Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2009 Red Hat, Inc. |
| 3 | * |
| 4 | * This work is licensed under the terms of the GNU GPL, version 2. See |
| 5 | * the COPYING file in the top-level directory. |
| 6 | */ |
| 7 | |
| 8 | #include <linux/mm.h> |
| 9 | #include <linux/sched.h> |
| 10 | #include <linux/highmem.h> |
| 11 | #include <linux/hugetlb.h> |
| 12 | #include <linux/mmu_notifier.h> |
| 13 | #include <linux/rmap.h> |
| 14 | #include <linux/swap.h> |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 15 | #include <linux/shrinker.h> |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 16 | #include <linux/mm_inline.h> |
| 17 | #include <linux/kthread.h> |
| 18 | #include <linux/khugepaged.h> |
Andrea Arcangeli | 878aee7 | 2011-01-13 15:47:10 -0800 | [diff] [blame] | 19 | #include <linux/freezer.h> |
Andrea Arcangeli | a664b2d | 2011-01-13 15:47:17 -0800 | [diff] [blame] | 20 | #include <linux/mman.h> |
Ralf Baechle | 325adeb | 2012-10-15 13:44:56 +0200 | [diff] [blame] | 21 | #include <linux/pagemap.h> |
Mel Gorman | 4daae3b | 2012-11-02 11:33:45 +0000 | [diff] [blame] | 22 | #include <linux/migrate.h> |
Sasha Levin | 43b5fbb | 2013-02-22 16:32:27 -0800 | [diff] [blame] | 23 | #include <linux/hashtable.h> |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 24 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 25 | #include <asm/tlb.h> |
| 26 | #include <asm/pgalloc.h> |
| 27 | #include "internal.h" |
| 28 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 29 | /* |
Jianguo Wu | 8bfa3f9 | 2013-11-12 15:07:16 -0800 | [diff] [blame] | 30 | * By default transparent hugepage support is disabled in order that avoid |
| 31 | * to risk increase the memory footprint of applications without a guaranteed |
| 32 | * benefit. When transparent hugepage support is enabled, is for all mappings, |
| 33 | * and khugepaged scans all mappings. |
| 34 | * Defrag is invoked by khugepaged hugepage allocations and by page faults |
| 35 | * for all hugepage allocations. |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 36 | */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 37 | unsigned long transparent_hugepage_flags __read_mostly = |
Andrea Arcangeli | 13ece88 | 2011-01-13 15:47:07 -0800 | [diff] [blame] | 38 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 39 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
Andrea Arcangeli | 13ece88 | 2011-01-13 15:47:07 -0800 | [diff] [blame] | 40 | #endif |
| 41 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE |
| 42 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| |
| 43 | #endif |
Andrea Arcangeli | d39d33c | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 44 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| |
Kirill A. Shutemov | 79da540 | 2012-12-12 13:51:12 -0800 | [diff] [blame] | 45 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
| 46 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 47 | |
| 48 | /* default scan 8*512 pte (or vmas) every 30 second */ |
| 49 | static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; |
| 50 | static unsigned int khugepaged_pages_collapsed; |
| 51 | static unsigned int khugepaged_full_scans; |
| 52 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; |
| 53 | /* during fragmentation poll the hugepage allocator once every minute */ |
| 54 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; |
| 55 | static struct task_struct *khugepaged_thread __read_mostly; |
| 56 | static DEFINE_MUTEX(khugepaged_mutex); |
| 57 | static DEFINE_SPINLOCK(khugepaged_mm_lock); |
| 58 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); |
| 59 | /* |
| 60 | * default collapse hugepages if there is at least one pte mapped like |
| 61 | * it would have happened if the vma was large enough during page |
| 62 | * fault. |
| 63 | */ |
| 64 | static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; |
| 65 | |
| 66 | static int khugepaged(void *none); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 67 | static int khugepaged_slab_init(void); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 68 | |
Sasha Levin | 43b5fbb | 2013-02-22 16:32:27 -0800 | [diff] [blame] | 69 | #define MM_SLOTS_HASH_BITS 10 |
| 70 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); |
| 71 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 72 | static struct kmem_cache *mm_slot_cache __read_mostly; |
| 73 | |
| 74 | /** |
| 75 | * struct mm_slot - hash lookup from mm to mm_slot |
| 76 | * @hash: hash collision list |
| 77 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head |
| 78 | * @mm: the mm that this information is valid for |
| 79 | */ |
| 80 | struct mm_slot { |
| 81 | struct hlist_node hash; |
| 82 | struct list_head mm_node; |
| 83 | struct mm_struct *mm; |
| 84 | }; |
| 85 | |
| 86 | /** |
| 87 | * struct khugepaged_scan - cursor for scanning |
| 88 | * @mm_head: the head of the mm list to scan |
| 89 | * @mm_slot: the current mm_slot we are scanning |
| 90 | * @address: the next address inside that to be scanned |
| 91 | * |
| 92 | * There is only the one khugepaged_scan instance of this cursor structure. |
| 93 | */ |
| 94 | struct khugepaged_scan { |
| 95 | struct list_head mm_head; |
| 96 | struct mm_slot *mm_slot; |
| 97 | unsigned long address; |
H Hartley Sweeten | 2f1da64 | 2011-10-31 17:09:25 -0700 | [diff] [blame] | 98 | }; |
| 99 | static struct khugepaged_scan khugepaged_scan = { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 100 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), |
| 101 | }; |
| 102 | |
Andrea Arcangeli | f000565 | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 103 | |
| 104 | static int set_recommended_min_free_kbytes(void) |
| 105 | { |
| 106 | struct zone *zone; |
| 107 | int nr_zones = 0; |
| 108 | unsigned long recommended_min; |
Andrea Arcangeli | f000565 | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 109 | |
Xiao Guangrong | 17c230a | 2012-10-08 16:29:56 -0700 | [diff] [blame] | 110 | if (!khugepaged_enabled()) |
Andrea Arcangeli | f000565 | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 111 | return 0; |
| 112 | |
| 113 | for_each_populated_zone(zone) |
| 114 | nr_zones++; |
| 115 | |
| 116 | /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */ |
| 117 | recommended_min = pageblock_nr_pages * nr_zones * 2; |
| 118 | |
| 119 | /* |
| 120 | * Make sure that on average at least two pageblocks are almost free |
| 121 | * of another type, one for a migratetype to fall back to and a |
| 122 | * second to avoid subsequent fallbacks of other types There are 3 |
| 123 | * MIGRATE_TYPES we care about. |
| 124 | */ |
| 125 | recommended_min += pageblock_nr_pages * nr_zones * |
| 126 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; |
| 127 | |
| 128 | /* don't ever allow to reserve more than 5% of the lowmem */ |
| 129 | recommended_min = min(recommended_min, |
| 130 | (unsigned long) nr_free_buffer_pages() / 20); |
| 131 | recommended_min <<= (PAGE_SHIFT-10); |
| 132 | |
Han Pingtian | 42aa83c | 2014-01-23 15:53:28 -0800 | [diff] [blame] | 133 | if (recommended_min > min_free_kbytes) { |
| 134 | if (user_min_free_kbytes >= 0) |
| 135 | pr_info("raising min_free_kbytes from %d to %lu " |
| 136 | "to help transparent hugepage allocations\n", |
| 137 | min_free_kbytes, recommended_min); |
| 138 | |
Andrea Arcangeli | f000565 | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 139 | min_free_kbytes = recommended_min; |
Han Pingtian | 42aa83c | 2014-01-23 15:53:28 -0800 | [diff] [blame] | 140 | } |
