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> 19b46e756fSKirill A. Shutemov #include <linux/swapops.h> 20f3f0e1d2SKirill A. Shutemov #include <linux/shmem_fs.h> 21b46e756fSKirill A. Shutemov 22b46e756fSKirill A. Shutemov #include <asm/tlb.h> 23b46e756fSKirill A. Shutemov #include <asm/pgalloc.h> 24b46e756fSKirill A. Shutemov #include "internal.h" 25b46e756fSKirill A. Shutemov 26b46e756fSKirill A. Shutemov enum scan_result { 27b46e756fSKirill A. Shutemov SCAN_FAIL, 28b46e756fSKirill A. Shutemov SCAN_SUCCEED, 29b46e756fSKirill A. Shutemov SCAN_PMD_NULL, 30b46e756fSKirill A. Shutemov SCAN_EXCEED_NONE_PTE, 31b46e756fSKirill A. Shutemov SCAN_PTE_NON_PRESENT, 32b46e756fSKirill A. Shutemov SCAN_PAGE_RO, 330db501f7SEbru Akagunduz SCAN_LACK_REFERENCED_PAGE, 34b46e756fSKirill A. Shutemov SCAN_PAGE_NULL, 35b46e756fSKirill A. Shutemov SCAN_SCAN_ABORT, 36b46e756fSKirill A. Shutemov SCAN_PAGE_COUNT, 37b46e756fSKirill A. Shutemov SCAN_PAGE_LRU, 38b46e756fSKirill A. Shutemov SCAN_PAGE_LOCK, 39b46e756fSKirill A. Shutemov SCAN_PAGE_ANON, 40b46e756fSKirill A. Shutemov SCAN_PAGE_COMPOUND, 41b46e756fSKirill A. Shutemov SCAN_ANY_PROCESS, 42b46e756fSKirill A. Shutemov SCAN_VMA_NULL, 43b46e756fSKirill A. Shutemov SCAN_VMA_CHECK, 44b46e756fSKirill A. Shutemov SCAN_ADDRESS_RANGE, 45b46e756fSKirill A. Shutemov SCAN_SWAP_CACHE_PAGE, 46b46e756fSKirill A. Shutemov SCAN_DEL_PAGE_LRU, 47b46e756fSKirill A. Shutemov SCAN_ALLOC_HUGE_PAGE_FAIL, 48b46e756fSKirill A. Shutemov SCAN_CGROUP_CHARGE_FAIL, 49f3f0e1d2SKirill A. Shutemov SCAN_EXCEED_SWAP_PTE, 50f3f0e1d2SKirill A. Shutemov SCAN_TRUNCATED, 51b46e756fSKirill A. Shutemov }; 52b46e756fSKirill A. Shutemov 53b46e756fSKirill A. Shutemov #define CREATE_TRACE_POINTS 54b46e756fSKirill A. Shutemov #include <trace/events/huge_memory.h> 55b46e756fSKirill A. Shutemov 56b46e756fSKirill A. Shutemov /* default scan 8*512 pte (or vmas) every 30 second */ 57b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_to_scan __read_mostly; 58b46e756fSKirill A. Shutemov static unsigned int khugepaged_pages_collapsed; 59b46e756fSKirill A. Shutemov static unsigned int khugepaged_full_scans; 60b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 61b46e756fSKirill A. Shutemov /* during fragmentation poll the hugepage allocator once every minute */ 62b46e756fSKirill A. Shutemov static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 63b46e756fSKirill A. Shutemov static unsigned long khugepaged_sleep_expire; 64b46e756fSKirill A. Shutemov static DEFINE_SPINLOCK(khugepaged_mm_lock); 65b46e756fSKirill A. Shutemov static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 66b46e756fSKirill A. Shutemov /* 67b46e756fSKirill A. Shutemov * default collapse hugepages if there is at least one pte mapped like 68b46e756fSKirill A. Shutemov * it would have happened if the vma was large enough during page 69b46e756fSKirill A. Shutemov * fault. 70b46e756fSKirill A. Shutemov */ 71b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_none __read_mostly; 72b46e756fSKirill A. Shutemov static unsigned int khugepaged_max_ptes_swap __read_mostly; 73b46e756fSKirill A. Shutemov 74b46e756fSKirill A. Shutemov #define MM_SLOTS_HASH_BITS 10 75b46e756fSKirill A. Shutemov static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 76b46e756fSKirill A. Shutemov 77b46e756fSKirill A. Shutemov static struct kmem_cache *mm_slot_cache __read_mostly; 78b46e756fSKirill A. Shutemov 79b46e756fSKirill A. Shutemov /** 80b46e756fSKirill A. Shutemov * struct mm_slot - hash lookup from mm to mm_slot 81b46e756fSKirill A. Shutemov * @hash: hash collision list 82b46e756fSKirill A. Shutemov * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 83b46e756fSKirill A. Shutemov * @mm: the mm that this information is valid for 84b46e756fSKirill A. Shutemov */ 85b46e756fSKirill A. Shutemov struct mm_slot { 86b46e756fSKirill A. Shutemov struct hlist_node hash; 87b46e756fSKirill A. Shutemov struct list_head mm_node; 88b46e756fSKirill A. Shutemov struct mm_struct *mm; 89b46e756fSKirill A. Shutemov }; 90b46e756fSKirill A. Shutemov 91b46e756fSKirill A. Shutemov /** 92b46e756fSKirill A. Shutemov * struct khugepaged_scan - cursor for scanning 93b46e756fSKirill A. Shutemov * @mm_head: the head of the mm list to scan 94b46e756fSKirill A. Shutemov * @mm_slot: the current mm_slot we are scanning 95b46e756fSKirill A. Shutemov * @address: the next address inside that to be scanned 96b46e756fSKirill A. Shutemov * 97b46e756fSKirill A. Shutemov * There is only the one khugepaged_scan instance of this cursor structure. 98b46e756fSKirill A. Shutemov */ 99b46e756fSKirill A. Shutemov struct khugepaged_scan { 100b46e756fSKirill A. Shutemov struct list_head mm_head; 101b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 102b46e756fSKirill A. Shutemov unsigned long address; 103b46e756fSKirill A. Shutemov }; 104b46e756fSKirill A. Shutemov 105b46e756fSKirill A. Shutemov static struct khugepaged_scan khugepaged_scan = { 106b46e756fSKirill A. Shutemov .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 107b46e756fSKirill A. Shutemov }; 108b46e756fSKirill A. Shutemov 109e1465d12SJérémy Lefaure #ifdef CONFIG_SYSFS 110b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 111b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 112b46e756fSKirill A. Shutemov char *buf) 113b46e756fSKirill A. Shutemov { 114b46e756fSKirill A. Shutemov return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); 115b46e756fSKirill A. Shutemov } 116b46e756fSKirill A. Shutemov 117b46e756fSKirill A. Shutemov static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 118b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 119b46e756fSKirill A. Shutemov const char *buf, size_t count) 120b46e756fSKirill A. Shutemov { 121b46e756fSKirill A. Shutemov unsigned long msecs; 122b46e756fSKirill A. Shutemov int err; 123b46e756fSKirill A. Shutemov 124b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &msecs); 125b46e756fSKirill A. Shutemov if (err || msecs > UINT_MAX) 126b46e756fSKirill A. Shutemov return -EINVAL; 127b46e756fSKirill A. Shutemov 128b46e756fSKirill A. Shutemov khugepaged_scan_sleep_millisecs = msecs; 129b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 130b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 131b46e756fSKirill A. Shutemov 132b46e756fSKirill A. Shutemov return count; 133b46e756fSKirill A. Shutemov } 134b46e756fSKirill A. Shutemov static struct kobj_attribute scan_sleep_millisecs_attr = 135b46e756fSKirill A. Shutemov __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 136b46e756fSKirill A. Shutemov scan_sleep_millisecs_store); 137b46e756fSKirill A. Shutemov 138b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 139b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 140b46e756fSKirill A. Shutemov char *buf) 141b46e756fSKirill A. Shutemov { 142b46e756fSKirill A. Shutemov return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 143b46e756fSKirill A. Shutemov } 144b46e756fSKirill A. Shutemov 145b46e756fSKirill A. Shutemov static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 146b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 147b46e756fSKirill A. Shutemov const char *buf, size_t count) 148b46e756fSKirill A. Shutemov { 149b46e756fSKirill A. Shutemov unsigned long msecs; 150b46e756fSKirill A. Shutemov int err; 151b46e756fSKirill A. Shutemov 152b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &msecs); 153b46e756fSKirill A. Shutemov if (err || msecs > UINT_MAX) 154b46e756fSKirill A. Shutemov return -EINVAL; 155b46e756fSKirill A. Shutemov 156b46e756fSKirill A. Shutemov khugepaged_alloc_sleep_millisecs = msecs; 157b46e756fSKirill A. Shutemov khugepaged_sleep_expire = 0; 158b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 159b46e756fSKirill A. Shutemov 160b46e756fSKirill A. Shutemov return count; 161b46e756fSKirill A. Shutemov } 162b46e756fSKirill A. Shutemov static struct kobj_attribute alloc_sleep_millisecs_attr = 163b46e756fSKirill A. Shutemov __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 164b46e756fSKirill A. Shutemov alloc_sleep_millisecs_store); 165b46e756fSKirill A. Shutemov 166b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_show(struct kobject *kobj, 167b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 168b46e756fSKirill A. Shutemov char *buf) 169b46e756fSKirill A. Shutemov { 170b46e756fSKirill A. Shutemov return sprintf(buf, "%u\n", khugepaged_pages_to_scan); 171b46e756fSKirill A. Shutemov } 172b46e756fSKirill A. Shutemov static ssize_t pages_to_scan_store(struct kobject *kobj, 173b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 174b46e756fSKirill A. Shutemov const char *buf, size_t count) 175b46e756fSKirill A. Shutemov { 176b46e756fSKirill A. Shutemov int err; 177b46e756fSKirill A. Shutemov unsigned long pages; 178b46e756fSKirill A. Shutemov 179b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &pages); 180b46e756fSKirill A. Shutemov if (err || !pages || pages > UINT_MAX) 181b46e756fSKirill A. Shutemov return -EINVAL; 182b46e756fSKirill A. Shutemov 183b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = pages; 184b46e756fSKirill A. Shutemov 185b46e756fSKirill A. Shutemov return count; 186b46e756fSKirill A. Shutemov } 187b46e756fSKirill A. Shutemov static struct kobj_attribute pages_to_scan_attr = 188b46e756fSKirill A. Shutemov __ATTR(pages_to_scan, 0644, pages_to_scan_show, 189b46e756fSKirill A. Shutemov pages_to_scan_store); 190b46e756fSKirill A. Shutemov 191b46e756fSKirill A. Shutemov static ssize_t pages_collapsed_show(struct kobject *kobj, 192b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 193b46e756fSKirill A. Shutemov char *buf) 194b46e756fSKirill A. Shutemov { 195b46e756fSKirill A. Shutemov return sprintf(buf, "%u\n", khugepaged_pages_collapsed); 196b46e756fSKirill A. Shutemov } 197b46e756fSKirill A. Shutemov static struct kobj_attribute pages_collapsed_attr = 198b46e756fSKirill A. Shutemov __ATTR_RO(pages_collapsed); 199b46e756fSKirill A. Shutemov 200b46e756fSKirill A. Shutemov static ssize_t full_scans_show(struct kobject *kobj, 201b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 202b46e756fSKirill A. Shutemov char *buf) 203b46e756fSKirill A. Shutemov { 204b46e756fSKirill A. Shutemov return sprintf(buf, "%u\n", khugepaged_full_scans); 205b46e756fSKirill A. Shutemov } 206b46e756fSKirill A. Shutemov static struct kobj_attribute full_scans_attr = 207b46e756fSKirill A. Shutemov __ATTR_RO(full_scans); 208b46e756fSKirill A. Shutemov 209b46e756fSKirill A. Shutemov static ssize_t khugepaged_defrag_show(struct kobject *kobj, 210b46e756fSKirill A. Shutemov struct kobj_attribute *attr, char *buf) 211b46e756fSKirill A. Shutemov { 212b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 213b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 214b46e756fSKirill A. Shutemov } 215b46e756fSKirill A. Shutemov static ssize_t khugepaged_defrag_store(struct kobject *kobj, 216b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 217b46e756fSKirill A. Shutemov const char *buf, size_t count) 218b46e756fSKirill A. Shutemov { 219b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 220b46e756fSKirill A. Shutemov TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 221b46e756fSKirill A. Shutemov } 222b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_defrag_attr = 223b46e756fSKirill A. Shutemov __ATTR(defrag, 0644, khugepaged_defrag_show, 224b46e756fSKirill A. Shutemov khugepaged_defrag_store); 225b46e756fSKirill A. Shutemov 226b46e756fSKirill A. Shutemov /* 227b46e756fSKirill A. Shutemov * max_ptes_none controls if khugepaged should collapse hugepages over 228b46e756fSKirill A. Shutemov * any unmapped ptes in turn potentially increasing the memory 229b46e756fSKirill A. Shutemov * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 230b46e756fSKirill A. Shutemov * reduce the available free memory in the system as it 231b46e756fSKirill A. Shutemov * runs. Increasing max_ptes_none will instead potentially reduce the 232b46e756fSKirill A. Shutemov * free memory in the system during the khugepaged scan. 233b46e756fSKirill A. Shutemov */ 234b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 235b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 236b46e756fSKirill A. Shutemov char *buf) 237b46e756fSKirill A. Shutemov { 238b46e756fSKirill A. Shutemov return sprintf(buf, "%u\n", khugepaged_max_ptes_none); 239b46e756fSKirill A. Shutemov } 240b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 241b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 242b46e756fSKirill A. Shutemov const char *buf, size_t count) 243b46e756fSKirill A. Shutemov { 244b46e756fSKirill A. Shutemov int err; 245b46e756fSKirill A. Shutemov unsigned long max_ptes_none; 246b46e756fSKirill A. Shutemov 247b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_none); 248b46e756fSKirill A. Shutemov if (err || max_ptes_none > HPAGE_PMD_NR-1) 249b46e756fSKirill A. Shutemov return -EINVAL; 250b46e756fSKirill A. Shutemov 251b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = max_ptes_none; 252b46e756fSKirill