A. Shutemov SCAN_VMA_CHECK, 54b46e756fSKirill A. Shutemov SCAN_ADDRESS_RANGE, 55b46e756fSKirill A. Shutemov SCAN_DEL_PAGE_LRU, 56b46e756fSKirill A. Shutemov SCAN_ALLOC_HUGE_PAGE_FAIL, 57b46e756fSKirill A. Shutemov SCAN_CGROUP_CHARGE_FAIL, 58f3f0e1d2SKirill A. Shutemov SCAN_TRUNCATED, 5999cb0dbdSSong Liu SCAN_PAGE_HAS_PRIVATE, 602ce0bdfeSIvan Orlov SCAN_STORE_FAILED, 6198c76c9fSJiaqi Yan SCAN_COPY_MC, 62ac492b9cSDavid Stevens SCAN_PAGE_FILLED, 63b46e756fSKirill A. Shutemov }; 64b46e756fSKirill A. Shutemov 65b46e756fSKirill A. Shutemov #define CREATE_TRACE_POINTS 66b46e756fSKirill A. Shutemov #include <trace/events/huge_memory.h> 67b46e756fSKirill A. Shutemov 684aab2be0SVijay Balakrishna static struct task_struct *khugepaged_thread __read_mostly; 694aab2be0SVijay Balakrishna static DEFINE_MUTEX(khugepaged_mutex); 704aab2be0SVijay Balakrishna 71b46e756fSKirill A. Shutemov /* default scan 8*512 pte (or vmas) every 30 second */ 72b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_to_scan __read_mostly; 73b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_collapsed; 74b46e756fSKirill A. Shutemov static unsigned int khugepaged_full_scans; 75b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 76b46e756fSKirill A. Shutemov /* during fragmentation poll the hugepage allocator once every minute */ 77b46e756fSKirill A. Shutemov static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 78b46e756fSKirill A. Shutemov static unsigned long khugepaged_sleep_expire; 79b46e756fSKirill A. Shutemov static DEFINE_SPINLOCK(khugepaged_mm_lock); 80b46e756fSKirill A. Shutemov static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 81b46e756fSKirill A. Shutemov /* 82b46e756fSKirill A. Shutemov * default collapse hugepages if there is at least one pte mapped like 83b46e756fSKirill A. Shutemov * it would have happened if the vma was large enough during page 84b46e756fSKirill A. Shutemov * fault. 85d8ea7cc8SZach O'Keefe * 86d8ea7cc8SZach O'Keefe * Note that these are only respected if collapse was initiated by khugepaged. 87b46e756fSKirill A. Shutemov */ 88b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_none __read_mostly; 89b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_swap __read_mostly; 9071a2c112SKirill A. Shutemov static unsigned int khugepaged_max_ptes_shared __read_mostly; 91b46e756fSKirill A. Shutemov 92b46e756fSKirill A. Shutemov #define MM_SLOTS_HASH_BITS 10 93e1ad3e66SNick Desaulniers static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 94b46e756fSKirill A. Shutemov 9568279f9cSAlexey Dobriyan static struct kmem_cache *mm_slot_cache __ro_after_init; 96b46e756fSKirill A. Shutemov 9734d6b470SZach O'Keefe struct collapse_control { 98d8ea7cc8SZach O'Keefe bool is_khugepaged; 99d8ea7cc8SZach O'Keefe 10034d6b470SZach O'Keefe /* Num pages scanned per node */ 10134d6b470SZach O'Keefe u32 node_load[MAX_NUMNODES]; 10234d6b470SZach O'Keefe 103e031ff96SYang Shi /* nodemask for allocation fallback */ 104e031ff96SYang Shi nodemask_t alloc_nmask; 10534d6b470SZach O'Keefe }; 10634d6b470SZach O'Keefe 107b46e756fSKirill A. Shutemov /** 108b26e2701SQi Zheng * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned 109b26e2701SQi Zheng * @slot: hash lookup from mm to mm_slot 110b46e756fSKirill A. Shutemov */ 111b26e2701SQi Zheng struct khugepaged_mm_slot { 112b26e2701SQi Zheng struct mm_slot slot; 113b46e756fSKirill A. Shutemov }; 114b46e756fSKirill A. Shutemov 115b46e756fSKirill A. Shutemov /** 116b46e756fSKirill A. Shutemov * struct khugepaged_scan - cursor for scanning 117b46e756fSKirill A. Shutemov * @mm_head: the head of the mm list to scan 118b46e756fSKirill A. Shutemov * @mm_slot: the current mm_slot we are scanning 119b46e756fSKirill A. Shutemov * @address: the next address inside that to be scanned 120b46e756fSKirill A. Shutemov * 121b46e756fSKirill A. Shutemov * There is only the one khugepaged_scan instance of this cursor structure. 122b46e756fSKirill A. Shutemov */ 123b46e756fSKirill A. Shutemov struct khugepaged_scan { 124b46e756fSKirill A. Shutemov struct list_head mm_head; 125b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 126b46e756fSKirill A. Shutemov unsigned long address; 127b46e756fSKirill A. Shutemov }; 128b46e756fSKirill A. Shutemov 129b46e756fSKirill A. Shutemov static struct khugepaged_scan khugepaged_scan = { 130b46e756fSKirill A. Shutemov .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 131b46e756fSKirill A. Shutemov }; 132b46e756fSKirill A. Shutemov 133e1465d12SJérémy Lefaure #ifdef CONFIG_SYSFS 134b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 135b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 136b46e756fSKirill A. Shutemov char *buf) 137b46e756fSKirill A. Shutemov { 138ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); 139b46e756fSKirill A. Shutemov } 140b46e756fSKirill A. Shutemov 141b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 142b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 143b46e756fSKirill A. Shutemov const char *buf, size_t count) 144b46e756fSKirill A. Shutemov { 145dfefd226SAlexey Dobriyan unsigned int msecs; 146b46e756fSKirill A. Shutemov int err; 147b46e756fSKirill A. Shutemov 148dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 149dfefd226SAlexey Dobriyan if (err) 150b46e756fSKirill A. Shutemov return -EINVAL; 151b46e756fSKirill A. Shutemov 152b46e756fSKirill A. Shutemov khugepaged_scan_sleep_millisecs = msecs; 153b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 154b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 155b46e756fSKirill A. Shutemov 156b46e756fSKirill A. Shutemov return count; 157b46e756fSKirill A. Shutemov } 158b46e756fSKirill A. Shutemov static struct kobj_attribute scan_sleep_millisecs_attr = 1596dcdc94dSMiaohe Lin __ATTR_RW(scan_sleep_millisecs); 160b46e756fSKirill A. Shutemov 161b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 162b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 163b46e756fSKirill A. Shutemov char *buf) 164b46e756fSKirill A. Shutemov { 165ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 166b46e756fSKirill A. Shutemov } 167b46e756fSKirill A. Shutemov 168b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 169b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 170b46e756fSKirill A. Shutemov const char *buf, size_t count) 171b46e756fSKirill A. Shutemov { 172dfefd226SAlexey Dobriyan unsigned int msecs; 173b46e756fSKirill A. Shutemov int err; 174b46e756fSKirill A. Shutemov 175dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 176dfefd226SAlexey Dobriyan if (err) 177b46e756fSKirill A. Shutemov return -EINVAL; 178b46e756fSKirill A. Shutemov 179b46e756fSKirill A. Shutemov khugepaged_alloc_sleep_millisecs = msecs; 180b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 181b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 182b46e756fSKirill A. Shutemov 183b46e756fSKirill A. Shutemov return count; 184b46e756fSKirill A. Shutemov } 185b46e756fSKirill A. Shutemov static struct kobj_attribute alloc_sleep_millisecs_attr = 1866dcdc94dSMiaohe Lin __ATTR_RW(alloc_sleep_millisecs); 187b46e756fSKirill A. Shutemov 188b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_show(struct kobject *kobj, 189b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 190b46e756fSKirill A. Shutemov char *buf) 191b46e756fSKirill A. Shutemov { 192ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); 193b46e756fSKirill A. Shutemov } 194b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_store(struct kobject *kobj, 195b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 196b46e756fSKirill A. Shutemov const char *buf, size_t count) 197b46e756fSKirill A. Shutemov { 198dfefd226SAlexey Dobriyan unsigned int pages; 199b46e756fSKirill A. Shutemov int err; 200b46e756fSKirill A. Shutemov 201dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &pages); 202dfefd226SAlexey Dobriyan if (err || !pages) 203b46e756fSKirill A. Shutemov return -EINVAL; 204b46e756fSKirill A. Shutemov 205b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = pages; 206b46e756fSKirill A. Shutemov 207b46e756fSKirill A. Shutemov return count; 208b46e756fSKirill A. Shutemov } 209b46e756fSKirill A. Shutemov static struct kobj_attribute pages_to_scan_attr = 2106dcdc94dSMiaohe Lin __ATTR_RW(pages_to_scan); 211b46e756fSKirill A. Shutemov 212b46e756fSKirill A. Shutemov static ssize_t pages_collapsed_show(struct kobject *kobj, 213b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 214b46e756fSKirill A. Shutemov char *buf) 215b46e756fSKirill A. Shutemov { 216ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); 217b46e756fSKirill A. Shutemov } 218b46e756fSKirill A. Shutemov static struct kobj_attribute pages_collapsed_attr = 219b46e756fSKirill A. Shutemov __ATTR_RO(pages_collapsed); 220b46e756fSKirill A. Shutemov 221b46e756fSKirill A. Shutemov static ssize_t full_scans_show(struct kobject *kobj, 222b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 223b46e756fSKirill A. Shutemov char *buf) 224b46e756fSKirill A. Shutemov { 225ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_full_scans); 226b46e756fSKirill A. Shutemov } 227b46e756fSKirill A. Shutemov static struct kobj_attribute full_scans_attr = 228b46e756fSKirill A. Shutemov __ATTR_RO(full_scans); 229b46e756fSKirill A. Shutemov 2306dcdc94dSMiaohe Lin static ssize_t defrag_show(struct kobject *kobj, 231b46e756fSKirill A. Shutemov struct kobj_attribute *attr, char *buf) 232b46e756fSKirill A. Shutemov { 233b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 234b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 235b46e756fSKirill A. Shutemov } 2366dcdc94dSMiaohe Lin static ssize_t defrag_store(struct kobject *kobj, 237b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 238b46e756fSKirill A. Shutemov const char *buf, size_t count) 239b46e756fSKirill A. Shutemov { 240b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 241b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 242b46e756fSKirill A. Shutemov } 243b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_defrag_attr = 2446dcdc94dSMiaohe Lin __ATTR_RW(defrag); 245b46e756fSKirill A. Shutemov 246b46e756fSKirill A. Shutemov /* 247b46e756fSKirill A. Shutemov * max_ptes_none controls if khugepaged should collapse hugepages over 248b46e756fSKirill A. Shutemov * any unmapped ptes in turn potentially increasing the memory 249b46e756fSKirill A. Shutemov * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 250b46e756fSKirill A. Shutemov * reduce the available free memory in the system as it 251b46e756fSKirill A. Shutemov * runs. Increasing max_ptes_none will instead potentially reduce the 252b46e756fSKirill A. Shutemov * free memory in the system during the khugepaged scan. 253b46e756fSKirill A. Shutemov */ 2546dcdc94dSMiaohe Lin static ssize_t max_ptes_none_show(struct kobject *kobj, 255b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 256b46e756fSKirill A. Shutemov char *buf) 257b46e756fSKirill A. Shutemov { 258ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); 259b46e756fSKirill A. Shutemov } 2606dcdc94dSMiaohe Lin static ssize_t max_ptes_none_store(struct kobject *kobj, 261b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 262b46e756fSKirill A. Shutemov const char *buf, size_t count) 263b46e756fSKirill A. Shutemov { 264b46e756fSKirill A. Shutemov int err; 265b46e756fSKirill A. Shutemov unsigned long max_ptes_none; 266b46e756fSKirill A. Shutemov 267b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_none); 268b46e756fSKirill A. Shutemov if (err || max_ptes_none > HPAGE_PMD_NR - 1) 269b46e756fSKirill A. Shutemov return -EINVAL; 270b46e756fSKirill A. Shutemov 271b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = max_ptes_none; 272b46e756fSKirill A. Shutemov 273b46e756fSKirill A. Shutemov return count; 274b46e756fSKirill A. Shutemov } 275b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_none_attr = 2766dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_none); 277b46e756fSKirill A. Shutemov 2786dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_show(struct kobject *kobj, 279b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 280b46e756fSKirill A. Shutemov char *buf) 281b46e756fSKirill A. Shutemov { 282ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); 283b46e756fSKirill A. Shutemov } 284b46e756fSKirill A. Shutemov 2856dcdc94dSMiaohe Lin static ssize_t max_ptes_swap_store(struct kobject *kobj, 286b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 287b46e756fSKirill A. Shutemov const char *buf, size_t count) 288b46e756fSKirill A. Shutemov { 289b46e756fSKirill A. Shutemov int err; 290b46e756fSKirill A. Shutemov unsigned long max_ptes_swap; 291b46e756fSKirill A. Shutemov 292b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_swap); 293b46e756fSKirill A. Shutemov if (err || max_ptes_swap > HPAGE_PMD_NR - 1) 294b46e756fSKirill A. Shutemov return -EINVAL; 295b46e756fSKirill A. Shutemov 296b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = max_ptes_swap; 297b46e756fSKirill A. Shutemov 298b46e756fSKirill A. Shutemov return count; 299b46e756fSKirill A. Shutemov } 300b46e756fSKirill A. Shutemov 301b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_swap_attr = 3026dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_swap); 303b46e756fSKirill A. Shutemov 3046dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_show(struct kobject *kobj, 30571a2c112SKirill A. Shutemov struct kobj_attribute *attr, 30671a2c112SKirill A. Shutemov char *buf) 30771a2c112SKirill A. Shutemov { 308ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); 30971a2c112SKirill A. Shutemov } 31071a2c112SKirill A. Shutemov 3116dcdc94dSMiaohe Lin static ssize_t max_ptes_shared_store(struct kobject *kobj, 31271a2c112SKirill A. Shutemov struct kobj_attribute *attr, 31371a2c112SKirill A. Shutemov const char *buf, size_t count) 31471a2c112SKirill A. Shutemov { 31571a2c112SKirill A. Shutemov int err; 31671a2c112SKirill A. Shutemov unsigned long max_ptes_shared; 31771a2c112SKirill A. Shutemov 31871a2c112SKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_shared); 31971a2c112SKirill A. Shutemov if (err || max_ptes_shared > HPAGE_PMD_NR - 1) 32071a2c112SKirill A. Shutemov return -EINVAL; 32171a2c112SKirill A. Shutemov 32271a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = max_ptes_shared; 32371a2c112SKirill A. Shutemov 32471a2c112SKirill A. Shutemov return count; 32571a2c112SKirill A. Shutemov } 32671a2c112SKirill A. Shutemov 32771a2c112SKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_shared_attr = 3286dcdc94dSMiaohe Lin __ATTR_RW(max_ptes_shared); 32971a2c112SKirill A. Shutemov 330b46e756fSKirill A. Shutemov static struct attribute *khugepaged_attr[] = { 331b46e756fSKirill A. Shutemov &khugepaged_defrag_attr.attr, 332b46e756fSKirill A. Shutemov &khugepaged_max_ptes_none_attr.attr, 33371a2c112SKirill A. Shutemov &khugepaged_max_ptes_swap_attr.attr, 33471a2c112SKirill A. Shutemov &khugepaged_max_ptes_shared_attr.attr, 335b46e756fSKirill A. Shutemov &pages_to_scan_attr.attr, 336b46e756fSKirill A. Shutemov &pages_collapsed_attr.attr, 337b46e756fSKirill A. Shutemov &full_scans_attr.attr, 338b46e756fSKirill A. Shutemov &scan_sleep_millisecs_attr.attr, 339b46e756fSKirill A. Shutemov &alloc_sleep_millisecs_attr.attr, 340b46e756fSKirill A. Shutemov NULL, 341b46e756fSKirill A. Shutemov }; 342b46e756fSKirill A. Shutemov 343b46e756fSKirill A. Shutemov struct attribute_group khugepaged_attr_group = { 344b46e756fSKirill A. Shutemov .attrs = khugepaged_attr, 345b46e756fSKirill A. Shutemov .name = "khugepaged", 346b46e756fSKirill A. Shutemov }; 347e1465d12SJérémy Lefaure #endif /* CONFIG_SYSFS */ 348b46e756fSKirill A. Shutemov 349b46e756fSKirill A. Shutemov int hugepage_madvise(struct vm_area_struct *vma, 350b46e756fSKirill A. Shutemov unsigned long *vm_flags, int advice) 351b46e756fSKirill A. Shutemov { 352b46e756fSKirill A. Shutemov switch (advice) { 353b46e756fSKirill A. Shutemov case MADV_HUGEPAGE: 354b46e756fSKirill A. Shutemov #ifdef CONFIG_S390 355b46e756fSKirill A. Shutemov /* 356b46e756fSKirill A. Shutemov * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 357b46e756fSKirill A. Shutemov * can't handle this properly after s390_enable_sie, so we simply 358b46e756fSKirill A. Shutemov * ignore the madvise to prevent qemu from causing a SIGSEGV. 