Andrea Arcangeli | f000565 | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 141 | setup_per_zone_wmarks(); |
| 142 | return 0; |
| 143 | } |
| 144 | late_initcall(set_recommended_min_free_kbytes); |
| 145 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 146 | static int start_khugepaged(void) |
| 147 | { |
| 148 | int err = 0; |
| 149 | if (khugepaged_enabled()) { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 150 | if (!khugepaged_thread) |
| 151 | khugepaged_thread = kthread_run(khugepaged, NULL, |
| 152 | "khugepaged"); |
| 153 | if (unlikely(IS_ERR(khugepaged_thread))) { |
| 154 | printk(KERN_ERR |
| 155 | "khugepaged: kthread_run(khugepaged) failed\n"); |
| 156 | err = PTR_ERR(khugepaged_thread); |
| 157 | khugepaged_thread = NULL; |
| 158 | } |
Xiao Guangrong | 911891a | 2012-10-08 16:29:41 -0700 | [diff] [blame] | 159 | |
| 160 | if (!list_empty(&khugepaged_scan.mm_head)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 161 | wake_up_interruptible(&khugepaged_wait); |
Andrea Arcangeli | f000565 | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 162 | |
| 163 | set_recommended_min_free_kbytes(); |
Xiao Guangrong | 911891a | 2012-10-08 16:29:41 -0700 | [diff] [blame] | 164 | } else if (khugepaged_thread) { |
Xiao Guangrong | 911891a | 2012-10-08 16:29:41 -0700 | [diff] [blame] | 165 | kthread_stop(khugepaged_thread); |
| 166 | khugepaged_thread = NULL; |
| 167 | } |
Xiao Guangrong | 637e3a2 | 2012-10-08 16:29:38 -0700 | [diff] [blame] | 168 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 169 | return err; |
| 170 | } |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 171 | |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 172 | static atomic_t huge_zero_refcount; |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 173 | static struct page *huge_zero_page __read_mostly; |
Kirill A. Shutemov | 4a6c129 | 2012-12-12 13:50:47 -0800 | [diff] [blame] | 174 | |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 175 | static inline bool is_huge_zero_page(struct page *page) |
Kirill A. Shutemov | 4a6c129 | 2012-12-12 13:50:47 -0800 | [diff] [blame] | 176 | { |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 177 | return ACCESS_ONCE(huge_zero_page) == page; |
Kirill A. Shutemov | 4a6c129 | 2012-12-12 13:50:47 -0800 | [diff] [blame] | 178 | } |
| 179 | |
| 180 | static inline bool is_huge_zero_pmd(pmd_t pmd) |
| 181 | { |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 182 | return is_huge_zero_page(pmd_page(pmd)); |
Kirill A. Shutemov | 4a6c129 | 2012-12-12 13:50:47 -0800 | [diff] [blame] | 183 | } |
| 184 | |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 185 | static struct page *get_huge_zero_page(void) |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 186 | { |
| 187 | struct page *zero_page; |
| 188 | retry: |
| 189 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 190 | return ACCESS_ONCE(huge_zero_page); |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 191 | |
| 192 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, |
| 193 | HPAGE_PMD_ORDER); |
Kirill A. Shutemov | d8a8e1f | 2012-12-12 13:51:09 -0800 | [diff] [blame] | 194 | if (!zero_page) { |
| 195 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 196 | return NULL; |
Kirill A. Shutemov | d8a8e1f | 2012-12-12 13:51:09 -0800 | [diff] [blame] | 197 | } |
| 198 | count_vm_event(THP_ZERO_PAGE_ALLOC); |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 199 | preempt_disable(); |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 200 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 201 | preempt_enable(); |
| 202 | __free_page(zero_page); |
| 203 | goto retry; |
| 204 | } |
| 205 | |
| 206 | /* We take additional reference here. It will be put back by shrinker */ |
| 207 | atomic_set(&huge_zero_refcount, 2); |
| 208 | preempt_enable(); |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 209 | return ACCESS_ONCE(huge_zero_page); |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 210 | } |
| 211 | |
| 212 | static void put_huge_zero_page(void) |
| 213 | { |
| 214 | /* |
| 215 | * Counter should never go to zero here. Only shrinker can put |
| 216 | * last reference. |
| 217 | */ |
| 218 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); |
| 219 | } |
| 220 | |
Glauber Costa | 4889646 | 2013-08-28 10:18:15 +1000 | [diff] [blame] | 221 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
| 222 | struct shrink_control *sc) |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 223 | { |
Glauber Costa | 4889646 | 2013-08-28 10:18:15 +1000 | [diff] [blame] | 224 | /* we can free zero page only if last reference remains */ |
| 225 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; |
| 226 | } |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 227 | |
Glauber Costa | 4889646 | 2013-08-28 10:18:15 +1000 | [diff] [blame] | 228 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
| 229 | struct shrink_control *sc) |
| 230 | { |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 231 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 232 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
| 233 | BUG_ON(zero_page == NULL); |
| 234 | __free_page(zero_page); |
Glauber Costa | 4889646 | 2013-08-28 10:18:15 +1000 | [diff] [blame] | 235 | return HPAGE_PMD_NR; |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 236 | } |
| 237 | |
| 238 | return 0; |
| 239 | } |
| 240 | |
| 241 | static struct shrinker huge_zero_page_shrinker = { |
Glauber Costa | 4889646 | 2013-08-28 10:18:15 +1000 | [diff] [blame] | 242 | .count_objects = shrink_huge_zero_page_count, |
| 243 | .scan_objects = shrink_huge_zero_page_scan, |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 244 | .seeks = DEFAULT_SEEKS, |
| 245 | }; |
| 246 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 247 | #ifdef CONFIG_SYSFS |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 248 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 249 | static ssize_t double_flag_show(struct kobject *kobj, |
| 250 | struct kobj_attribute *attr, char *buf, |
| 251 | enum transparent_hugepage_flag enabled, |
| 252 | enum transparent_hugepage_flag req_madv) |
| 253 | { |
| 254 | if (test_bit(enabled, &transparent_hugepage_flags)) { |
| 255 | VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); |
| 256 | return sprintf(buf, "[always] madvise never\n"); |
| 257 | } else if (test_bit(req_madv, &transparent_hugepage_flags)) |
| 258 | return sprintf(buf, "always [madvise] never\n"); |
| 259 | else |
| 260 | return sprintf(buf, "always madvise [never]\n"); |
| 261 | } |
| 262 | static ssize_t double_flag_store(struct kobject *kobj, |
| 263 | struct kobj_attribute *attr, |
| 264 | const char *buf, size_t count, |
| 265 | enum transparent_hugepage_flag enabled, |
| 266 | enum transparent_hugepage_flag req_madv) |
| 267 | { |
| 268 | if (!memcmp("always", buf, |
| 269 | min(sizeof("always")-1, count))) { |
| 270 | set_bit(enabled, &transparent_hugepage_flags); |
| 271 | clear_bit(req_madv, &transparent_hugepage_flags); |
| 272 | } else if (!memcmp("madvise", buf, |
| 273 | min(sizeof("madvise")-1, count))) { |
| 274 | clear_bit(enabled, &transparent_hugepage_flags); |
| 275 | set_bit(req_madv, &transparent_hugepage_flags); |
| 276 | } else if (!memcmp("never", buf, |
| 277 | min(sizeof("never")-1, count))) { |
| 278 | clear_bit(enabled, &transparent_hugepage_flags); |
| 279 | clear_bit(req_madv, &transparent_hugepage_flags); |
| 280 | } else |
| 281 | return -EINVAL; |
| 282 | |