A. Shutemov 253b46e756fSKirill A. Shutemov return count; 254b46e756fSKirill A. Shutemov } 255b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_none_attr = 256b46e756fSKirill A. Shutemov __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 257b46e756fSKirill A. Shutemov khugepaged_max_ptes_none_store); 258b46e756fSKirill A. Shutemov 259b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj, 260b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 261b46e756fSKirill A. Shutemov char *buf) 262b46e756fSKirill A. Shutemov { 263b46e756fSKirill A. Shutemov return sprintf(buf, "%u\n", khugepaged_max_ptes_swap); 264b46e756fSKirill A. Shutemov } 265b46e756fSKirill A. Shutemov 266b46e756fSKirill A. Shutemov static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj, 267b46e756fSKirill A. Shutemov struct kobj_attribute *attr, 268b46e756fSKirill A. Shutemov const char *buf, size_t count) 269b46e756fSKirill A. Shutemov { 270b46e756fSKirill A. Shutemov int err; 271b46e756fSKirill A. Shutemov unsigned long max_ptes_swap; 272b46e756fSKirill A. Shutemov 273b46e756fSKirill A. Shutemov err = kstrtoul(buf, 10, &max_ptes_swap); 274b46e756fSKirill A. Shutemov if (err || max_ptes_swap > HPAGE_PMD_NR-1) 275b46e756fSKirill A. Shutemov return -EINVAL; 276b46e756fSKirill A. Shutemov 277b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = max_ptes_swap; 278b46e756fSKirill A. Shutemov 279b46e756fSKirill A. Shutemov return count; 280b46e756fSKirill A. Shutemov } 281b46e756fSKirill A. Shutemov 282b46e756fSKirill A. Shutemov static struct kobj_attribute khugepaged_max_ptes_swap_attr = 283b46e756fSKirill A. Shutemov __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show, 284b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap_store); 285b46e756fSKirill A. Shutemov 286b46e756fSKirill A. Shutemov static struct attribute *khugepaged_attr[] = { 287b46e756fSKirill A. Shutemov &khugepaged_defrag_attr.attr, 288b46e756fSKirill A. Shutemov &khugepaged_max_ptes_none_attr.attr, 289b46e756fSKirill A. Shutemov &pages_to_scan_attr.attr, 290b46e756fSKirill A. Shutemov &pages_collapsed_attr.attr, 291b46e756fSKirill A. Shutemov &full_scans_attr.attr, 292b46e756fSKirill A. Shutemov &scan_sleep_millisecs_attr.attr, 293b46e756fSKirill A. Shutemov &alloc_sleep_millisecs_attr.attr, 294b46e756fSKirill A. Shutemov &khugepaged_max_ptes_swap_attr.attr, 295b46e756fSKirill A. Shutemov NULL, 296b46e756fSKirill A. Shutemov }; 297b46e756fSKirill A. Shutemov 298b46e756fSKirill A. Shutemov struct attribute_group khugepaged_attr_group = { 299b46e756fSKirill A. Shutemov .attrs = khugepaged_attr, 300b46e756fSKirill A. Shutemov .name = "khugepaged", 301b46e756fSKirill A. Shutemov }; 302e1465d12SJérémy Lefaure #endif /* CONFIG_SYSFS */ 303b46e756fSKirill A. Shutemov 304f3f0e1d2SKirill A. Shutemov #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB) 305b46e756fSKirill A. Shutemov 306b46e756fSKirill A. Shutemov int hugepage_madvise(struct vm_area_struct *vma, 307b46e756fSKirill A. Shutemov unsigned long *vm_flags, int advice) 308b46e756fSKirill A. Shutemov { 309b46e756fSKirill A. Shutemov switch (advice) { 310b46e756fSKirill A. Shutemov case MADV_HUGEPAGE: 311b46e756fSKirill A. Shutemov #ifdef CONFIG_S390 312b46e756fSKirill A. Shutemov /* 313b46e756fSKirill A. Shutemov * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 314b46e756fSKirill A. Shutemov * can't handle this properly after s390_enable_sie, so we simply 315b46e756fSKirill A. Shutemov * ignore the madvise to prevent qemu from causing a SIGSEGV. 316b46e756fSKirill A. Shutemov */ 317b46e756fSKirill A. Shutemov if (mm_has_pgste(vma->vm_mm)) 318b46e756fSKirill A. Shutemov return 0; 319b46e756fSKirill A. Shutemov #endif 320b46e756fSKirill A. Shutemov *vm_flags &= ~VM_NOHUGEPAGE; 321b46e756fSKirill A. Shutemov *vm_flags |= VM_HUGEPAGE; 322b46e756fSKirill A. Shutemov /* 323b46e756fSKirill A. Shutemov * If the vma become good for khugepaged to scan, 324b46e756fSKirill A. Shutemov * register it here without waiting a page fault that 325b46e756fSKirill A. Shutemov * may not happen any time soon. 326b46e756fSKirill A. Shutemov */ 327b46e756fSKirill A. Shutemov if (!(*vm_flags & VM_NO_KHUGEPAGED) && 328b46e756fSKirill A. Shutemov khugepaged_enter_vma_merge(vma, *vm_flags)) 329b46e756fSKirill A. Shutemov return -ENOMEM; 330b46e756fSKirill A. Shutemov break; 331b46e756fSKirill A. Shutemov case MADV_NOHUGEPAGE: 332b46e756fSKirill A. Shutemov *vm_flags &= ~VM_HUGEPAGE; 333b46e756fSKirill A. Shutemov *vm_flags |= VM_NOHUGEPAGE; 334b46e756fSKirill A. Shutemov /* 335b46e756fSKirill A. Shutemov * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 336b46e756fSKirill A. Shutemov * this vma even if we leave the mm registered in khugepaged if 337b46e756fSKirill A. Shutemov * it got registered before VM_NOHUGEPAGE was set. 338b46e756fSKirill A. Shutemov */ 339b46e756fSKirill A. Shutemov break; 340b46e756fSKirill A. Shutemov } 341b46e756fSKirill A. Shutemov 342b46e756fSKirill A. Shutemov return 0; 343b46e756fSKirill A. Shutemov } 344b46e756fSKirill A. Shutemov 345b46e756fSKirill A. Shutemov int __init khugepaged_init(void) 346b46e756fSKirill A. Shutemov { 347b46e756fSKirill A. Shutemov mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 348b46e756fSKirill A. Shutemov sizeof(struct mm_slot), 349b46e756fSKirill A. Shutemov __alignof__(struct mm_slot), 0, NULL); 350b46e756fSKirill A. Shutemov if (!mm_slot_cache) 351b46e756fSKirill A. Shutemov return -ENOMEM; 352b46e756fSKirill A. Shutemov 353b46e756fSKirill A. Shutemov khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 354b46e756fSKirill A. Shutemov khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 355b46e756fSKirill A. Shutemov khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 356b46e756fSKirill A. Shutemov 357b46e756fSKirill A. Shutemov return 0; 358b46e756fSKirill A. Shutemov } 359b46e756fSKirill A. Shutemov 360b46e756fSKirill A. Shutemov void __init khugepaged_destroy(void) 361b46e756fSKirill A. Shutemov { 362b46e756fSKirill A. Shutemov kmem_cache_destroy(mm_slot_cache); 363b46e756fSKirill A. Shutemov } 364b46e756fSKirill A. Shutemov 365b46e756fSKirill A. Shutemov static inline struct mm_slot *alloc_mm_slot(void) 366b46e756fSKirill A. Shutemov { 367b46e756fSKirill A. Shutemov if (!mm_slot_cache) /* initialization failed */ 368b46e756fSKirill A. Shutemov return NULL; 369b46e756fSKirill A. Shutemov return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 370b46e756fSKirill A. Shutemov } 371b46e756fSKirill A. Shutemov 372b46e756fSKirill A. Shutemov static inline void free_mm_slot(struct mm_slot *mm_slot) 373b46e756fSKirill A. Shutemov { 374b46e756fSKirill A. Shutemov kmem_cache_free(mm_slot_cache, mm_slot); 375b46e756fSKirill A. Shutemov } 376b46e756fSKirill A. Shutemov 377b46e756fSKirill A. Shutemov static struct mm_slot *get_mm_slot(struct mm_struct *mm) 378b46e756fSKirill A. Shutemov { 379b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 380b46e756fSKirill A. Shutemov 381b46e756fSKirill A. Shutemov hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 382b46e756fSKirill A. Shutemov if (mm == mm_slot->mm) 383b46e756fSKirill A. Shutemov return mm_slot; 384b46e756fSKirill A. Shutemov 385b46e756fSKirill A. Shutemov return NULL; 386b46e756fSKirill A. Shutemov } 387b46e756fSKirill A. Shutemov 388b46e756fSKirill A. Shutemov static void insert_to_mm_slots_hash(struct mm_struct *mm, 389b46e756fSKirill A. Shutemov struct mm_slot *mm_slot) 390b46e756fSKirill A. Shutemov { 391b46e756fSKirill A. Shutemov mm_slot->mm = mm; 392b46e756fSKirill A. Shutemov hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 393b46e756fSKirill A. Shutemov } 394b46e756fSKirill A. Shutemov 395b46e756fSKirill A. Shutemov static inline int khugepaged_test_exit(struct mm_struct *mm) 396b46e756fSKirill A. Shutemov { 397b46e756fSKirill A. Shutemov return atomic_read(&mm->mm_users) == 0; 398b46e756fSKirill A. Shutemov } 399b46e756fSKirill A. Shutemov 40050f8b92fSSong Liu static bool hugepage_vma_check(struct vm_area_struct *vma, 40150f8b92fSSong Liu unsigned long vm_flags) 402c2231020SYang Shi { 40350f8b92fSSong Liu if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || 40450f8b92fSSong Liu (vm_flags & VM_NOHUGEPAGE) || 405c2231020SYang Shi test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)) 406c2231020SYang Shi return false; 407c2231020SYang Shi if (shmem_file(vma->vm_file)) { 408c2231020SYang Shi if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) 409c2231020SYang Shi return false; 410c2231020SYang Shi return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff, 411c2231020SYang Shi HPAGE_PMD_NR); 412c2231020SYang Shi } 413c2231020SYang Shi if (!vma->anon_vma || vma->vm_ops) 414c2231020SYang Shi return false; 415c2231020SYang Shi if (is_vma_temporary_stack(vma)) 416c2231020SYang Shi return false; 41750f8b92fSSong Liu return !(vm_flags & VM_NO_KHUGEPAGED); 418c2231020SYang Shi } 419c2231020SYang Shi 420b46e756fSKirill A. Shutemov int __khugepaged_enter(struct mm_struct *mm) 421b46e756fSKirill A. Shutemov { 422b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 423b46e756fSKirill A. Shutemov int wakeup; 424b46e756fSKirill A. Shutemov 425b46e756fSKirill A. Shutemov mm_slot = alloc_mm_slot(); 426b46e756fSKirill A. Shutemov if (!mm_slot) 427b46e756fSKirill A. Shutemov return -ENOMEM; 428b46e756fSKirill A. Shutemov 429b46e756fSKirill A. Shutemov /* __khugepaged_exit() must not run from under us */ 430b46e756fSKirill A. Shutemov VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); 431b46e756fSKirill A. Shutemov if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 432b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 433b46e756fSKirill A. Shutemov return 0; 434b46e756fSKirill A. Shutemov } 435b46e756fSKirill A. Shutemov 436b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 437b46e756fSKirill A. Shutemov insert_to_mm_slots_hash(mm, 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); 443b46e756fSKirill A. Shutemov list_add_tail(&mm_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 return 0; 451b46e756fSKirill A. Shutemov } 452b46e756fSKirill A. Shutemov 453b46e756fSKirill A. Shutemov int khugepaged_enter_vma_merge(struct vm_area_struct *vma, 454b46e756fSKirill A. Shutemov unsigned long vm_flags) 455b46e756fSKirill A. Shutemov { 456b46e756fSKirill A. Shutemov unsigned long hstart, hend; 457c2231020SYang Shi 458b46e756fSKirill A. Shutemov /* 459c2231020SYang Shi * khugepaged does not yet work on non-shmem files or special 460c2231020SYang Shi * mappings. And file-private shmem THP is not supported. 461b46e756fSKirill A. Shutemov */ 46250f8b92fSSong Liu if (!hugepage_vma_check(vma, vm_flags)) 463b46e756fSKirill A. Shutemov return 0; 464c2231020SYang Shi 465b46e756fSKirill A. Shutemov hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 466b46e756fSKirill A. Shutemov hend = vma->vm_end & HPAGE_PMD_MASK; 467b46e756fSKirill A. Shutemov if (hstart < hend) 468b46e756fSKirill A. Shutemov return khugepaged_enter(vma, vm_flags); 469b46e756fSKirill A. Shutemov return 0; 470b46e756fSKirill A. Shutemov } 471b46e756fSKirill A. Shutemov 472b46e756fSKirill A. Shutemov void __khugepaged_exit(struct mm_struct *mm) 473b46e756fSKirill A. Shutemov { 474b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 475b46e756fSKirill A. Shutemov int free = 0; 476b46e756fSKirill A. Shutemov 477b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 478b46e756fSKirill A. Shutemov mm_slot = get_mm_slot(mm); 479b46e756fSKirill A. Shutemov if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 480b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 481b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 482b46e756fSKirill A. Shutemov free = 1; 483b46e756fSKirill A. Shutemov } 484b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 485b46e756fSKirill A. Shutemov 486b46e756fSKirill A. Shutemov if (free) { 487b46e756fSKirill A. Shutemov clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 488b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 489b46e756fSKirill A. Shutemov mmdrop(mm); 490b46e756fSKirill A. Shutemov } else if (mm_slot) { 491b46e756fSKirill A. Shutemov /* 492b46e756fSKirill A. Shutemov * This is required to serialize against 493b46e756fSKirill A. Shutemov * khugepaged_test_exit() (which is guaranteed to run 494b46e756fSKirill A. Shutemov * under mmap sem read mode). Stop here (after we 495b46e756fSKirill A. Shutemov * return all pagetables will be destroyed) until 496b46e756fSKirill A. Shutemov * khugepaged has finished working on the pagetables 497b46e756fSKirill A. Shutemov * under the mmap_sem. 498b46e756fSKirill A. Shutemov */ 499b46e756fSKirill A. Shutemov down_write(&mm->mmap_sem); 500b46e756fSKirill A. Shutemov up_write(&mm->mmap_sem); 501b46e756fSKirill A. Shutemov } 502b46e756fSKirill A. Shutemov } 503b46e756fSKirill A. Shutemov 504b46e756fSKirill A. Shutemov static void release_pte_page(struct page *page) 505b46e756fSKirill A. Shutemov { 506d44d363fSShaohua Li dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page)); 