359b46e756fSKirill A. Shutemov */ 360b46e756fSKirill A. Shutemov if (mm_has_pgste(vma->vm_mm)) 361b46e756fSKirill A. Shutemov return 0; 362b46e756fSKirill A. Shutemov #endif 363b46e756fSKirill A. Shutemov *vm_flags &= ~VM_NOHUGEPAGE; 364b46e756fSKirill A. Shutemov *vm_flags |= VM_HUGEPAGE; 365b46e756fSKirill A. Shutemov /* 366b46e756fSKirill A. Shutemov * If the vma become good for khugepaged to scan, 367b46e756fSKirill A. Shutemov * register it here without waiting a page fault that 368b46e756fSKirill A. Shutemov * may not happen any time soon. 369b46e756fSKirill A. Shutemov */ 370c791576cSYang Shi khugepaged_enter_vma(vma, *vm_flags); 371b46e756fSKirill A. Shutemov break; 372b46e756fSKirill A. Shutemov case MADV_NOHUGEPAGE: 373b46e756fSKirill A. Shutemov *vm_flags &= ~VM_HUGEPAGE; 374b46e756fSKirill A. Shutemov *vm_flags |= VM_NOHUGEPAGE; 375b46e756fSKirill A. Shutemov /* 376b46e756fSKirill A. Shutemov * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 377b46e756fSKirill A. Shutemov * this vma even if we leave the mm registered in khugepaged if 378b46e756fSKirill A. Shutemov * it got registered before VM_NOHUGEPAGE was set. 379b46e756fSKirill A. Shutemov */ 380b46e756fSKirill A. Shutemov break; 381b46e756fSKirill A. Shutemov } 382b46e756fSKirill A. Shutemov 383b46e756fSKirill A. Shutemov return 0; 384b46e756fSKirill A. Shutemov } 385b46e756fSKirill A. Shutemov 386b46e756fSKirill A. Shutemov int __init khugepaged_init(void) 387b46e756fSKirill A. Shutemov { 388b46e756fSKirill A. Shutemov mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 389b26e2701SQi Zheng sizeof(struct khugepaged_mm_slot), 390b26e2701SQi Zheng __alignof__(struct khugepaged_mm_slot), 391b26e2701SQi Zheng 0, NULL); 392b46e756fSKirill A. Shutemov if (!mm_slot_cache) 393b46e756fSKirill A. Shutemov return -ENOMEM; 394b46e756fSKirill A. Shutemov 395b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 396b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 397b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 39871a2c112SKirill A. Shutemov khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; 399b46e756fSKirill A. Shutemov 400b46e756fSKirill A. Shutemov return 0; 401b46e756fSKirill A. Shutemov } 402b46e756fSKirill A. Shutemov 403b46e756fSKirill A. Shutemov void __init khugepaged_destroy(void) 404b46e756fSKirill A. Shutemov { 405b46e756fSKirill A. Shutemov kmem_cache_destroy(mm_slot_cache); 406b46e756fSKirill A. Shutemov } 407b46e756fSKirill A. Shutemov 4087d2c4385SZach O'Keefe static inline int hpage_collapse_test_exit(struct mm_struct *mm) 409b46e756fSKirill A. Shutemov { 4104d45e75aSJann Horn return atomic_read(&mm->mm_users) == 0; 411b46e756fSKirill A. Shutemov } 412b46e756fSKirill A. Shutemov 413879c6000SLance Yang static inline int hpage_collapse_test_exit_or_disable(struct mm_struct *mm) 414879c6000SLance Yang { 415879c6000SLance Yang return hpage_collapse_test_exit(mm) || 416879c6000SLance Yang test_bit(MMF_DISABLE_THP, &mm->flags); 417879c6000SLance Yang } 418879c6000SLance Yang 419d2081b2bSYang Shi void __khugepaged_enter(struct mm_struct *mm) 420b46e756fSKirill A. Shutemov { 421b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 422b26e2701SQi Zheng struct mm_slot *slot; 423b46e756fSKirill A. Shutemov int wakeup; 424b46e756fSKirill A. Shutemov 42516618670SXin Hao /* __khugepaged_exit() must not run from under us */ 42616618670SXin Hao VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm); 42716618670SXin Hao if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) 42816618670SXin Hao return; 42916618670SXin Hao 430b26e2701SQi Zheng mm_slot = mm_slot_alloc(mm_slot_cache); 431b46e756fSKirill A. Shutemov if (!mm_slot) 432d2081b2bSYang Shi return; 433b46e756fSKirill A. Shutemov 434b26e2701SQi Zheng slot = &mm_slot->slot; 435b26e2701SQi Zheng 436b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 437b26e2701SQi Zheng mm_slot_insert(mm_slots_hash, mm, slot); 438b46e756fSKirill A. Shutemov /* 439b46e756fSKirill A. Shutemov * Insert just behind the scanning cursor, to let the area settle 440b46e756fSKirill A. Shutemov * down a little. 441b46e756fSKirill A. Shutemov */ 442b46e756fSKirill A. Shutemov wakeup = list_empty(&khugepaged_scan.mm_head); 443b26e2701SQi Zheng list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head); 444b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 445b46e756fSKirill A. Shutemov 446f1f10076SVegard Nossum mmgrab(mm); 447b46e756fSKirill A. Shutemov if (wakeup) 448b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 449b46e756fSKirill A. Shutemov } 450b46e756fSKirill A. Shutemov 451c791576cSYang Shi void khugepaged_enter_vma(struct vm_area_struct *vma, 452b46e756fSKirill A. Shutemov unsigned long vm_flags) 453b46e756fSKirill A. Shutemov { 4542647d11bSYang Shi if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && 4551064026bSYang Shi hugepage_flags_enabled()) { 4563485b883SRyan Roberts if (thp_vma_allowable_order(vma, vm_flags, false, false, true, 4573485b883SRyan Roberts PMD_ORDER)) 4582647d11bSYang Shi __khugepaged_enter(vma->vm_mm); 4592647d11bSYang Shi } 460b46e756fSKirill A. Shutemov } 461b46e756fSKirill A. Shutemov 462b46e756fSKirill A. Shutemov void __khugepaged_exit(struct mm_struct *mm) 463b46e756fSKirill A. Shutemov { 464b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 465b26e2701SQi Zheng struct mm_slot *slot; 466b46e756fSKirill A. Shutemov int free = 0; 467b46e756fSKirill A. Shutemov 468b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 469b26e2701SQi Zheng slot = mm_slot_lookup(mm_slots_hash, mm); 470b26e2701SQi Zheng mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 471b46e756fSKirill A. Shutemov if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 472b26e2701SQi Zheng hash_del(&slot->hash); 473b26e2701SQi Zheng list_del(&slot->mm_node); 474b46e756fSKirill A. Shutemov free = 1; 475b46e756fSKirill A. Shutemov } 476b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 477b46e756fSKirill A. Shutemov 478b46e756fSKirill A. Shutemov if (free) { 479b46e756fSKirill A. Shutemov clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 480b26e2701SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 481b46e756fSKirill A. Shutemov mmdrop(mm); 482b46e756fSKirill A. Shutemov } else if (mm_slot) { 483b46e756fSKirill A. Shutemov /* 484b46e756fSKirill A. Shutemov * This is required to serialize against 4857d2c4385SZach O'Keefe * hpage_collapse_test_exit() (which is guaranteed to run 4867d2c4385SZach O'Keefe * under mmap sem read mode). Stop here (after we return all 4877d2c4385SZach O'Keefe * pagetables will be destroyed) until khugepaged has finished 4887d2c4385SZach O'Keefe * working on the pagetables under the mmap_lock. 489b46e756fSKirill A. Shutemov */ 490d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 491d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 492b46e756fSKirill A. Shutemov } 493b46e756fSKirill A. Shutemov } 494b46e756fSKirill A. Shutemov 49592644f58SVishal Moola (Oracle) static void release_pte_folio(struct folio *folio) 49692644f58SVishal Moola (Oracle) { 49792644f58SVishal Moola (Oracle) node_stat_mod_folio(folio, 49892644f58SVishal Moola (Oracle) NR_ISOLATED_ANON + folio_is_file_lru(folio), 49992644f58SVishal Moola (Oracle) -folio_nr_pages(folio)); 50092644f58SVishal Moola (Oracle) folio_unlock(folio); 50192644f58SVishal Moola (Oracle) folio_putback_lru(folio); 50292644f58SVishal Moola (Oracle) } 50392644f58SVishal Moola (Oracle) 5045503fbf2SKirill A. Shutemov static void release_pte_pages(pte_t *pte, pte_t *_pte, 5055503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 506b46e756fSKirill A. Shutemov { 5079bdfeea4SVishal Moola (Oracle) struct folio *folio, *tmp; 5085503fbf2SKirill A. Shutemov 509b46e756fSKirill A. Shutemov while (--_pte >= pte) { 510c33c7948SRyan Roberts pte_t pteval = ptep_get(_pte); 511f528260bSVishal Moola (Oracle) unsigned long pfn; 5125503fbf2SKirill A. Shutemov 513f528260bSVishal Moola (Oracle) if (pte_none(pteval)) 514f528260bSVishal Moola (Oracle) continue; 515f528260bSVishal Moola (Oracle) pfn = pte_pfn(pteval); 516f528260bSVishal Moola (Oracle) if (is_zero_pfn(pfn)) 517f528260bSVishal Moola (Oracle) continue; 518f528260bSVishal Moola (Oracle) folio = pfn_folio(pfn); 519f528260bSVishal Moola (Oracle) if (folio_test_large(folio)) 520f528260bSVishal Moola (Oracle) continue; 5219bdfeea4SVishal Moola (Oracle) release_pte_folio(folio); 5225503fbf2SKirill A. Shutemov } 5235503fbf2SKirill A. Shutemov 5249bdfeea4SVishal Moola (Oracle) list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) { 5259bdfeea4SVishal Moola (Oracle) list_del(&folio->lru); 5269bdfeea4SVishal Moola (Oracle) release_pte_folio(folio); 527b46e756fSKirill A. Shutemov } 528b46e756fSKirill A. Shutemov } 529b46e756fSKirill A. Shutemov 530dbf85c21SVishal Moola (Oracle) static bool is_refcount_suitable(struct folio *folio) 5319445689fSKirill A. Shutemov { 5329445689fSKirill A. Shutemov int expected_refcount; 5339445689fSKirill A. Shutemov 534dbf85c21SVishal Moola (Oracle) expected_refcount = folio_mapcount(folio); 535dbf85c21SVishal Moola (Oracle) if (folio_test_swapcache(folio)) 536dbf85c21SVishal Moola (Oracle) expected_refcount += folio_nr_pages(folio); 5379445689fSKirill A. Shutemov 538dbf85c21SVishal Moola (Oracle) return folio_ref_count(folio) == expected_refcount; 5399445689fSKirill A. Shutemov } 5409445689fSKirill A. Shutemov 541b46e756fSKirill A. Shutemov static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 542b46e756fSKirill A. Shutemov unsigned long address, 5435503fbf2SKirill A. Shutemov pte_t *pte, 544d8ea7cc8SZach O'Keefe struct collapse_control *cc, 5455503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 546b46e756fSKirill A. Shutemov { 547b46e756fSKirill A. Shutemov struct page *page = NULL; 5488dd1e896SVishal Moola (Oracle) struct folio *folio = NULL; 549b46e756fSKirill A. Shutemov pte_t *_pte; 55050ad2f24SZach O'Keefe int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0; 5510db501f7SEbru Akagunduz bool writable = false; 552b46e756fSKirill A. Shutemov 553b46e756fSKirill A. Shutemov for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 554b46e756fSKirill A. Shutemov _pte++, address += PAGE_SIZE) { 555c33c7948SRyan Roberts pte_t pteval = ptep_get(_pte); 556b46e756fSKirill A. Shutemov if (pte_none(pteval) || (pte_present(pteval) && 557b46e756fSKirill A. Shutemov is_zero_pfn(pte_pfn(pteval)))) { 558d8ea7cc8SZach O'Keefe ++none_or_zero; 559b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 560d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 561d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 562b46e756fSKirill A. Shutemov continue; 563b46e756fSKirill A. Shutemov } else { 564b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 565e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 566b46e756fSKirill A. Shutemov goto out; 567b46e756fSKirill A. Shutemov } 568b46e756fSKirill A. Shutemov } 569b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 570b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 571b46e756fSKirill A. Shutemov goto out; 572b46e756fSKirill A. Shutemov } 573dd47ac42SPeter Xu if (pte_uffd_wp(pteval)) { 574dd47ac42SPeter Xu result = SCAN_PTE_UFFD_WP; 575dd47ac42SPeter Xu goto out; 576dd47ac42SPeter Xu } 577b46e756fSKirill A. Shutemov page = vm_normal_page(vma, address, pteval); 5783218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 579b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 580b46e756fSKirill A. Shutemov goto out; 581b46e756fSKirill A. Shutemov } 582b46e756fSKirill A. Shutemov 5838dd1e896SVishal Moola (Oracle) folio = page_folio(page); 5848dd1e896SVishal Moola (Oracle) VM_BUG_ON_FOLIO(!folio_test_anon(folio), folio); 585b46e756fSKirill A. Shutemov 586d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 587d8ea7cc8SZach O'Keefe ++shared; 588d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 589d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 59071a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 591e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 59271a2c112SKirill A. Shutemov goto out; 59371a2c112SKirill A. Shutemov } 594d8ea7cc8SZach O'Keefe } 59571a2c112SKirill A. Shutemov 5968dd1e896SVishal Moola (Oracle) if (folio_test_large(folio)) { 5978dd1e896SVishal Moola (Oracle) struct folio *f; 5985503fbf2SKirill A. Shutemov 5995503fbf2SKirill A. Shutemov /* 6005503fbf2SKirill A. Shutemov * Check if we have dealt with the compound page 6015503fbf2SKirill A. Shutemov * already 6025503fbf2SKirill A. Shutemov */ 6038dd1e896SVishal Moola (Oracle) list_for_each_entry(f, compound_pagelist, lru) { 6048dd1e896SVishal Moola (Oracle) if (folio == f) 6055503fbf2SKirill A. Shutemov goto next; 6065503fbf2SKirill A. Shutemov } 6075503fbf2SKirill A. Shutemov } 6085503fbf2SKirill A. Shutemov 609b46e756fSKirill A. Shutemov /* 610b46e756fSKirill A. Shutemov * We can do it before isolate_lru_page because the 611b46e756fSKirill A. Shutemov * page can't be freed from under us. NOTE: PG_lock 612b46e756fSKirill A. Shutemov * is needed to serialize against split_huge_page 613b46e756fSKirill A. Shutemov * when invoked from the VM. 614b46e756fSKirill A. Shutemov */ 6158dd1e896SVishal Moola (Oracle) if (!folio_trylock(folio)) { 616b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 617b46e756fSKirill A. Shutemov goto out; 618b46e756fSKirill A. Shutemov } 619b46e756fSKirill A. Shutemov 620b46e756fSKirill A. Shutemov /* 6219445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 6229445689fSKirill A. Shutemov * 6239445689fSKirill A. Shutemov * The page table that maps the page has been already unlinked 6249445689fSKirill A. Shutemov * from the page table tree and this process cannot get 625f0953a1bSIngo Molnar * an additional pin on the page. 6269445689fSKirill A. Shutemov * 6279445689fSKirill A. Shutemov * New pins can come later if the page is shared across fork, 6289445689fSKirill A. Shutemov * but not from this process. The other process cannot write to 6299445689fSKirill A. Shutemov * the page, only trigger CoW. 630b46e756fSKirill A. Shutemov */ 631dbf85c21SVishal Moola (Oracle) if (!is_refcount_suitable(folio)) { 6328dd1e896SVishal Moola (Oracle) folio_unlock(folio); 633b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 634b46e756fSKirill A. Shutemov goto out; 635b46e756fSKirill A. Shutemov } 636b46e756fSKirill A. Shutemov 637b46e756fSKirill A. Shutemov /* 638b46e756fSKirill A. Shutemov * Isolate the page to avoid collapsing an hugepage 639b46e756fSKirill A. Shutemov * currently in use by the VM. 640b46e756fSKirill A. Shutemov */ 6418dd1e896SVishal Moola (Oracle) if (!folio_isolate_lru(folio)) { 6428dd1e896SVishal Moola (Oracle) folio_unlock(folio); 643b46e756fSKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 644b46e756fSKirill A. Shutemov goto out; 645b46e756fSKirill A. Shutemov } 6468dd1e896SVishal Moola (Oracle) node_stat_mod_folio(folio, 6478dd1e896SVishal Moola (Oracle) NR_ISOLATED_ANON + folio_is_file_lru(folio), 6488dd1e896SVishal Moola (Oracle) folio_nr_pages(folio)); 6498dd1e896SVishal Moola (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 6508dd1e896SVishal Moola (Oracle) VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); 651b46e756fSKirill A. Shutemov 6528dd1e896SVishal Moola (Oracle) if (folio_test_large(folio)) 6538dd1e896SVishal Moola (Oracle) list_add_tail(&folio->lru, compound_pagelist); 6545503fbf2SKirill A. Shutemov next: 655d8ea7cc8SZach O'Keefe /* 656d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 657d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 658d8ea7cc8SZach O'Keefe */ 659d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 6608dd1e896SVishal Moola (Oracle) (pte_young(pteval) || folio_test_young(folio) || 6618dd1e896SVishal Moola (Oracle) folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm, 662d8ea7cc8SZach O'Keefe address))) 6630db501f7SEbru Akagunduz referenced++; 6645503fbf2SKirill A. Shutemov 6655503fbf2SKirill A. Shutemov if (pte_write(pteval)) 6665503fbf2SKirill A. Shutemov writable = true; 667b46e756fSKirill A. Shutemov } 66874e579bfSMiaohe Lin 66974e579bfSMiaohe Lin if (unlikely(!writable)) { 67074e579bfSMiaohe Lin result = SCAN_PAGE_RO; 671d8ea7cc8SZach O'Keefe } else if (unlikely(cc->is_khugepaged && !referenced)) { 67274e579bfSMiaohe Lin result = SCAN_LACK_REFERENCED_PAGE; 67374e579bfSMiaohe Lin } else { 674b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 6758dd1e896SVishal Moola (Oracle) trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero, 676b46e756fSKirill A. Shutemov referenced, writable, result); 67750ad2f24SZach O'Keefe return result; 678b46e756fSKirill A. Shutemov } 679b46e756fSKirill A. Shutemov out: 6805503fbf2SKirill A. Shutemov release_pte_pages(pte, _pte, compound_pagelist); 6818dd1e896SVishal Moola (Oracle) trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero, 682b46e756fSKirill A. Shutemov referenced, writable, result); 68350ad2f24SZach O'Keefe return result; 684b46e756fSKirill A. Shutemov } 685b46e756fSKirill A. Shutemov 68698c76c9fSJiaqi Yan static void __collapse_huge_page_copy_succeeded(pte_t *pte, 687b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 688b46e756fSKirill A. Shutemov unsigned long address, 6895503fbf2SKirill A. Shutemov spinlock_t *ptl, 6905503fbf2SKirill A. Shutemov struct list_head *compound_pagelist) 691b46e756fSKirill A. Shutemov { 692d4111eecSMatthew Wilcox (Oracle) struct folio *src, *tmp; 693b46e756fSKirill A. Shutemov pte_t *_pte; 69498c76c9fSJiaqi Yan pte_t pteval; 695b46e756fSKirill A. Shutemov 69698c76c9fSJiaqi Yan for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 69798c76c9fSJiaqi Yan _pte++, address += PAGE_SIZE) { 698c33c7948SRyan Roberts pteval = ptep_get(_pte); 699b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 700b46e756fSKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 701b46e756fSKirill A. Shutemov if (is_zero_pfn(pte_pfn(pteval))) { 702b46e756fSKirill A. Shutemov /* 703b46e756fSKirill A. Shutemov * ptl mostly unnecessary. 704b46e756fSKirill A. Shutemov */ 705b46e756fSKirill A. Shutemov spin_lock(ptl); 70608d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 707b46e756fSKirill A. Shutemov spin_unlock(ptl); 7086080d19fSxu xin ksm_might_unmap_zero_page(vma->vm_mm, pteval); 709b46e756fSKirill A. Shutemov } 710b46e756fSKirill A. Shutemov } else { 711d4111eecSMatthew Wilcox (Oracle) struct page *src_page = pte_page(pteval); 712d4111eecSMatthew Wilcox (Oracle) 713d4111eecSMatthew Wilcox (Oracle) src = page_folio(src_page); 714d4111eecSMatthew Wilcox (Oracle) if (!folio_test_large(src)) 715d4111eecSMatthew Wilcox (Oracle) release_pte_folio(src); 716b46e756fSKirill A. Shutemov /* 717b46e756fSKirill A. Shutemov * ptl mostly unnecessary, but preempt has to 718b46e756fSKirill A. Shutemov * be disabled to update the per-cpu stats 71935668a43SDavid Hildenbrand * inside folio_remove_rmap_pte(). 