| 283 | return count; |
| 284 | } |
| 285 | |
| 286 | static ssize_t enabled_show(struct kobject *kobj, |
| 287 | struct kobj_attribute *attr, char *buf) |
| 288 | { |
| 289 | return double_flag_show(kobj, attr, buf, |
| 290 | TRANSPARENT_HUGEPAGE_FLAG, |
| 291 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); |
| 292 | } |
| 293 | static ssize_t enabled_store(struct kobject *kobj, |
| 294 | struct kobj_attribute *attr, |
| 295 | const char *buf, size_t count) |
| 296 | { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 297 | ssize_t ret; |
| 298 | |
| 299 | ret = double_flag_store(kobj, attr, buf, count, |
| 300 | TRANSPARENT_HUGEPAGE_FLAG, |
| 301 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); |
| 302 | |
| 303 | if (ret > 0) { |
Xiao Guangrong | 911891a | 2012-10-08 16:29:41 -0700 | [diff] [blame] | 304 | int err; |
| 305 | |
| 306 | mutex_lock(&khugepaged_mutex); |
| 307 | err = start_khugepaged(); |
| 308 | mutex_unlock(&khugepaged_mutex); |
| 309 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 310 | if (err) |
| 311 | ret = err; |
| 312 | } |
| 313 | |
| 314 | return ret; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 315 | } |
| 316 | static struct kobj_attribute enabled_attr = |
| 317 | __ATTR(enabled, 0644, enabled_show, enabled_store); |
| 318 | |
| 319 | static ssize_t single_flag_show(struct kobject *kobj, |
| 320 | struct kobj_attribute *attr, char *buf, |
| 321 | enum transparent_hugepage_flag flag) |
| 322 | { |
Ben Hutchings | e27e615 | 2011-04-14 15:22:21 -0700 | [diff] [blame] | 323 | return sprintf(buf, "%d\n", |
| 324 | !!test_bit(flag, &transparent_hugepage_flags)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 325 | } |
Ben Hutchings | e27e615 | 2011-04-14 15:22:21 -0700 | [diff] [blame] | 326 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 327 | static ssize_t single_flag_store(struct kobject *kobj, |
| 328 | struct kobj_attribute *attr, |
| 329 | const char *buf, size_t count, |
| 330 | enum transparent_hugepage_flag flag) |
| 331 | { |
Ben Hutchings | e27e615 | 2011-04-14 15:22:21 -0700 | [diff] [blame] | 332 | unsigned long value; |
| 333 | int ret; |
| 334 | |
| 335 | ret = kstrtoul(buf, 10, &value); |
| 336 | if (ret < 0) |
| 337 | return ret; |
| 338 | if (value > 1) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 339 | return -EINVAL; |
| 340 | |
Ben Hutchings | e27e615 | 2011-04-14 15:22:21 -0700 | [diff] [blame] | 341 | if (value) |
| 342 | set_bit(flag, &transparent_hugepage_flags); |
| 343 | else |
| 344 | clear_bit(flag, &transparent_hugepage_flags); |
| 345 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 346 | return count; |
| 347 | } |
| 348 | |
| 349 | /* |
| 350 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind |
| 351 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of |
| 352 | * memory just to allocate one more hugepage. |
| 353 | */ |
| 354 | static ssize_t defrag_show(struct kobject *kobj, |
| 355 | struct kobj_attribute *attr, char *buf) |
| 356 | { |
| 357 | return double_flag_show(kobj, attr, buf, |
| 358 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, |
| 359 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); |
| 360 | } |
| 361 | static ssize_t defrag_store(struct kobject *kobj, |
| 362 | struct kobj_attribute *attr, |
| 363 | const char *buf, size_t count) |
| 364 | { |
| 365 | return double_flag_store(kobj, attr, buf, count, |
| 366 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, |
| 367 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); |
| 368 | } |
| 369 | static struct kobj_attribute defrag_attr = |
| 370 | __ATTR(defrag, 0644, defrag_show, defrag_store); |
| 371 | |
Kirill A. Shutemov | 79da540 | 2012-12-12 13:51:12 -0800 | [diff] [blame] | 372 | static ssize_t use_zero_page_show(struct kobject *kobj, |
| 373 | struct kobj_attribute *attr, char *buf) |
| 374 | { |
| 375 | return single_flag_show(kobj, attr, buf, |
| 376 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
| 377 | } |
| 378 | static ssize_t use_zero_page_store(struct kobject *kobj, |
| 379 | struct kobj_attribute *attr, const char *buf, size_t count) |
| 380 | { |
| 381 | return single_flag_store(kobj, attr, buf, count, |
| 382 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
| 383 | } |
| 384 | static struct kobj_attribute use_zero_page_attr = |
| 385 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 386 | #ifdef CONFIG_DEBUG_VM |
| 387 | static ssize_t debug_cow_show(struct kobject *kobj, |
| 388 | struct kobj_attribute *attr, char *buf) |
| 389 | { |
| 390 | return single_flag_show(kobj, attr, buf, |
| 391 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); |
| 392 | } |
| 393 | static ssize_t debug_cow_store(struct kobject *kobj, |
| 394 | struct kobj_attribute *attr, |
| 395 | const char *buf, size_t count) |
| 396 | { |
| 397 | return single_flag_store(kobj, attr, buf, count, |
| 398 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); |
| 399 | } |
| 400 | static struct kobj_attribute debug_cow_attr = |
| 401 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); |
| 402 | #endif /* CONFIG_DEBUG_VM */ |
| 403 | |
| 404 | static struct attribute *hugepage_attr[] = { |
| 405 | &enabled_attr.attr, |
| 406 | &defrag_attr.attr, |
Kirill A. Shutemov | 79da540 | 2012-12-12 13:51:12 -0800 | [diff] [blame] | 407 | &use_zero_page_attr.attr, |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 408 | #ifdef CONFIG_DEBUG_VM |
| 409 | &debug_cow_attr.attr, |
| 410 | #endif |
| 411 | NULL, |
| 412 | }; |
| 413 | |
| 414 | static struct attribute_group hugepage_attr_group = { |
| 415 | .attrs = hugepage_attr, |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 416 | }; |
| 417 | |
| 418 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, |
| 419 | struct kobj_attribute *attr, |
| 420 | char *buf) |
| 421 | { |
| 422 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); |
| 423 | } |
| 424 | |
| 425 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, |
| 426 | struct kobj_attribute *attr, |
| 427 | const char *buf, size_t count) |
| 428 | { |
| 429 | unsigned long msecs; |
| 430 | int err; |
| 431 | |
Jingoo Han | 3dbb95f | 2013-09-11 14:20:25 -0700 | [diff] [blame] | 432 | err = kstrtoul(buf, 10, &msecs); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 433 | if (err || msecs > UINT_MAX) |
| 434 | return -EINVAL; |
| 435 | |
| 436 | khugepaged_scan_sleep_millisecs = msecs; |
| 437 | wake_up_interruptible(&khugepaged_wait); |
| 438 | |
| 439 | return count; |
| 440 | } |
| 441 | static struct kobj_attribute scan_sleep_millisecs_attr = |
| 442 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, |
| 443 | scan_sleep_millisecs_store); |
| 444 | |
| 445 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, |
| 446 | struct kobj_attribute *attr, |
| 447 | char *buf) |
| 448 | { |
| 449 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); |
| 450 | } |
| 451 | |
| 452 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, |
| 453 | struct kobj_attribute *attr, |
| 454 | const char *buf, size_t count) |
| 455 | { |
| 456 | unsigned long msecs; |
| 457 | int err; |
| 458 | |
Jingoo Han | 3dbb95f | 2013-09-11 14:20:25 -0700 | [diff] [blame] | 459 | err = kstrtoul(buf, 10, &msecs); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 460 | if (err || msecs > UINT_MAX) |
| 461 | return -EINVAL; |
| 462 | |
| 463 | khugepaged_alloc_sleep_millisecs = msecs; |
| 464 | wake_up_interruptible(&khugepaged_wait); |
| 465 | |
| 466 | return count; |
| 467 | } |