507b46e756fSKirill A. Shutemov unlock_page(page); 508b46e756fSKirill A. Shutemov putback_lru_page(page); 509b46e756fSKirill A. Shutemov } 510b46e756fSKirill A. Shutemov 511b46e756fSKirill A. Shutemov static void release_pte_pages(pte_t *pte, pte_t *_pte) 512b46e756fSKirill A. Shutemov { 513b46e756fSKirill A. Shutemov while (--_pte >= pte) { 514b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 515b46e756fSKirill A. Shutemov if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) 516b46e756fSKirill A. Shutemov release_pte_page(pte_page(pteval)); 517b46e756fSKirill A. Shutemov } 518b46e756fSKirill A. Shutemov } 519b46e756fSKirill A. Shutemov 520b46e756fSKirill A. Shutemov static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 521b46e756fSKirill A. Shutemov unsigned long address, 522b46e756fSKirill A. Shutemov pte_t *pte) 523b46e756fSKirill A. Shutemov { 524b46e756fSKirill A. Shutemov struct page *page = NULL; 525b46e756fSKirill A. Shutemov pte_t *_pte; 5260db501f7SEbru Akagunduz int none_or_zero = 0, result = 0, referenced = 0; 5270db501f7SEbru Akagunduz bool writable = false; 528b46e756fSKirill A. Shutemov 529b46e756fSKirill A. Shutemov for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 530b46e756fSKirill A. Shutemov _pte++, address += PAGE_SIZE) { 531b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 532b46e756fSKirill A. Shutemov if (pte_none(pteval) || (pte_present(pteval) && 533b46e756fSKirill A. Shutemov is_zero_pfn(pte_pfn(pteval)))) { 534b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 535b46e756fSKirill A. Shutemov ++none_or_zero <= khugepaged_max_ptes_none) { 536b46e756fSKirill A. Shutemov continue; 537b46e756fSKirill A. Shutemov } else { 538b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 539b46e756fSKirill A. Shutemov goto out; 540b46e756fSKirill A. Shutemov } 541b46e756fSKirill A. Shutemov } 542b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 543b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 544b46e756fSKirill A. Shutemov goto out; 545b46e756fSKirill A. Shutemov } 546b46e756fSKirill A. Shutemov page = vm_normal_page(vma, address, pteval); 547b46e756fSKirill A. Shutemov if (unlikely(!page)) { 548b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 549b46e756fSKirill A. Shutemov goto out; 550b46e756fSKirill A. Shutemov } 551b46e756fSKirill A. Shutemov 552fece2029SKirill A. Shutemov /* TODO: teach khugepaged to collapse THP mapped with pte */ 553fece2029SKirill A. Shutemov if (PageCompound(page)) { 554fece2029SKirill A. Shutemov result = SCAN_PAGE_COMPOUND; 555fece2029SKirill A. Shutemov goto out; 556fece2029SKirill A. Shutemov } 557fece2029SKirill A. Shutemov 558b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageAnon(page), page); 559b46e756fSKirill A. Shutemov 560b46e756fSKirill A. Shutemov /* 561b46e756fSKirill A. Shutemov * We can do it before isolate_lru_page because the 562b46e756fSKirill A. Shutemov * page can't be freed from under us. NOTE: PG_lock 563b46e756fSKirill A. Shutemov * is needed to serialize against split_huge_page 564b46e756fSKirill A. Shutemov * when invoked from the VM. 565b46e756fSKirill A. Shutemov */ 566b46e756fSKirill A. Shutemov if (!trylock_page(page)) { 567b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 568b46e756fSKirill A. Shutemov goto out; 569b46e756fSKirill A. Shutemov } 570b46e756fSKirill A. Shutemov 571b46e756fSKirill A. Shutemov /* 572b46e756fSKirill A. Shutemov * cannot use mapcount: can't collapse if there's a gup pin. 573b46e756fSKirill A. Shutemov * The page must only be referenced by the scanned process 574b46e756fSKirill A. Shutemov * and page swap cache. 575b46e756fSKirill A. Shutemov */ 5762948be5aSMinchan Kim if (page_count(page) != 1 + PageSwapCache(page)) { 577b46e756fSKirill A. Shutemov unlock_page(page); 578b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 579b46e756fSKirill A. Shutemov goto out; 580b46e756fSKirill A. Shutemov } 581b46e756fSKirill A. Shutemov if (pte_write(pteval)) { 582b46e756fSKirill A. Shutemov writable = true; 583b46e756fSKirill A. Shutemov } else { 584b46e756fSKirill A. Shutemov if (PageSwapCache(page) && 585b46e756fSKirill A. Shutemov !reuse_swap_page(page, NULL)) { 586b46e756fSKirill A. Shutemov unlock_page(page); 587b46e756fSKirill A. Shutemov result = SCAN_SWAP_CACHE_PAGE; 588b46e756fSKirill A. Shutemov goto out; 589b46e756fSKirill A. Shutemov } 590b46e756fSKirill A. Shutemov /* 591b46e756fSKirill A. Shutemov * Page is not in the swap cache. It can be collapsed 592b46e756fSKirill A. Shutemov * into a THP. 593b46e756fSKirill A. Shutemov */ 594b46e756fSKirill A. Shutemov } 595b46e756fSKirill A. Shutemov 596b46e756fSKirill A. Shutemov /* 597b46e756fSKirill A. Shutemov * Isolate the page to avoid collapsing an hugepage 598b46e756fSKirill A. Shutemov * currently in use by the VM. 599b46e756fSKirill A. Shutemov */ 600b46e756fSKirill A. Shutemov if (isolate_lru_page(page)) { 601b46e756fSKirill A. Shutemov unlock_page(page); 602b46e756fSKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 603b46e756fSKirill A. Shutemov goto out; 604b46e756fSKirill A. Shutemov } 605d44d363fSShaohua Li inc_node_page_state(page, 606d44d363fSShaohua Li NR_ISOLATED_ANON + page_is_file_cache(page)); 607b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 608b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(PageLRU(page), page); 609b46e756fSKirill A. Shutemov 6100db501f7SEbru Akagunduz /* There should be enough young pte to collapse the page */ 611b46e756fSKirill A. Shutemov if (pte_young(pteval) || 612b46e756fSKirill A. Shutemov page_is_young(page) || PageReferenced(page) || 613b46e756fSKirill A. Shutemov mmu_notifier_test_young(vma->vm_mm, address)) 6140db501f7SEbru Akagunduz referenced++; 615b46e756fSKirill A. Shutemov } 616b46e756fSKirill A. Shutemov if (likely(writable)) { 617b46e756fSKirill A. Shutemov if (likely(referenced)) { 618b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 619b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 620b46e756fSKirill A. Shutemov referenced, writable, result); 621b46e756fSKirill A. Shutemov return 1; 622b46e756fSKirill A. Shutemov } 623b46e756fSKirill A. Shutemov } else { 624b46e756fSKirill A. Shutemov result = SCAN_PAGE_RO; 625b46e756fSKirill A. Shutemov } 626b46e756fSKirill A. Shutemov 627b46e756fSKirill A. Shutemov out: 628b46e756fSKirill A. Shutemov release_pte_pages(pte, _pte); 629b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page_isolate(page, none_or_zero, 630b46e756fSKirill A. Shutemov referenced, writable, result); 631b46e756fSKirill A. Shutemov return 0; 632b46e756fSKirill A. Shutemov } 633b46e756fSKirill A. Shutemov 634b46e756fSKirill A. Shutemov static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 635b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 636b46e756fSKirill A. Shutemov unsigned long address, 637b46e756fSKirill A. Shutemov spinlock_t *ptl) 638b46e756fSKirill A. Shutemov { 639b46e756fSKirill A. Shutemov pte_t *_pte; 640338a16baSDavid Rientjes for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 641338a16baSDavid Rientjes _pte++, page++, address += PAGE_SIZE) { 642b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 643b46e756fSKirill A. Shutemov struct page *src_page; 644b46e756fSKirill A. Shutemov 645b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 646b46e756fSKirill A. Shutemov clear_user_highpage(page, address); 647b46e756fSKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 648b46e756fSKirill A. Shutemov if (is_zero_pfn(pte_pfn(pteval))) { 649b46e756fSKirill A. Shutemov /* 650b46e756fSKirill A. Shutemov * ptl mostly unnecessary. 651b46e756fSKirill A. Shutemov */ 652b46e756fSKirill A. Shutemov spin_lock(ptl); 653b46e756fSKirill A. Shutemov /* 654b46e756fSKirill A. Shutemov * paravirt calls inside pte_clear here are 655b46e756fSKirill A. Shutemov * superfluous. 656b46e756fSKirill A. Shutemov */ 657b46e756fSKirill A. Shutemov pte_clear(vma->vm_mm, address, _pte); 658b46e756fSKirill A. Shutemov spin_unlock(ptl); 659b46e756fSKirill A. Shutemov } 660b46e756fSKirill A. Shutemov } else { 661b46e756fSKirill A. Shutemov src_page = pte_page(pteval); 662b46e756fSKirill A. Shutemov copy_user_highpage(page, src_page, address, vma); 663b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); 664b46e756fSKirill A. Shutemov release_pte_page(src_page); 665b46e756fSKirill A. Shutemov /* 666b46e756fSKirill A. Shutemov * ptl mostly unnecessary, but preempt has to 667b46e756fSKirill A. Shutemov * be disabled to update the per-cpu stats 668b46e756fSKirill A. Shutemov * inside page_remove_rmap(). 669b46e756fSKirill A. Shutemov */ 670b46e756fSKirill A. Shutemov spin_lock(ptl); 671b46e756fSKirill A. Shutemov /* 672b46e756fSKirill A. Shutemov * paravirt calls inside pte_clear here are 673b46e756fSKirill A. Shutemov * superfluous. 674b46e756fSKirill A. Shutemov */ 675b46e756fSKirill A. Shutemov pte_clear(vma->vm_mm, address, _pte); 676b46e756fSKirill A. Shutemov page_remove_rmap(src_page, false); 677b46e756fSKirill A. Shutemov spin_unlock(ptl); 678b46e756fSKirill A. Shutemov free_page_and_swap_cache(src_page); 679b46e756fSKirill A. Shutemov } 680b46e756fSKirill A. Shutemov } 681b46e756fSKirill A. Shutemov } 682b46e756fSKirill A. Shutemov 683b46e756fSKirill A. Shutemov static void khugepaged_alloc_sleep(void) 684b46e756fSKirill A. Shutemov { 685b46e756fSKirill A. Shutemov DEFINE_WAIT(wait); 686b46e756fSKirill A. Shutemov 687b46e756fSKirill A. Shutemov add_wait_queue(&khugepaged_wait, &wait); 688b46e756fSKirill A. Shutemov freezable_schedule_timeout_interruptible( 689b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 690b46e756fSKirill A. Shutemov remove_wait_queue(&khugepaged_wait, &wait); 691b46e756fSKirill A. Shutemov } 692b46e756fSKirill A. Shutemov 693b46e756fSKirill A. Shutemov static int khugepaged_node_load[MAX_NUMNODES]; 694b46e756fSKirill A. Shutemov 695b46e756fSKirill A. Shutemov static bool khugepaged_scan_abort(int nid) 696b46e756fSKirill A. Shutemov { 697b46e756fSKirill A. Shutemov int i; 698b46e756fSKirill A. Shutemov 699b46e756fSKirill A. Shutemov /* 700a5f5f91dSMel Gorman * If node_reclaim_mode is disabled, then no extra effort is made to 701b46e756fSKirill A. Shutemov * allocate memory locally. 702b46e756fSKirill A. Shutemov */ 703a5f5f91dSMel Gorman if (!node_reclaim_mode) 704b46e756fSKirill A. Shutemov return false; 705b46e756fSKirill A. Shutemov 706b46e756fSKirill A. Shutemov /* If there is a count for this node already, it must be acceptable */ 707b46e756fSKirill A. Shutemov if (khugepaged_node_load[nid]) 708b46e756fSKirill A. Shutemov return false; 709b46e756fSKirill A. Shutemov 710b46e756fSKirill A. Shutemov for (i = 0; i < MAX_NUMNODES; i++) { 711b46e756fSKirill A. Shutemov if (!khugepaged_node_load[i]) 712b46e756fSKirill A. Shutemov continue; 713b46e756fSKirill A. Shutemov if (node_distance(nid, i) > RECLAIM_DISTANCE) 714b46e756fSKirill A. Shutemov return true; 715b46e756fSKirill A. Shutemov } 716b46e756fSKirill A. Shutemov return false; 717b46e756fSKirill A. Shutemov } 718b46e756fSKirill A. Shutemov 719b46e756fSKirill A. Shutemov /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 720b46e756fSKirill A. Shutemov static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 721b46e756fSKirill A. Shutemov { 72225160354SVlastimil Babka return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 723b46e756fSKirill A. Shutemov } 724b46e756fSKirill A. Shutemov 725b46e756fSKirill A. Shutemov #ifdef CONFIG_NUMA 726b46e756fSKirill A. Shutemov static int khugepaged_find_target_node(void) 727b46e756fSKirill A. Shutemov { 728b46e756fSKirill A. Shutemov static int last_khugepaged_target_node = NUMA_NO_NODE; 729b46e756fSKirill A. Shutemov int nid, target_node = 0, max_value = 0; 730b46e756fSKirill A. Shutemov 731b46e756fSKirill A. Shutemov /* find first node with max normal pages hit */ 732b46e756fSKirill A. Shutemov for (nid = 0; nid < MAX_NUMNODES; nid++) 733b46e756fSKirill A. Shutemov if (khugepaged_node_load[nid] > max_value) { 734b46e756fSKirill A. Shutemov max_value = khugepaged_node_load[nid]; 735b46e756fSKirill A. Shutemov target_node = nid; 736b46e756fSKirill A. Shutemov } 737b46e756fSKirill A. Shutemov 738b46e756fSKirill A. Shutemov /* do some balance if several nodes have the same hit record */ 739b46e756fSKirill A. Shutemov if (target_node <= last_khugepaged_target_node) 740b46e756fSKirill A. Shutemov for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; 741b46e756fSKirill A. Shutemov nid++) 742b46e756fSKirill A. Shutemov if (max_value == khugepaged_node_load[nid]) { 743b46e756fSKirill A. Shutemov target_node = nid; 744b46e756fSKirill A. Shutemov break; 745b46e756fSKirill A. Shutemov } 746b46e756fSKirill A. Shutemov 747b46e756fSKirill A. Shutemov last_khugepaged_target_node = target_node; 748b46e756fSKirill A. Shutemov return target_node; 749b46e756fSKirill A. Shutemov } 750b46e756fSKirill A. Shutemov 751b46e756fSKirill A. Shutemov static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 752b46e756fSKirill A. Shutemov { 753b46e756fSKirill A. Shutemov if (IS_ERR(*hpage)) { 754b46e756fSKirill A. Shutemov if (!