720b46e756fSKirill A. Shutemov */ 721b46e756fSKirill A. Shutemov spin_lock(ptl); 72208d5b29eSPasha Tatashin ptep_clear(vma->vm_mm, address, _pte); 723d4111eecSMatthew Wilcox (Oracle) folio_remove_rmap_pte(src, src_page, vma); 724b46e756fSKirill A. Shutemov spin_unlock(ptl); 725b46e756fSKirill A. Shutemov free_page_and_swap_cache(src_page); 726b46e756fSKirill A. Shutemov } 727b46e756fSKirill A. Shutemov } 7285503fbf2SKirill A. Shutemov 729d4111eecSMatthew Wilcox (Oracle) list_for_each_entry_safe(src, tmp, compound_pagelist, lru) { 730d4111eecSMatthew Wilcox (Oracle) list_del(&src->lru); 731d4111eecSMatthew Wilcox (Oracle) node_stat_sub_folio(src, NR_ISOLATED_ANON + 732d4111eecSMatthew Wilcox (Oracle) folio_is_file_lru(src)); 733d4111eecSMatthew Wilcox (Oracle) folio_unlock(src); 734*63b77499SMatthew Wilcox (Oracle) free_swap_cache(src); 735d4111eecSMatthew Wilcox (Oracle) folio_putback_lru(src); 7365503fbf2SKirill A. Shutemov } 737b46e756fSKirill A. Shutemov } 738b46e756fSKirill A. Shutemov 73998c76c9fSJiaqi Yan static void __collapse_huge_page_copy_failed(pte_t *pte, 74098c76c9fSJiaqi Yan pmd_t *pmd, 74198c76c9fSJiaqi Yan pmd_t orig_pmd, 74298c76c9fSJiaqi Yan struct vm_area_struct *vma, 74398c76c9fSJiaqi Yan struct list_head *compound_pagelist) 74498c76c9fSJiaqi Yan { 74598c76c9fSJiaqi Yan spinlock_t *pmd_ptl; 74698c76c9fSJiaqi Yan 74798c76c9fSJiaqi Yan /* 74898c76c9fSJiaqi Yan * Re-establish the PMD to point to the original page table 74998c76c9fSJiaqi Yan * entry. Restoring PMD needs to be done prior to releasing 75098c76c9fSJiaqi Yan * pages. Since pages are still isolated and locked here, 75198c76c9fSJiaqi Yan * acquiring anon_vma_lock_write is unnecessary. 75298c76c9fSJiaqi Yan */ 75398c76c9fSJiaqi Yan pmd_ptl = pmd_lock(vma->vm_mm, pmd); 75498c76c9fSJiaqi Yan pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd)); 75598c76c9fSJiaqi Yan spin_unlock(pmd_ptl); 75698c76c9fSJiaqi Yan /* 75798c76c9fSJiaqi Yan * Release both raw and compound pages isolated 75898c76c9fSJiaqi Yan * in __collapse_huge_page_isolate. 75998c76c9fSJiaqi Yan */ 76098c76c9fSJiaqi Yan release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist); 76198c76c9fSJiaqi Yan } 76298c76c9fSJiaqi Yan 76398c76c9fSJiaqi Yan /* 76498c76c9fSJiaqi Yan * __collapse_huge_page_copy - attempts to copy memory contents from raw 76598c76c9fSJiaqi Yan * pages to a hugepage. Cleans up the raw pages if copying succeeds; 76698c76c9fSJiaqi Yan * otherwise restores the original page table and releases isolated raw pages. 76798c76c9fSJiaqi Yan * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC. 76898c76c9fSJiaqi Yan * 76998c76c9fSJiaqi Yan * @pte: starting of the PTEs to copy from 77098c76c9fSJiaqi Yan * @page: the new hugepage to copy contents to 77198c76c9fSJiaqi Yan * @pmd: pointer to the new hugepage's PMD 77298c76c9fSJiaqi Yan * @orig_pmd: the original raw pages' PMD 77398c76c9fSJiaqi Yan * @vma: the original raw pages' virtual memory area 77498c76c9fSJiaqi Yan * @address: starting address to copy 77598c76c9fSJiaqi Yan * @ptl: lock on raw pages' PTEs 77698c76c9fSJiaqi Yan * @compound_pagelist: list that stores compound pages 77798c76c9fSJiaqi Yan */ 77898c76c9fSJiaqi Yan static int __collapse_huge_page_copy(pte_t *pte, 77998c76c9fSJiaqi Yan struct page *page, 78098c76c9fSJiaqi Yan pmd_t *pmd, 78198c76c9fSJiaqi Yan pmd_t orig_pmd, 78298c76c9fSJiaqi Yan struct vm_area_struct *vma, 78398c76c9fSJiaqi Yan unsigned long address, 78498c76c9fSJiaqi Yan spinlock_t *ptl, 78598c76c9fSJiaqi Yan struct list_head *compound_pagelist) 78698c76c9fSJiaqi Yan { 78798c76c9fSJiaqi Yan struct page *src_page; 78898c76c9fSJiaqi Yan pte_t *_pte; 78998c76c9fSJiaqi Yan pte_t pteval; 79098c76c9fSJiaqi Yan unsigned long _address; 79198c76c9fSJiaqi Yan int result = SCAN_SUCCEED; 79298c76c9fSJiaqi Yan 79398c76c9fSJiaqi Yan /* 79498c76c9fSJiaqi Yan * Copying pages' contents is subject to memory poison at any iteration. 79598c76c9fSJiaqi Yan */ 79698c76c9fSJiaqi Yan for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; 79798c76c9fSJiaqi Yan _pte++, page++, _address += PAGE_SIZE) { 798c33c7948SRyan Roberts pteval = ptep_get(_pte); 79998c76c9fSJiaqi Yan if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 80098c76c9fSJiaqi Yan clear_user_highpage(page, _address); 80198c76c9fSJiaqi Yan continue; 80298c76c9fSJiaqi Yan } 80398c76c9fSJiaqi Yan src_page = pte_page(pteval); 80498c76c9fSJiaqi Yan if (copy_mc_user_highpage(page, src_page, _address, vma) > 0) { 80598c76c9fSJiaqi Yan result = SCAN_COPY_MC; 80698c76c9fSJiaqi Yan break; 80798c76c9fSJiaqi Yan } 80898c76c9fSJiaqi Yan } 80998c76c9fSJiaqi Yan 81098c76c9fSJiaqi Yan if (likely(result == SCAN_SUCCEED)) 81198c76c9fSJiaqi Yan __collapse_huge_page_copy_succeeded(pte, vma, address, ptl, 81298c76c9fSJiaqi Yan compound_pagelist); 81398c76c9fSJiaqi Yan else 81498c76c9fSJiaqi Yan __collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma, 81598c76c9fSJiaqi Yan compound_pagelist); 81698c76c9fSJiaqi Yan 81798c76c9fSJiaqi Yan return result; 81898c76c9fSJiaqi Yan } 81998c76c9fSJiaqi Yan 820b46e756fSKirill A. Shutemov static void khugepaged_alloc_sleep(void) 821b46e756fSKirill A. Shutemov { 822b46e756fSKirill A. Shutemov DEFINE_WAIT(wait); 823b46e756fSKirill A. Shutemov 824b46e756fSKirill A. Shutemov add_wait_queue(&khugepaged_wait, &wait); 825f5d39b02SPeter Zijlstra __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE); 826f5d39b02SPeter Zijlstra schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 827b46e756fSKirill A. Shutemov remove_wait_queue(&khugepaged_wait, &wait); 828b46e756fSKirill A. Shutemov } 829b46e756fSKirill A. Shutemov 83034d6b470SZach O'Keefe struct collapse_control khugepaged_collapse_control = { 831d8ea7cc8SZach O'Keefe .is_khugepaged = true, 83234d6b470SZach O'Keefe }; 833b46e756fSKirill A. Shutemov 8347d2c4385SZach O'Keefe static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc) 835b46e756fSKirill A. Shutemov { 836b46e756fSKirill A. Shutemov int i; 837b46e756fSKirill A. Shutemov 838b46e756fSKirill A. Shutemov /* 839a5f5f91dSMel Gorman * If node_reclaim_mode is disabled, then no extra effort is made to 840b46e756fSKirill A. Shutemov * allocate memory locally. 841b46e756fSKirill A. Shutemov */ 842202e35dbSDave Hansen if (!node_reclaim_enabled()) 843b46e756fSKirill A. Shutemov return false; 844b46e756fSKirill A. Shutemov 845b46e756fSKirill A. Shutemov /* If there is a count for this node already, it must be acceptable */ 84634d6b470SZach O'Keefe if (cc->node_load[nid]) 847b46e756fSKirill A. Shutemov return false; 848b46e756fSKirill A. Shutemov 849b46e756fSKirill A. Shutemov for (i = 0; i < MAX_NUMNODES; i++) { 85034d6b470SZach O'Keefe if (!cc->node_load[i]) 851b46e756fSKirill A. Shutemov continue; 852a55c7454SMatt Fleming if (node_distance(nid, i) > node_reclaim_distance) 853b46e756fSKirill A. Shutemov return true; 854b46e756fSKirill A. Shutemov } 855b46e756fSKirill A. Shutemov return false; 856b46e756fSKirill A. Shutemov } 857b46e756fSKirill A. Shutemov 8581064026bSYang Shi #define khugepaged_defrag() \ 8591064026bSYang Shi (transparent_hugepage_flags & \ 8601064026bSYang Shi (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)) 8611064026bSYang Shi 862b46e756fSKirill A. Shutemov /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 863b46e756fSKirill A. Shutemov static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 864b46e756fSKirill A. Shutemov { 86525160354SVlastimil Babka return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 866b46e756fSKirill A. Shutemov } 867b46e756fSKirill A. Shutemov 868b46e756fSKirill A. Shutemov #ifdef CONFIG_NUMA 8697d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 870b46e756fSKirill A. Shutemov { 871b46e756fSKirill A. Shutemov int nid, target_node = 0, max_value = 0; 872b46e756fSKirill A. Shutemov 873b46e756fSKirill A. Shutemov /* find first node with max normal pages hit */ 874b46e756fSKirill A. Shutemov for (nid = 0; nid < MAX_NUMNODES; nid++) 87534d6b470SZach O'Keefe if (cc->node_load[nid] > max_value) { 87634d6b470SZach O'Keefe max_value = cc->node_load[nid]; 877b46e756fSKirill A. Shutemov target_node = nid; 878b46e756fSKirill A. Shutemov } 879b46e756fSKirill A. Shutemov 880e031ff96SYang Shi for_each_online_node(nid) { 881e031ff96SYang Shi if (max_value == cc->node_load[nid]) 882e031ff96SYang Shi node_set(nid, cc->alloc_nmask); 883b46e756fSKirill A. Shutemov } 884b46e756fSKirill A. Shutemov 885b46e756fSKirill A. Shutemov return target_node; 886b46e756fSKirill A. Shutemov } 887c6a7f445SYang Shi #else 8887d2c4385SZach O'Keefe static int hpage_collapse_find_target_node(struct collapse_control *cc) 889b46e756fSKirill A. Shutemov { 890c6a7f445SYang Shi return 0; 891b46e756fSKirill A. Shutemov } 892c6a7f445SYang Shi #endif 893b46e756fSKirill A. Shutemov 894b455f39dSVishal Moola (Oracle) static bool hpage_collapse_alloc_folio(struct folio **folio, gfp_t gfp, int node, 895e031ff96SYang Shi nodemask_t *nmask) 896b46e756fSKirill A. Shutemov { 897b455f39dSVishal Moola (Oracle) *folio = __folio_alloc(gfp, HPAGE_PMD_ORDER, node, nmask); 898b455f39dSVishal Moola (Oracle) 899b455f39dSVishal Moola (Oracle) if (unlikely(!*folio)) { 900b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 9019710a78aSZach O'Keefe return false; 902b46e756fSKirill A. Shutemov } 903b46e756fSKirill A. Shutemov 904b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 905b46e756fSKirill A. Shutemov return true; 906b46e756fSKirill A. Shutemov } 907b46e756fSKirill A. Shutemov 908b46e756fSKirill A. Shutemov /* 909c1e8d7c6SMichel Lespinasse * If mmap_lock temporarily dropped, revalidate vma 910c1e8d7c6SMichel Lespinasse * before taking mmap_lock. 91150ad2f24SZach O'Keefe * Returns enum scan_result value. 912b46e756fSKirill A. Shutemov */ 913b46e756fSKirill A. Shutemov 914c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 91534488399SZach O'Keefe bool expect_anon, 916a7f4e6e4SZach O'Keefe struct vm_area_struct **vmap, 917a7f4e6e4SZach O'Keefe struct collapse_control *cc) 918b46e756fSKirill A. Shutemov { 919b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 920b46e756fSKirill A. Shutemov 9215dad6048SLance Yang if (unlikely(hpage_collapse_test_exit_or_disable(mm))) 922b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 923b46e756fSKirill A. Shutemov 924c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 925b46e756fSKirill A. Shutemov if (!vma) 926b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 927b46e756fSKirill A. Shutemov 9283485b883SRyan Roberts if (!thp_vma_suitable_order(vma, address, PMD_ORDER)) 929b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 9303485b883SRyan Roberts if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, 9313485b883SRyan Roberts cc->is_khugepaged, PMD_ORDER)) 932b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 933f707fa49SYang Shi /* 934f707fa49SYang Shi * Anon VMA expected, the address may be unmapped then 935f707fa49SYang Shi * remapped to file after khugepaged reaquired the mmap_lock. 936f707fa49SYang Shi * 9373485b883SRyan Roberts * thp_vma_allowable_order may return true for qualified file 938f707fa49SYang Shi * vmas. 939f707fa49SYang Shi */ 94034488399SZach O'Keefe if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) 94134488399SZach O'Keefe return SCAN_PAGE_ANON; 94250ad2f24SZach O'Keefe return SCAN_SUCCEED; 943b46e756fSKirill A. Shutemov } 944b46e756fSKirill A. Shutemov 94550722804SZach O'Keefe static int find_pmd_or_thp_or_none(struct mm_struct *mm, 94650722804SZach O'Keefe unsigned long address, 94750722804SZach O'Keefe pmd_t **pmd) 94850722804SZach O'Keefe { 94950722804SZach O'Keefe pmd_t pmde; 95050722804SZach O'Keefe 95150722804SZach O'Keefe *pmd = mm_find_pmd(mm, address); 95250722804SZach O'Keefe if (!*pmd) 95350722804SZach O'Keefe return SCAN_PMD_NULL; 95450722804SZach O'Keefe 955dab6e717SPeter Zijlstra pmde = pmdp_get_lockless(*pmd); 95634488399SZach O'Keefe if (pmd_none(pmde)) 95734488399SZach O'Keefe return SCAN_PMD_NONE; 958edb5d0cfSZach O'Keefe if (!pmd_present(pmde)) 959edb5d0cfSZach O'Keefe return SCAN_PMD_NULL; 96050722804SZach O'Keefe if (pmd_trans_huge(pmde)) 96150722804SZach O'Keefe return SCAN_PMD_MAPPED; 962edb5d0cfSZach O'Keefe if (pmd_devmap(pmde)) 963edb5d0cfSZach O'Keefe return SCAN_PMD_NULL; 96450722804SZach O'Keefe if (pmd_bad(pmde)) 96550722804SZach O'Keefe return SCAN_PMD_NULL; 96650722804SZach O'Keefe return SCAN_SUCCEED; 96750722804SZach O'Keefe } 96850722804SZach O'Keefe 96950722804SZach O'Keefe static int check_pmd_still_valid(struct mm_struct *mm, 97050722804SZach O'Keefe unsigned long address, 97150722804SZach O'Keefe pmd_t *pmd) 97250722804SZach O'Keefe { 97350722804SZach O'Keefe pmd_t *new_pmd; 97450722804SZach O'Keefe int result = find_pmd_or_thp_or_none(mm, address, &new_pmd); 97550722804SZach O'Keefe 97650722804SZach O'Keefe if (result != SCAN_SUCCEED) 97750722804SZach O'Keefe return result; 97850722804SZach O'Keefe if (new_pmd != pmd) 97950722804SZach O'Keefe return SCAN_FAIL; 98050722804SZach O'Keefe return SCAN_SUCCEED; 981b46e756fSKirill A. Shutemov } 982b46e756fSKirill A. Shutemov 983b46e756fSKirill A. Shutemov /* 984b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 9857d2c4385SZach O'Keefe * Only done if hpage_collapse_scan_pmd believes it is worthwhile. 986b46e756fSKirill A. Shutemov * 9874d928e20SMiaohe Lin * Called and returns without pte mapped or spinlocks held. 988895f5ee4SHugh Dickins * Returns result: if not SCAN_SUCCEED, mmap_lock has been released. 989b46e756fSKirill A. Shutemov */ 99050ad2f24SZach O'Keefe static int __collapse_huge_page_swapin(struct mm_struct *mm, 991b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 9922b635dd3SWill Deacon unsigned long haddr, pmd_t *pmd, 9930db501f7SEbru Akagunduz int referenced) 994b46e756fSKirill A. Shutemov { 9952b740303SSouptick Joarder int swapped_in = 0; 9962b740303SSouptick Joarder vm_fault_t ret = 0; 9972b635dd3SWill Deacon unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); 998895f5ee4SHugh Dickins int result; 999895f5ee4SHugh Dickins pte_t *pte = NULL; 1000c7ad0880SHugh Dickins spinlock_t *ptl; 10012b635dd3SWill Deacon 10022b635dd3SWill Deacon for (address = haddr; address < end; address += PAGE_SIZE) { 100382b0f8c3SJan Kara struct vm_fault vmf = { 1004b46e756fSKirill A. Shutemov .vma = vma, 1005b46e756fSKirill A. Shutemov .address = address, 1006895f5ee4SHugh Dickins .pgoff = linear_page_index(vma, address), 1007b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 1008b46e756fSKirill A. Shutemov .pmd = pmd, 1009b46e756fSKirill A. Shutemov }; 1010b46e756fSKirill A. Shutemov 1011895f5ee4SHugh Dickins if (!pte++) { 1012c7ad0880SHugh Dickins pte = pte_offset_map_nolock(mm, pmd, address, &ptl); 1013895f5ee4SHugh Dickins if (!pte) { 1014895f5ee4SHugh Dickins mmap_read_unlock(mm); 1015895f5ee4SHugh Dickins result = SCAN_PMD_NULL; 1016895f5ee4SHugh Dickins goto out; 10172b635dd3SWill Deacon } 1018895f5ee4SHugh Dickins } 1019895f5ee4SHugh Dickins 1020c7ad0880SHugh Dickins vmf.orig_pte = ptep_get_lockless(pte); 1021895f5ee4SHugh Dickins if (!is_swap_pte(vmf.orig_pte)) 1022895f5ee4SHugh Dickins continue; 1023895f5ee4SHugh Dickins 1024895f5ee4SHugh Dickins vmf.pte = pte; 1025c7ad0880SHugh Dickins vmf.ptl = ptl; 10262994302bSJan Kara ret = do_swap_page(&vmf); 1027895f5ee4SHugh Dickins /* Which unmaps pte (after perhaps re-checking the entry) */ 1028895f5ee4SHugh Dickins pte = NULL; 10290db501f7SEbru Akagunduz 10304d928e20SMiaohe Lin /* 10314d928e20SMiaohe Lin * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. 10324d928e20SMiaohe Lin * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because 10334d928e20SMiaohe Lin * we do not retry here and swap entry will remain in pagetable 10344d928e20SMiaohe Lin * resulting in later failure. 