| 468 | static struct kobj_attribute alloc_sleep_millisecs_attr = |
| 469 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, |
| 470 | alloc_sleep_millisecs_store); |
| 471 | |
| 472 | static ssize_t pages_to_scan_show(struct kobject *kobj, |
| 473 | struct kobj_attribute *attr, |
| 474 | char *buf) |
| 475 | { |
| 476 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); |
| 477 | } |
| 478 | static ssize_t pages_to_scan_store(struct kobject *kobj, |
| 479 | struct kobj_attribute *attr, |
| 480 | const char *buf, size_t count) |
| 481 | { |
| 482 | int err; |
| 483 | unsigned long pages; |
| 484 | |
Jingoo Han | 3dbb95f | 2013-09-11 14:20:25 -0700 | [diff] [blame] | 485 | err = kstrtoul(buf, 10, &pages); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 486 | if (err || !pages || pages > UINT_MAX) |
| 487 | return -EINVAL; |
| 488 | |
| 489 | khugepaged_pages_to_scan = pages; |
| 490 | |
| 491 | return count; |
| 492 | } |
| 493 | static struct kobj_attribute pages_to_scan_attr = |
| 494 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, |
| 495 | pages_to_scan_store); |
| 496 | |
| 497 | static ssize_t pages_collapsed_show(struct kobject *kobj, |
| 498 | struct kobj_attribute *attr, |
| 499 | char *buf) |
| 500 | { |
| 501 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); |
| 502 | } |
| 503 | static struct kobj_attribute pages_collapsed_attr = |
| 504 | __ATTR_RO(pages_collapsed); |
| 505 | |
| 506 | static ssize_t full_scans_show(struct kobject *kobj, |
| 507 | struct kobj_attribute *attr, |
| 508 | char *buf) |
| 509 | { |
| 510 | return sprintf(buf, "%u\n", khugepaged_full_scans); |
| 511 | } |
| 512 | static struct kobj_attribute full_scans_attr = |
| 513 | __ATTR_RO(full_scans); |
| 514 | |
| 515 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, |
| 516 | struct kobj_attribute *attr, char *buf) |
| 517 | { |
| 518 | return single_flag_show(kobj, attr, buf, |
| 519 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); |
| 520 | } |
| 521 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, |
| 522 | struct kobj_attribute *attr, |
| 523 | const char *buf, size_t count) |
| 524 | { |
| 525 | return single_flag_store(kobj, attr, buf, count, |
| 526 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); |
| 527 | } |
| 528 | static struct kobj_attribute khugepaged_defrag_attr = |
| 529 | __ATTR(defrag, 0644, khugepaged_defrag_show, |
| 530 | khugepaged_defrag_store); |
| 531 | |
| 532 | /* |
| 533 | * max_ptes_none controls if khugepaged should collapse hugepages over |
| 534 | * any unmapped ptes in turn potentially increasing the memory |
| 535 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not |
| 536 | * reduce the available free memory in the system as it |
| 537 | * runs. Increasing max_ptes_none will instead potentially reduce the |
| 538 | * free memory in the system during the khugepaged scan. |
| 539 | */ |
| 540 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, |
| 541 | struct kobj_attribute *attr, |
| 542 | char *buf) |
| 543 | { |
| 544 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); |
| 545 | } |
| 546 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, |
| 547 | struct kobj_attribute *attr, |
| 548 | const char *buf, size_t count) |
| 549 | { |
| 550 | int err; |
| 551 | unsigned long max_ptes_none; |
| 552 | |
Jingoo Han | 3dbb95f | 2013-09-11 14:20:25 -0700 | [diff] [blame] | 553 | err = kstrtoul(buf, 10, &max_ptes_none); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 554 | if (err || max_ptes_none > HPAGE_PMD_NR-1) |
| 555 | return -EINVAL; |
| 556 | |
| 557 | khugepaged_max_ptes_none = max_ptes_none; |
| 558 | |
| 559 | return count; |
| 560 | } |
| 561 | static struct kobj_attribute khugepaged_max_ptes_none_attr = |
| 562 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, |
| 563 | khugepaged_max_ptes_none_store); |
| 564 | |
| 565 | static struct attribute *khugepaged_attr[] = { |
| 566 | &khugepaged_defrag_attr.attr, |
| 567 | &khugepaged_max_ptes_none_attr.attr, |
| 568 | &pages_to_scan_attr.attr, |
| 569 | &pages_collapsed_attr.attr, |
| 570 | &full_scans_attr.attr, |
| 571 | &scan_sleep_millisecs_attr.attr, |
| 572 | &alloc_sleep_millisecs_attr.attr, |
| 573 | NULL, |
| 574 | }; |
| 575 | |
| 576 | static struct attribute_group khugepaged_attr_group = { |
| 577 | .attrs = khugepaged_attr, |
| 578 | .name = "khugepaged", |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 579 | }; |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 580 | |
| 581 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
| 582 | { |
| 583 | int err; |
| 584 | |
| 585 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
| 586 | if (unlikely(!*hugepage_kobj)) { |
Jeremy Eder | 2c79737 | 2012-12-20 15:05:32 -0800 | [diff] [blame] | 587 | printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n"); |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 588 | return -ENOMEM; |
| 589 | } |
| 590 | |
| 591 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
| 592 | if (err) { |
Jeremy Eder | 2c79737 | 2012-12-20 15:05:32 -0800 | [diff] [blame] | 593 | printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n"); |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 594 | goto delete_obj; |
| 595 | } |
| 596 | |
| 597 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
| 598 | if (err) { |
Jeremy Eder | 2c79737 | 2012-12-20 15:05:32 -0800 | [diff] [blame] | 599 | printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n"); |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 600 | goto remove_hp_group; |
| 601 | } |
| 602 | |
| 603 | return 0; |
| 604 | |
| 605 | remove_hp_group: |
| 606 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); |
| 607 | delete_obj: |
| 608 | kobject_put(*hugepage_kobj); |
| 609 | return err; |
| 610 | } |
| 611 | |
| 612 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) |
| 613 | { |
| 614 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); |
| 615 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); |
| 616 | kobject_put(hugepage_kobj); |
| 617 | } |
| 618 | #else |
| 619 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) |
| 620 | { |
| 621 | return 0; |
| 622 | } |
| 623 | |
| 624 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) |
| 625 | { |
| 626 | } |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 627 | #endif /* CONFIG_SYSFS */ |
| 628 | |
| 629 | static int __init hugepage_init(void) |
| 630 | { |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 631 | int err; |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 632 | struct kobject *hugepage_kobj; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 633 | |
Andrea Arcangeli | 4b7167b | 2011-01-13 15:47:09 -0800 | [diff] [blame] | 634 | if (!has_transparent_hugepage()) { |
| 635 | transparent_hugepage_flags = 0; |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 636 | return -EINVAL; |
Andrea Arcangeli | 4b7167b | 2011-01-13 15:47:09 -0800 | [diff] [blame] | 637 | } |
| 638 | |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 639 | err = hugepage_init_sysfs(&hugepage_kobj); |
| 640 | if (err) |
| 641 | return err; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 642 | |
| 643 | err = khugepaged_slab_init(); |
| 644 | if (err) |
| 645 | goto out; |
| 646 | |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 647 | register_shrinker(&huge_zero_page_shrinker); |
| 648 | |
Rik van Riel | 97562cd | 2011-01-13 15:47:12 -0800 | [diff] [blame] | 649 | /* |
| 650 | * By default disable transparent hugepages on smaller systems, |