*wait) 755b46e756fSKirill A. Shutemov return false; 756b46e756fSKirill A. Shutemov 757b46e756fSKirill A. Shutemov *wait = false; 758b46e756fSKirill A. Shutemov *hpage = NULL; 759b46e756fSKirill A. Shutemov khugepaged_alloc_sleep(); 760b46e756fSKirill A. Shutemov } else if (*hpage) { 761b46e756fSKirill A. Shutemov put_page(*hpage); 762b46e756fSKirill A. Shutemov *hpage = NULL; 763b46e756fSKirill A. Shutemov } 764b46e756fSKirill A. Shutemov 765b46e756fSKirill A. Shutemov return true; 766b46e756fSKirill A. Shutemov } 767b46e756fSKirill A. Shutemov 768b46e756fSKirill A. Shutemov static struct page * 769988ddb71SKirill A. Shutemov khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) 770b46e756fSKirill A. Shutemov { 771b46e756fSKirill A. Shutemov VM_BUG_ON_PAGE(*hpage, *hpage); 772b46e756fSKirill A. Shutemov 773b46e756fSKirill A. Shutemov *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); 774b46e756fSKirill A. Shutemov if (unlikely(!*hpage)) { 775b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 776b46e756fSKirill A. Shutemov *hpage = ERR_PTR(-ENOMEM); 777b46e756fSKirill A. Shutemov return NULL; 778b46e756fSKirill A. Shutemov } 779b46e756fSKirill A. Shutemov 780b46e756fSKirill A. Shutemov prep_transhuge_page(*hpage); 781b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 782b46e756fSKirill A. Shutemov return *hpage; 783b46e756fSKirill A. Shutemov } 784b46e756fSKirill A. Shutemov #else 785b46e756fSKirill A. Shutemov static int khugepaged_find_target_node(void) 786b46e756fSKirill A. Shutemov { 787b46e756fSKirill A. Shutemov return 0; 788b46e756fSKirill A. Shutemov } 789b46e756fSKirill A. Shutemov 790b46e756fSKirill A. Shutemov static inline struct page *alloc_khugepaged_hugepage(void) 791b46e756fSKirill A. Shutemov { 792b46e756fSKirill A. Shutemov struct page *page; 793b46e756fSKirill A. Shutemov 794b46e756fSKirill A. Shutemov page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(), 795b46e756fSKirill A. Shutemov HPAGE_PMD_ORDER); 796b46e756fSKirill A. Shutemov if (page) 797b46e756fSKirill A. Shutemov prep_transhuge_page(page); 798b46e756fSKirill A. Shutemov return page; 799b46e756fSKirill A. Shutemov } 800b46e756fSKirill A. Shutemov 801b46e756fSKirill A. Shutemov static struct page *khugepaged_alloc_hugepage(bool *wait) 802b46e756fSKirill A. Shutemov { 803b46e756fSKirill A. Shutemov struct page *hpage; 804b46e756fSKirill A. Shutemov 805b46e756fSKirill A. Shutemov do { 806b46e756fSKirill A. Shutemov hpage = alloc_khugepaged_hugepage(); 807b46e756fSKirill A. Shutemov if (!hpage) { 808b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 809b46e756fSKirill A. Shutemov if (!*wait) 810b46e756fSKirill A. Shutemov return NULL; 811b46e756fSKirill A. Shutemov 812b46e756fSKirill A. Shutemov *wait = false; 813b46e756fSKirill A. Shutemov khugepaged_alloc_sleep(); 814b46e756fSKirill A. Shutemov } else 815b46e756fSKirill A. Shutemov count_vm_event(THP_COLLAPSE_ALLOC); 816b46e756fSKirill A. Shutemov } while (unlikely(!hpage) && likely(khugepaged_enabled())); 817b46e756fSKirill A. Shutemov 818b46e756fSKirill A. Shutemov return hpage; 819b46e756fSKirill A. Shutemov } 820b46e756fSKirill A. Shutemov 821b46e756fSKirill A. Shutemov static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 822b46e756fSKirill A. Shutemov { 823b46e756fSKirill A. Shutemov if (!*hpage) 824b46e756fSKirill A. Shutemov *hpage = khugepaged_alloc_hugepage(wait); 825b46e756fSKirill A. Shutemov 826b46e756fSKirill A. Shutemov if (unlikely(!*hpage)) 827b46e756fSKirill A. Shutemov return false; 828b46e756fSKirill A. Shutemov 829b46e756fSKirill A. Shutemov return true; 830b46e756fSKirill A. Shutemov } 831b46e756fSKirill A. Shutemov 832b46e756fSKirill A. Shutemov static struct page * 833988ddb71SKirill A. Shutemov khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) 834b46e756fSKirill A. Shutemov { 835b46e756fSKirill A. Shutemov VM_BUG_ON(!*hpage); 836b46e756fSKirill A. Shutemov 837b46e756fSKirill A. Shutemov return *hpage; 838b46e756fSKirill A. Shutemov } 839b46e756fSKirill A. Shutemov #endif 840b46e756fSKirill A. Shutemov 841b46e756fSKirill A. Shutemov /* 842b46e756fSKirill A. Shutemov * If mmap_sem temporarily dropped, revalidate vma 843b46e756fSKirill A. Shutemov * before taking mmap_sem. 844b46e756fSKirill A. Shutemov * Return 0 if succeeds, otherwise return none-zero 845b46e756fSKirill A. Shutemov * value (scan code). 846b46e756fSKirill A. Shutemov */ 847b46e756fSKirill A. Shutemov 848c131f751SKirill A. Shutemov static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 849c131f751SKirill A. Shutemov struct vm_area_struct **vmap) 850b46e756fSKirill A. Shutemov { 851b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 852b46e756fSKirill A. Shutemov unsigned long hstart, hend; 853b46e756fSKirill A. Shutemov 854b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 855b46e756fSKirill A. Shutemov return SCAN_ANY_PROCESS; 856b46e756fSKirill A. Shutemov 857c131f751SKirill A. Shutemov *vmap = vma = find_vma(mm, address); 858b46e756fSKirill A. Shutemov if (!vma) 859b46e756fSKirill A. Shutemov return SCAN_VMA_NULL; 860b46e756fSKirill A. Shutemov 861b46e756fSKirill A. Shutemov hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 862b46e756fSKirill A. Shutemov hend = vma->vm_end & HPAGE_PMD_MASK; 863b46e756fSKirill A. Shutemov if (address < hstart || address + HPAGE_PMD_SIZE > hend) 864b46e756fSKirill A. Shutemov return SCAN_ADDRESS_RANGE; 86550f8b92fSSong Liu if (!hugepage_vma_check(vma, vma->vm_flags)) 866b46e756fSKirill A. Shutemov return SCAN_VMA_CHECK; 867b46e756fSKirill A. Shutemov return 0; 868b46e756fSKirill A. Shutemov } 869b46e756fSKirill A. Shutemov 870b46e756fSKirill A. Shutemov /* 871b46e756fSKirill A. Shutemov * Bring missing pages in from swap, to complete THP collapse. 872b46e756fSKirill A. Shutemov * Only done if khugepaged_scan_pmd believes it is worthwhile. 873b46e756fSKirill A. Shutemov * 874b46e756fSKirill A. Shutemov * Called and returns without pte mapped or spinlocks held, 875b46e756fSKirill A. Shutemov * but with mmap_sem held to protect against vma changes. 876b46e756fSKirill A. Shutemov */ 877b46e756fSKirill A. Shutemov 878b46e756fSKirill A. Shutemov static bool __collapse_huge_page_swapin(struct mm_struct *mm, 879b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 8800db501f7SEbru Akagunduz unsigned long address, pmd_t *pmd, 8810db501f7SEbru Akagunduz int referenced) 882b46e756fSKirill A. Shutemov { 8832b740303SSouptick Joarder int swapped_in = 0; 8842b740303SSouptick Joarder vm_fault_t ret = 0; 88582b0f8c3SJan Kara struct vm_fault vmf = { 886b46e756fSKirill A. Shutemov .vma = vma, 887b46e756fSKirill A. Shutemov .address = address, 888b46e756fSKirill A. Shutemov .flags = FAULT_FLAG_ALLOW_RETRY, 889b46e756fSKirill A. Shutemov .pmd = pmd, 8900721ec8bSJan Kara .pgoff = linear_page_index(vma, address), 891b46e756fSKirill A. Shutemov }; 892b46e756fSKirill A. Shutemov 893982785c6SEbru Akagunduz /* we only decide to swapin, if there is enough young ptes */ 894982785c6SEbru Akagunduz if (referenced < HPAGE_PMD_NR/2) { 895982785c6SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 896982785c6SEbru Akagunduz return false; 897982785c6SEbru Akagunduz } 89882b0f8c3SJan Kara vmf.pte = pte_offset_map(pmd, address); 89982b0f8c3SJan Kara for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE; 90082b0f8c3SJan Kara vmf.pte++, vmf.address += PAGE_SIZE) { 9012994302bSJan Kara vmf.orig_pte = *vmf.pte; 9022994302bSJan Kara if (!is_swap_pte(vmf.orig_pte)) 903b46e756fSKirill A. Shutemov continue; 904b46e756fSKirill A. Shutemov swapped_in++; 9052994302bSJan Kara ret = do_swap_page(&vmf); 9060db501f7SEbru Akagunduz 907b46e756fSKirill A. Shutemov /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */ 908b46e756fSKirill A. Shutemov if (ret & VM_FAULT_RETRY) { 909b46e756fSKirill A. Shutemov down_read(&mm->mmap_sem); 91082b0f8c3SJan Kara if (hugepage_vma_revalidate(mm, address, &vmf.vma)) { 911b46e756fSKirill A. Shutemov /* vma is no longer available, don't continue to swapin */ 9120db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 913b46e756fSKirill A. Shutemov return false; 91447f863eaSEbru Akagunduz } 915b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 916835152a2SSeongJae Park if (mm_find_pmd(mm, address) != pmd) { 917835152a2SSeongJae Park trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 918b46e756fSKirill A. Shutemov return false; 919b46e756fSKirill A. Shutemov } 920835152a2SSeongJae Park } 921b46e756fSKirill A. Shutemov if (ret & VM_FAULT_ERROR) { 9220db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 923b46e756fSKirill A. Shutemov return false; 924b46e756fSKirill A. Shutemov } 925b46e756fSKirill A. Shutemov /* pte is unmapped now, we need to map it */ 92682b0f8c3SJan Kara vmf.pte = pte_offset_map(pmd, vmf.address); 927b46e756fSKirill A. Shutemov } 92882b0f8c3SJan Kara vmf.pte--; 92982b0f8c3SJan Kara pte_unmap(vmf.pte); 9300db501f7SEbru Akagunduz trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); 931b46e756fSKirill A. Shutemov return true; 932b46e756fSKirill A. Shutemov } 933b46e756fSKirill A. Shutemov 934b46e756fSKirill A. Shutemov static void collapse_huge_page(struct mm_struct *mm, 935b46e756fSKirill A. Shutemov unsigned long address, 936b46e756fSKirill A. Shutemov struct page **hpage, 9370db501f7SEbru Akagunduz int node, int referenced) 938b46e756fSKirill A. Shutemov { 939b46e756fSKirill A. Shutemov pmd_t *pmd, _pmd; 940b46e756fSKirill A. Shutemov pte_t *pte; 941b46e756fSKirill A. Shutemov pgtable_t pgtable; 942b46e756fSKirill A. Shutemov struct page *new_page; 943b46e756fSKirill A. Shutemov spinlock_t *pmd_ptl, *pte_ptl; 944b46e756fSKirill A. Shutemov int isolated = 0, result = 0; 945b46e756fSKirill A. Shutemov struct mem_cgroup *memcg; 946c131f751SKirill A. Shutemov struct vm_area_struct *vma; 947b46e756fSKirill A. Shutemov unsigned long mmun_start; /* For mmu_notifiers */ 948b46e756fSKirill A. Shutemov unsigned long mmun_end; /* For mmu_notifiers */ 949b46e756fSKirill A. Shutemov gfp_t gfp; 950b46e756fSKirill A. Shutemov 951b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 952b46e756fSKirill A. Shutemov 953b46e756fSKirill A. Shutemov /* Only allocate from the target node */ 95441b6167eSMichal Hocko gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 955b46e756fSKirill A. Shutemov 956988ddb71SKirill A. Shutemov /* 957988ddb71SKirill A. Shutemov * Before allocating the hugepage, release the mmap_sem read lock. 958988ddb71SKirill A. Shutemov * The allocation can take potentially a long time if it involves 959988ddb71SKirill A. Shutemov * sync compaction, and we do not need to hold the mmap_sem during 960988ddb71SKirill A. Shutemov * that. We will recheck the vma after taking it again in write mode. 961988ddb71SKirill A. Shutemov */ 962988ddb71SKirill A. Shutemov up_read(&mm->mmap_sem); 963988ddb71SKirill A. Shutemov new_page = khugepaged_alloc_page(hpage, gfp, node); 964b46e756fSKirill A. Shutemov if (!new_page) { 965b46e756fSKirill A. Shutemov result = SCAN_ALLOC_HUGE_PAGE_FAIL; 966b46e756fSKirill A. Shutemov goto out_nolock; 967b46e756fSKirill A. Shutemov } 968b46e756fSKirill A. Shutemov 9692a70f6a7SMichal Hocko if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { 970b46e756fSKirill A. Shutemov result = SCAN_CGROUP_CHARGE_FAIL; 971b46e756fSKirill A. Shutemov goto out_nolock; 972b46e756fSKirill A. Shutemov } 973b46e756fSKirill A. Shutemov 974b46e756fSKirill A. Shutemov down_read(&mm->mmap_sem); 975c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 976b46e756fSKirill A. Shutemov if (result) { 977b46e756fSKirill A. Shutemov mem_cgroup_cancel_charge(new_page, memcg, true); 978b46e756fSKirill A. Shutemov up_read(&mm->mmap_sem); 979b46e756fSKirill A. Shutemov goto out_nolock; 980b46e756fSKirill A. Shutemov } 981b46e756fSKirill A. Shutemov 982b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 983b46e756fSKirill A. Shutemov if (!pmd) { 984b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 985b46e756fSKirill A. Shutemov mem_cgroup_cancel_charge(new_page, memcg, true); 986b46e756fSKirill A. Shutemov up_read(&mm->mmap_sem); 987b46e756fSKirill A. Shutemov goto out_nolock; 988b46e756fSKirill A. Shutemov } 989b46e756fSKirill A. Shutemov 990b46e756fSKirill A. Shutemov /* 991b46e756fSKirill A. Shutemov * __collapse_huge_page_swapin always returns with mmap_sem locked. 99247f863eaSEbru Akagunduz * If it fails, we release mmap_sem and jump out_nolock. 993b46e756fSKirill A. Shutemov * Continuing to collapse causes inconsistency. 