10354d928e20SMiaohe Lin */ 1036b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 103750ad2f24SZach O'Keefe /* Likely, but not guaranteed, that page lock failed */ 1038895f5ee4SHugh Dickins result = SCAN_PAGE_LOCK; 1039895f5ee4SHugh Dickins goto out; 104047f863eaSEbru Akagunduz } 1041b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 10424d928e20SMiaohe Lin mmap_read_unlock(mm); 1043895f5ee4SHugh Dickins result = SCAN_FAIL; 1044895f5ee4SHugh Dickins goto out; 1045b46e756fSKirill A. Shutemov } 10464d928e20SMiaohe Lin swapped_in++; 1047b46e756fSKirill A. Shutemov } 1048ae2c5d80SKirill A. Shutemov 1049895f5ee4SHugh Dickins if (pte) 1050895f5ee4SHugh Dickins pte_unmap(pte); 1051895f5ee4SHugh Dickins 10521fec6890SMatthew Wilcox (Oracle) /* Drain LRU cache to remove extra pin on the swapped in pages */ 1053ae2c5d80SKirill A. Shutemov if (swapped_in) 1054ae2c5d80SKirill A. Shutemov lru_add_drain(); 1055ae2c5d80SKirill A. Shutemov 1056895f5ee4SHugh Dickins result = SCAN_SUCCEED; 1057895f5ee4SHugh Dickins out: 1058895f5ee4SHugh Dickins trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result); 1059895f5ee4SHugh Dickins return result; 1060b46e756fSKirill A. Shutemov } 1061b46e756fSKirill A. Shutemov 10629710a78aSZach O'Keefe static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, 10639710a78aSZach O'Keefe struct collapse_control *cc) 10649710a78aSZach O'Keefe { 10657d8faaf1SZach O'Keefe gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() : 1066e031ff96SYang Shi GFP_TRANSHUGE); 10677d2c4385SZach O'Keefe int node = hpage_collapse_find_target_node(cc); 106894c02ad7SPeter Xu struct folio *folio; 10699710a78aSZach O'Keefe 1070b455f39dSVishal Moola (Oracle) if (!hpage_collapse_alloc_folio(&folio, gfp, node, &cc->alloc_nmask)) { 1071b455f39dSVishal Moola (Oracle) *hpage = NULL; 10729710a78aSZach O'Keefe return SCAN_ALLOC_HUGE_PAGE_FAIL; 1073b455f39dSVishal Moola (Oracle) } 107494c02ad7SPeter Xu 107594c02ad7SPeter Xu if (unlikely(mem_cgroup_charge(folio, mm, gfp))) { 107694c02ad7SPeter Xu folio_put(folio); 107794c02ad7SPeter Xu *hpage = NULL; 10789710a78aSZach O'Keefe return SCAN_CGROUP_CHARGE_FAIL; 107994c02ad7SPeter Xu } 108094c02ad7SPeter Xu 1081b455f39dSVishal Moola (Oracle) count_memcg_folio_events(folio, THP_COLLAPSE_ALLOC, 1); 1082b455f39dSVishal Moola (Oracle) 1083b455f39dSVishal Moola (Oracle) *hpage = folio_page(folio, 0); 10849710a78aSZach O'Keefe return SCAN_SUCCEED; 10859710a78aSZach O'Keefe } 10869710a78aSZach O'Keefe 108750ad2f24SZach O'Keefe static int collapse_huge_page(struct mm_struct *mm, unsigned long address, 108850ad2f24SZach O'Keefe int referenced, int unmapped, 108950ad2f24SZach O'Keefe struct collapse_control *cc) 1090b46e756fSKirill A. Shutemov { 10915503fbf2SKirill A. Shutemov LIST_HEAD(compound_pagelist); 1092b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 1093b46e756fSKirill A. Shutemov pte_t *pte; 1094b46e756fSKirill A. Shutemov pgtable_t pgtable; 109554327268SMatthew Wilcox (Oracle) struct folio *folio; 109650ad2f24SZach O'Keefe struct page *hpage; 1097b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 109850ad2f24SZach O'Keefe int result = SCAN_FAIL; 1099c131f751SKirill A. Shutemov struct vm_area_struct *vma; 1100ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1101b46e756fSKirill A. Shutemov 1102b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1103b46e756fSKirill A. Shutemov 1104988ddb71SKirill A. Shutemov /* 1105c1e8d7c6SMichel Lespinasse * Before allocating the hugepage, release the mmap_lock read lock. 1106988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 1107c1e8d7c6SMichel Lespinasse * sync compaction, and we do not need to hold the mmap_lock during 1108988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 1109988ddb71SKirill A. Shutemov */ 1110d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1111b46e756fSKirill A. Shutemov 111250ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 11139710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1114b46e756fSKirill A. Shutemov goto out_nolock; 1115b46e756fSKirill A. Shutemov 1116d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 111734488399SZach O'Keefe result = hugepage_vma_revalidate(mm, address, true, &vma, cc); 111850ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) { 1119d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1120b46e756fSKirill A. Shutemov goto out_nolock; 1121b46e756fSKirill A. Shutemov } 1122b46e756fSKirill A. Shutemov 112350722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 112450722804SZach O'Keefe if (result != SCAN_SUCCEED) { 1125d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1126b46e756fSKirill A. Shutemov goto out_nolock; 1127b46e756fSKirill A. Shutemov } 1128b46e756fSKirill A. Shutemov 112950ad2f24SZach O'Keefe if (unmapped) { 1130b46e756fSKirill A. Shutemov /* 113150ad2f24SZach O'Keefe * __collapse_huge_page_swapin will return with mmap_lock 113250ad2f24SZach O'Keefe * released when it fails. So we jump out_nolock directly in 113350ad2f24SZach O'Keefe * that case. Continuing to collapse causes inconsistency. 1134b46e756fSKirill A. Shutemov */ 113550ad2f24SZach O'Keefe result = __collapse_huge_page_swapin(mm, vma, address, pmd, 113650ad2f24SZach O'Keefe referenced); 113750ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 1138b46e756fSKirill A. Shutemov goto out_nolock; 1139b46e756fSKirill A. Shutemov } 1140b46e756fSKirill A. Shutemov 1141d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1142b46e756fSKirill A. Shutemov /* 1143b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 1144b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 1145b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 1146adef4406SAndrea Arcangeli * 1147adef4406SAndrea Arcangeli * UFFDIO_MOVE is prevented to race as well thanks to the 1148adef4406SAndrea Arcangeli * mmap_lock. 1149b46e756fSKirill A. Shutemov */ 1150d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 115134488399SZach O'Keefe result = hugepage_vma_revalidate(mm, address, true, &vma, cc); 115250ad2f24SZach O'Keefe if (result != SCAN_SUCCEED) 115318d24a7cSMiaohe Lin goto out_up_write; 1154b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 115550722804SZach O'Keefe result = check_pmd_still_valid(mm, address, pmd); 115650722804SZach O'Keefe if (result != SCAN_SUCCEED) 115718d24a7cSMiaohe Lin goto out_up_write; 1158b46e756fSKirill A. Shutemov 115955fd6fccSSuren Baghdasaryan vma_start_write(vma); 1160b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1161b46e756fSKirill A. Shutemov 11627d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address, 11637d4a8be0SAlistair Popple address + HPAGE_PMD_SIZE); 1164ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1165ec649c9dSVille Syrjälä 1166b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1167b46e756fSKirill A. Shutemov /* 116870cbc3ccSYang Shi * This removes any huge TLB entry from the CPU so we won't allow 116970cbc3ccSYang Shi * huge and small TLB entries for the same virtual address to 117070cbc3ccSYang Shi * avoid the risk of CPU bugs in that area. 117170cbc3ccSYang Shi * 117270cbc3ccSYang Shi * Parallel fast GUP is fine since fast GUP will back off when 117370cbc3ccSYang Shi * it detects PMD is changed. 1174b46e756fSKirill A. Shutemov */ 1175b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1176b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1177ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 11782ba99c5eSJann Horn tlb_remove_table_sync_one(); 1179b46e756fSKirill A. Shutemov 1180895f5ee4SHugh Dickins pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl); 1181895f5ee4SHugh Dickins if (pte) { 1182d8ea7cc8SZach O'Keefe result = __collapse_huge_page_isolate(vma, address, pte, cc, 11835503fbf2SKirill A. Shutemov &compound_pagelist); 1184b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1185895f5ee4SHugh Dickins } else { 1186895f5ee4SHugh Dickins result = SCAN_PMD_NULL; 1187895f5ee4SHugh Dickins } 1188b46e756fSKirill A. Shutemov 118950ad2f24SZach O'Keefe if (unlikely(result != SCAN_SUCCEED)) { 1190895f5ee4SHugh Dickins if (pte) 1191b46e756fSKirill A. Shutemov pte_unmap(pte); 1192b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1193b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1194b46e756fSKirill A. Shutemov /* 1195b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1196b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1197b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1198b46e756fSKirill A. Shutemov */ 1199b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1200b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1201b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 120218d24a7cSMiaohe Lin goto out_up_write; 1203b46e756fSKirill A. Shutemov } 1204b46e756fSKirill A. Shutemov 1205b46e756fSKirill A. Shutemov /* 1206b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1207b46e756fSKirill A. Shutemov * can't run anymore. 1208b46e756fSKirill A. Shutemov */ 1209b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1210b46e756fSKirill A. Shutemov 121198c76c9fSJiaqi Yan result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd, 121298c76c9fSJiaqi Yan vma, address, pte_ptl, 12135503fbf2SKirill A. Shutemov &compound_pagelist); 1214b46e756fSKirill A. Shutemov pte_unmap(pte); 121598c76c9fSJiaqi Yan if (unlikely(result != SCAN_SUCCEED)) 121698c76c9fSJiaqi Yan goto out_up_write; 121798c76c9fSJiaqi Yan 121854327268SMatthew Wilcox (Oracle) folio = page_folio(hpage); 1219588d01f9SMiaohe Lin /* 122054327268SMatthew Wilcox (Oracle) * The smp_wmb() inside __folio_mark_uptodate() ensures the 122154327268SMatthew Wilcox (Oracle) * copy_huge_page writes become visible before the set_pmd_at() 122254327268SMatthew Wilcox (Oracle) * write. 1223588d01f9SMiaohe Lin */ 122454327268SMatthew Wilcox (Oracle) __folio_mark_uptodate(folio); 1225b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1226b46e756fSKirill A. Shutemov 122750ad2f24SZach O'Keefe _pmd = mk_huge_pmd(hpage, vma->vm_page_prot); 1228f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1229b46e756fSKirill A. Shutemov 1230b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1231b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 123254327268SMatthew Wilcox (Oracle) folio_add_new_anon_rmap(folio, vma, address); 123354327268SMatthew Wilcox (Oracle) folio_add_lru_vma(folio, vma); 1234b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1235b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1236b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1237b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1238b46e756fSKirill A. Shutemov 123950ad2f24SZach O'Keefe hpage = NULL; 1240b46e756fSKirill A. Shutemov 1241b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1242b46e756fSKirill A. Shutemov out_up_write: 1243d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 1244b46e756fSKirill A. Shutemov out_nolock: 12457cb1d7efSPeter Xu if (hpage) 124650ad2f24SZach O'Keefe put_page(hpage); 124750ad2f24SZach O'Keefe trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); 124850ad2f24SZach O'Keefe return result; 1249b46e756fSKirill A. Shutemov } 1250b46e756fSKirill A. Shutemov 12517d2c4385SZach O'Keefe static int hpage_collapse_scan_pmd(struct mm_struct *mm, 1252b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 125350ad2f24SZach O'Keefe unsigned long address, bool *mmap_locked, 125434d6b470SZach O'Keefe struct collapse_control *cc) 1255b46e756fSKirill A. Shutemov { 1256b46e756fSKirill A. Shutemov pmd_t *pmd; 1257b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 125850ad2f24SZach O'Keefe int result = SCAN_FAIL, referenced = 0; 125971a2c112SKirill A. Shutemov int none_or_zero = 0, shared = 0; 1260b46e756fSKirill A. Shutemov struct page *page = NULL; 12615c07ebb3SVishal Moola (Oracle) struct folio *folio = NULL; 1262b46e756fSKirill A. Shutemov unsigned long _address; 1263b46e756fSKirill A. Shutemov spinlock_t *ptl; 1264b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 12650db501f7SEbru Akagunduz bool writable = false; 1266b46e756fSKirill A. Shutemov 1267b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1268b46e756fSKirill A. Shutemov 126950722804SZach O'Keefe result = find_pmd_or_thp_or_none(mm, address, &pmd); 127050722804SZach O'Keefe if (result != SCAN_SUCCEED) 1271b46e756fSKirill A. Shutemov goto out; 1272b46e756fSKirill A. Shutemov 127334d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 1274e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 1275b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1276895f5ee4SHugh Dickins if (!pte) { 1277895f5ee4SHugh Dickins result = SCAN_PMD_NULL; 1278895f5ee4SHugh Dickins goto out; 1279895f5ee4SHugh Dickins } 1280895f5ee4SHugh Dickins 1281b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; 1282b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1283c33c7948SRyan Roberts pte_t pteval = ptep_get(_pte); 1284b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1285d8ea7cc8SZach O'Keefe ++unmapped; 1286d8ea7cc8SZach O'Keefe if (!cc->is_khugepaged || 1287d8ea7cc8SZach O'Keefe unmapped <= khugepaged_max_ptes_swap) { 1288e1e267c7SPeter Xu /* 1289e1e267c7SPeter Xu * Always be strict with uffd-wp 1290e1e267c7SPeter Xu * enabled swap entries. Please see 1291e1e267c7SPeter Xu * comment below for pte_uffd_wp(). 1292e1e267c7SPeter Xu */ 12932bad466cSPeter Xu if (pte_swp_uffd_wp_any(pteval)) { 1294e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1295e1e267c7SPeter Xu goto out_unmap; 1296e1e267c7SPeter Xu } 1297b46e756fSKirill A. Shutemov continue; 1298b46e756fSKirill A. Shutemov } else { 1299b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1300e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 1301b46e756fSKirill A. Shutemov goto out_unmap; 1302b46e756fSKirill A. Shutemov } 1303b46e756fSKirill A. Shutemov } 1304b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1305d8ea7cc8SZach O'Keefe ++none_or_zero; 1306b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1307d8ea7cc8SZach O'Keefe (!cc->is_khugepaged || 1308d8ea7cc8SZach O'Keefe none_or_zero <= khugepaged_max_ptes_none)) { 1309b46e756fSKirill A. Shutemov continue; 1310b46e756fSKirill A. Shutemov } else { 1311b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1312e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 1313b46e756fSKirill A. Shutemov goto out_unmap; 1314b46e756fSKirill A. Shutemov } 1315b46e756fSKirill A. Shutemov } 1316e1e267c7SPeter Xu if (pte_uffd_wp(pteval)) { 1317e1e267c7SPeter Xu /* 1318e1e267c7SPeter Xu * Don't collapse the page if any of the small 1319e1e267c7SPeter Xu * PTEs are armed with uffd write protection. 1320e1e267c7SPeter Xu * Here we can also mark the new huge pmd as 1321e1e267c7SPeter Xu * write protected if any of the small ones is 13228958b249SHaitao Shi * marked but that could bring unknown 1323e1e267c7SPeter Xu * userfault messages that falls outside of 1324e1e267c7SPeter Xu * the registered range. So, just be simple. 1325e1e267c7SPeter Xu */ 1326e1e267c7SPeter Xu result = SCAN_PTE_UFFD_WP; 1327e1e267c7SPeter Xu goto out_unmap; 1328e1e267c7SPeter Xu } 1329b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1330b46e756fSKirill A. Shutemov writable = true; 1331b46e756fSKirill A. Shutemov 1332b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 13333218f871SAlex Sierra if (unlikely(!page) || unlikely(is_zone_device_page(page))) { 1334b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1335b46e756fSKirill A. Shutemov goto out_unmap; 1336b46e756fSKirill A. Shutemov } 1337b46e756fSKirill A. Shutemov 1338d8ea7cc8SZach O'Keefe if (page_mapcount(page) > 1) { 1339d8ea7cc8SZach O'Keefe ++shared; 1340d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 1341d8ea7cc8SZach O'Keefe shared > khugepaged_max_ptes_shared) { 134271a2c112SKirill A. Shutemov result = SCAN_EXCEED_SHARED_PTE; 1343e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); 134471a2c112SKirill A. Shutemov goto out_unmap; 134571a2c112SKirill A. Shutemov } 1346d8ea7cc8SZach O'Keefe } 134771a2c112SKirill A. Shutemov 13485c07ebb3SVishal Moola (Oracle) folio = page_folio(page); 1349b46e756fSKirill A. Shutemov /* 1350b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 135134d6b470SZach O'Keefe * information to cc->node_load[]. 13520b8f0d87SQuanfa Fu * Khugepaged will allocate hugepage from the node has the max 1353b46e756fSKirill A. Shutemov * hit record. 1354b46e756fSKirill A. Shutemov */ 13555c07ebb3SVishal Moola (Oracle) node = folio_nid(folio); 13567d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 1357b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1358b46e756fSKirill A. Shutemov goto out_unmap; 1359b46e756fSKirill A. Shutemov } 136034d6b470SZach O'Keefe cc->node_load[node]++; 13615c07ebb3SVishal Moola (Oracle) if (!folio_test_lru(folio)) { 1362b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1363b46e756fSKirill A. Shutemov goto out_unmap; 1364b46e756fSKirill A. Shutemov } 13655c07ebb3SVishal Moola (Oracle) if (folio_test_locked(folio)) { 1366b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1367b46e756fSKirill A. Shutemov goto out_unmap; 1368b46e756fSKirill A. Shutemov } 13695c07ebb3SVishal Moola (Oracle) if (!folio_test_anon(folio)) { 1370b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1371b46e756fSKirill A. Shutemov goto out_unmap; 1372b46e756fSKirill A. Shutemov } 1373b46e756fSKirill A. Shutemov 1374b46e756fSKirill A. Shutemov /* 13759445689fSKirill A. Shutemov * Check if the page has any GUP (or other external) pins. 13769445689fSKirill A. Shutemov * 1377cb67f428SHugh Dickins * Here the check may be racy: 1378cb67f428SHugh Dickins * it may see total_mapcount > refcount in some cases? 