| 651 | * where the extra memory used could hurt more than TLB overhead |
| 652 | * is likely to save. The admin can still enable it through /sys. |
| 653 | */ |
| 654 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) |
| 655 | transparent_hugepage_flags = 0; |
| 656 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 657 | start_khugepaged(); |
| 658 | |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 659 | return 0; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 660 | out: |
Shaohua Li | 569e559 | 2012-01-12 17:19:11 -0800 | [diff] [blame] | 661 | hugepage_exit_sysfs(hugepage_kobj); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 662 | return err; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 663 | } |
Paul Gortmaker | a64fb3c | 2014-01-23 15:53:30 -0800 | [diff] [blame] | 664 | subsys_initcall(hugepage_init); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 665 | |
| 666 | static int __init setup_transparent_hugepage(char *str) |
| 667 | { |
| 668 | int ret = 0; |
| 669 | if (!str) |
| 670 | goto out; |
| 671 | if (!strcmp(str, "always")) { |
| 672 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, |
| 673 | &transparent_hugepage_flags); |
| 674 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, |
| 675 | &transparent_hugepage_flags); |
| 676 | ret = 1; |
| 677 | } else if (!strcmp(str, "madvise")) { |
| 678 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, |
| 679 | &transparent_hugepage_flags); |
| 680 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, |
| 681 | &transparent_hugepage_flags); |
| 682 | ret = 1; |
| 683 | } else if (!strcmp(str, "never")) { |
| 684 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, |
| 685 | &transparent_hugepage_flags); |
| 686 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, |
| 687 | &transparent_hugepage_flags); |
| 688 | ret = 1; |
| 689 | } |
| 690 | out: |
| 691 | if (!ret) |
| 692 | printk(KERN_WARNING |
| 693 | "transparent_hugepage= cannot parse, ignored\n"); |
| 694 | return ret; |
| 695 | } |
| 696 | __setup("transparent_hugepage=", setup_transparent_hugepage); |
| 697 | |
Mel Gorman | b32967f | 2012-11-19 12:35:47 +0000 | [diff] [blame] | 698 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 699 | { |
| 700 | if (likely(vma->vm_flags & VM_WRITE)) |
| 701 | pmd = pmd_mkwrite(pmd); |
| 702 | return pmd; |
| 703 | } |
| 704 | |
Kirill A. Shutemov | 3122359 | 2013-09-12 15:14:01 -0700 | [diff] [blame] | 705 | static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) |
Bob Liu | b3092b3 | 2012-12-11 16:00:41 -0800 | [diff] [blame] | 706 | { |
| 707 | pmd_t entry; |
Kirill A. Shutemov | 3122359 | 2013-09-12 15:14:01 -0700 | [diff] [blame] | 708 | entry = mk_pmd(page, prot); |
Bob Liu | b3092b3 | 2012-12-11 16:00:41 -0800 | [diff] [blame] | 709 | entry = pmd_mkhuge(entry); |
| 710 | return entry; |
| 711 | } |
| 712 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 713 | static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, |
| 714 | struct vm_area_struct *vma, |
| 715 | unsigned long haddr, pmd_t *pmd, |
| 716 | struct page *page) |
| 717 | { |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 718 | pgtable_t pgtable; |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 719 | spinlock_t *ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 720 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 721 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 722 | pgtable = pte_alloc_one(mm, haddr); |
David Rientjes | edad9d2 | 2012-05-29 15:06:17 -0700 | [diff] [blame] | 723 | if (unlikely(!pgtable)) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 724 | return VM_FAULT_OOM; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 725 | |
| 726 | clear_huge_page(page, haddr, HPAGE_PMD_NR); |
Minchan Kim | 52f3762 | 2013-04-29 15:08:15 -0700 | [diff] [blame] | 727 | /* |
| 728 | * The memory barrier inside __SetPageUptodate makes sure that |
| 729 | * clear_huge_page writes become visible before the set_pmd_at() |
| 730 | * write. |
| 731 | */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 732 | __SetPageUptodate(page); |
| 733 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 734 | ptl = pmd_lock(mm, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 735 | if (unlikely(!pmd_none(*pmd))) { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 736 | spin_unlock(ptl); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 737 | mem_cgroup_uncharge_page(page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 738 | put_page(page); |
| 739 | pte_free(mm, pgtable); |
| 740 | } else { |
| 741 | pmd_t entry; |
Kirill A. Shutemov | 3122359 | 2013-09-12 15:14:01 -0700 | [diff] [blame] | 742 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
| 743 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 744 | page_add_new_anon_rmap(page, vma, haddr); |
Aneesh Kumar K.V | 6b0b50b | 2013-06-05 17:14:02 -0700 | [diff] [blame] | 745 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 746 | set_pmd_at(mm, haddr, pmd, entry); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 747 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
Kirill A. Shutemov | e1f56c8 | 2013-11-14 14:30:48 -0800 | [diff] [blame] | 748 | atomic_long_inc(&mm->nr_ptes); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 749 | spin_unlock(ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 750 | } |
| 751 | |
David Rientjes | aa2e878 | 2012-05-29 15:06:17 -0700 | [diff] [blame] | 752 | return 0; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 753 | } |
| 754 | |
Andi Kleen | cc5d462 | 2011-03-22 16:33:13 -0700 | [diff] [blame] | 755 | static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 756 | { |
Andi Kleen | cc5d462 | 2011-03-22 16:33:13 -0700 | [diff] [blame] | 757 | return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp; |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 758 | } |
| 759 | |
| 760 | static inline struct page *alloc_hugepage_vma(int defrag, |
| 761 | struct vm_area_struct *vma, |
Andi Kleen | cc5d462 | 2011-03-22 16:33:13 -0700 | [diff] [blame] | 762 | unsigned long haddr, int nd, |
| 763 | gfp_t extra_gfp) |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 764 | { |
Andi Kleen | cc5d462 | 2011-03-22 16:33:13 -0700 | [diff] [blame] | 765 | return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp), |
Andi Kleen | 5c4b4be | 2011-03-04 17:36:32 -0800 | [diff] [blame] | 766 | HPAGE_PMD_ORDER, vma, haddr, nd); |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 767 | } |
| 768 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 769 | /* Caller must hold page table lock. */ |
Kirill A. Shutemov | 3ea41e6 | 2012-12-12 13:51:14 -0800 | [diff] [blame] | 770 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 771 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 772 | struct page *zero_page) |
Kirill A. Shutemov | fc9fe82 | 2012-12-12 13:50:51 -0800 | [diff] [blame] | 773 | { |
| 774 | pmd_t entry; |
Kirill A. Shutemov | 3ea41e6 | 2012-12-12 13:51:14 -0800 | [diff] [blame] | 775 | if (!pmd_none(*pmd)) |
| 776 | return false; |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 777 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
Kirill A. Shutemov | fc9fe82 | 2012-12-12 13:50:51 -0800 | [diff] [blame] | 778 | entry = pmd_wrprotect(entry); |
| 779 | entry = pmd_mkhuge(entry); |
Aneesh Kumar K.V | 6b0b50b | 2013-06-05 17:14:02 -0700 | [diff] [blame] | 780 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