994b46e756fSKirill A. Shutemov */ 9950db501f7SEbru Akagunduz if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) { 996b46e756fSKirill A. Shutemov mem_cgroup_cancel_charge(new_page, memcg, true); 997b46e756fSKirill A. Shutemov up_read(&mm->mmap_sem); 998b46e756fSKirill A. Shutemov goto out_nolock; 999b46e756fSKirill A. Shutemov } 1000b46e756fSKirill A. Shutemov 1001b46e756fSKirill A. Shutemov up_read(&mm->mmap_sem); 1002b46e756fSKirill A. Shutemov /* 1003b46e756fSKirill A. Shutemov * Prevent all access to pagetables with the exception of 1004b46e756fSKirill A. Shutemov * gup_fast later handled by the ptep_clear_flush and the VM 1005b46e756fSKirill A. Shutemov * handled by the anon_vma lock + PG_lock. 1006b46e756fSKirill A. Shutemov */ 1007b46e756fSKirill A. Shutemov down_write(&mm->mmap_sem); 1008c131f751SKirill A. Shutemov result = hugepage_vma_revalidate(mm, address, &vma); 1009b46e756fSKirill A. Shutemov if (result) 1010b46e756fSKirill A. Shutemov goto out; 1011b46e756fSKirill A. Shutemov /* check if the pmd is still valid */ 1012b46e756fSKirill A. Shutemov if (mm_find_pmd(mm, address) != pmd) 1013b46e756fSKirill A. Shutemov goto out; 1014b46e756fSKirill A. Shutemov 1015b46e756fSKirill A. Shutemov anon_vma_lock_write(vma->anon_vma); 1016b46e756fSKirill A. Shutemov 1017b46e756fSKirill A. Shutemov pte = pte_offset_map(pmd, address); 1018b46e756fSKirill A. Shutemov pte_ptl = pte_lockptr(mm, pmd); 1019b46e756fSKirill A. Shutemov 1020b46e756fSKirill A. Shutemov mmun_start = address; 1021b46e756fSKirill A. Shutemov mmun_end = address + HPAGE_PMD_SIZE; 1022b46e756fSKirill A. Shutemov mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1023b46e756fSKirill A. Shutemov pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1024b46e756fSKirill A. Shutemov /* 1025b46e756fSKirill A. Shutemov * After this gup_fast can't run anymore. This also removes 1026b46e756fSKirill A. Shutemov * any huge TLB entry from the CPU so we won't allow 1027b46e756fSKirill A. Shutemov * huge and small TLB entries for the same virtual address 1028b46e756fSKirill A. Shutemov * to avoid the risk of CPU bugs in that area. 1029b46e756fSKirill A. Shutemov */ 1030b46e756fSKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, address, pmd); 1031b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1032b46e756fSKirill A. Shutemov mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1033b46e756fSKirill A. Shutemov 1034b46e756fSKirill A. Shutemov spin_lock(pte_ptl); 1035b46e756fSKirill A. Shutemov isolated = __collapse_huge_page_isolate(vma, address, pte); 1036b46e756fSKirill A. Shutemov spin_unlock(pte_ptl); 1037b46e756fSKirill A. Shutemov 1038b46e756fSKirill A. Shutemov if (unlikely(!isolated)) { 1039b46e756fSKirill A. Shutemov pte_unmap(pte); 1040b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1041b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1042b46e756fSKirill A. Shutemov /* 1043b46e756fSKirill A. Shutemov * We can only use set_pmd_at when establishing 1044b46e756fSKirill A. Shutemov * hugepmds and never for establishing regular pmds that 1045b46e756fSKirill A. Shutemov * points to regular pagetables. Use pmd_populate for that 1046b46e756fSKirill A. Shutemov */ 1047b46e756fSKirill A. Shutemov pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1048b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1049b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1050b46e756fSKirill A. Shutemov result = SCAN_FAIL; 1051b46e756fSKirill A. Shutemov goto out; 1052b46e756fSKirill A. Shutemov } 1053b46e756fSKirill A. Shutemov 1054b46e756fSKirill A. Shutemov /* 1055b46e756fSKirill A. Shutemov * All pages are isolated and locked so anon_vma rmap 1056b46e756fSKirill A. Shutemov * can't run anymore. 1057b46e756fSKirill A. Shutemov */ 1058b46e756fSKirill A. Shutemov anon_vma_unlock_write(vma->anon_vma); 1059b46e756fSKirill A. Shutemov 1060b46e756fSKirill A. Shutemov __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); 1061b46e756fSKirill A. Shutemov pte_unmap(pte); 1062b46e756fSKirill A. Shutemov __SetPageUptodate(new_page); 1063b46e756fSKirill A. Shutemov pgtable = pmd_pgtable(_pmd); 1064b46e756fSKirill A. Shutemov 1065b46e756fSKirill A. Shutemov _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); 1066f55e1014SLinus Torvalds _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1067b46e756fSKirill A. Shutemov 1068b46e756fSKirill A. Shutemov /* 1069b46e756fSKirill A. Shutemov * spin_lock() below is not the equivalent of smp_wmb(), so 1070b46e756fSKirill A. Shutemov * this is needed to avoid the copy_huge_page writes to become 1071b46e756fSKirill A. Shutemov * visible after the set_pmd_at() write. 1072b46e756fSKirill A. Shutemov */ 1073b46e756fSKirill A. Shutemov smp_wmb(); 1074b46e756fSKirill A. Shutemov 1075b46e756fSKirill A. Shutemov spin_lock(pmd_ptl); 1076b46e756fSKirill A. Shutemov BUG_ON(!pmd_none(*pmd)); 1077b46e756fSKirill A. Shutemov page_add_new_anon_rmap(new_page, vma, address, true); 1078b46e756fSKirill A. Shutemov mem_cgroup_commit_charge(new_page, memcg, false, true); 1079b46e756fSKirill A. Shutemov lru_cache_add_active_or_unevictable(new_page, vma); 1080b46e756fSKirill A. Shutemov pgtable_trans_huge_deposit(mm, pmd, pgtable); 1081b46e756fSKirill A. Shutemov set_pmd_at(mm, address, pmd, _pmd); 1082b46e756fSKirill A. Shutemov update_mmu_cache_pmd(vma, address, pmd); 1083b46e756fSKirill A. Shutemov spin_unlock(pmd_ptl); 1084b46e756fSKirill A. Shutemov 1085b46e756fSKirill A. Shutemov *hpage = NULL; 1086b46e756fSKirill A. Shutemov 1087b46e756fSKirill A. Shutemov khugepaged_pages_collapsed++; 1088b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1089b46e756fSKirill A. Shutemov out_up_write: 1090b46e756fSKirill A. Shutemov up_write(&mm->mmap_sem); 1091b46e756fSKirill A. Shutemov out_nolock: 1092b46e756fSKirill A. Shutemov trace_mm_collapse_huge_page(mm, isolated, result); 1093b46e756fSKirill A. Shutemov return; 1094b46e756fSKirill A. Shutemov out: 1095b46e756fSKirill A. Shutemov mem_cgroup_cancel_charge(new_page, memcg, true); 1096b46e756fSKirill A. Shutemov goto out_up_write; 1097b46e756fSKirill A. Shutemov } 1098b46e756fSKirill A. Shutemov 1099b46e756fSKirill A. Shutemov static int khugepaged_scan_pmd(struct mm_struct *mm, 1100b46e756fSKirill A. Shutemov struct vm_area_struct *vma, 1101b46e756fSKirill A. Shutemov unsigned long address, 1102b46e756fSKirill A. Shutemov struct page **hpage) 1103b46e756fSKirill A. Shutemov { 1104b46e756fSKirill A. Shutemov pmd_t *pmd; 1105b46e756fSKirill A. Shutemov pte_t *pte, *_pte; 11060db501f7SEbru Akagunduz int ret = 0, none_or_zero = 0, result = 0, referenced = 0; 1107b46e756fSKirill A. Shutemov struct page *page = NULL; 1108b46e756fSKirill A. Shutemov unsigned long _address; 1109b46e756fSKirill A. Shutemov spinlock_t *ptl; 1110b46e756fSKirill A. Shutemov int node = NUMA_NO_NODE, unmapped = 0; 11110db501f7SEbru Akagunduz bool writable = false; 1112b46e756fSKirill A. Shutemov 1113b46e756fSKirill A. Shutemov VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1114b46e756fSKirill A. Shutemov 1115b46e756fSKirill A. Shutemov pmd = mm_find_pmd(mm, address); 1116b46e756fSKirill A. Shutemov if (!pmd) { 1117b46e756fSKirill A. Shutemov result = SCAN_PMD_NULL; 1118b46e756fSKirill A. Shutemov goto out; 1119b46e756fSKirill A. Shutemov } 1120b46e756fSKirill A. Shutemov 1121b46e756fSKirill A. Shutemov memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 1122b46e756fSKirill A. Shutemov pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1123b46e756fSKirill A. Shutemov for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 1124b46e756fSKirill A. Shutemov _pte++, _address += PAGE_SIZE) { 1125b46e756fSKirill A. Shutemov pte_t pteval = *_pte; 1126b46e756fSKirill A. Shutemov if (is_swap_pte(pteval)) { 1127b46e756fSKirill A. Shutemov if (++unmapped <= khugepaged_max_ptes_swap) { 1128b46e756fSKirill A. Shutemov continue; 1129b46e756fSKirill A. Shutemov } else { 1130b46e756fSKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1131b46e756fSKirill A. Shutemov goto out_unmap; 1132b46e756fSKirill A. Shutemov } 1133b46e756fSKirill A. Shutemov } 1134b46e756fSKirill A. Shutemov if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1135b46e756fSKirill A. Shutemov if (!userfaultfd_armed(vma) && 1136b46e756fSKirill A. Shutemov ++none_or_zero <= khugepaged_max_ptes_none) { 1137b46e756fSKirill A. Shutemov continue; 1138b46e756fSKirill A. Shutemov } else { 1139b46e756fSKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1140b46e756fSKirill A. Shutemov goto out_unmap; 1141b46e756fSKirill A. Shutemov } 1142b46e756fSKirill A. Shutemov } 1143b46e756fSKirill A. Shutemov if (!pte_present(pteval)) { 1144b46e756fSKirill A. Shutemov result = SCAN_PTE_NON_PRESENT; 1145b46e756fSKirill A. Shutemov goto out_unmap; 1146b46e756fSKirill A. Shutemov } 1147b46e756fSKirill A. Shutemov if (pte_write(pteval)) 1148b46e756fSKirill A. Shutemov writable = true; 1149b46e756fSKirill A. Shutemov 1150b46e756fSKirill A. Shutemov page = vm_normal_page(vma, _address, pteval); 1151b46e756fSKirill A. Shutemov if (unlikely(!page)) { 1152b46e756fSKirill A. Shutemov result = SCAN_PAGE_NULL; 1153b46e756fSKirill A. Shutemov goto out_unmap; 1154b46e756fSKirill A. Shutemov } 1155b46e756fSKirill A. Shutemov 1156b46e756fSKirill A. Shutemov /* TODO: teach khugepaged to collapse THP mapped with pte */ 1157b46e756fSKirill A. Shutemov if (PageCompound(page)) { 1158b46e756fSKirill A. Shutemov result = SCAN_PAGE_COMPOUND; 1159b46e756fSKirill A. Shutemov goto out_unmap; 1160b46e756fSKirill A. Shutemov } 1161b46e756fSKirill A. Shutemov 1162b46e756fSKirill A. Shutemov /* 1163b46e756fSKirill A. Shutemov * Record which node the original page is from and save this 1164b46e756fSKirill A. Shutemov * information to khugepaged_node_load[]. 1165b46e756fSKirill A. Shutemov * Khupaged will allocate hugepage from the node has the max 1166b46e756fSKirill A. Shutemov * hit record. 1167b46e756fSKirill A. Shutemov */ 1168b46e756fSKirill A. Shutemov node = page_to_nid(page); 1169b46e756fSKirill A. Shutemov if (khugepaged_scan_abort(node)) { 1170b46e756fSKirill A. Shutemov result = SCAN_SCAN_ABORT; 1171b46e756fSKirill A. Shutemov goto out_unmap; 1172b46e756fSKirill A. Shutemov } 1173b46e756fSKirill A. Shutemov khugepaged_node_load[node]++; 1174b46e756fSKirill A. Shutemov if (!PageLRU(page)) { 1175b46e756fSKirill A. Shutemov result = SCAN_PAGE_LRU; 1176b46e756fSKirill A. Shutemov goto out_unmap; 1177b46e756fSKirill A. Shutemov } 1178b46e756fSKirill A. Shutemov if (PageLocked(page)) { 1179b46e756fSKirill A. Shutemov result = SCAN_PAGE_LOCK; 1180b46e756fSKirill A. Shutemov goto out_unmap; 1181b46e756fSKirill A. Shutemov } 1182b46e756fSKirill A. Shutemov if (!PageAnon(page)) { 1183b46e756fSKirill A. Shutemov result = SCAN_PAGE_ANON; 1184b46e756fSKirill A. Shutemov goto out_unmap; 1185b46e756fSKirill A. Shutemov } 1186b46e756fSKirill A. Shutemov 1187b46e756fSKirill A. Shutemov /* 1188b46e756fSKirill A. Shutemov * cannot use mapcount: can't collapse if there's a gup pin. 1189b46e756fSKirill A. Shutemov * The page must only be referenced by the scanned process 1190b46e756fSKirill A. Shutemov * and page swap cache. 1191b46e756fSKirill A. Shutemov */ 11922948be5aSMinchan Kim if (page_count(page) != 1 + PageSwapCache(page)) { 1193b46e756fSKirill A. Shutemov result = SCAN_PAGE_COUNT; 1194b46e756fSKirill A. Shutemov goto out_unmap; 1195b46e756fSKirill A. Shutemov } 1196b46e756fSKirill A. Shutemov if (pte_young(pteval) || 1197b46e756fSKirill A. Shutemov page_is_young(page) || PageReferenced(page) || 1198b46e756fSKirill A. Shutemov mmu_notifier_test_young(vma->vm_mm, address)) 11990db501f7SEbru Akagunduz referenced++; 1200b46e756fSKirill A. Shutemov } 1201b46e756fSKirill A. Shutemov if (writable) { 1202b46e756fSKirill A. Shutemov if (referenced) { 1203b46e756fSKirill A. Shutemov result = SCAN_SUCCEED; 1204b46e756fSKirill A. Shutemov ret = 1; 1205b46e756fSKirill A. Shutemov } else { 12060db501f7SEbru Akagunduz result = SCAN_LACK_REFERENCED_PAGE; 1207b46e756fSKirill A. Shutemov } 1208b46e756fSKirill A. Shutemov } else { 1209b46e756fSKirill A. Shutemov result = SCAN_PAGE_RO; 1210b46e756fSKirill A. Shutemov } 1211b46e756fSKirill A. Shutemov out_unmap: 1212b46e756fSKirill A. Shutemov pte_unmap_unlock(pte, ptl); 1213b46e756fSKirill A. Shutemov if (ret) { 1214b46e756fSKirill A. Shutemov node = khugepaged_find_target_node(); 1215b46e756fSKirill A. Shutemov /* collapse_huge_page will return with the mmap_sem released */ 1216c131f751SKirill A. Shutemov collapse_huge_page(mm, address, hpage, node, referenced); 1217b46e756fSKirill A. Shutemov } 1218b46e756fSKirill A. Shutemov out: 1219b46e756fSKirill A. Shutemov trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, 1220b46e756fSKirill A. Shutemov none_or_zero, result, unmapped); 1221b46e756fSKirill A. Shutemov return ret; 1222b46e756fSKirill A. Shutemov } 1223b46e756fSKirill A. Shutemov 1224b46e756fSKirill A. Shutemov static void collect_mm_slot(struct mm_slot *mm_slot) 1225b46e756fSKirill A. Shutemov { 1226b46e756fSKirill A. Shutemov struct mm_struct *mm = mm_slot->mm; 1227b46e756fSKirill A. Shutemov 1228b46e756fSKirill A. Shutemov VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 1229b46e756fSKirill A. Shutemov 1230b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm)) { 1231b46e756fSKirill A. Shutemov /* free mm_slot */ 1232b46e756fSKirill A. Shutemov hash_del(&mm_slot->hash); 1233b46e756fSKirill A. Shutemov list_del(&mm_slot->mm_node); 1234b46e756fSKirill A. Shutemov 1235b46e756fSKirill A. Shutemov /* 1236b46e756fSKirill A. Shutemov * Not strictly needed because the mm exited already. 