13799445689fSKirill A. Shutemov * But such case is ephemeral we could always retry collapse 13809445689fSKirill A. Shutemov * later. However it may report false positive if the page 13819445689fSKirill A. Shutemov * has excessive GUP pins (i.e. 512). Anyway the same check 13829445689fSKirill A. Shutemov * will be done again later the risk seems low. 1383b46e756fSKirill A. Shutemov */ 1384dbf85c21SVishal Moola (Oracle) if (!is_refcount_suitable(folio)) { 1385b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1386b46e756fSKirill A. Shutemov goto out_unmap; 1387b46e756fSKirill A. Shutemov } 1388d8ea7cc8SZach O'Keefe 1389d8ea7cc8SZach O'Keefe /* 1390d8ea7cc8SZach O'Keefe * If collapse was initiated by khugepaged, check that there is 1391d8ea7cc8SZach O'Keefe * enough young pte to justify collapsing the page 1392d8ea7cc8SZach O'Keefe */ 1393d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 13945c07ebb3SVishal Moola (Oracle) (pte_young(pteval) || folio_test_young(folio) || 13955c07ebb3SVishal Moola (Oracle) folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm, 1396d8ea7cc8SZach O'Keefe address))) 13970db501f7SEbru Akagunduz referenced++; 1398b46e756fSKirill A. Shutemov } 1399ffe945e6SKirill A. Shutemov if (!writable) { 1400ffe945e6SKirill A. Shutemov result = SCAN_PAGE_RO; 1401d8ea7cc8SZach O'Keefe } else if (cc->is_khugepaged && 1402d8ea7cc8SZach O'Keefe (!referenced || 1403d8ea7cc8SZach O'Keefe (unmapped && referenced < HPAGE_PMD_NR / 2))) { 1404ffe945e6SKirill A. Shutemov result = SCAN_LACK_REFERENCED_PAGE; 1405ffe945e6SKirill A. Shutemov } else { 1406b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1407b46e756fSKirill A. Shutemov } 1408b46e756fSKirill A. Shutemov out_unmap: 1409b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 141050ad2f24SZach O'Keefe if (result == SCAN_SUCCEED) { 141150ad2f24SZach O'Keefe result = collapse_huge_page(mm, address, referenced, 141250ad2f24SZach O'Keefe unmapped, cc); 1413c1e8d7c6SMichel Lespinasse /* collapse_huge_page will return with the mmap_lock released */ 141450ad2f24SZach O'Keefe *mmap_locked = false; 1415b46e756fSKirill A. Shutemov } 1416b46e756fSKirill A. Shutemov out: 14175c07ebb3SVishal Moola (Oracle) trace_mm_khugepaged_scan_pmd(mm, &folio->page, writable, referenced, 1418b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 141950ad2f24SZach O'Keefe return result; 1420b46e756fSKirill A. Shutemov } 1421b46e756fSKirill A. Shutemov 1422b26e2701SQi Zheng static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot) 1423b46e756fSKirill A. Shutemov { 1424b26e2701SQi Zheng struct mm_slot *slot = &mm_slot->slot; 1425b26e2701SQi Zheng struct mm_struct *mm = slot->mm; 1426b46e756fSKirill A. Shutemov 142735f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 1428b46e756fSKirill A. Shutemov 14295dad6048SLance Yang if (hpage_collapse_test_exit(mm)) { 1430b46e756fSKirill A. Shutemov /* free mm_slot */ 1431b26e2701SQi Zheng hash_del(&slot->hash); 1432b26e2701SQi Zheng list_del(&slot->mm_node); 1433b46e756fSKirill A. Shutemov 1434b46e756fSKirill A. Shutemov /* 1435b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1436b46e756fSKirill A. Shutemov * 1437b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1438b46e756fSKirill A. Shutemov */ 1439b46e756fSKirill A. Shutemov 1440b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1441b26e2701SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 1442b46e756fSKirill A. Shutemov mmdrop(mm); 1443b46e756fSKirill A. Shutemov } 1444b46e756fSKirill A. Shutemov } 1445b46e756fSKirill A. Shutemov 1446396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 14471043173eSHugh Dickins /* hpage must be locked, and mmap_lock must be held */ 144834488399SZach O'Keefe static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, 144934488399SZach O'Keefe pmd_t *pmdp, struct page *hpage) 145034488399SZach O'Keefe { 145134488399SZach O'Keefe struct vm_fault vmf = { 145234488399SZach O'Keefe .vma = vma, 145334488399SZach O'Keefe .address = addr, 145434488399SZach O'Keefe .flags = 0, 145534488399SZach O'Keefe .pmd = pmdp, 145634488399SZach O'Keefe }; 145734488399SZach O'Keefe 145834488399SZach O'Keefe VM_BUG_ON(!PageTransHuge(hpage)); 14591043173eSHugh Dickins mmap_assert_locked(vma->vm_mm); 146034488399SZach O'Keefe 146134488399SZach O'Keefe if (do_set_pmd(&vmf, hpage)) 146234488399SZach O'Keefe return SCAN_FAIL; 146334488399SZach O'Keefe 146434488399SZach O'Keefe get_page(hpage); 146534488399SZach O'Keefe return SCAN_SUCCEED; 146627e1f827SSong Liu } 146727e1f827SSong Liu 146827e1f827SSong Liu /** 1469336e6b53SAlex Shi * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at 1470336e6b53SAlex Shi * address haddr. 1471336e6b53SAlex Shi * 1472336e6b53SAlex Shi * @mm: process address space where collapse happens 1473336e6b53SAlex Shi * @addr: THP collapse address 147434488399SZach O'Keefe * @install_pmd: If a huge PMD should be installed 147527e1f827SSong Liu * 147627e1f827SSong Liu * This function checks whether all the PTEs in the PMD are pointing to the 147727e1f827SSong Liu * right THP. If so, retract the page table so the THP can refault in with 147834488399SZach O'Keefe * as pmd-mapped. Possibly install a huge PMD mapping the THP. 147927e1f827SSong Liu */ 148034488399SZach O'Keefe int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, 148134488399SZach O'Keefe bool install_pmd) 148227e1f827SSong Liu { 14831043173eSHugh Dickins struct mmu_notifier_range range; 14841043173eSHugh Dickins bool notified = false; 148527e1f827SSong Liu unsigned long haddr = addr & HPAGE_PMD_MASK; 148694d815b2SLiam R. Howlett struct vm_area_struct *vma = vma_lookup(mm, haddr); 148798b32d29SVishal Moola (Oracle) struct folio *folio; 148827e1f827SSong Liu pte_t *start_pte, *pte; 14891043173eSHugh Dickins pmd_t *pmd, pgt_pmd; 1490a9846049SHugh Dickins spinlock_t *pml = NULL, *ptl; 14911043173eSHugh Dickins int nr_ptes = 0, result = SCAN_FAIL; 149227e1f827SSong Liu int i; 149327e1f827SSong Liu 14941043173eSHugh Dickins mmap_assert_locked(mm); 14951043173eSHugh Dickins 14961043173eSHugh Dickins /* First check VMA found, in case page tables are being torn down */ 14971043173eSHugh Dickins if (!vma || !vma->vm_file || 14981043173eSHugh Dickins !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) 14991043173eSHugh Dickins return SCAN_VMA_CHECK; 150058ac9a89SZach O'Keefe 150134488399SZach O'Keefe /* Fast check before locking page if already PMD-mapped */ 150258ac9a89SZach O'Keefe result = find_pmd_or_thp_or_none(mm, haddr, &pmd); 150334488399SZach O'Keefe if (result == SCAN_PMD_MAPPED) 150434488399SZach O'Keefe return result; 150558ac9a89SZach O'Keefe 150627e1f827SSong Liu /* 1507a7f4e6e4SZach O'Keefe * If we are here, we've succeeded in replacing all the native pages 1508a7f4e6e4SZach O'Keefe * in the page cache with a single hugepage. If a mm were to fault-in 1509a7f4e6e4SZach O'Keefe * this memory (mapped by a suitably aligned VMA), we'd get the hugepage 1510a7f4e6e4SZach O'Keefe * and map it by a PMD, regardless of sysfs THP settings. As such, let's 1511a7f4e6e4SZach O'Keefe * analogously elide sysfs THP settings here. 151227e1f827SSong Liu */ 15133485b883SRyan Roberts if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, false, 15143485b883SRyan Roberts PMD_ORDER)) 151534488399SZach O'Keefe return SCAN_VMA_CHECK; 151627e1f827SSong Liu 1517deb4c93aSPeter Xu /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ 1518deb4c93aSPeter Xu if (userfaultfd_wp(vma)) 151934488399SZach O'Keefe return SCAN_PTE_UFFD_WP; 1520deb4c93aSPeter Xu 152198b32d29SVishal Moola (Oracle) folio = filemap_lock_folio(vma->vm_file->f_mapping, 1522119a5fc1SHugh Dickins linear_page_index(vma, haddr)); 152398b32d29SVishal Moola (Oracle) if (IS_ERR(folio)) 152434488399SZach O'Keefe return SCAN_PAGE_NULL; 1525119a5fc1SHugh Dickins 152698b32d29SVishal Moola (Oracle) if (folio_order(folio) != HPAGE_PMD_ORDER) { 152734488399SZach O'Keefe result = SCAN_PAGE_COMPOUND; 152898b32d29SVishal Moola (Oracle) goto drop_folio; 152934488399SZach O'Keefe } 1530780a4b6fSZach O'Keefe 15311043173eSHugh Dickins result = find_pmd_or_thp_or_none(mm, haddr, &pmd); 153234488399SZach O'Keefe switch (result) { 153334488399SZach O'Keefe case SCAN_SUCCEED: 153434488399SZach O'Keefe break; 153534488399SZach O'Keefe case SCAN_PMD_NONE: 153634488399SZach O'Keefe /* 15371d65b771SHugh Dickins * All pte entries have been removed and pmd cleared. 15381d65b771SHugh Dickins * Skip all the pte checks and just update the pmd mapping. 153934488399SZach O'Keefe */ 154034488399SZach O'Keefe goto maybe_install_pmd; 154134488399SZach O'Keefe default: 154298b32d29SVishal Moola (Oracle) goto drop_folio; 154334488399SZach O'Keefe } 154427e1f827SSong Liu 154534488399SZach O'Keefe result = SCAN_FAIL; 1546895f5ee4SHugh Dickins start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 15471043173eSHugh Dickins if (!start_pte) /* mmap_lock + page lock should prevent this */ 154898b32d29SVishal Moola (Oracle) goto drop_folio; 154927e1f827SSong Liu 155027e1f827SSong Liu /* step 1: check all mapped PTEs are to the right huge page */ 155127e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 155227e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 155327e1f827SSong Liu struct page *page; 1554c33c7948SRyan Roberts pte_t ptent = ptep_get(pte); 155527e1f827SSong Liu 155627e1f827SSong Liu /* empty pte, skip */ 1557c33c7948SRyan Roberts if (pte_none(ptent)) 155827e1f827SSong Liu continue; 155927e1f827SSong Liu 156027e1f827SSong Liu /* page swapped out, abort */ 1561c33c7948SRyan Roberts if (!pte_present(ptent)) { 156234488399SZach O'Keefe result = SCAN_PTE_NON_PRESENT; 156327e1f827SSong Liu goto abort; 156434488399SZach O'Keefe } 156527e1f827SSong Liu 1566c33c7948SRyan Roberts page = vm_normal_page(vma, addr, ptent); 15673218f871SAlex Sierra if (WARN_ON_ONCE(page && is_zone_device_page(page))) 15683218f871SAlex Sierra page = NULL; 156927e1f827SSong Liu /* 1570119a5fc1SHugh Dickins * Note that uprobe, debugger, or MAP_PRIVATE may change the 1571119a5fc1SHugh Dickins * page table, but the new page will not be a subpage of hpage. 157227e1f827SSong Liu */ 157398b32d29SVishal Moola (Oracle) if (folio_page(folio, i) != page) 157427e1f827SSong Liu goto abort; 157527e1f827SSong Liu } 157627e1f827SSong Liu 15771043173eSHugh Dickins pte_unmap_unlock(start_pte, ptl); 15781043173eSHugh Dickins mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 15791043173eSHugh Dickins haddr, haddr + HPAGE_PMD_SIZE); 15801043173eSHugh Dickins mmu_notifier_invalidate_range_start(&range); 15811043173eSHugh Dickins notified = true; 1582a9846049SHugh Dickins 1583a9846049SHugh Dickins /* 1584a9846049SHugh Dickins * pmd_lock covers a wider range than ptl, and (if split from mm's 1585a9846049SHugh Dickins * page_table_lock) ptl nests inside pml. The less time we hold pml, 1586a9846049SHugh Dickins * the better; but userfaultfd's mfill_atomic_pte() on a private VMA 1587a9846049SHugh Dickins * inserts a valid as-if-COWed PTE without even looking up page cache. 158898b32d29SVishal Moola (Oracle) * So page lock of folio does not protect from it, so we must not drop 1589a9846049SHugh Dickins * ptl before pgt_pmd is removed, so uffd private needs pml taken now. 1590a9846049SHugh Dickins */ 1591a9846049SHugh Dickins if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED)) 1592a9846049SHugh Dickins pml = pmd_lock(mm, pmd); 1593a9846049SHugh Dickins 1594a9846049SHugh Dickins start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl); 15951043173eSHugh Dickins if (!start_pte) /* mmap_lock + page lock should prevent this */ 15961043173eSHugh Dickins goto abort; 1597a9846049SHugh Dickins if (!pml) 1598a9846049SHugh Dickins spin_lock(ptl); 1599a9846049SHugh Dickins else if (ptl != pml) 1600a9846049SHugh Dickins spin_lock_nested(ptl, SINGLE_DEPTH_NESTING); 16011043173eSHugh Dickins 16021043173eSHugh Dickins /* step 2: clear page table and adjust rmap */ 160327e1f827SSong Liu for (i = 0, addr = haddr, pte = start_pte; 160427e1f827SSong Liu i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 160527e1f827SSong Liu struct page *page; 1606c33c7948SRyan Roberts pte_t ptent = ptep_get(pte); 160727e1f827SSong Liu 1608c33c7948SRyan Roberts if (pte_none(ptent)) 160927e1f827SSong Liu continue; 16101043173eSHugh Dickins /* 16111043173eSHugh Dickins * We dropped ptl after the first scan, to do the mmu_notifier: 161298b32d29SVishal Moola (Oracle) * page lock stops more PTEs of the folio being faulted in, but 16131043173eSHugh Dickins * does not stop write faults COWing anon copies from existing 16141043173eSHugh Dickins * PTEs; and does not stop those being swapped out or migrated. 16151043173eSHugh Dickins */ 16161043173eSHugh Dickins if (!pte_present(ptent)) { 16171043173eSHugh Dickins result = SCAN_PTE_NON_PRESENT; 16183218f871SAlex Sierra goto abort; 16191043173eSHugh Dickins } 16201043173eSHugh Dickins page = vm_normal_page(vma, addr, ptent); 162198b32d29SVishal Moola (Oracle) if (folio_page(folio, i) != page) 16221043173eSHugh Dickins goto abort; 16231043173eSHugh Dickins 16241043173eSHugh Dickins /* 16251043173eSHugh Dickins * Must clear entry, or a racing truncate may re-remove it. 16261043173eSHugh Dickins * TLB flush can be left until pmdp_collapse_flush() does it. 16271043173eSHugh Dickins * PTE dirty? Shmem page is already dirty; file is read-only. 16281043173eSHugh Dickins */ 16291043173eSHugh Dickins ptep_clear(mm, addr, pte); 163035668a43SDavid Hildenbrand folio_remove_rmap_pte(folio, page, vma); 16311043173eSHugh Dickins nr_ptes++; 163227e1f827SSong Liu } 163327e1f827SSong Liu 1634a9846049SHugh Dickins pte_unmap(start_pte); 1635a9846049SHugh Dickins if (!pml) 1636a9846049SHugh Dickins spin_unlock(ptl); 163727e1f827SSong Liu 163827e1f827SSong Liu /* step 3: set proper refcount and mm_counters. */ 16391043173eSHugh Dickins if (nr_ptes) { 164098b32d29SVishal Moola (Oracle) folio_ref_sub(folio, nr_ptes); 16416b27cc6cSKefeng Wang add_mm_counter(mm, mm_counter_file(folio), -nr_ptes); 164227e1f827SSong Liu } 164327e1f827SSong Liu 1644a9846049SHugh Dickins /* step 4: remove empty page table */ 1645a9846049SHugh Dickins if (!pml) { 16461043173eSHugh Dickins pml = pmd_lock(mm, pmd); 16471043173eSHugh Dickins if (ptl != pml) 16481043173eSHugh Dickins spin_lock_nested(ptl, SINGLE_DEPTH_NESTING); 1649a9846049SHugh Dickins } 16501043173eSHugh Dickins pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd); 16511043173eSHugh Dickins pmdp_get_lockless_sync(); 16521043173eSHugh Dickins if (ptl != pml) 16531043173eSHugh Dickins spin_unlock(ptl); 16541043173eSHugh Dickins spin_unlock(pml); 1655ab0c3f12SHugh Dickins 16561043173eSHugh Dickins mmu_notifier_invalidate_range_end(&range); 165734488399SZach O'Keefe 16581043173eSHugh Dickins mm_dec_nr_ptes(mm); 16591043173eSHugh Dickins page_table_check_pte_clear_range(mm, haddr, pgt_pmd); 16601043173eSHugh Dickins pte_free_defer(mm, pmd_pgtable(pgt_pmd)); 16618d3c106eSJann Horn 166234488399SZach O'Keefe maybe_install_pmd: 166334488399SZach O'Keefe /* step 5: install pmd entry */ 166434488399SZach O'Keefe result = install_pmd 166598b32d29SVishal Moola (Oracle) ? set_huge_pmd(vma, haddr, pmd, &folio->page) 166634488399SZach O'Keefe : SCAN_SUCCEED; 166798b32d29SVishal Moola (Oracle) goto drop_folio; 16681043173eSHugh Dickins abort: 16691043173eSHugh Dickins if (nr_ptes) { 16701043173eSHugh Dickins flush_tlb_mm(mm); 167198b32d29SVishal Moola (Oracle) folio_ref_sub(folio, nr_ptes); 16726b27cc6cSKefeng Wang add_mm_counter(mm, mm_counter_file(folio), -nr_ptes); 16731043173eSHugh Dickins } 16741043173eSHugh Dickins if (start_pte) 16751043173eSHugh Dickins pte_unmap_unlock(start_pte, ptl); 1676a9846049SHugh Dickins if (pml && pml != ptl) 1677a9846049SHugh Dickins spin_unlock(pml); 16781043173eSHugh Dickins if (notified) 16791043173eSHugh Dickins mmu_notifier_invalidate_range_end(&range); 168098b32d29SVishal Moola (Oracle) drop_folio: 168198b32d29SVishal Moola (Oracle) folio_unlock(folio); 168298b32d29SVishal Moola (Oracle) folio_put(folio); 168334488399SZach O'Keefe return result; 168427e1f827SSong Liu } 168527e1f827SSong Liu 16861d65b771SHugh Dickins static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1687f3f0e1d2SKirill A. Shutemov { 1688f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 1689f3f0e1d2SKirill A. Shutemov 16901d65b771SHugh Dickins i_mmap_lock_read(mapping); 1691f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 16921d65b771SHugh Dickins struct mmu_notifier_range range; 16931d65b771SHugh Dickins struct mm_struct *mm; 16941d65b771SHugh Dickins unsigned long addr; 16951d65b771SHugh Dickins pmd_t *pmd, pgt_pmd; 16961d65b771SHugh Dickins spinlock_t *pml; 16971d65b771SHugh Dickins spinlock_t *ptl; 16981d65b771SHugh Dickins bool skipped_uffd = false; 169934488399SZach O'Keefe 170027e1f827SSong Liu /* 170127e1f827SSong Liu * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 17021d65b771SHugh Dickins * got written to. These VMAs are likely not worth removing 17031d65b771SHugh Dickins * page tables from, as PMD-mapping is likely to be split later. 170427e1f827SSong Liu */ 17051d65b771SHugh Dickins if (READ_ONCE(vma->anon_vma)) 17061d65b771SHugh Dickins continue; 17071d65b771SHugh Dickins 1708f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 170934488399SZach O'Keefe if (addr & ~HPAGE_PMD_MASK || 17101d65b771SHugh Dickins vma->vm_end < addr + HPAGE_PMD_SIZE) 171134488399SZach O'Keefe continue; 17121d65b771SHugh Dickins 17131d65b771SHugh Dickins mm = vma->vm_mm; 17141d65b771SHugh Dickins if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED) 17151d65b771SHugh Dickins continue; 17161d65b771SHugh Dickins 17171d65b771SHugh Dickins if (hpage_collapse_test_exit(mm)) 17181d65b771SHugh Dickins continue; 17191d65b771SHugh Dickins /* 17201d65b771SHugh Dickins * When a vma is registered with uffd-wp, we cannot recycle 17211d65b771SHugh Dickins * the page table because there may be pte markers installed. 17221d65b771SHugh Dickins * Other vmas can still have the same file mapped hugely, but 17231d65b771SHugh Dickins * skip this one: it will always be mapped in small page size 17241d65b771SHugh Dickins * for uffd-wp registered ranges. 17251d65b771SHugh Dickins */ 17261d65b771SHugh Dickins if (userfaultfd_wp(vma)) 17271d65b771SHugh Dickins continue; 17281d65b771SHugh Dickins 17291d65b771SHugh Dickins /* PTEs were notified when unmapped; but now for the PMD? */ 17301d65b771SHugh Dickins mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 17311d65b771SHugh Dickins addr, addr + HPAGE_PMD_SIZE); 17321d65b771SHugh Dickins mmu_notifier_invalidate_range_start(&range); 17331d65b771SHugh Dickins 17341d65b771SHugh Dickins pml = pmd_lock(mm, pmd); 17351d65b771SHugh Dickins ptl = pte_lockptr(mm, pmd); 17361d65b771SHugh Dickins if (ptl != pml) 17371d65b771SHugh Dickins spin_lock_nested(ptl, SINGLE_DEPTH_NESTING); 17381d65b771SHugh Dickins 17391d65b771SHugh Dickins /* 17401d65b771SHugh Dickins * Huge page lock is still held, so normally the page table 17411d65b771SHugh Dickins * must remain empty; and we have already skipped anon_vma 17421d65b771SHugh Dickins * and userfaultfd_wp() vmas. But since the mmap_lock is not 17431d65b771SHugh Dickins * held, it is still possible for a racing userfaultfd_ioctl() 17441d65b771SHugh Dickins * to have inserted ptes or markers. Now that we hold ptlock, 17451d65b771SHugh Dickins * repeating the anon_vma check protects from one category, 17461d65b771SHugh Dickins * and repeating the userfaultfd_wp() check from another. 