Kirill A. Shutemov | fc9fe82 | 2012-12-12 13:50:51 -0800 | [diff] [blame] | 781 | set_pmd_at(mm, haddr, pmd, entry); |
Kirill A. Shutemov | e1f56c8 | 2013-11-14 14:30:48 -0800 | [diff] [blame] | 782 | atomic_long_inc(&mm->nr_ptes); |
Kirill A. Shutemov | 3ea41e6 | 2012-12-12 13:51:14 -0800 | [diff] [blame] | 783 | return true; |
Kirill A. Shutemov | fc9fe82 | 2012-12-12 13:50:51 -0800 | [diff] [blame] | 784 | } |
| 785 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 786 | int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, |
| 787 | unsigned long address, pmd_t *pmd, |
| 788 | unsigned int flags) |
| 789 | { |
| 790 | struct page *page; |
| 791 | unsigned long haddr = address & HPAGE_PMD_MASK; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 792 | |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 793 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
Kirill A. Shutemov | c029255 | 2013-09-12 15:14:05 -0700 | [diff] [blame] | 794 | return VM_FAULT_FALLBACK; |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 795 | if (unlikely(anon_vma_prepare(vma))) |
| 796 | return VM_FAULT_OOM; |
| 797 | if (unlikely(khugepaged_enter(vma))) |
| 798 | return VM_FAULT_OOM; |
| 799 | if (!(flags & FAULT_FLAG_WRITE) && |
| 800 | transparent_hugepage_use_zero_page()) { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 801 | spinlock_t *ptl; |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 802 | pgtable_t pgtable; |
| 803 | struct page *zero_page; |
| 804 | bool set; |
| 805 | pgtable = pte_alloc_one(mm, haddr); |
| 806 | if (unlikely(!pgtable)) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 807 | return VM_FAULT_OOM; |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 808 | zero_page = get_huge_zero_page(); |
| 809 | if (unlikely(!zero_page)) { |
| 810 | pte_free(mm, pgtable); |
Andi Kleen | 81ab420 | 2011-04-14 15:22:06 -0700 | [diff] [blame] | 811 | count_vm_event(THP_FAULT_FALLBACK); |
Kirill A. Shutemov | c029255 | 2013-09-12 15:14:05 -0700 | [diff] [blame] | 812 | return VM_FAULT_FALLBACK; |
Andi Kleen | 81ab420 | 2011-04-14 15:22:06 -0700 | [diff] [blame] | 813 | } |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 814 | ptl = pmd_lock(mm, pmd); |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 815 | set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd, |
| 816 | zero_page); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 817 | spin_unlock(ptl); |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 818 | if (!set) { |
| 819 | pte_free(mm, pgtable); |
| 820 | put_huge_zero_page(); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 821 | } |
David Rientjes | edad9d2 | 2012-05-29 15:06:17 -0700 | [diff] [blame] | 822 | return 0; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 823 | } |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 824 | page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), |
| 825 | vma, haddr, numa_node_id(), 0); |
| 826 | if (unlikely(!page)) { |
| 827 | count_vm_event(THP_FAULT_FALLBACK); |
Kirill A. Shutemov | c029255 | 2013-09-12 15:14:05 -0700 | [diff] [blame] | 828 | return VM_FAULT_FALLBACK; |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 829 | } |
Michal Hocko | d715ae0 | 2014-04-07 15:37:46 -0700 | [diff] [blame] | 830 | if (unlikely(mem_cgroup_charge_anon(page, mm, GFP_KERNEL))) { |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 831 | put_page(page); |
David Rientjes | 17766dd | 2013-09-12 15:14:06 -0700 | [diff] [blame] | 832 | count_vm_event(THP_FAULT_FALLBACK); |
Kirill A. Shutemov | c029255 | 2013-09-12 15:14:05 -0700 | [diff] [blame] | 833 | return VM_FAULT_FALLBACK; |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 834 | } |
| 835 | if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) { |
| 836 | mem_cgroup_uncharge_page(page); |
| 837 | put_page(page); |
David Rientjes | 17766dd | 2013-09-12 15:14:06 -0700 | [diff] [blame] | 838 | count_vm_event(THP_FAULT_FALLBACK); |
Kirill A. Shutemov | c029255 | 2013-09-12 15:14:05 -0700 | [diff] [blame] | 839 | return VM_FAULT_FALLBACK; |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 840 | } |
| 841 | |
David Rientjes | 17766dd | 2013-09-12 15:14:06 -0700 | [diff] [blame] | 842 | count_vm_event(THP_FAULT_ALLOC); |
Kirill A. Shutemov | 128ec03 | 2013-09-12 15:14:03 -0700 | [diff] [blame] | 843 | return 0; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 844 | } |
| 845 | |
| 846 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
| 847 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, |
| 848 | struct vm_area_struct *vma) |
| 849 | { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 850 | spinlock_t *dst_ptl, *src_ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 851 | struct page *src_page; |
| 852 | pmd_t pmd; |
| 853 | pgtable_t pgtable; |
| 854 | int ret; |
| 855 | |
| 856 | ret = -ENOMEM; |
| 857 | pgtable = pte_alloc_one(dst_mm, addr); |
| 858 | if (unlikely(!pgtable)) |
| 859 | goto out; |
| 860 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 861 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
| 862 | src_ptl = pmd_lockptr(src_mm, src_pmd); |
| 863 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 864 | |
| 865 | ret = -EAGAIN; |
| 866 | pmd = *src_pmd; |
| 867 | if (unlikely(!pmd_trans_huge(pmd))) { |
| 868 | pte_free(dst_mm, pgtable); |
| 869 | goto out_unlock; |
| 870 | } |
Kirill A. Shutemov | fc9fe82 | 2012-12-12 13:50:51 -0800 | [diff] [blame] | 871 | /* |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 872 | * When page table lock is held, the huge zero pmd should not be |
Kirill A. Shutemov | fc9fe82 | 2012-12-12 13:50:51 -0800 | [diff] [blame] | 873 | * under splitting since we don't split the page itself, only pmd to |
| 874 | * a page table. |
| 875 | */ |
| 876 | if (is_huge_zero_pmd(pmd)) { |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 877 | struct page *zero_page; |
Kirill A. Shutemov | 3ea41e6 | 2012-12-12 13:51:14 -0800 | [diff] [blame] | 878 | bool set; |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 879 | /* |
| 880 | * get_huge_zero_page() will never allocate a new page here, |
| 881 | * since we already have a zero page to copy. It just takes a |
| 882 | * reference. |
| 883 | */ |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 884 | zero_page = get_huge_zero_page(); |
Kirill A. Shutemov | 3ea41e6 | 2012-12-12 13:51:14 -0800 | [diff] [blame] | 885 | set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 886 | zero_page); |
Kirill A. Shutemov | 3ea41e6 | 2012-12-12 13:51:14 -0800 | [diff] [blame] | 887 | BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */ |
Kirill A. Shutemov | fc9fe82 | 2012-12-12 13:50:51 -0800 | [diff] [blame] | 888 | ret = 0; |
| 889 | goto out_unlock; |
| 890 | } |
Mel Gorman | de466bd | 2013-12-18 17:08:42 -0800 | [diff] [blame] | 891 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 892 | if (unlikely(pmd_trans_splitting(pmd))) { |
| 893 | /* split huge page running from under us */ |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 894 | spin_unlock(src_ptl); |
| 895 | spin_unlock(dst_ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 896 | pte_free(dst_mm, pgtable); |
| 897 | |
| 898 | wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ |
| 899 | goto out; |
| 900 | } |
| 901 | src_page = pmd_page(pmd); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 902 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 903 | get_page(src_page); |
| 904 | page_dup_rmap(src_page); |