1237b46e756fSKirill A. Shutemov * 1238b46e756fSKirill A. Shutemov * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1239b46e756fSKirill A. Shutemov */ 1240b46e756fSKirill A. Shutemov 1241b46e756fSKirill A. Shutemov /* khugepaged_mm_lock actually not necessary for the below */ 1242b46e756fSKirill A. Shutemov free_mm_slot(mm_slot); 1243b46e756fSKirill A. Shutemov mmdrop(mm); 1244b46e756fSKirill A. Shutemov } 1245b46e756fSKirill A. Shutemov } 1246b46e756fSKirill A. Shutemov 1247e496cf3dSKirill A. Shutemov #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) 1248f3f0e1d2SKirill A. Shutemov static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1249f3f0e1d2SKirill A. Shutemov { 1250f3f0e1d2SKirill A. Shutemov struct vm_area_struct *vma; 1251f3f0e1d2SKirill A. Shutemov unsigned long addr; 1252f3f0e1d2SKirill A. Shutemov pmd_t *pmd, _pmd; 1253f3f0e1d2SKirill A. Shutemov 1254f3f0e1d2SKirill A. Shutemov i_mmap_lock_write(mapping); 1255f3f0e1d2SKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 1256f3f0e1d2SKirill A. Shutemov /* probably overkill */ 1257f3f0e1d2SKirill A. Shutemov if (vma->anon_vma) 1258f3f0e1d2SKirill A. Shutemov continue; 1259f3f0e1d2SKirill A. Shutemov addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 1260f3f0e1d2SKirill A. Shutemov if (addr & ~HPAGE_PMD_MASK) 1261f3f0e1d2SKirill A. Shutemov continue; 1262f3f0e1d2SKirill A. Shutemov if (vma->vm_end < addr + HPAGE_PMD_SIZE) 1263f3f0e1d2SKirill A. Shutemov continue; 1264f3f0e1d2SKirill A. Shutemov pmd = mm_find_pmd(vma->vm_mm, addr); 1265f3f0e1d2SKirill A. Shutemov if (!pmd) 1266f3f0e1d2SKirill A. Shutemov continue; 1267f3f0e1d2SKirill A. Shutemov /* 1268f3f0e1d2SKirill A. Shutemov * We need exclusive mmap_sem to retract page table. 1269f3f0e1d2SKirill A. Shutemov * If trylock fails we would end up with pte-mapped THP after 1270f3f0e1d2SKirill A. Shutemov * re-fault. Not ideal, but it's more important to not disturb 1271f3f0e1d2SKirill A. Shutemov * the system too much. 1272f3f0e1d2SKirill A. Shutemov */ 1273f3f0e1d2SKirill A. Shutemov if (down_write_trylock(&vma->vm_mm->mmap_sem)) { 1274f3f0e1d2SKirill A. Shutemov spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd); 1275f3f0e1d2SKirill A. Shutemov /* assume page table is clear */ 1276f3f0e1d2SKirill A. Shutemov _pmd = pmdp_collapse_flush(vma, addr, pmd); 1277f3f0e1d2SKirill A. Shutemov spin_unlock(ptl); 1278f3f0e1d2SKirill A. Shutemov up_write(&vma->vm_mm->mmap_sem); 1279c4812909SKirill A. Shutemov mm_dec_nr_ptes(vma->vm_mm); 1280f3f0e1d2SKirill A. Shutemov pte_free(vma->vm_mm, pmd_pgtable(_pmd)); 1281f3f0e1d2SKirill A. Shutemov } 1282f3f0e1d2SKirill A. Shutemov } 1283f3f0e1d2SKirill A. Shutemov i_mmap_unlock_write(mapping); 1284f3f0e1d2SKirill A. Shutemov } 1285f3f0e1d2SKirill A. Shutemov 1286f3f0e1d2SKirill A. Shutemov /** 1287f3f0e1d2SKirill A. Shutemov * collapse_shmem - collapse small tmpfs/shmem pages into huge one. 1288f3f0e1d2SKirill A. Shutemov * 1289f3f0e1d2SKirill A. Shutemov * Basic scheme is simple, details are more complex: 1290f3f0e1d2SKirill A. Shutemov * - allocate and freeze a new huge page; 1291f3f0e1d2SKirill A. Shutemov * - scan over radix tree replacing old pages the new one 1292f3f0e1d2SKirill A. Shutemov * + swap in pages if necessary; 1293f3f0e1d2SKirill A. Shutemov * + fill in gaps; 1294f3f0e1d2SKirill A. Shutemov * + keep old pages around in case if rollback is required; 1295f3f0e1d2SKirill A. Shutemov * - if replacing succeed: 1296f3f0e1d2SKirill A. Shutemov * + copy data over; 1297f3f0e1d2SKirill A. Shutemov * + free old pages; 1298f3f0e1d2SKirill A. Shutemov * + unfreeze huge page; 1299f3f0e1d2SKirill A. Shutemov * - if replacing failed; 1300f3f0e1d2SKirill A. Shutemov * + put all pages back and unfreeze them; 1301f3f0e1d2SKirill A. Shutemov * + restore gaps in the radix-tree; 1302f3f0e1d2SKirill A. Shutemov * + free huge page; 1303f3f0e1d2SKirill A. Shutemov */ 1304f3f0e1d2SKirill A. Shutemov static void collapse_shmem(struct mm_struct *mm, 1305f3f0e1d2SKirill A. Shutemov struct address_space *mapping, pgoff_t start, 1306f3f0e1d2SKirill A. Shutemov struct page **hpage, int node) 1307f3f0e1d2SKirill A. Shutemov { 1308f3f0e1d2SKirill A. Shutemov gfp_t gfp; 1309f3f0e1d2SKirill A. Shutemov struct page *page, *new_page, *tmp; 1310f3f0e1d2SKirill A. Shutemov struct mem_cgroup *memcg; 1311f3f0e1d2SKirill A. Shutemov pgoff_t index, end = start + HPAGE_PMD_NR; 1312f3f0e1d2SKirill A. Shutemov LIST_HEAD(pagelist); 1313f3f0e1d2SKirill A. Shutemov struct radix_tree_iter iter; 1314f3f0e1d2SKirill A. Shutemov void **slot; 1315f3f0e1d2SKirill A. Shutemov int nr_none = 0, result = SCAN_SUCCEED; 1316f3f0e1d2SKirill A. Shutemov 1317f3f0e1d2SKirill A. Shutemov VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1318f3f0e1d2SKirill A. Shutemov 1319f3f0e1d2SKirill A. Shutemov /* Only allocate from the target node */ 132041b6167eSMichal Hocko gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 1321f3f0e1d2SKirill A. Shutemov 1322f3f0e1d2SKirill A. Shutemov new_page = khugepaged_alloc_page(hpage, gfp, node); 1323f3f0e1d2SKirill A. Shutemov if (!new_page) { 1324f3f0e1d2SKirill A. Shutemov result = SCAN_ALLOC_HUGE_PAGE_FAIL; 1325f3f0e1d2SKirill A. Shutemov goto out; 1326f3f0e1d2SKirill A. Shutemov } 1327f3f0e1d2SKirill A. Shutemov 13282a70f6a7SMichal Hocko if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { 1329f3f0e1d2SKirill A. Shutemov result = SCAN_CGROUP_CHARGE_FAIL; 1330f3f0e1d2SKirill A. Shutemov goto out; 1331f3f0e1d2SKirill A. Shutemov } 1332f3f0e1d2SKirill A. Shutemov 1333f3f0e1d2SKirill A. Shutemov new_page->index = start; 1334f3f0e1d2SKirill A. Shutemov new_page->mapping = mapping; 1335f3f0e1d2SKirill A. Shutemov __SetPageSwapBacked(new_page); 1336f3f0e1d2SKirill A. Shutemov __SetPageLocked(new_page); 1337f3f0e1d2SKirill A. Shutemov BUG_ON(!page_ref_freeze(new_page, 1)); 1338f3f0e1d2SKirill A. Shutemov 1339f3f0e1d2SKirill A. Shutemov 1340f3f0e1d2SKirill A. Shutemov /* 1341f3f0e1d2SKirill A. Shutemov * At this point the new_page is 'frozen' (page_count() is zero), locked 1342f3f0e1d2SKirill A. Shutemov * and not up-to-date. It's safe to insert it into radix tree, because 1343f3f0e1d2SKirill A. Shutemov * nobody would be able to map it or use it in other way until we 1344f3f0e1d2SKirill A. Shutemov * unfreeze it. 1345f3f0e1d2SKirill A. Shutemov */ 1346f3f0e1d2SKirill A. Shutemov 1347f3f0e1d2SKirill A. Shutemov index = start; 1348b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 1349b93b0163SMatthew Wilcox radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) { 1350f3f0e1d2SKirill A. Shutemov int n = min(iter.index, end) - index; 1351f3f0e1d2SKirill A. Shutemov 1352f3f0e1d2SKirill A. Shutemov /* 1353f3f0e1d2SKirill A. Shutemov * Handle holes in the radix tree: charge it from shmem and 1354f3f0e1d2SKirill A. Shutemov * insert relevant subpage of new_page into the radix-tree. 1355f3f0e1d2SKirill A. Shutemov */ 1356f3f0e1d2SKirill A. Shutemov if (n && !shmem_charge(mapping->host, n)) { 1357f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1358f3f0e1d2SKirill A. Shutemov break; 1359f3f0e1d2SKirill A. Shutemov } 1360f3f0e1d2SKirill A. Shutemov nr_none += n; 1361f3f0e1d2SKirill A. Shutemov for (; index < min(iter.index, end); index++) { 1362b93b0163SMatthew Wilcox radix_tree_insert(&mapping->i_pages, index, 1363f3f0e1d2SKirill A. Shutemov new_page + (index % HPAGE_PMD_NR)); 1364f3f0e1d2SKirill A. Shutemov } 1365f3f0e1d2SKirill A. Shutemov 1366f3f0e1d2SKirill A. Shutemov /* We are done. */ 1367f3f0e1d2SKirill A. Shutemov if (index >= end) 1368f3f0e1d2SKirill A. Shutemov break; 1369f3f0e1d2SKirill A. Shutemov 1370f3f0e1d2SKirill A. Shutemov page = radix_tree_deref_slot_protected(slot, 1371b93b0163SMatthew Wilcox &mapping->i_pages.xa_lock); 1372*3159f943SMatthew Wilcox if (xa_is_value(page) || !PageUptodate(page)) { 1373b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 1374f3f0e1d2SKirill A. Shutemov /* swap in or instantiate fallocated page */ 1375f3f0e1d2SKirill A. Shutemov if (shmem_getpage(mapping->host, index, &page, 1376f3f0e1d2SKirill A. Shutemov SGP_NOHUGE)) { 1377f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1378f3f0e1d2SKirill A. Shutemov goto tree_unlocked; 1379f3f0e1d2SKirill A. Shutemov } 1380b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 1381f3f0e1d2SKirill A. Shutemov } else if (trylock_page(page)) { 1382f3f0e1d2SKirill A. Shutemov get_page(page); 1383f3f0e1d2SKirill A. Shutemov } else { 1384f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LOCK; 1385f3f0e1d2SKirill A. Shutemov break; 1386f3f0e1d2SKirill A. Shutemov } 1387f3f0e1d2SKirill A. Shutemov 1388f3f0e1d2SKirill A. Shutemov /* 1389b93b0163SMatthew Wilcox * The page must be locked, so we can drop the i_pages lock 1390f3f0e1d2SKirill A. Shutemov * without racing with truncate. 1391f3f0e1d2SKirill A. Shutemov */ 1392f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 1393f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(!PageUptodate(page), page); 1394f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(PageTransCompound(page), page); 1395f3f0e1d2SKirill A. Shutemov 1396f3f0e1d2SKirill A. Shutemov if (page_mapping(page) != mapping) { 1397f3f0e1d2SKirill A. Shutemov result = SCAN_TRUNCATED; 1398f3f0e1d2SKirill A. Shutemov goto out_unlock; 1399f3f0e1d2SKirill A. Shutemov } 1400b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 1401f3f0e1d2SKirill A. Shutemov 1402f3f0e1d2SKirill A. Shutemov if (isolate_lru_page(page)) { 1403f3f0e1d2SKirill A. Shutemov result = SCAN_DEL_PAGE_LRU; 1404f3f0e1d2SKirill A. Shutemov goto out_isolate_failed; 1405f3f0e1d2SKirill A. Shutemov } 1406f3f0e1d2SKirill A. Shutemov 1407f3f0e1d2SKirill A. Shutemov if (page_mapped(page)) 1408977fbdcdSMatthew Wilcox unmap_mapping_pages(mapping, index, 1, false); 1409f3f0e1d2SKirill A. Shutemov 1410b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 1411f3f0e1d2SKirill A. Shutemov 1412b93b0163SMatthew Wilcox slot = radix_tree_lookup_slot(&mapping->i_pages, index); 141391a45f71SJohannes Weiner VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot, 1414b93b0163SMatthew Wilcox &mapping->i_pages.xa_lock), page); 1415f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(page_mapped(page), page); 1416f3f0e1d2SKirill A. Shutemov 1417f3f0e1d2SKirill A. Shutemov /* 1418f3f0e1d2SKirill A. Shutemov * The page is expected to have page_count() == 3: 1419f3f0e1d2SKirill A. Shutemov * - we hold a pin on it; 1420f3f0e1d2SKirill A. Shutemov * - one reference from radix tree; 1421f3f0e1d2SKirill A. Shutemov * - one from isolate_lru_page; 1422f3f0e1d2SKirill A. Shutemov */ 1423f3f0e1d2SKirill A. Shutemov if (!page_ref_freeze(page, 3)) { 1424f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1425f3f0e1d2SKirill A. Shutemov goto out_lru; 1426f3f0e1d2SKirill A. Shutemov } 1427f3f0e1d2SKirill A. Shutemov 1428f3f0e1d2SKirill A. Shutemov /* 1429f3f0e1d2SKirill A. Shutemov * Add the page to the list to be able to undo the collapse if 1430f3f0e1d2SKirill A. Shutemov * something go wrong. 1431f3f0e1d2SKirill A. Shutemov */ 1432f3f0e1d2SKirill A. Shutemov list_add_tail(&page->lru, &pagelist); 1433f3f0e1d2SKirill A. Shutemov 1434f3f0e1d2SKirill A. Shutemov /* Finally, replace with the new page. */ 1435b93b0163SMatthew Wilcox radix_tree_replace_slot(&mapping->i_pages, slot, 1436f3f0e1d2SKirill A. Shutemov new_page + (index % HPAGE_PMD_NR)); 1437f3f0e1d2SKirill A. Shutemov 1438148deab2SMatthew Wilcox slot = radix_tree_iter_resume(slot, &iter); 1439f3f0e1d2SKirill A. Shutemov index++; 1440f3f0e1d2SKirill A. Shutemov continue; 1441f3f0e1d2SKirill A. Shutemov out_lru: 1442b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 1443f3f0e1d2SKirill A. Shutemov putback_lru_page(page); 1444f3f0e1d2SKirill A. Shutemov out_isolate_failed: 1445f3f0e1d2SKirill A. Shutemov unlock_page(page); 1446f3f0e1d2SKirill A. Shutemov put_page(page); 1447f3f0e1d2SKirill A. Shutemov goto tree_unlocked; 1448f3f0e1d2SKirill A. Shutemov out_unlock: 1449f3f0e1d2SKirill A. Shutemov unlock_page(page); 1450f3f0e1d2SKirill A. Shutemov put_page(page); 1451f3f0e1d2SKirill A. Shutemov break; 1452f3f0e1d2SKirill A. Shutemov } 1453f3f0e1d2SKirill A. Shutemov 1454f3f0e1d2SKirill A. Shutemov /* 1455f3f0e1d2SKirill A. Shutemov * Handle hole in radix tree at the end of the range. 1456f3f0e1d2SKirill A. Shutemov * This code only triggers if there's nothing in radix tree 1457f3f0e1d2SKirill A. Shutemov * beyond 'end'. 