17471d65b771SHugh Dickins */ 17481d65b771SHugh Dickins if (unlikely(vma->anon_vma || userfaultfd_wp(vma))) { 17491d65b771SHugh Dickins skipped_uffd = true; 17501d65b771SHugh Dickins } else { 17511d65b771SHugh Dickins pgt_pmd = pmdp_collapse_flush(vma, addr, pmd); 17521d65b771SHugh Dickins pmdp_get_lockless_sync(); 175334488399SZach O'Keefe } 17541d65b771SHugh Dickins 17551d65b771SHugh Dickins if (ptl != pml) 17561d65b771SHugh Dickins spin_unlock(ptl); 17571d65b771SHugh Dickins spin_unlock(pml); 17581d65b771SHugh Dickins 17591d65b771SHugh Dickins mmu_notifier_invalidate_range_end(&range); 17601d65b771SHugh Dickins 17611d65b771SHugh Dickins if (!skipped_uffd) { 17621d65b771SHugh Dickins mm_dec_nr_ptes(mm); 17631d65b771SHugh Dickins page_table_check_pte_clear_range(mm, addr, pgt_pmd); 17641d65b771SHugh Dickins pte_free_defer(mm, pmd_pgtable(pgt_pmd)); 1765f3f0e1d2SKirill A. Shutemov } 17661d65b771SHugh Dickins } 17671d65b771SHugh Dickins i_mmap_unlock_read(mapping); 1768f3f0e1d2SKirill A. Shutemov } 1769f3f0e1d2SKirill A. Shutemov 1770f3f0e1d2SKirill A. Shutemov /** 177199cb0dbdSSong Liu * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1772f3f0e1d2SKirill A. Shutemov * 1773336e6b53SAlex Shi * @mm: process address space where collapse happens 177434488399SZach O'Keefe * @addr: virtual collapse start address 1775336e6b53SAlex Shi * @file: file that collapse on 1776336e6b53SAlex Shi * @start: collapse start address 17779710a78aSZach O'Keefe * @cc: collapse context and scratchpad 1778336e6b53SAlex Shi * 1779f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 178087c460a0SHugh Dickins * - allocate and lock a new huge page; 1781a2e17cc2SDavid Stevens * - scan page cache, locking old pages 178299cb0dbdSSong Liu * + swap/gup in pages if necessary; 1783a2e17cc2SDavid Stevens * - copy data to new page 1784a2e17cc2SDavid Stevens * - handle shmem holes 1785a2e17cc2SDavid Stevens * + re-validate that holes weren't filled by someone else 1786a2e17cc2SDavid Stevens * + check for userfaultfd 1787ac492b9cSDavid Stevens * - finalize updates to the page cache; 178877da9389SMatthew Wilcox * - if replacing succeeds: 178987c460a0SHugh Dickins * + unlock huge page; 1790a2e17cc2SDavid Stevens * + free old pages; 1791f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1792a2e17cc2SDavid Stevens * + unlock old pages 179387c460a0SHugh Dickins * + unlock and free huge page; 1794f3f0e1d2SKirill A. Shutemov */ 179534488399SZach O'Keefe static int collapse_file(struct mm_struct *mm, unsigned long addr, 1796579c571eSSong Liu struct file *file, pgoff_t start, 179734488399SZach O'Keefe struct collapse_control *cc) 1798f3f0e1d2SKirill A. Shutemov { 1799579c571eSSong Liu struct address_space *mapping = file->f_mapping; 180050ad2f24SZach O'Keefe struct page *hpage; 180112904d95SJiaqi Yan struct page *page; 180212904d95SJiaqi Yan struct page *tmp; 180312904d95SJiaqi Yan struct folio *folio; 18044c9473e8SGautam Menghani pgoff_t index = 0, end = start + HPAGE_PMD_NR; 1805f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 180677da9389SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1807f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 180899cb0dbdSSong Liu bool is_shmem = shmem_file(file); 18094c9473e8SGautam Menghani int nr = 0; 1810f3f0e1d2SKirill A. Shutemov 181199cb0dbdSSong Liu VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1812f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1813f3f0e1d2SKirill A. Shutemov 181450ad2f24SZach O'Keefe result = alloc_charge_hpage(&hpage, mm, cc); 18159710a78aSZach O'Keefe if (result != SCAN_SUCCEED) 1816f3f0e1d2SKirill A. Shutemov goto out; 1817f3f0e1d2SKirill A. Shutemov 1818cae106ddSDavid Stevens __SetPageLocked(hpage); 1819cae106ddSDavid Stevens if (is_shmem) 1820cae106ddSDavid Stevens __SetPageSwapBacked(hpage); 1821cae106ddSDavid Stevens hpage->index = start; 1822cae106ddSDavid Stevens hpage->mapping = mapping; 1823cae106ddSDavid Stevens 18246b24ca4aSMatthew Wilcox (Oracle) /* 18256b24ca4aSMatthew Wilcox (Oracle) * Ensure we have slots for all the pages in the range. This is 18266b24ca4aSMatthew Wilcox (Oracle) * almost certainly a no-op because most of the pages must be present 18276b24ca4aSMatthew Wilcox (Oracle) */ 182895feeabbSHugh Dickins do { 182995feeabbSHugh Dickins xas_lock_irq(&xas); 183095feeabbSHugh Dickins xas_create_range(&xas); 183195feeabbSHugh Dickins if (!xas_error(&xas)) 183295feeabbSHugh Dickins break; 183395feeabbSHugh Dickins xas_unlock_irq(&xas); 183495feeabbSHugh Dickins if (!xas_nomem(&xas, GFP_KERNEL)) { 183595feeabbSHugh Dickins result = SCAN_FAIL; 1836cae106ddSDavid Stevens goto rollback; 183795feeabbSHugh Dickins } 183895feeabbSHugh Dickins } while (1); 183995feeabbSHugh Dickins 184077da9389SMatthew Wilcox for (index = start; index < end; index++) { 1841e8c716bcSHugh Dickins xas_set(&xas, index); 1842e8c716bcSHugh Dickins page = xas_load(&xas); 184377da9389SMatthew Wilcox 184477da9389SMatthew Wilcox VM_BUG_ON(index != xas.xa_index); 184599cb0dbdSSong Liu if (is_shmem) { 184677da9389SMatthew Wilcox if (!page) { 1847701270faSHugh Dickins /* 184899cb0dbdSSong Liu * Stop if extent has been truncated or 184999cb0dbdSSong Liu * hole-punched, and is now completely 185099cb0dbdSSong Liu * empty. 1851701270faSHugh Dickins */ 1852701270faSHugh Dickins if (index == start) { 1853701270faSHugh Dickins if (!xas_next_entry(&xas, end - 1)) { 1854701270faSHugh Dickins result = SCAN_TRUNCATED; 1855042a3082SHugh Dickins goto xa_locked; 1856701270faSHugh Dickins } 1857701270faSHugh Dickins } 185877da9389SMatthew Wilcox nr_none++; 185977da9389SMatthew Wilcox continue; 1860f3f0e1d2SKirill A. Shutemov } 1861f3f0e1d2SKirill A. Shutemov 18623159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 186377da9389SMatthew Wilcox xas_unlock_irq(&xas); 1864f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 18657459c149SMatthew Wilcox (Oracle) if (shmem_get_folio(mapping->host, index, 18667459c149SMatthew Wilcox (Oracle) &folio, SGP_NOALLOC)) { 1867f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 186877da9389SMatthew Wilcox goto xa_unlocked; 1869f3f0e1d2SKirill A. Shutemov } 18701fec6890SMatthew Wilcox (Oracle) /* drain lru cache to help isolate_lru_page() */ 1871efa3d814SDavid Stevens lru_add_drain(); 18727459c149SMatthew Wilcox (Oracle) page = folio_file_page(folio, index); 1873f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1874f3f0e1d2SKirill A. Shutemov get_page(page); 1875042a3082SHugh Dickins xas_unlock_irq(&xas); 1876f3f0e1d2SKirill A. Shutemov } else { 1877f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1878042a3082SHugh Dickins goto xa_locked; 1879f3f0e1d2SKirill A. Shutemov } 188099cb0dbdSSong Liu } else { /* !is_shmem */ 188199cb0dbdSSong Liu if (!page || xa_is_value(page)) { 188299cb0dbdSSong Liu xas_unlock_irq(&xas); 188399cb0dbdSSong Liu page_cache_sync_readahead(mapping, &file->f_ra, 188499cb0dbdSSong Liu file, index, 1885e5a59d30SDavid Howells end - index); 18861fec6890SMatthew Wilcox (Oracle) /* drain lru cache to help isolate_lru_page() */ 188799cb0dbdSSong Liu lru_add_drain(); 188899cb0dbdSSong Liu page = find_lock_page(mapping, index); 188999cb0dbdSSong Liu if (unlikely(page == NULL)) { 189099cb0dbdSSong Liu result = SCAN_FAIL; 189199cb0dbdSSong Liu goto xa_unlocked; 189299cb0dbdSSong Liu } 189375f36069SSong Liu } else if (PageDirty(page)) { 189475f36069SSong Liu /* 189575f36069SSong Liu * khugepaged only works on read-only fd, 189675f36069SSong Liu * so this page is dirty because it hasn't 189775f36069SSong Liu * been flushed since first write. There 189875f36069SSong Liu * won't be new dirty pages. 189975f36069SSong Liu * 190075f36069SSong Liu * Trigger async flush here and hope the 190175f36069SSong Liu * writeback is done when khugepaged 190275f36069SSong Liu * revisits this page. 190375f36069SSong Liu * 190475f36069SSong Liu * This is a one-off situation. We are not 190575f36069SSong Liu * forcing writeback in loop. 190675f36069SSong Liu */ 190775f36069SSong Liu xas_unlock_irq(&xas); 190875f36069SSong Liu filemap_flush(mapping); 190975f36069SSong Liu result = SCAN_FAIL; 191075f36069SSong Liu goto xa_unlocked; 191174c42e1bSRongwei Wang } else if (PageWriteback(page)) { 191274c42e1bSRongwei Wang xas_unlock_irq(&xas); 191374c42e1bSRongwei Wang result = SCAN_FAIL; 191474c42e1bSRongwei Wang goto xa_unlocked; 191599cb0dbdSSong Liu } else if (trylock_page(page)) { 191699cb0dbdSSong Liu get_page(page); 191799cb0dbdSSong Liu xas_unlock_irq(&xas); 191899cb0dbdSSong Liu } else { 191999cb0dbdSSong Liu result = SCAN_PAGE_LOCK; 192099cb0dbdSSong Liu goto xa_locked; 192199cb0dbdSSong Liu } 192299cb0dbdSSong Liu } 1923f3f0e1d2SKirill A. Shutemov 1924f3f0e1d2SKirill A. Shutemov /* 1925b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1926f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1927f3f0e1d2SKirill A. Shutemov */ 1928f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 19294655e5e5SSong Liu 19304655e5e5SSong Liu /* make sure the page is up to date */ 19314655e5e5SSong Liu if (unlikely(!PageUptodate(page))) { 19324655e5e5SSong Liu result = SCAN_FAIL; 19334655e5e5SSong Liu goto out_unlock; 19344655e5e5SSong Liu } 193506a5e126SHugh Dickins 193606a5e126SHugh Dickins /* 193706a5e126SHugh Dickins * If file was truncated then extended, or hole-punched, before 193806a5e126SHugh Dickins * we locked the first page, then a THP might be there already. 193958ac9a89SZach O'Keefe * This will be discovered on the first iteration. 194006a5e126SHugh Dickins */ 194106a5e126SHugh Dickins if (PageTransCompound(page)) { 194258ac9a89SZach O'Keefe struct page *head = compound_head(page); 194358ac9a89SZach O'Keefe 194458ac9a89SZach O'Keefe result = compound_order(head) == HPAGE_PMD_ORDER && 194558ac9a89SZach O'Keefe head->index == start 194658ac9a89SZach O'Keefe /* Maybe PMD-mapped */ 194758ac9a89SZach O'Keefe ? SCAN_PTE_MAPPED_HUGEPAGE 194858ac9a89SZach O'Keefe : SCAN_PAGE_COMPOUND; 194906a5e126SHugh Dickins goto out_unlock; 195006a5e126SHugh Dickins } 1951f3f0e1d2SKirill A. Shutemov 195264ab3195SVishal Moola (Oracle) folio = page_folio(page); 195364ab3195SVishal Moola (Oracle) 195464ab3195SVishal Moola (Oracle) if (folio_mapping(folio) != mapping) { 1955f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1956f3f0e1d2SKirill A. Shutemov goto out_unlock; 1957f3f0e1d2SKirill A. Shutemov } 1958f3f0e1d2SKirill A. Shutemov 195964ab3195SVishal Moola (Oracle) if (!is_shmem && (folio_test_dirty(folio) || 196064ab3195SVishal Moola (Oracle) folio_test_writeback(folio))) { 19614655e5e5SSong Liu /* 19624655e5e5SSong Liu * khugepaged only works on read-only fd, so this 19634655e5e5SSong Liu * page is dirty because it hasn't been flushed 19644655e5e5SSong Liu * since first write. 19654655e5e5SSong Liu */ 19664655e5e5SSong Liu result = SCAN_FAIL; 19674655e5e5SSong Liu goto out_unlock; 19684655e5e5SSong Liu } 19694655e5e5SSong Liu 1970be2d5756SBaolin Wang if (!folio_isolate_lru(folio)) { 1971f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1972042a3082SHugh Dickins goto out_unlock; 1973f3f0e1d2SKirill A. Shutemov } 1974f3f0e1d2SKirill A. Shutemov 19750201ebf2SDavid Howells if (!filemap_release_folio(folio, GFP_KERNEL)) { 197699cb0dbdSSong Liu result = SCAN_PAGE_HAS_PRIVATE; 197764ab3195SVishal Moola (Oracle) folio_putback_lru(folio); 197899cb0dbdSSong Liu goto out_unlock; 197999cb0dbdSSong Liu } 198099cb0dbdSSong Liu 198164ab3195SVishal Moola (Oracle) if (folio_mapped(folio)) 198264ab3195SVishal Moola (Oracle) try_to_unmap(folio, 1983869f7ee6SMatthew Wilcox (Oracle) TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); 1984f3f0e1d2SKirill A. Shutemov 198577da9389SMatthew Wilcox xas_lock_irq(&xas); 1986f3f0e1d2SKirill A. Shutemov 1987e8c716bcSHugh Dickins VM_BUG_ON_PAGE(page != xa_load(xas.xa, index), page); 1988f3f0e1d2SKirill A. Shutemov 1989f3f0e1d2SKirill A. Shutemov /* 1990a2e17cc2SDavid Stevens * We control three references to the page: 1991f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 199277da9389SMatthew Wilcox * - one reference from page cache; 1993f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1994a2e17cc2SDavid Stevens * If those are the only references, then any new usage of the 1995a2e17cc2SDavid Stevens * page will have to fetch it from the page cache. That requires 1996a2e17cc2SDavid Stevens * locking the page to handle truncate, so any new usage will be 1997a2e17cc2SDavid Stevens * blocked until we unlock page after collapse/during rollback. 1998f3f0e1d2SKirill A. Shutemov */ 1999a2e17cc2SDavid Stevens if (page_count(page) != 3) { 2000f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 2001042a3082SHugh Dickins xas_unlock_irq(&xas); 2002042a3082SHugh Dickins putback_lru_page(page); 2003042a3082SHugh Dickins goto out_unlock; 2004f3f0e1d2SKirill A. Shutemov } 2005f3f0e1d2SKirill A. Shutemov 2006f3f0e1d2SKirill A. Shutemov /* 2007a2e17cc2SDavid Stevens * Accumulate the pages that are being collapsed. 2008f3f0e1d2SKirill A. Shutemov */ 2009f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 2010f3f0e1d2SKirill A. Shutemov continue; 2011f3f0e1d2SKirill A. Shutemov out_unlock: 2012f3f0e1d2SKirill A. Shutemov unlock_page(page); 2013f3f0e1d2SKirill A. Shutemov put_page(page); 2014042a3082SHugh Dickins goto xa_unlocked; 2015f3f0e1d2SKirill A. Shutemov } 2016f3f0e1d2SKirill A. Shutemov 201712904d95SJiaqi Yan if (!is_shmem) { 201809d91cdaSSong Liu filemap_nr_thps_inc(mapping); 2019eb6ecbedSCollin Fijalkovich /* 2020eb6ecbedSCollin Fijalkovich * Paired with smp_mb() in do_dentry_open() to ensure 2021eb6ecbedSCollin Fijalkovich * i_writecount is up to date and the update to nr_thps is 2022eb6ecbedSCollin Fijalkovich * visible. Ensures the page cache will be truncated if the 2023eb6ecbedSCollin Fijalkovich * file is opened writable. 2024eb6ecbedSCollin Fijalkovich */ 2025eb6ecbedSCollin Fijalkovich smp_mb(); 2026eb6ecbedSCollin Fijalkovich if (inode_is_open_for_write(mapping->host)) { 2027eb6ecbedSCollin Fijalkovich result = SCAN_FAIL; 2028eb6ecbedSCollin Fijalkovich filemap_nr_thps_dec(mapping); 2029eb6ecbedSCollin Fijalkovich } 203009d91cdaSSong Liu } 203199cb0dbdSSong Liu 2032042a3082SHugh Dickins xa_locked: 2033042a3082SHugh Dickins xas_unlock_irq(&xas); 203477da9389SMatthew Wilcox xa_unlocked: 2035042a3082SHugh Dickins 20366d9df8a5SHugh Dickins /* 20376d9df8a5SHugh Dickins * If collapse is successful, flush must be done now before copying. 20386d9df8a5SHugh Dickins * If collapse is unsuccessful, does flush actually need to be done? 20396d9df8a5SHugh Dickins * Do it anyway, to clear the state. 20406d9df8a5SHugh Dickins */ 20416d9df8a5SHugh Dickins try_to_unmap_flush(); 20426d9df8a5SHugh Dickins 2043509f0069SHugh Dickins if (result == SCAN_SUCCEED && nr_none && 2044509f0069SHugh Dickins !shmem_charge(mapping->host, nr_none)) 2045509f0069SHugh Dickins result = SCAN_FAIL; 2046509f0069SHugh Dickins if (result != SCAN_SUCCEED) { 2047509f0069SHugh Dickins nr_none = 0; 2048cae106ddSDavid Stevens goto rollback; 2049509f0069SHugh Dickins } 2050cae106ddSDavid Stevens 2051f3f0e1d2SKirill A. Shutemov /* 2052a2e17cc2SDavid Stevens * The old pages are locked, so they won't change anymore. 2053f3f0e1d2SKirill A. Shutemov */ 20542af8ff29SHugh Dickins index = start; 205512904d95SJiaqi Yan list_for_each_entry(page, &pagelist, lru) { 20562af8ff29SHugh Dickins while (index < page->index) { 205750ad2f24SZach O'Keefe clear_highpage(hpage + (index % HPAGE_PMD_NR)); 20582af8ff29SHugh Dickins index++; 20592af8ff29SHugh Dickins } 2060cae106ddSDavid Stevens if (copy_mc_highpage(hpage + (page->index % HPAGE_PMD_NR), page) > 0) { 206112904d95SJiaqi Yan result = SCAN_COPY_MC; 2062cae106ddSDavid Stevens goto rollback; 206312904d95SJiaqi Yan } 206412904d95SJiaqi Yan index++; 206512904d95SJiaqi Yan } 2066cae106ddSDavid Stevens while (index < end) { 206712904d95SJiaqi Yan clear_highpage(hpage + (index % HPAGE_PMD_NR)); 206812904d95SJiaqi Yan index++; 206912904d95SJiaqi Yan } 207012904d95SJiaqi Yan 2071ac492b9cSDavid Stevens if (nr_none) { 2072ac492b9cSDavid Stevens struct vm_area_struct *vma; 2073ac492b9cSDavid Stevens int nr_none_check = 0; 2074ac492b9cSDavid Stevens 2075ac492b9cSDavid Stevens i_mmap_lock_read(mapping); 2076ac492b9cSDavid Stevens xas_lock_irq(&xas); 2077ac492b9cSDavid Stevens 2078ac492b9cSDavid Stevens xas_set(&xas, start); 2079ac492b9cSDavid Stevens for (index = start; index < end; index++) { 2080ac492b9cSDavid Stevens if (!xas_next(&xas)) { 2081ac492b9cSDavid Stevens xas_store(&xas, XA_RETRY_ENTRY); 2082ac492b9cSDavid Stevens if (xas_error(&xas)) { 2083ac492b9cSDavid Stevens result = SCAN_STORE_FAILED; 2084ac492b9cSDavid Stevens goto immap_locked; 2085ac492b9cSDavid Stevens } 2086ac492b9cSDavid Stevens nr_none_check++; 2087ac492b9cSDavid Stevens } 2088ac492b9cSDavid Stevens } 2089ac492b9cSDavid Stevens 2090ac492b9cSDavid Stevens if (nr_none != nr_none_check) { 2091ac492b9cSDavid Stevens result = SCAN_PAGE_FILLED; 2092ac492b9cSDavid Stevens goto immap_locked; 2093ac492b9cSDavid Stevens } 2094ac492b9cSDavid Stevens 209512904d95SJiaqi Yan /* 2096ac492b9cSDavid Stevens * If userspace observed a missing page in a VMA with a MODE_MISSING 2097ac492b9cSDavid Stevens * userfaultfd, then it might expect a UFFD_EVENT_PAGEFAULT for that 2098ac492b9cSDavid Stevens * page. If so, we need to roll back to avoid suppressing such an 2099ac492b9cSDavid Stevens * event. Since wp/minor userfaultfds don't give userspace any 2100ac492b9cSDavid Stevens * guarantees that the kernel doesn't fill a missing page with a zero 2101ac492b9cSDavid Stevens * page, so they don't matter here. 2102ac492b9cSDavid Stevens * 2103ac492b9cSDavid Stevens * Any userfaultfds registered after this point will not be able to 2104ac492b9cSDavid Stevens * observe any missing pages due to the previously inserted retry 2105ac492b9cSDavid Stevens * entries. 