| 905 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
| 906 | |
| 907 | pmdp_set_wrprotect(src_mm, addr, src_pmd); |
| 908 | pmd = pmd_mkold(pmd_wrprotect(pmd)); |
Aneesh Kumar K.V | 6b0b50b | 2013-06-05 17:14:02 -0700 | [diff] [blame] | 909 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 910 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
Kirill A. Shutemov | e1f56c8 | 2013-11-14 14:30:48 -0800 | [diff] [blame] | 911 | atomic_long_inc(&dst_mm->nr_ptes); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 912 | |
| 913 | ret = 0; |
| 914 | out_unlock: |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 915 | spin_unlock(src_ptl); |
| 916 | spin_unlock(dst_ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 917 | out: |
| 918 | return ret; |
| 919 | } |
| 920 | |
Will Deacon | a1dd450 | 2012-12-11 16:01:27 -0800 | [diff] [blame] | 921 | void huge_pmd_set_accessed(struct mm_struct *mm, |
| 922 | struct vm_area_struct *vma, |
| 923 | unsigned long address, |
| 924 | pmd_t *pmd, pmd_t orig_pmd, |
| 925 | int dirty) |
| 926 | { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 927 | spinlock_t *ptl; |
Will Deacon | a1dd450 | 2012-12-11 16:01:27 -0800 | [diff] [blame] | 928 | pmd_t entry; |
| 929 | unsigned long haddr; |
| 930 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 931 | ptl = pmd_lock(mm, pmd); |
Will Deacon | a1dd450 | 2012-12-11 16:01:27 -0800 | [diff] [blame] | 932 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
| 933 | goto unlock; |
| 934 | |
| 935 | entry = pmd_mkyoung(orig_pmd); |
| 936 | haddr = address & HPAGE_PMD_MASK; |
| 937 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) |
| 938 | update_mmu_cache_pmd(vma, address, pmd); |
| 939 | |
| 940 | unlock: |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 941 | spin_unlock(ptl); |
Will Deacon | a1dd450 | 2012-12-11 16:01:27 -0800 | [diff] [blame] | 942 | } |
| 943 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 944 | static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, |
| 945 | struct vm_area_struct *vma, |
| 946 | unsigned long address, |
| 947 | pmd_t *pmd, pmd_t orig_pmd, |
| 948 | struct page *page, |
| 949 | unsigned long haddr) |
| 950 | { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 951 | spinlock_t *ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 952 | pgtable_t pgtable; |
| 953 | pmd_t _pmd; |
| 954 | int ret = 0, i; |
| 955 | struct page **pages; |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 956 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 957 | unsigned long mmun_end; /* For mmu_notifiers */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 958 | |
| 959 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, |
| 960 | GFP_KERNEL); |
| 961 | if (unlikely(!pages)) { |
| 962 | ret |= VM_FAULT_OOM; |
| 963 | goto out; |
| 964 | } |
| 965 | |
| 966 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
Andi Kleen | cc5d462 | 2011-03-22 16:33:13 -0700 | [diff] [blame] | 967 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | |
| 968 | __GFP_OTHER_NODE, |
Andi Kleen | 19ee151 | 2011-03-04 17:36:31 -0800 | [diff] [blame] | 969 | vma, address, page_to_nid(page)); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 970 | if (unlikely(!pages[i] || |
Michal Hocko | d715ae0 | 2014-04-07 15:37:46 -0700 | [diff] [blame] | 971 | mem_cgroup_charge_anon(pages[i], mm, |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 972 | GFP_KERNEL))) { |
| 973 | if (pages[i]) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 974 | put_page(pages[i]); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 975 | mem_cgroup_uncharge_start(); |
| 976 | while (--i >= 0) { |
| 977 | mem_cgroup_uncharge_page(pages[i]); |
| 978 | put_page(pages[i]); |
| 979 | } |
| 980 | mem_cgroup_uncharge_end(); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 981 | kfree(pages); |
| 982 | ret |= VM_FAULT_OOM; |
| 983 | goto out; |
| 984 | } |
| 985 | } |
| 986 | |
| 987 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
| 988 | copy_user_highpage(pages[i], page + i, |
Hillf Danton | 0089e48 | 2011-10-31 17:09:38 -0700 | [diff] [blame] | 989 | haddr + PAGE_SIZE * i, vma); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 990 | __SetPageUptodate(pages[i]); |
| 991 | cond_resched(); |
| 992 | } |
| 993 | |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 994 | mmun_start = haddr; |
| 995 | mmun_end = haddr + HPAGE_PMD_SIZE; |
| 996 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
| 997 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 998 | ptl = pmd_lock(mm, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 999 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
| 1000 | goto out_free_pages; |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1001 | VM_BUG_ON_PAGE(!PageHead(page), page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1002 | |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1003 | pmdp_clear_flush(vma, haddr, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1004 | /* leave pmd empty until pte is filled */ |
| 1005 | |
Aneesh Kumar K.V | 6b0b50b | 2013-06-05 17:14:02 -0700 | [diff] [blame] | 1006 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1007 | pmd_populate(mm, &_pmd, pgtable); |
| 1008 | |
| 1009 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { |
| 1010 | pte_t *pte, entry; |
| 1011 | entry = mk_pte(pages[i], vma->vm_page_prot); |
| 1012 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
| 1013 | page_add_new_anon_rmap(pages[i], vma, haddr); |
| 1014 | pte = pte_offset_map(&_pmd, haddr); |
| 1015 | VM_BUG_ON(!pte_none(*pte)); |
| 1016 | set_pte_at(mm, haddr, pte, entry); |
| 1017 | pte_unmap(pte); |
| 1018 | } |
| 1019 | kfree(pages); |
| 1020 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1021 | smp_wmb(); /* make pte visible before pmd */ |
| 1022 | pmd_populate(mm, pmd, pgtable); |
| 1023 | page_remove_rmap(page); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1024 | spin_unlock(ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1025 | |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1026 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
| 1027 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1028 | ret |= VM_FAULT_WRITE; |
| 1029 | put_page(page); |
| 1030 | |
| 1031 | out: |
| 1032 | return ret; |
| 1033 | |
| 1034 | out_free_pages: |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1035 | spin_unlock(ptl); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1036 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 1037 | mem_cgroup_uncharge_start(); |
| 1038 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
| 1039 | mem_cgroup_uncharge_page(pages[i]); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1040 | put_page(pages[i]); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 1041 | } |
| 1042 | mem_cgroup_uncharge_end(); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1043 | kfree(pages); |
| 1044 | goto out; |
| 1045 | } |
| 1046 | |
| 1047 | int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, |
| 1048 | unsigned long address, pmd_t *pmd, pmd_t orig_pmd) |