1458f3f0e1d2SKirill A. Shutemov */ 1459f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED && index < end) { 1460f3f0e1d2SKirill A. Shutemov int n = end - index; 1461f3f0e1d2SKirill A. Shutemov 1462f3f0e1d2SKirill A. Shutemov if (!shmem_charge(mapping->host, n)) { 1463f3f0e1d2SKirill A. Shutemov result = SCAN_FAIL; 1464f3f0e1d2SKirill A. Shutemov goto tree_locked; 1465f3f0e1d2SKirill A. Shutemov } 1466f3f0e1d2SKirill A. Shutemov 1467f3f0e1d2SKirill A. Shutemov for (; index < end; index++) { 1468b93b0163SMatthew Wilcox radix_tree_insert(&mapping->i_pages, index, 1469f3f0e1d2SKirill A. Shutemov new_page + (index % HPAGE_PMD_NR)); 1470f3f0e1d2SKirill A. Shutemov } 1471f3f0e1d2SKirill A. Shutemov nr_none += n; 1472f3f0e1d2SKirill A. Shutemov } 1473f3f0e1d2SKirill A. Shutemov 1474f3f0e1d2SKirill A. Shutemov tree_locked: 1475b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 1476f3f0e1d2SKirill A. Shutemov tree_unlocked: 1477f3f0e1d2SKirill A. Shutemov 1478f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 1479f3f0e1d2SKirill A. Shutemov unsigned long flags; 1480f3f0e1d2SKirill A. Shutemov struct zone *zone = page_zone(new_page); 1481f3f0e1d2SKirill A. Shutemov 1482f3f0e1d2SKirill A. Shutemov /* 1483f3f0e1d2SKirill A. Shutemov * Replacing old pages with new one has succeed, now we need to 1484f3f0e1d2SKirill A. Shutemov * copy the content and free old pages. 1485f3f0e1d2SKirill A. Shutemov */ 1486f3f0e1d2SKirill A. Shutemov list_for_each_entry_safe(page, tmp, &pagelist, lru) { 1487f3f0e1d2SKirill A. Shutemov copy_highpage(new_page + (page->index % HPAGE_PMD_NR), 1488f3f0e1d2SKirill A. Shutemov page); 1489f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1490f3f0e1d2SKirill A. Shutemov unlock_page(page); 1491f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(page, 1); 1492f3f0e1d2SKirill A. Shutemov page->mapping = NULL; 1493f3f0e1d2SKirill A. Shutemov ClearPageActive(page); 1494f3f0e1d2SKirill A. Shutemov ClearPageUnevictable(page); 1495f3f0e1d2SKirill A. Shutemov put_page(page); 1496f3f0e1d2SKirill A. Shutemov } 1497f3f0e1d2SKirill A. Shutemov 1498f3f0e1d2SKirill A. Shutemov local_irq_save(flags); 149911fb9989SMel Gorman __inc_node_page_state(new_page, NR_SHMEM_THPS); 1500f3f0e1d2SKirill A. Shutemov if (nr_none) { 150111fb9989SMel Gorman __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none); 150211fb9989SMel Gorman __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none); 1503f3f0e1d2SKirill A. Shutemov } 1504f3f0e1d2SKirill A. Shutemov local_irq_restore(flags); 1505f3f0e1d2SKirill A. Shutemov 1506f3f0e1d2SKirill A. Shutemov /* 1507f3f0e1d2SKirill A. Shutemov * Remove pte page tables, so we can re-faulti 1508f3f0e1d2SKirill A. Shutemov * the page as huge. 1509f3f0e1d2SKirill A. Shutemov */ 1510f3f0e1d2SKirill A. Shutemov retract_page_tables(mapping, start); 1511f3f0e1d2SKirill A. Shutemov 1512f3f0e1d2SKirill A. Shutemov /* Everything is ready, let's unfreeze the new_page */ 1513f3f0e1d2SKirill A. Shutemov set_page_dirty(new_page); 1514f3f0e1d2SKirill A. Shutemov SetPageUptodate(new_page); 1515f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(new_page, HPAGE_PMD_NR); 1516f3f0e1d2SKirill A. Shutemov mem_cgroup_commit_charge(new_page, memcg, false, true); 1517f3f0e1d2SKirill A. Shutemov lru_cache_add_anon(new_page); 1518f3f0e1d2SKirill A. Shutemov unlock_page(new_page); 1519f3f0e1d2SKirill A. Shutemov 1520f3f0e1d2SKirill A. Shutemov *hpage = NULL; 152187aa7529SYang Shi 152287aa7529SYang Shi khugepaged_pages_collapsed++; 1523f3f0e1d2SKirill A. Shutemov } else { 1524f3f0e1d2SKirill A. Shutemov /* Something went wrong: rollback changes to the radix-tree */ 1525f3f0e1d2SKirill A. Shutemov shmem_uncharge(mapping->host, nr_none); 1526b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 1527b93b0163SMatthew Wilcox radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) { 1528f3f0e1d2SKirill A. Shutemov if (iter.index >= end) 1529f3f0e1d2SKirill A. Shutemov break; 1530f3f0e1d2SKirill A. Shutemov page = list_first_entry_or_null(&pagelist, 1531f3f0e1d2SKirill A. Shutemov struct page, lru); 1532f3f0e1d2SKirill A. Shutemov if (!page || iter.index < page->index) { 1533f3f0e1d2SKirill A. Shutemov if (!nr_none) 1534f3f0e1d2SKirill A. Shutemov break; 1535f3f0e1d2SKirill A. Shutemov nr_none--; 153659749e6cSJohannes Weiner /* Put holes back where they were */ 1537b93b0163SMatthew Wilcox radix_tree_delete(&mapping->i_pages, iter.index); 1538f3f0e1d2SKirill A. Shutemov continue; 1539f3f0e1d2SKirill A. Shutemov } 1540f3f0e1d2SKirill A. Shutemov 1541f3f0e1d2SKirill A. Shutemov VM_BUG_ON_PAGE(page->index != iter.index, page); 1542f3f0e1d2SKirill A. Shutemov 1543f3f0e1d2SKirill A. Shutemov /* Unfreeze the page. */ 1544f3f0e1d2SKirill A. Shutemov list_del(&page->lru); 1545f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(page, 2); 1546b93b0163SMatthew Wilcox radix_tree_replace_slot(&mapping->i_pages, slot, page); 1547148deab2SMatthew Wilcox slot = radix_tree_iter_resume(slot, &iter); 1548b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 1549f3f0e1d2SKirill A. Shutemov putback_lru_page(page); 1550f3f0e1d2SKirill A. Shutemov unlock_page(page); 1551b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 1552f3f0e1d2SKirill A. Shutemov } 1553f3f0e1d2SKirill A. Shutemov VM_BUG_ON(nr_none); 1554b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 1555f3f0e1d2SKirill A. Shutemov 1556f3f0e1d2SKirill A. Shutemov /* Unfreeze new_page, caller would take care about freeing it */ 1557f3f0e1d2SKirill A. Shutemov page_ref_unfreeze(new_page, 1); 1558f3f0e1d2SKirill A. Shutemov mem_cgroup_cancel_charge(new_page, memcg, true); 1559f3f0e1d2SKirill A. Shutemov unlock_page(new_page); 1560f3f0e1d2SKirill A. Shutemov new_page->mapping = NULL; 1561f3f0e1d2SKirill A. Shutemov } 1562f3f0e1d2SKirill A. Shutemov out: 1563f3f0e1d2SKirill A. Shutemov VM_BUG_ON(!list_empty(&pagelist)); 1564f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 1565f3f0e1d2SKirill A. Shutemov } 1566f3f0e1d2SKirill A. Shutemov 1567f3f0e1d2SKirill A. Shutemov static void khugepaged_scan_shmem(struct mm_struct *mm, 1568f3f0e1d2SKirill A. Shutemov struct address_space *mapping, 1569f3f0e1d2SKirill A. Shutemov pgoff_t start, struct page **hpage) 1570f3f0e1d2SKirill A. Shutemov { 1571f3f0e1d2SKirill A. Shutemov struct page *page = NULL; 1572f3f0e1d2SKirill A. Shutemov struct radix_tree_iter iter; 1573f3f0e1d2SKirill A. Shutemov void **slot; 1574f3f0e1d2SKirill A. Shutemov int present, swap; 1575f3f0e1d2SKirill A. Shutemov int node = NUMA_NO_NODE; 1576f3f0e1d2SKirill A. Shutemov int result = SCAN_SUCCEED; 1577f3f0e1d2SKirill A. Shutemov 1578f3f0e1d2SKirill A. Shutemov present = 0; 1579f3f0e1d2SKirill A. Shutemov swap = 0; 1580f3f0e1d2SKirill A. Shutemov memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 1581f3f0e1d2SKirill A. Shutemov rcu_read_lock(); 1582b93b0163SMatthew Wilcox radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) { 1583f3f0e1d2SKirill A. Shutemov if (iter.index >= start + HPAGE_PMD_NR) 1584f3f0e1d2SKirill A. Shutemov break; 1585f3f0e1d2SKirill A. Shutemov 1586f3f0e1d2SKirill A. Shutemov page = radix_tree_deref_slot(slot); 1587f3f0e1d2SKirill A. Shutemov if (radix_tree_deref_retry(page)) { 1588f3f0e1d2SKirill A. Shutemov slot = radix_tree_iter_retry(&iter); 1589f3f0e1d2SKirill A. Shutemov continue; 1590f3f0e1d2SKirill A. Shutemov } 1591f3f0e1d2SKirill A. Shutemov 1592f3f0e1d2SKirill A. Shutemov if (radix_tree_exception(page)) { 1593f3f0e1d2SKirill A. Shutemov if (++swap > khugepaged_max_ptes_swap) { 1594f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_SWAP_PTE; 1595f3f0e1d2SKirill A. Shutemov break; 1596f3f0e1d2SKirill A. Shutemov } 1597f3f0e1d2SKirill A. Shutemov continue; 1598f3f0e1d2SKirill A. Shutemov } 1599f3f0e1d2SKirill A. Shutemov 1600f3f0e1d2SKirill A. Shutemov if (PageTransCompound(page)) { 1601f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COMPOUND; 1602f3f0e1d2SKirill A. Shutemov break; 1603f3f0e1d2SKirill A. Shutemov } 1604f3f0e1d2SKirill A. Shutemov 1605f3f0e1d2SKirill A. Shutemov node = page_to_nid(page); 1606f3f0e1d2SKirill A. Shutemov if (khugepaged_scan_abort(node)) { 1607f3f0e1d2SKirill A. Shutemov result = SCAN_SCAN_ABORT; 1608f3f0e1d2SKirill A. Shutemov break; 1609f3f0e1d2SKirill A. Shutemov } 1610f3f0e1d2SKirill A. Shutemov khugepaged_node_load[node]++; 1611f3f0e1d2SKirill A. Shutemov 1612f3f0e1d2SKirill A. Shutemov if (!PageLRU(page)) { 1613f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_LRU; 1614f3f0e1d2SKirill A. Shutemov break; 1615f3f0e1d2SKirill A. Shutemov } 1616f3f0e1d2SKirill A. Shutemov 1617f3f0e1d2SKirill A. Shutemov if (page_count(page) != 1 + page_mapcount(page)) { 1618f3f0e1d2SKirill A. Shutemov result = SCAN_PAGE_COUNT; 1619f3f0e1d2SKirill A. Shutemov break; 1620f3f0e1d2SKirill A. Shutemov } 1621f3f0e1d2SKirill A. Shutemov 1622f3f0e1d2SKirill A. Shutemov /* 1623f3f0e1d2SKirill A. Shutemov * We probably should check if the page is referenced here, but 1624f3f0e1d2SKirill A. Shutemov * nobody would transfer pte_young() to PageReferenced() for us. 1625f3f0e1d2SKirill A. Shutemov * And rmap walk here is just too costly... 1626f3f0e1d2SKirill A. Shutemov */ 1627f3f0e1d2SKirill A. Shutemov 1628f3f0e1d2SKirill A. Shutemov present++; 1629f3f0e1d2SKirill A. Shutemov 1630f3f0e1d2SKirill A. Shutemov if (need_resched()) { 1631148deab2SMatthew Wilcox slot = radix_tree_iter_resume(slot, &iter); 1632f3f0e1d2SKirill A. Shutemov cond_resched_rcu(); 1633f3f0e1d2SKirill A. Shutemov } 1634f3f0e1d2SKirill A. Shutemov } 1635f3f0e1d2SKirill A. Shutemov rcu_read_unlock(); 1636f3f0e1d2SKirill A. Shutemov 1637f3f0e1d2SKirill A. Shutemov if (result == SCAN_SUCCEED) { 1638f3f0e1d2SKirill A. Shutemov if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 1639f3f0e1d2SKirill A. Shutemov result = SCAN_EXCEED_NONE_PTE; 1640f3f0e1d2SKirill A. Shutemov } else { 1641f3f0e1d2SKirill A. Shutemov node = khugepaged_find_target_node(); 1642f3f0e1d2SKirill A. Shutemov collapse_shmem(mm, mapping, start, hpage, node); 1643f3f0e1d2SKirill A. Shutemov } 1644f3f0e1d2SKirill A. Shutemov } 1645f3f0e1d2SKirill A. Shutemov 1646f3f0e1d2SKirill A. Shutemov /* TODO: tracepoints */ 1647f3f0e1d2SKirill A. Shutemov } 1648f3f0e1d2SKirill A. Shutemov #else 1649f3f0e1d2SKirill A. Shutemov static void khugepaged_scan_shmem(struct mm_struct *mm, 1650f3f0e1d2SKirill A. Shutemov struct address_space *mapping, 1651f3f0e1d2SKirill A. Shutemov pgoff_t start, struct page **hpage) 1652f3f0e1d2SKirill A. Shutemov { 1653f3f0e1d2SKirill A. Shutemov BUILD_BUG(); 1654f3f0e1d2SKirill A. Shutemov } 1655f3f0e1d2SKirill A. Shutemov #endif 1656f3f0e1d2SKirill A. Shutemov 1657b46e756fSKirill A. Shutemov static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 1658b46e756fSKirill A. Shutemov struct page **hpage) 1659b46e756fSKirill A. Shutemov __releases(&khugepaged_mm_lock) 1660b46e756fSKirill A. Shutemov __acquires(&khugepaged_mm_lock) 1661b46e756fSKirill A. Shutemov { 1662b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 1663b46e756fSKirill A. Shutemov struct mm_struct *mm; 1664b46e756fSKirill A. Shutemov struct vm_area_struct *vma; 1665b46e756fSKirill A. Shutemov int progress = 0; 1666b46e756fSKirill A. Shutemov 1667b46e756fSKirill A. Shutemov VM_BUG_ON(!pages); 1668b46e756fSKirill A. Shutemov VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 1669b46e756fSKirill A. Shutemov 1670b46e756fSKirill A. Shutemov if (khugepaged_scan.mm_slot) 1671b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 1672b46e756fSKirill A. Shutemov else { 1673b46e756fSKirill A. Shutemov mm_slot = list_entry(khugepaged_scan.mm_head.next, 1674b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 1675b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 1676b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = mm_slot; 1677b46e756fSKirill A. Shutemov } 1678b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 1679b46e756fSKirill A. Shutemov 1680b46e756fSKirill A. Shutemov mm = mm_slot->mm; 16813b454ad3SYang Shi /* 16823b454ad3SYang Shi * Don't wait for semaphore (to avoid long wait times). Just move to 16833b454ad3SYang Shi * the next mm on the list. 16843b454ad3SYang Shi */ 1685b46e756fSKirill A. Shutemov vma = NULL; 16863b454ad3SYang Shi if (unlikely(!down_read_trylock(&mm->mmap_sem))) 16873b454ad3SYang Shi goto breakouterloop_mmap_sem; 16883b454ad3SYang Shi if (likely(!khugepaged_test_exit(mm))) 1689b46e756fSKirill A. Shutemov vma = find_vma(mm, khugepaged_scan.address); 1690b46e756fSKirill A. Shutemov 1691b46e756fSKirill A. Shutemov progress++; 1692b46e756fSKirill A. Shutemov for (; vma; vma = vma->vm_next) { 1693b46e756fSKirill A. Shutemov unsigned long hstart, hend; 1694b46e756fSKirill A. Shutemov 1695b46e756fSKirill A. Shutemov cond_resched(); 1696b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) { 1697b46e756fSKirill A. Shutemov progress++; 1698b46e756fSKirill A. Shutemov break; 1699b46e756fSKirill A. Shutemov } 170050f8b92fSSong Liu if (!hugepage_vma_check(vma, vma->vm_flags)) { 1701b46e756fSKirill A. Shutemov skip: 1702b46e756fSKirill A. Shutemov progress++; 1703b46e756fSKirill A. Shutemov continue; 1704b46e756fSKirill A. Shutemov } 1705b46e756fSKirill A. Shutemov hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 1706b46e756fSKirill A. Shutemov hend = vma->vm_end & HPAGE_PMD_MASK; 1707b46e756fSKirill A. Shutemov if (hstart >= hend) 1708b46e756fSKirill A. Shutemov goto skip; 1709b46e756fSKirill A. Shutemov if (khugepaged_scan.address > hend) 1710b46e756fSKirill