210612904d95SJiaqi Yan */ 2107ac492b9cSDavid Stevens vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) { 2108ac492b9cSDavid Stevens if (userfaultfd_missing(vma)) { 2109ac492b9cSDavid Stevens result = SCAN_EXCEED_NONE_PTE; 2110ac492b9cSDavid Stevens goto immap_locked; 2111ac492b9cSDavid Stevens } 2112ac492b9cSDavid Stevens } 2113ac492b9cSDavid Stevens 2114ac492b9cSDavid Stevens immap_locked: 2115ac492b9cSDavid Stevens i_mmap_unlock_read(mapping); 2116ac492b9cSDavid Stevens if (result != SCAN_SUCCEED) { 2117ac492b9cSDavid Stevens xas_set(&xas, start); 2118ac492b9cSDavid Stevens for (index = start; index < end; index++) { 2119ac492b9cSDavid Stevens if (xas_next(&xas) == XA_RETRY_ENTRY) 2120ac492b9cSDavid Stevens xas_store(&xas, NULL); 2121ac492b9cSDavid Stevens } 2122ac492b9cSDavid Stevens 2123ac492b9cSDavid Stevens xas_unlock_irq(&xas); 2124ac492b9cSDavid Stevens goto rollback; 2125ac492b9cSDavid Stevens } 2126ac492b9cSDavid Stevens } else { 2127ac492b9cSDavid Stevens xas_lock_irq(&xas); 21282af8ff29SHugh Dickins } 212912904d95SJiaqi Yan 2130b54d60b1SMatthew Wilcox (Oracle) folio = page_folio(hpage); 2131b54d60b1SMatthew Wilcox (Oracle) nr = folio_nr_pages(folio); 213212904d95SJiaqi Yan if (is_shmem) 2133b54d60b1SMatthew Wilcox (Oracle) __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr); 213412904d95SJiaqi Yan else 2135b54d60b1SMatthew Wilcox (Oracle) __lruvec_stat_mod_folio(folio, NR_FILE_THPS, nr); 213612904d95SJiaqi Yan 213712904d95SJiaqi Yan if (nr_none) { 2138b54d60b1SMatthew Wilcox (Oracle) __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr_none); 213912904d95SJiaqi Yan /* nr_none is always 0 for non-shmem. */ 2140b54d60b1SMatthew Wilcox (Oracle) __lruvec_stat_mod_folio(folio, NR_SHMEM, nr_none); 2141f3f0e1d2SKirill A. Shutemov } 2142f3f0e1d2SKirill A. Shutemov 2143a2e17cc2SDavid Stevens /* 2144a2e17cc2SDavid Stevens * Mark hpage as uptodate before inserting it into the page cache so 2145a2e17cc2SDavid Stevens * that it isn't mistaken for an fallocated but unwritten page. 2146a2e17cc2SDavid Stevens */ 2147284a344eSVishal Moola (Oracle) folio_mark_uptodate(folio); 2148284a344eSVishal Moola (Oracle) folio_ref_add(folio, HPAGE_PMD_NR - 1); 2149284a344eSVishal Moola (Oracle) 21506058eaecSJohannes Weiner if (is_shmem) 2151284a344eSVishal Moola (Oracle) folio_mark_dirty(folio); 2152284a344eSVishal Moola (Oracle) folio_add_lru(folio); 2153f3f0e1d2SKirill A. Shutemov 2154a2e17cc2SDavid Stevens /* Join all the small entries into a single multi-index entry. */ 2155a2e17cc2SDavid Stevens xas_set_order(&xas, start, HPAGE_PMD_ORDER); 2156b54d60b1SMatthew Wilcox (Oracle) xas_store(&xas, folio); 21570175ab61SHugh Dickins WARN_ON_ONCE(xas_error(&xas)); 2158a2e17cc2SDavid Stevens xas_unlock_irq(&xas); 2159a2e17cc2SDavid Stevens 2160042a3082SHugh Dickins /* 2161042a3082SHugh Dickins * Remove pte page tables, so we can re-fault the page as huge. 21621d65b771SHugh Dickins * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp(). 2163042a3082SHugh Dickins */ 21641d65b771SHugh Dickins retract_page_tables(mapping, start); 21651d65b771SHugh Dickins if (cc && !cc->is_khugepaged) 21661d65b771SHugh Dickins result = SCAN_PTE_MAPPED_HUGEPAGE; 2167b54d60b1SMatthew Wilcox (Oracle) folio_unlock(folio); 2168ac492b9cSDavid Stevens 2169ac492b9cSDavid Stevens /* 2170ac492b9cSDavid Stevens * The collapse has succeeded, so free the old pages. 2171ac492b9cSDavid Stevens */ 2172ac492b9cSDavid Stevens list_for_each_entry_safe(page, tmp, &pagelist, lru) { 2173ac492b9cSDavid Stevens list_del(&page->lru); 2174ac492b9cSDavid Stevens page->mapping = NULL; 2175ac492b9cSDavid Stevens ClearPageActive(page); 2176ac492b9cSDavid Stevens ClearPageUnevictable(page); 2177ac492b9cSDavid Stevens unlock_page(page); 2178a2e17cc2SDavid Stevens folio_put_refs(page_folio(page), 3); 2179ac492b9cSDavid Stevens } 2180ac492b9cSDavid Stevens 2181cae106ddSDavid Stevens goto out; 2182cae106ddSDavid Stevens 2183cae106ddSDavid Stevens rollback: 218477da9389SMatthew Wilcox /* Something went wrong: roll back page cache changes */ 21852f55f070SMiaohe Lin if (nr_none) { 2186a2e17cc2SDavid Stevens xas_lock_irq(&xas); 2187aaa52e34SHugh Dickins mapping->nrpages -= nr_none; 218877da9389SMatthew Wilcox xas_unlock_irq(&xas); 2189509f0069SHugh Dickins shmem_uncharge(mapping->host, nr_none); 2190a2e17cc2SDavid Stevens } 2191a2e17cc2SDavid Stevens 2192a2e17cc2SDavid Stevens list_for_each_entry_safe(page, tmp, &pagelist, lru) { 2193a2e17cc2SDavid Stevens list_del(&page->lru); 2194f3f0e1d2SKirill A. Shutemov unlock_page(page); 2195042a3082SHugh Dickins putback_lru_page(page); 2196a2e17cc2SDavid Stevens put_page(page); 2197f3f0e1d2SKirill A. Shutemov } 219812904d95SJiaqi Yan /* 219912904d95SJiaqi Yan * Undo the updates of filemap_nr_thps_inc for non-SHMEM 220012904d95SJiaqi Yan * file only. This undo is not needed unless failure is 220112904d95SJiaqi Yan * due to SCAN_COPY_MC. 220212904d95SJiaqi Yan */ 220312904d95SJiaqi Yan if (!is_shmem && result == SCAN_COPY_MC) { 220412904d95SJiaqi Yan filemap_nr_thps_dec(mapping); 220512904d95SJiaqi Yan /* 220612904d95SJiaqi Yan * Paired with smp_mb() in do_dentry_open() to 220712904d95SJiaqi Yan * ensure the update to nr_thps is visible. 220812904d95SJiaqi Yan */ 220912904d95SJiaqi Yan smp_mb(); 221012904d95SJiaqi Yan } 221112904d95SJiaqi Yan 221250ad2f24SZach O'Keefe hpage->mapping = NULL; 2213042a3082SHugh Dickins 221450ad2f24SZach O'Keefe unlock_page(hpage); 2215cae106ddSDavid Stevens put_page(hpage); 2216f3f0e1d2SKirill A. Shutemov out: 2217f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 22184c9473e8SGautam Menghani trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result); 221950ad2f24SZach O'Keefe return result; 2220f3f0e1d2SKirill A. Shutemov } 2221f3f0e1d2SKirill A. Shutemov 222234488399SZach O'Keefe static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, 222334488399SZach O'Keefe struct file *file, pgoff_t start, 222434488399SZach O'Keefe struct collapse_control *cc) 2225f3f0e1d2SKirill A. Shutemov { 2226f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 2227579c571eSSong Liu struct address_space *mapping = file->f_mapping; 222885b392dbSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, start); 2229f3f0e1d2SKirill A. Shutemov int present, swap; 2230f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 2231f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 2232f3f0e1d2SKirill A. Shutemov 2233f3f0e1d2SKirill A. Shutemov present = 0; 2234f3f0e1d2SKirill A. Shutemov swap = 0; 223534d6b470SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 2236e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 2237f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 223885b392dbSMatthew Wilcox xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 223985b392dbSMatthew Wilcox if (xas_retry(&xas, page)) 2240f3f0e1d2SKirill A. Shutemov continue; 2241f3f0e1d2SKirill A. Shutemov 224285b392dbSMatthew Wilcox if (xa_is_value(page)) { 2243d8ea7cc8SZach O'Keefe ++swap; 2244d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 2245d8ea7cc8SZach O'Keefe swap > khugepaged_max_ptes_swap) { 2246f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 2247e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); 2248f3f0e1d2SKirill A. Shutemov break; 2249f3f0e1d2SKirill A. Shutemov } 2250f3f0e1d2SKirill A. Shutemov continue; 2251f3f0e1d2SKirill A. Shutemov } 2252f3f0e1d2SKirill A. Shutemov 22536b24ca4aSMatthew Wilcox (Oracle) /* 225458ac9a89SZach O'Keefe * TODO: khugepaged should compact smaller compound pages 22556b24ca4aSMatthew Wilcox (Oracle) * into a PMD sized page 22566b24ca4aSMatthew Wilcox (Oracle) */ 2257f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 225858ac9a89SZach O'Keefe struct page *head = compound_head(page); 225958ac9a89SZach O'Keefe 226058ac9a89SZach O'Keefe result = compound_order(head) == HPAGE_PMD_ORDER && 226158ac9a89SZach O'Keefe head->index == start 226258ac9a89SZach O'Keefe /* Maybe PMD-mapped */ 226358ac9a89SZach O'Keefe ? SCAN_PTE_MAPPED_HUGEPAGE 226458ac9a89SZach O'Keefe : SCAN_PAGE_COMPOUND; 226558ac9a89SZach O'Keefe /* 226658ac9a89SZach O'Keefe * For SCAN_PTE_MAPPED_HUGEPAGE, further processing 226758ac9a89SZach O'Keefe * by the caller won't touch the page cache, and so 226858ac9a89SZach O'Keefe * it's safe to skip LRU and refcount checks before 226958ac9a89SZach O'Keefe * returning. 227058ac9a89SZach O'Keefe */ 2271f3f0e1d2SKirill A. Shutemov break; 2272f3f0e1d2SKirill A. Shutemov } 2273f3f0e1d2SKirill A. Shutemov 2274f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 22757d2c4385SZach O'Keefe if (hpage_collapse_scan_abort(node, cc)) { 2276f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 2277f3f0e1d2SKirill A. Shutemov break; 2278f3f0e1d2SKirill A. Shutemov } 227934d6b470SZach O'Keefe cc->node_load[node]++; 2280f3f0e1d2SKirill A. Shutemov 2281f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 2282f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 2283f3f0e1d2SKirill A. Shutemov break; 2284f3f0e1d2SKirill A. Shutemov } 2285f3f0e1d2SKirill A. Shutemov 228699cb0dbdSSong Liu if (page_count(page) != 228799cb0dbdSSong Liu 1 + page_mapcount(page) + page_has_private(page)) { 2288f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 2289f3f0e1d2SKirill A. Shutemov break; 2290f3f0e1d2SKirill A. Shutemov } 2291f3f0e1d2SKirill A. Shutemov 2292f3f0e1d2SKirill A. Shutemov /* 2293f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 2294f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 2295f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 2296f3f0e1d2SKirill A. Shutemov */ 2297f3f0e1d2SKirill A. Shutemov 2298f3f0e1d2SKirill A. Shutemov present++; 2299f3f0e1d2SKirill A. Shutemov 2300f3f0e1d2SKirill A. Shutemov if (need_resched()) { 230185b392dbSMatthew Wilcox xas_pause(&xas); 2302f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 2303f3f0e1d2SKirill A. Shutemov } 2304f3f0e1d2SKirill A. Shutemov } 2305f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 2306f3f0e1d2SKirill A. Shutemov 2307f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 2308d8ea7cc8SZach O'Keefe if (cc->is_khugepaged && 2309d8ea7cc8SZach O'Keefe present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 2310f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 2311e9ea874aSYang Yang count_vm_event(THP_SCAN_EXCEED_NONE_PTE); 2312f3f0e1d2SKirill A. Shutemov } else { 231334488399SZach O'Keefe result = collapse_file(mm, addr, file, start, cc); 2314f3f0e1d2SKirill A. Shutemov } 2315f3f0e1d2SKirill A. Shutemov } 2316f3f0e1d2SKirill A. Shutemov 2317045634ffSGautam Menghani trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result); 231850ad2f24SZach O'Keefe return result; 2319f3f0e1d2SKirill A. Shutemov } 2320f3f0e1d2SKirill A. Shutemov #else 232134488399SZach O'Keefe static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, 232234488399SZach O'Keefe struct file *file, pgoff_t start, 232334488399SZach O'Keefe struct collapse_control *cc) 2324f3f0e1d2SKirill A. Shutemov { 2325f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 2326f3f0e1d2SKirill A. Shutemov } 2327f3f0e1d2SKirill A. Shutemov #endif 2328f3f0e1d2SKirill A. Shutemov 232950ad2f24SZach O'Keefe static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, 233034d6b470SZach O'Keefe struct collapse_control *cc) 2331b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 2332b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 2333b46e756fSKirill A. Shutemov { 233468540502SMatthew Wilcox (Oracle) struct vma_iterator vmi; 2335b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 2336b26e2701SQi Zheng struct mm_slot *slot; 2337b46e756fSKirill A. Shutemov struct mm_struct *mm; 2338b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 2339b46e756fSKirill A. Shutemov int progress = 0; 2340b46e756fSKirill A. Shutemov 2341b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 234235f3aa39SLance Roy lockdep_assert_held(&khugepaged_mm_lock); 234350ad2f24SZach O'Keefe *result = SCAN_FAIL; 2344b46e756fSKirill A. Shutemov 2345b26e2701SQi Zheng if (khugepaged_scan.mm_slot) { 2346b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2347b26e2701SQi Zheng slot = &mm_slot->slot; 2348b26e2701SQi Zheng } else { 2349b26e2701SQi Zheng slot = list_entry(khugepaged_scan.mm_head.next, 2350b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2351b26e2701SQi Zheng mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 2352b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2353b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 2354b46e756fSKirill A. Shutemov } 2355b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2356b46e756fSKirill A. Shutemov 2357b26e2701SQi Zheng mm = slot->mm; 23583b454ad3SYang Shi /* 23593b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 23603b454ad3SYang Shi * the next mm on the list. 23613b454ad3SYang Shi */ 2362b46e756fSKirill A. Shutemov vma = NULL; 2363d8ed45c5SMichel Lespinasse if (unlikely(!mmap_read_trylock(mm))) 2364c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2365b46e756fSKirill A. Shutemov 2366b46e756fSKirill A. Shutemov progress++; 2367879c6000SLance Yang if (unlikely(hpage_collapse_test_exit_or_disable(mm))) 236868540502SMatthew Wilcox (Oracle) goto breakouterloop; 236968540502SMatthew Wilcox (Oracle) 237068540502SMatthew Wilcox (Oracle) vma_iter_init(&vmi, mm, khugepaged_scan.address); 237168540502SMatthew Wilcox (Oracle) for_each_vma(vmi, vma) { 2372b46e756fSKirill A. Shutemov unsigned long hstart, hend; 2373b46e756fSKirill A. Shutemov 2374b46e756fSKirill A. Shutemov cond_resched(); 2375879c6000SLance Yang if (unlikely(hpage_collapse_test_exit_or_disable(mm))) { 2376b46e756fSKirill A. Shutemov progress++; 2377b46e756fSKirill A. Shutemov break; 2378b46e756fSKirill A. Shutemov } 23793485b883SRyan Roberts if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, 23803485b883SRyan Roberts true, PMD_ORDER)) { 2381b46e756fSKirill A. Shutemov skip: 2382b46e756fSKirill A. Shutemov progress++; 2383b46e756fSKirill A. Shutemov continue; 2384b46e756fSKirill A. Shutemov } 23854fa6893fSYang Shi hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); 23864fa6893fSYang Shi hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); 2387b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 2388b46e756fSKirill A. Shutemov goto skip; 2389b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 2390b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 2391b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2392b46e756fSKirill A. Shutemov 2393b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 239450ad2f24SZach O'Keefe bool mmap_locked = true; 239550ad2f24SZach O'Keefe 2396b46e756fSKirill A. Shutemov cond_resched(); 2397879c6000SLance Yang if (unlikely(hpage_collapse_test_exit_or_disable(mm))) 2398b46e756fSKirill A. Shutemov goto breakouterloop; 2399b46e756fSKirill A. Shutemov 2400b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 2401b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 2402b46e756fSKirill A. Shutemov hend); 240399cb0dbdSSong Liu if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 2404396bcc52SMatthew Wilcox (Oracle) struct file *file = get_file(vma->vm_file); 2405f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 2406f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 240799cb0dbdSSong Liu 2408d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 240950ad2f24SZach O'Keefe mmap_locked = false; 2410d50791c2SHugh Dickins *result = hpage_collapse_scan_file(mm, 2411d50791c2SHugh Dickins khugepaged_scan.address, file, pgoff, cc); 2412f3f0e1d2SKirill A. Shutemov fput(file); 2413d50791c2SHugh Dickins if (*result == SCAN_PTE_MAPPED_HUGEPAGE) { 2414d50791c2SHugh Dickins mmap_read_lock(mm); 2415879c6000SLance Yang if (hpage_collapse_test_exit_or_disable(mm)) 2416d50791c2SHugh Dickins goto breakouterloop; 2417d50791c2SHugh Dickins *result = collapse_pte_mapped_thp(mm, 2418d50791c2SHugh Dickins khugepaged_scan.address, false); 2419d50791c2SHugh Dickins if (*result == SCAN_PMD_MAPPED) 2420d50791c2SHugh Dickins *result = SCAN_SUCCEED; 2421d50791c2SHugh Dickins mmap_read_unlock(mm); 2422d50791c2SHugh Dickins } 2423f3f0e1d2SKirill A. Shutemov } else { 24247d2c4385SZach O'Keefe *result = hpage_collapse_scan_pmd(mm, vma, 2425d50791c2SHugh Dickins khugepaged_scan.address, &mmap_locked, cc); 2426f3f0e1d2SKirill A. Shutemov } 242758ac9a89SZach O'Keefe 2428d50791c2SHugh Dickins if (*result == SCAN_SUCCEED) 242950ad2f24SZach O'Keefe ++khugepaged_pages_collapsed; 243058ac9a89SZach O'Keefe 2431b46e756fSKirill A. Shutemov /* move to next address */ 2432b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 2433b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 243450ad2f24SZach O'Keefe if (!mmap_locked) 243550ad2f24SZach O'Keefe /* 243650ad2f24SZach O'Keefe * We released mmap_lock so break loop. Note 243750ad2f24SZach O'Keefe * that we drop mmap_lock before all hugepage 243850ad2f24SZach O'Keefe * allocations, so if allocation fails, we are 243950ad2f24SZach O'Keefe * guaranteed to break here and report the 244050ad2f24SZach O'Keefe * correct result back to caller. 