A. Shutemov goto skip; 1711b46e756fSKirill A. Shutemov if (khugepaged_scan.address < hstart) 1712b46e756fSKirill A. Shutemov khugepaged_scan.address = hstart; 1713b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 1714b46e756fSKirill A. Shutemov 1715b46e756fSKirill A. Shutemov while (khugepaged_scan.address < hend) { 1716b46e756fSKirill A. Shutemov int ret; 1717b46e756fSKirill A. Shutemov cond_resched(); 1718b46e756fSKirill A. Shutemov if (unlikely(khugepaged_test_exit(mm))) 1719b46e756fSKirill A. Shutemov goto breakouterloop; 1720b46e756fSKirill A. Shutemov 1721b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.address < hstart || 1722b46e756fSKirill A. Shutemov khugepaged_scan.address + HPAGE_PMD_SIZE > 1723b46e756fSKirill A. Shutemov hend); 1724f3f0e1d2SKirill A. Shutemov if (shmem_file(vma->vm_file)) { 1725e496cf3dSKirill A. Shutemov struct file *file; 1726f3f0e1d2SKirill A. Shutemov pgoff_t pgoff = linear_page_index(vma, 1727f3f0e1d2SKirill A. Shutemov khugepaged_scan.address); 1728e496cf3dSKirill A. Shutemov if (!shmem_huge_enabled(vma)) 1729e496cf3dSKirill A. Shutemov goto skip; 1730e496cf3dSKirill A. Shutemov file = get_file(vma->vm_file); 1731f3f0e1d2SKirill A. Shutemov up_read(&mm->mmap_sem); 1732f3f0e1d2SKirill A. Shutemov ret = 1; 1733f3f0e1d2SKirill A. Shutemov khugepaged_scan_shmem(mm, file->f_mapping, 1734f3f0e1d2SKirill A. Shutemov pgoff, hpage); 1735f3f0e1d2SKirill A. Shutemov fput(file); 1736f3f0e1d2SKirill A. Shutemov } else { 1737b46e756fSKirill A. Shutemov ret = khugepaged_scan_pmd(mm, vma, 1738b46e756fSKirill A. Shutemov khugepaged_scan.address, 1739b46e756fSKirill A. Shutemov hpage); 1740f3f0e1d2SKirill A. Shutemov } 1741b46e756fSKirill A. Shutemov /* move to next address */ 1742b46e756fSKirill A. Shutemov khugepaged_scan.address += HPAGE_PMD_SIZE; 1743b46e756fSKirill A. Shutemov progress += HPAGE_PMD_NR; 1744b46e756fSKirill A. Shutemov if (ret) 1745b46e756fSKirill A. Shutemov /* we released mmap_sem so break loop */ 1746b46e756fSKirill A. Shutemov goto breakouterloop_mmap_sem; 1747b46e756fSKirill A. Shutemov if (progress >= pages) 1748b46e756fSKirill A. Shutemov goto breakouterloop; 1749b46e756fSKirill A. Shutemov } 1750b46e756fSKirill A. Shutemov } 1751b46e756fSKirill A. Shutemov breakouterloop: 1752b46e756fSKirill A. Shutemov up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ 1753b46e756fSKirill A. Shutemov breakouterloop_mmap_sem: 1754b46e756fSKirill A. Shutemov 1755b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 1756b46e756fSKirill A. Shutemov VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 1757b46e756fSKirill A. Shutemov /* 1758b46e756fSKirill A. Shutemov * Release the current mm_slot if this mm is about to die, or 1759b46e756fSKirill A. Shutemov * if we scanned all vmas of this mm. 1760b46e756fSKirill A. Shutemov */ 1761b46e756fSKirill A. Shutemov if (khugepaged_test_exit(mm) || !vma) { 1762b46e756fSKirill A. Shutemov /* 1763b46e756fSKirill A. Shutemov * Make sure that if mm_users is reaching zero while 1764b46e756fSKirill A. Shutemov * khugepaged runs here, khugepaged_exit will find 1765b46e756fSKirill A. Shutemov * mm_slot not pointing to the exiting mm. 1766b46e756fSKirill A. Shutemov */ 1767b46e756fSKirill A. Shutemov if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 1768b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = list_entry( 1769b46e756fSKirill A. Shutemov mm_slot->mm_node.next, 1770b46e756fSKirill A. Shutemov struct mm_slot, mm_node); 1771b46e756fSKirill A. Shutemov khugepaged_scan.address = 0; 1772b46e756fSKirill A. Shutemov } else { 1773b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 1774b46e756fSKirill A. Shutemov khugepaged_full_scans++; 1775b46e756fSKirill A. Shutemov } 1776b46e756fSKirill A. Shutemov 1777b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 1778b46e756fSKirill A. Shutemov } 1779b46e756fSKirill A. Shutemov 1780b46e756fSKirill A. Shutemov return progress; 1781b46e756fSKirill A. Shutemov } 1782b46e756fSKirill A. Shutemov 1783b46e756fSKirill A. Shutemov static int khugepaged_has_work(void) 1784b46e756fSKirill A. Shutemov { 1785b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) && 1786b46e756fSKirill A. Shutemov khugepaged_enabled(); 1787b46e756fSKirill A. Shutemov } 1788b46e756fSKirill A. Shutemov 1789b46e756fSKirill A. Shutemov static int khugepaged_wait_event(void) 1790b46e756fSKirill A. Shutemov { 1791b46e756fSKirill A. Shutemov return !list_empty(&khugepaged_scan.mm_head) || 1792b46e756fSKirill A. Shutemov kthread_should_stop(); 1793b46e756fSKirill A. Shutemov } 1794b46e756fSKirill A. Shutemov 1795b46e756fSKirill A. Shutemov static void khugepaged_do_scan(void) 1796b46e756fSKirill A. Shutemov { 1797b46e756fSKirill A. Shutemov struct page *hpage = NULL; 1798b46e756fSKirill A. Shutemov unsigned int progress = 0, pass_through_head = 0; 1799b46e756fSKirill A. Shutemov unsigned int pages = khugepaged_pages_to_scan; 1800b46e756fSKirill A. Shutemov bool wait = true; 1801b46e756fSKirill A. Shutemov 1802b46e756fSKirill A. Shutemov barrier(); /* write khugepaged_pages_to_scan to local stack */ 1803b46e756fSKirill A. Shutemov 1804b46e756fSKirill A. Shutemov while (progress < pages) { 1805b46e756fSKirill A. Shutemov if (!khugepaged_prealloc_page(&hpage, &wait)) 1806b46e756fSKirill A. Shutemov break; 1807b46e756fSKirill A. Shutemov 1808b46e756fSKirill A. Shutemov cond_resched(); 1809b46e756fSKirill A. Shutemov 1810b46e756fSKirill A. Shutemov if (unlikely(kthread_should_stop() || try_to_freeze())) 1811b46e756fSKirill A. Shutemov break; 1812b46e756fSKirill A. Shutemov 1813b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 1814b46e756fSKirill A. Shutemov if (!khugepaged_scan.mm_slot) 1815b46e756fSKirill A. Shutemov pass_through_head++; 1816b46e756fSKirill A. Shutemov if (khugepaged_has_work() && 1817b46e756fSKirill A. Shutemov pass_through_head < 2) 1818b46e756fSKirill A. Shutemov progress += khugepaged_scan_mm_slot(pages - progress, 1819b46e756fSKirill A. Shutemov &hpage); 1820b46e756fSKirill A. Shutemov else 1821b46e756fSKirill A. Shutemov progress = pages; 1822b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 1823b46e756fSKirill A. Shutemov } 1824b46e756fSKirill A. Shutemov 1825b46e756fSKirill A. Shutemov if (!IS_ERR_OR_NULL(hpage)) 1826b46e756fSKirill A. Shutemov put_page(hpage); 1827b46e756fSKirill A. Shutemov } 1828b46e756fSKirill A. Shutemov 1829b46e756fSKirill A. Shutemov static bool khugepaged_should_wakeup(void) 1830b46e756fSKirill A. Shutemov { 1831b46e756fSKirill A. Shutemov return kthread_should_stop() || 1832b46e756fSKirill A. Shutemov time_after_eq(jiffies, khugepaged_sleep_expire); 1833b46e756fSKirill A. Shutemov } 1834b46e756fSKirill A. Shutemov 1835b46e756fSKirill A. Shutemov static void khugepaged_wait_work(void) 1836b46e756fSKirill A. Shutemov { 1837b46e756fSKirill A. Shutemov if (khugepaged_has_work()) { 1838b46e756fSKirill A. Shutemov const unsigned long scan_sleep_jiffies = 1839b46e756fSKirill A. Shutemov msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 1840b46e756fSKirill A. Shutemov 1841b46e756fSKirill A. Shutemov if (!scan_sleep_jiffies) 1842b46e756fSKirill A. Shutemov return; 1843b46e756fSKirill A. Shutemov 1844b46e756fSKirill A. Shutemov khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 1845b46e756fSKirill A. Shutemov wait_event_freezable_timeout(khugepaged_wait, 1846b46e756fSKirill A. Shutemov khugepaged_should_wakeup(), 1847b46e756fSKirill A. Shutemov scan_sleep_jiffies); 1848b46e756fSKirill A. Shutemov return; 1849b46e756fSKirill A. Shutemov } 1850b46e756fSKirill A. Shutemov 1851b46e756fSKirill A. Shutemov if (khugepaged_enabled()) 1852b46e756fSKirill A. Shutemov wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 1853b46e756fSKirill A. Shutemov } 1854b46e756fSKirill A. Shutemov 1855b46e756fSKirill A. Shutemov static int khugepaged(void *none) 1856b46e756fSKirill A. Shutemov { 1857b46e756fSKirill A. Shutemov struct mm_slot *mm_slot; 1858b46e756fSKirill A. Shutemov 1859b46e756fSKirill A. Shutemov set_freezable(); 1860b46e756fSKirill A. Shutemov set_user_nice(current, MAX_NICE); 1861b46e756fSKirill A. Shutemov 1862b46e756fSKirill A. Shutemov while (!kthread_should_stop()) { 1863b46e756fSKirill A. Shutemov khugepaged_do_scan(); 1864b46e756fSKirill A. Shutemov khugepaged_wait_work(); 1865b46e756fSKirill A. Shutemov } 1866b46e756fSKirill A. Shutemov 1867b46e756fSKirill A. Shutemov spin_lock(&khugepaged_mm_lock); 1868b46e756fSKirill A. Shutemov mm_slot = khugepaged_scan.mm_slot; 1869b46e756fSKirill A. Shutemov khugepaged_scan.mm_slot = NULL; 1870b46e756fSKirill A. Shutemov if (mm_slot) 1871b46e756fSKirill A. Shutemov collect_mm_slot(mm_slot); 1872b46e756fSKirill A. Shutemov spin_unlock(&khugepaged_mm_lock); 1873b46e756fSKirill A. Shutemov return 0; 1874b46e756fSKirill A. Shutemov } 1875b46e756fSKirill A. Shutemov 1876b46e756fSKirill A. Shutemov static void set_recommended_min_free_kbytes(void) 1877b46e756fSKirill A. Shutemov { 1878b46e756fSKirill A. Shutemov struct zone *zone; 1879b46e756fSKirill A. Shutemov int nr_zones = 0; 1880b46e756fSKirill A. Shutemov unsigned long recommended_min; 1881b46e756fSKirill A. Shutemov 1882b7d349c7SJoonsoo Kim for_each_populated_zone(zone) { 1883b7d349c7SJoonsoo Kim /* 1884b7d349c7SJoonsoo Kim * We don't need to worry about fragmentation of 1885b7d349c7SJoonsoo Kim * ZONE_MOVABLE since it only has movable pages. 1886b7d349c7SJoonsoo Kim */ 1887b7d349c7SJoonsoo Kim if (zone_idx(zone) > gfp_zone(GFP_USER)) 1888b7d349c7SJoonsoo Kim continue; 1889b7d349c7SJoonsoo Kim 1890b46e756fSKirill A. Shutemov nr_zones++; 1891b7d349c7SJoonsoo Kim } 1892b46e756fSKirill A. Shutemov 1893b46e756fSKirill A. Shutemov /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 1894b46e756fSKirill A. Shutemov recommended_min = pageblock_nr_pages * nr_zones * 2; 1895b46e756fSKirill A. Shutemov 1896b46e756fSKirill A. Shutemov /* 1897b46e756fSKirill A. Shutemov * Make sure that on average at least two pageblocks are almost free 1898b46e756fSKirill A. Shutemov * of another type, one for a migratetype to fall back to and a 1899b46e756fSKirill A. Shutemov * second to avoid subsequent fallbacks of other types There are 3 1900b46e756fSKirill A. Shutemov * MIGRATE_TYPES we care about. 1901b46e756fSKirill A. Shutemov */ 1902b46e756fSKirill A. Shutemov recommended_min += pageblock_nr_pages * nr_zones * 1903b46e756fSKirill A. Shutemov MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 1904b46e756fSKirill A. Shutemov 1905b46e756fSKirill A. Shutemov /* don't ever allow to reserve more than 5% of the lowmem */ 1906b46e756fSKirill A. Shutemov recommended_min = min(recommended_min, 1907b46e756fSKirill A. Shutemov (unsigned long) nr_free_buffer_pages() / 20); 1908b46e756fSKirill A. Shutemov recommended_min <<= (PAGE_SHIFT-10); 1909b46e756fSKirill A. Shutemov 1910b46e756fSKirill A. Shutemov if (recommended_min > min_free_kbytes) { 1911b46e756fSKirill A. Shutemov if (user_min_free_kbytes >= 0) 1912b46e756fSKirill A. Shutemov pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 1913b46e756fSKirill A. Shutemov min_free_kbytes, recommended_min); 1914b46e756fSKirill A. Shutemov 1915b46e756fSKirill A. Shutemov min_free_kbytes = recommended_min; 1916b46e756fSKirill A. Shutemov } 1917b46e756fSKirill A. Shutemov setup_per_zone_wmarks(); 1918b46e756fSKirill A. Shutemov } 1919b46e756fSKirill A. Shutemov 1920b46e756fSKirill A. Shutemov int start_stop_khugepaged(void) 1921b46e756fSKirill A. Shutemov { 1922b46e756fSKirill A. Shutemov static struct task_struct *khugepaged_thread __read_mostly; 1923b46e756fSKirill A. Shutemov static DEFINE_MUTEX(khugepaged_mutex); 1924b46e756fSKirill A. Shutemov int err = 0; 1925b46e756fSKirill A. Shutemov 1926b46e756fSKirill A. Shutemov mutex_lock(&khugepaged_mutex); 1927b46e756fSKirill A. Shutemov if (khugepaged_enabled()) { 1928b46e756fSKirill A. Shutemov if (!khugepaged_thread) 1929b46e756fSKirill A. Shutemov khugepaged_thread = kthread_run(khugepaged, NULL, 1930b46e756fSKirill A. Shutemov "khugepaged"); 1931b46e756fSKirill A. Shutemov if (IS_ERR(khugepaged_thread)) { 1932b46e756fSKirill A. Shutemov pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 1933b46e756fSKirill A. Shutemov err = PTR_ERR(khugepaged_thread); 1934b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 1935b46e756fSKirill A. Shutemov goto fail; 1936b46e756fSKirill A. Shutemov } 1937b46e756fSKirill A. Shutemov 1938b46e756fSKirill A. Shutemov if (!list_empty(&khugepaged_scan.mm_head)) 1939b46e756fSKirill A. Shutemov wake_up_interruptible(&khugepaged_wait); 1940b46e756fSKirill A. Shutemov 1941b46e756fSKirill A. Shutemov set_recommended_min_free_kbytes(); 1942b46e756fSKirill A. Shutemov } else if (khugepaged_thread) { 1943b46e756fSKirill A. Shutemov kthread_stop(khugepaged_thread); 1944b46e756fSKirill A. Shutemov khugepaged_thread = NULL; 1945b46e756fSKirill A. Shutemov } 1946b46e756fSKirill A. Shutemov fail: 1947b46e756fSKirill A. Shutemov mutex_unlock(&khugepaged_mutex); 1948b46e756fSKirill A. Shutemov return err; 1949b46e756fSKirill A. Shutemov } 1950