244150ad2f24SZach O'Keefe */ 2442c1e8d7c6SMichel Lespinasse goto breakouterloop_mmap_lock; 2443b46e756fSKirill A. Shutemov if (progress >= pages) 2444b46e756fSKirill A. Shutemov goto breakouterloop; 2445b46e756fSKirill A. Shutemov } 2446b46e756fSKirill A. Shutemov } 2447b46e756fSKirill A. Shutemov breakouterloop: 2448d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ 2449c1e8d7c6SMichel Lespinasse breakouterloop_mmap_lock: 2450b46e756fSKirill A. Shutemov 2451b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2452b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2453b46e756fSKirill A. Shutemov /* 2454b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 2455b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 2456b46e756fSKirill A. Shutemov */ 24575dad6048SLance Yang if (hpage_collapse_test_exit(mm) || !vma) { 2458b46e756fSKirill A. Shutemov /* 2459b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 2460b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 2461b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 2462b46e756fSKirill A. Shutemov */ 2463b26e2701SQi Zheng if (slot->mm_node.next != &khugepaged_scan.mm_head) { 2464b26e2701SQi Zheng slot = list_entry(slot->mm_node.next, 2465b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 2466b26e2701SQi Zheng khugepaged_scan.mm_slot = 2467b26e2701SQi Zheng mm_slot_entry(slot, struct khugepaged_mm_slot, slot); 2468b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 2469b46e756fSKirill A. Shutemov } else { 2470b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2471b46e756fSKirill A. Shutemov khugepaged_full_scans++; 2472b46e756fSKirill A. Shutemov } 2473b46e756fSKirill A. Shutemov 2474b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2475b46e756fSKirill A. Shutemov } 2476b46e756fSKirill A. Shutemov 2477b46e756fSKirill A. Shutemov return progress; 2478b46e756fSKirill A. Shutemov } 2479b46e756fSKirill A. Shutemov 2480b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 2481b46e756fSKirill A. Shutemov { 2482b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 24831064026bSYang Shi hugepage_flags_enabled(); 2484b46e756fSKirill A. Shutemov } 2485b46e756fSKirill A. Shutemov 2486b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 2487b46e756fSKirill A. Shutemov { 2488b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 2489b46e756fSKirill A. Shutemov kthread_should_stop(); 2490b46e756fSKirill A. Shutemov } 2491b46e756fSKirill A. Shutemov 249234d6b470SZach O'Keefe static void khugepaged_do_scan(struct collapse_control *cc) 2493b46e756fSKirill A. Shutemov { 2494b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 249589dc6a96SYanfei Xu unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); 2496b46e756fSKirill A. Shutemov bool wait = true; 249750ad2f24SZach O'Keefe int result = SCAN_SUCCEED; 2498b46e756fSKirill A. Shutemov 2499a980df33SKirill A. Shutemov lru_add_drain_all(); 2500a980df33SKirill A. Shutemov 2501c6a7f445SYang Shi while (true) { 2502b46e756fSKirill A. Shutemov cond_resched(); 2503b46e756fSKirill A. Shutemov 2504b39ca208SKevin Hao if (unlikely(kthread_should_stop())) 2505b46e756fSKirill A. Shutemov break; 2506b46e756fSKirill A. Shutemov 2507b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2508b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 2509b46e756fSKirill A. Shutemov pass_through_head++; 2510b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 2511b46e756fSKirill A. Shutemov pass_through_head < 2) 2512b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 251350ad2f24SZach O'Keefe &result, cc); 2514b46e756fSKirill A. Shutemov else 2515b46e756fSKirill A. Shutemov progress = pages; 2516b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2517b46e756fSKirill A. Shutemov 2518c6a7f445SYang Shi if (progress >= pages) 2519c6a7f445SYang Shi break; 2520c6a7f445SYang Shi 252150ad2f24SZach O'Keefe if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) { 2522c6a7f445SYang Shi /* 2523c6a7f445SYang Shi * If fail to allocate the first time, try to sleep for 2524c6a7f445SYang Shi * a while. When hit again, cancel the scan. 2525c6a7f445SYang Shi */ 2526c6a7f445SYang Shi if (!wait) 2527c6a7f445SYang Shi break; 2528c6a7f445SYang Shi wait = false; 2529c6a7f445SYang Shi khugepaged_alloc_sleep(); 2530c6a7f445SYang Shi } 2531c6a7f445SYang Shi } 2532b46e756fSKirill A. Shutemov } 2533b46e756fSKirill A. Shutemov 2534b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 2535b46e756fSKirill A. Shutemov { 2536b46e756fSKirill A. Shutemov return kthread_should_stop() || 2537b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 2538b46e756fSKirill A. Shutemov } 2539b46e756fSKirill A. Shutemov 2540b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 2541b46e756fSKirill A. Shutemov { 2542b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 2543b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 2544b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2545b46e756fSKirill A. Shutemov 2546b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 2547b46e756fSKirill A. Shutemov return; 2548b46e756fSKirill A. Shutemov 2549b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2550b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 2551b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 2552b46e756fSKirill A. Shutemov scan_sleep_jiffies); 2553b46e756fSKirill A. Shutemov return; 2554b46e756fSKirill A. Shutemov } 2555b46e756fSKirill A. Shutemov 25561064026bSYang Shi if (hugepage_flags_enabled()) 2557b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2558b46e756fSKirill A. Shutemov } 2559b46e756fSKirill A. Shutemov 2560b46e756fSKirill A. Shutemov static int khugepaged(void *none) 2561b46e756fSKirill A. Shutemov { 2562b26e2701SQi Zheng struct khugepaged_mm_slot *mm_slot; 2563b46e756fSKirill A. Shutemov 2564b46e756fSKirill A. Shutemov set_freezable(); 2565b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 2566b46e756fSKirill A. Shutemov 2567b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 256834d6b470SZach O'Keefe khugepaged_do_scan(&khugepaged_collapse_control); 2569b46e756fSKirill A. Shutemov khugepaged_wait_work(); 2570b46e756fSKirill A. Shutemov } 2571b46e756fSKirill A. Shutemov 2572b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 2573b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 2574b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 2575b46e756fSKirill A. Shutemov if (mm_slot) 2576b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 2577b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 2578b46e756fSKirill A. Shutemov return 0; 2579b46e756fSKirill A. Shutemov } 2580b46e756fSKirill A. Shutemov 2581b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 2582b46e756fSKirill A. Shutemov { 2583b46e756fSKirill A. Shutemov struct zone *zone; 2584b46e756fSKirill A. Shutemov int nr_zones = 0; 2585b46e756fSKirill A. Shutemov unsigned long recommended_min; 2586b46e756fSKirill A. Shutemov 25871064026bSYang Shi if (!hugepage_flags_enabled()) { 2588bd3400eaSLiangcai Fan calculate_min_free_kbytes(); 2589bd3400eaSLiangcai Fan goto update_wmarks; 2590bd3400eaSLiangcai Fan } 2591bd3400eaSLiangcai Fan 2592b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 2593b7d349c7SJoonsoo Kim /* 2594b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 2595b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 2596b7d349c7SJoonsoo Kim */ 2597b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 2598b7d349c7SJoonsoo Kim continue; 2599b7d349c7SJoonsoo Kim 2600b46e756fSKirill A. Shutemov nr_zones++; 2601b7d349c7SJoonsoo Kim } 2602b46e756fSKirill A. Shutemov 2603b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2604b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 2605b46e756fSKirill A. Shutemov 2606b46e756fSKirill A. Shutemov /* 2607b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 2608b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 2609b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 2610b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 2611b46e756fSKirill A. Shutemov */ 2612b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 2613b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2614b46e756fSKirill A. Shutemov 2615b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 2616b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 2617b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 2618b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 2619b46e756fSKirill A. Shutemov 2620b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 2621b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 2622b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2623b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 2624b46e756fSKirill A. Shutemov 2625b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 2626b46e756fSKirill A. Shutemov } 2627bd3400eaSLiangcai Fan 2628bd3400eaSLiangcai Fan update_wmarks: 2629b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 2630b46e756fSKirill A. Shutemov } 2631b46e756fSKirill A. Shutemov 2632b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 2633b46e756fSKirill A. Shutemov { 2634b46e756fSKirill A. Shutemov int err = 0; 2635b46e756fSKirill A. Shutemov 2636b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 26371064026bSYang Shi if (hugepage_flags_enabled()) { 2638b46e756fSKirill A. Shutemov if (!khugepaged_thread) 2639b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 2640b46e756fSKirill A. Shutemov "khugepaged"); 2641b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 2642b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2643b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 2644b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2645b46e756fSKirill A. Shutemov goto fail; 2646b46e756fSKirill A. Shutemov } 2647b46e756fSKirill A. Shutemov 2648b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 2649b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 2650b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 2651b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 2652b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 2653b46e756fSKirill A. Shutemov } 2654bd3400eaSLiangcai Fan set_recommended_min_free_kbytes(); 2655b46e756fSKirill A. Shutemov fail: 2656b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 2657b46e756fSKirill A. Shutemov return err; 2658b46e756fSKirill A. Shutemov } 26594aab2be0SVijay Balakrishna 26604aab2be0SVijay Balakrishna void khugepaged_min_free_kbytes_update(void) 26614aab2be0SVijay Balakrishna { 26624aab2be0SVijay Balakrishna mutex_lock(&khugepaged_mutex); 26631064026bSYang Shi if (hugepage_flags_enabled() && khugepaged_thread) 26644aab2be0SVijay Balakrishna set_recommended_min_free_kbytes(); 26654aab2be0SVijay Balakrishna mutex_unlock(&khugepaged_mutex); 26664aab2be0SVijay Balakrishna } 26677d8faaf1SZach O'Keefe 266857e9cc50SJohannes Weiner bool current_is_khugepaged(void) 266957e9cc50SJohannes Weiner { 267057e9cc50SJohannes Weiner return kthread_func(current) == khugepaged; 267157e9cc50SJohannes Weiner } 267257e9cc50SJohannes Weiner 26737d8faaf1SZach O'Keefe static int madvise_collapse_errno(enum scan_result r) 26747d8faaf1SZach O'Keefe { 26757d8faaf1SZach O'Keefe /* 26767d8faaf1SZach O'Keefe * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide 26777d8faaf1SZach O'Keefe * actionable feedback to caller, so they may take an appropriate 26787d8faaf1SZach O'Keefe * fallback measure depending on the nature of the failure. 26797d8faaf1SZach O'Keefe */ 26807d8faaf1SZach O'Keefe switch (r) { 26817d8faaf1SZach O'Keefe case SCAN_ALLOC_HUGE_PAGE_FAIL: 26827d8faaf1SZach O'Keefe return -ENOMEM; 26837d8faaf1SZach O'Keefe case SCAN_CGROUP_CHARGE_FAIL: 2684ac492b9cSDavid Stevens case SCAN_EXCEED_NONE_PTE: 26857d8faaf1SZach O'Keefe return -EBUSY; 26867d8faaf1SZach O'Keefe /* Resource temporary unavailable - trying again might succeed */ 2687ae63c898SZach O'Keefe case SCAN_PAGE_COUNT: 26887d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 26897d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 26900f3e2a2cSZach O'Keefe case SCAN_DEL_PAGE_LRU: 2691ac492b9cSDavid Stevens case SCAN_PAGE_FILLED: 26927d8faaf1SZach O'Keefe return -EAGAIN; 26937d8faaf1SZach O'Keefe /* 26947d8faaf1SZach O'Keefe * Other: Trying again likely not to succeed / error intrinsic to 26957d8faaf1SZach O'Keefe * specified memory range. khugepaged likely won't be able to collapse 26967d8faaf1SZach O'Keefe * either. 26977d8faaf1SZach O'Keefe */ 26987d8faaf1SZach O'Keefe default: 26997d8faaf1SZach O'Keefe return -EINVAL; 27007d8faaf1SZach O'Keefe } 27017d8faaf1SZach O'Keefe } 27027d8faaf1SZach O'Keefe 27037d8faaf1SZach O'Keefe int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev, 27047d8faaf1SZach O'Keefe unsigned long start, unsigned long end) 27057d8faaf1SZach O'Keefe { 27067d8faaf1SZach O'Keefe struct collapse_control *cc; 27077d8faaf1SZach O'Keefe struct mm_struct *mm = vma->vm_mm; 27087d8faaf1SZach O'Keefe unsigned long hstart, hend, addr; 27097d8faaf1SZach O'Keefe int thps = 0, last_fail = SCAN_FAIL; 27107d8faaf1SZach O'Keefe bool mmap_locked = true; 27117d8faaf1SZach O'Keefe 27127d8faaf1SZach O'Keefe BUG_ON(vma->vm_start > start); 27137d8faaf1SZach O'Keefe BUG_ON(vma->vm_end < end); 27147d8faaf1SZach O'Keefe 27157d8faaf1SZach O'Keefe *prev = vma; 27167d8faaf1SZach O'Keefe 27173485b883SRyan Roberts if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, false, 27183485b883SRyan Roberts PMD_ORDER)) 27197d8faaf1SZach O'Keefe return -EINVAL; 27207d8faaf1SZach O'Keefe 27217d8faaf1SZach O'Keefe cc = kmalloc(sizeof(*cc), GFP_KERNEL); 27227d8faaf1SZach O'Keefe if (!cc) 27237d8faaf1SZach O'Keefe return -ENOMEM; 27247d8faaf1SZach O'Keefe cc->is_khugepaged = false; 27257d8faaf1SZach O'Keefe 27267d8faaf1SZach O'Keefe mmgrab(mm); 27277d8faaf1SZach O'Keefe lru_add_drain_all(); 27287d8faaf1SZach O'Keefe 27297d8faaf1SZach O'Keefe hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 27307d8faaf1SZach O'Keefe hend = end & HPAGE_PMD_MASK; 27317d8faaf1SZach O'Keefe 27327d8faaf1SZach O'Keefe for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) { 27337d8faaf1SZach O'Keefe int result = SCAN_FAIL; 27347d8faaf1SZach O'Keefe 27357d8faaf1SZach O'Keefe if (!mmap_locked) { 27367d8faaf1SZach O'Keefe cond_resched(); 27377d8faaf1SZach O'Keefe mmap_read_lock(mm); 27387d8faaf1SZach O'Keefe mmap_locked = true; 273934488399SZach O'Keefe result = hugepage_vma_revalidate(mm, addr, false, &vma, 274034488399SZach O'Keefe cc); 27417d8faaf1SZach O'Keefe if (result != SCAN_SUCCEED) { 27427d8faaf1SZach O'Keefe last_fail = result; 27437d8faaf1SZach O'Keefe goto out_nolock; 27447d8faaf1SZach O'Keefe } 27454d24de94SYang Shi 274652dc0310SZach O'Keefe hend = min(hend, vma->vm_end & HPAGE_PMD_MASK); 27477d8faaf1SZach O'Keefe } 27487d8faaf1SZach O'Keefe mmap_assert_locked(mm); 27497d8faaf1SZach O'Keefe memset(cc->node_load, 0, sizeof(cc->node_load)); 2750e031ff96SYang Shi nodes_clear(cc->alloc_nmask); 275134488399SZach O'Keefe if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 275234488399SZach O'Keefe struct file *file = get_file(vma->vm_file); 275334488399SZach O'Keefe pgoff_t pgoff = linear_page_index(vma, addr); 275434488399SZach O'Keefe 275534488399SZach O'Keefe mmap_read_unlock(mm); 275634488399SZach O'Keefe mmap_locked = false; 275734488399SZach O'Keefe result = hpage_collapse_scan_file(mm, addr, file, pgoff, 27587d2c4385SZach O'Keefe cc); 275934488399SZach O'Keefe fput(file); 276034488399SZach O'Keefe } else { 276134488399SZach O'Keefe result = hpage_collapse_scan_pmd(mm, vma, addr, 276234488399SZach O'Keefe &mmap_locked, cc); 276334488399SZach O'Keefe } 27647d8faaf1SZach O'Keefe if (!mmap_locked) 27657d8faaf1SZach O'Keefe *prev = NULL; /* Tell caller we dropped mmap_lock */ 27667d8faaf1SZach O'Keefe 276734488399SZach O'Keefe handle_result: 27687d8faaf1SZach O'Keefe switch (result) { 27697d8faaf1SZach O'Keefe case SCAN_SUCCEED: 27707d8faaf1SZach O'Keefe case SCAN_PMD_MAPPED: 27717d8faaf1SZach O'Keefe ++thps; 27727d8faaf1SZach O'Keefe break; 277334488399SZach O'Keefe case SCAN_PTE_MAPPED_HUGEPAGE: 277434488399SZach O'Keefe BUG_ON(mmap_locked); 277534488399SZach O'Keefe BUG_ON(*prev); 27761043173eSHugh Dickins mmap_read_lock(mm); 277734488399SZach O'Keefe result = collapse_pte_mapped_thp(mm, addr, true); 27781043173eSHugh Dickins mmap_read_unlock(mm); 277934488399SZach O'Keefe goto handle_result; 27807d8faaf1SZach O'Keefe /* Whitelisted set of results where continuing OK */ 27817d8faaf1SZach O'Keefe case SCAN_PMD_NULL: 27827d8faaf1SZach O'Keefe case SCAN_PTE_NON_PRESENT: 27837d8faaf1SZach O'Keefe case SCAN_PTE_UFFD_WP: 27847d8faaf1SZach O'Keefe case SCAN_PAGE_RO: 27857d8faaf1SZach O'Keefe case SCAN_LACK_REFERENCED_PAGE: 27867d8faaf1SZach O'Keefe case SCAN_PAGE_NULL: 27877d8faaf1SZach O'Keefe case SCAN_PAGE_COUNT: 27887d8faaf1SZach O'Keefe case SCAN_PAGE_LOCK: 27897d8faaf1SZach O'Keefe case SCAN_PAGE_COMPOUND: 27907d8faaf1SZach O'Keefe case SCAN_PAGE_LRU: 27910f3e2a2cSZach O'Keefe case SCAN_DEL_PAGE_LRU: 27927d8faaf1SZach O'Keefe last_fail = result; 27937d8faaf1SZach O'Keefe break; 27947d8faaf1SZach O'Keefe default: 27957d8faaf1SZach O'Keefe last_fail = result; 27967d8faaf1SZach O'Keefe /* Other error, exit */ 27977d8faaf1SZach O'Keefe goto out_maybelock; 27987d8faaf1SZach O'Keefe } 27997d8faaf1SZach O'Keefe } 28007d8faaf1SZach O'Keefe 28017d8faaf1SZach O'Keefe out_maybelock: 28027d8faaf1SZach O'Keefe /* Caller expects us to hold mmap_lock on return */ 28037d8faaf1SZach O'Keefe if (!mmap_locked) 28047d8faaf1SZach O'Keefe mmap_read_lock(mm); 28057d8faaf1SZach O'Keefe out_nolock: 28067d8faaf1SZach O'Keefe mmap_assert_locked(mm); 28077d8faaf1SZach O'Keefe mmdrop(mm); 28087d8faaf1SZach O'Keefe kfree(cc); 28097d8faaf1SZach O'Keefe 28107d8faaf1SZach O'Keefe return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0 28117d8faaf1SZach O'Keefe : madvise_collapse_errno(last_fail); 28127d8faaf1SZach O'Keefe } 2813