| 1049 | { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1050 | spinlock_t *ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1051 | int ret = 0; |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1052 | struct page *page = NULL, *new_page; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1053 | unsigned long haddr; |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1054 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 1055 | unsigned long mmun_end; /* For mmu_notifiers */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1056 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1057 | ptl = pmd_lockptr(mm, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1058 | VM_BUG_ON(!vma->anon_vma); |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1059 | haddr = address & HPAGE_PMD_MASK; |
| 1060 | if (is_huge_zero_pmd(orig_pmd)) |
| 1061 | goto alloc; |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1062 | spin_lock(ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1063 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
| 1064 | goto out_unlock; |
| 1065 | |
| 1066 | page = pmd_page(orig_pmd); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1067 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1068 | if (page_mapcount(page) == 1) { |
| 1069 | pmd_t entry; |
| 1070 | entry = pmd_mkyoung(orig_pmd); |
| 1071 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
| 1072 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) |
David Miller | b113da6 | 2012-10-08 16:34:25 -0700 | [diff] [blame] | 1073 | update_mmu_cache_pmd(vma, address, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1074 | ret |= VM_FAULT_WRITE; |
| 1075 | goto out_unlock; |
| 1076 | } |
| 1077 | get_page(page); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1078 | spin_unlock(ptl); |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1079 | alloc: |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1080 | if (transparent_hugepage_enabled(vma) && |
| 1081 | !transparent_hugepage_debug_cow()) |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 1082 | new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), |
Andi Kleen | cc5d462 | 2011-03-22 16:33:13 -0700 | [diff] [blame] | 1083 | vma, haddr, numa_node_id(), 0); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1084 | else |
| 1085 | new_page = NULL; |
| 1086 | |
| 1087 | if (unlikely(!new_page)) { |
Hugh Dickins | eecc1e4 | 2014-01-12 01:25:21 -0800 | [diff] [blame] | 1088 | if (!page) { |
Kirill A. Shutemov | e9b71ca | 2014-04-03 14:48:17 -0700 | [diff] [blame] | 1089 | split_huge_page_pmd(vma, address, pmd); |
| 1090 | ret |= VM_FAULT_FALLBACK; |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1091 | } else { |
| 1092 | ret = do_huge_pmd_wp_page_fallback(mm, vma, address, |
| 1093 | pmd, orig_pmd, page, haddr); |
Kirill A. Shutemov | 9845cbb | 2014-02-25 15:01:42 -0800 | [diff] [blame] | 1094 | if (ret & VM_FAULT_OOM) { |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1095 | split_huge_page(page); |
Kirill A. Shutemov | 9845cbb | 2014-02-25 15:01:42 -0800 | [diff] [blame] | 1096 | ret |= VM_FAULT_FALLBACK; |
| 1097 | } |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1098 | put_page(page); |
| 1099 | } |
David Rientjes | 17766dd | 2013-09-12 15:14:06 -0700 | [diff] [blame] | 1100 | count_vm_event(THP_FAULT_FALLBACK); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1101 | goto out; |
| 1102 | } |
| 1103 | |
Michal Hocko | d715ae0 | 2014-04-07 15:37:46 -0700 | [diff] [blame] | 1104 | if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))) { |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 1105 | put_page(new_page); |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1106 | if (page) { |
| 1107 | split_huge_page(page); |
| 1108 | put_page(page); |
Kirill A. Shutemov | 9845cbb | 2014-02-25 15:01:42 -0800 | [diff] [blame] | 1109 | } else |
| 1110 | split_huge_page_pmd(vma, address, pmd); |
| 1111 | ret |= VM_FAULT_FALLBACK; |
David Rientjes | 17766dd | 2013-09-12 15:14:06 -0700 | [diff] [blame] | 1112 | count_vm_event(THP_FAULT_FALLBACK); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 1113 | goto out; |
| 1114 | } |
| 1115 | |
David Rientjes | 17766dd | 2013-09-12 15:14:06 -0700 | [diff] [blame] | 1116 | count_vm_event(THP_FAULT_ALLOC); |
| 1117 | |
Hugh Dickins | eecc1e4 | 2014-01-12 01:25:21 -0800 | [diff] [blame] | 1118 | if (!page) |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1119 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); |
| 1120 | else |
| 1121 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1122 | __SetPageUptodate(new_page); |
| 1123 | |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1124 | mmun_start = haddr; |
| 1125 | mmun_end = haddr + HPAGE_PMD_SIZE; |
| 1126 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
| 1127 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1128 | spin_lock(ptl); |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1129 | if (page) |
| 1130 | put_page(page); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 1131 | if (unlikely(!pmd_same(*pmd, orig_pmd))) { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1132 | spin_unlock(ptl); |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 1133 | mem_cgroup_uncharge_page(new_page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1134 | put_page(new_page); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1135 | goto out_mn; |
Andrea Arcangeli | b9bbfbe | 2011-01-13 15:46:57 -0800 | [diff] [blame] | 1136 | } else { |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1137 | pmd_t entry; |
Kirill A. Shutemov | 3122359 | 2013-09-12 15:14:01 -0700 | [diff] [blame] | 1138 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
| 1139 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1140 | pmdp_clear_flush(vma, haddr, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1141 | page_add_new_anon_rmap(new_page, vma, haddr); |
| 1142 | set_pmd_at(mm, haddr, pmd, entry); |
David Miller | b113da6 | 2012-10-08 16:34:25 -0700 | [diff] [blame] | 1143 | update_mmu_cache_pmd(vma, address, pmd); |
Hugh Dickins | eecc1e4 | 2014-01-12 01:25:21 -0800 | [diff] [blame] | 1144 | if (!page) { |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1145 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 1146 | put_huge_zero_page(); |
| 1147 | } else { |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1148 | VM_BUG_ON_PAGE(!PageHead(page), page); |
Kirill A. Shutemov | 93b4796 | 2012-12-12 13:50:54 -0800 | [diff] [blame] | 1149 | page_remove_rmap(page); |
| 1150 | put_page(page); |
| 1151 | } |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1152 | ret |= VM_FAULT_WRITE; |
| 1153 | } |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1154 | spin_unlock(ptl); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1155 | out_mn: |
| 1156 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
| 1157 | out: |
| 1158 | return ret; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1159 | out_unlock: |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1160 | spin_unlock(ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1161 | return ret; |
| 1162 | } |
| 1163 | |
David Rientjes | b676b29 | 2012-10-08 16:34:03 -0700 | [diff] [blame] | 1164 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1165 | unsigned long addr, |
| 1166 | pmd_t *pmd, |
| 1167 | unsigned int flags) |
| 1168 | { |
David Rientjes | b676b29 | 2012-10-08 16:34:03 -0700 | [diff] [blame] | 1169 | struct mm_struct *mm = vma->vm_mm; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1170 | struct page *page = NULL; |
| 1171 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1172 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1173 | |
| 1174 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) |
| 1175 | goto out; |
| 1176 | |
Kirill A. Shutemov | 85facf2 | 2013-02-04 14:28:42 -0800 | [diff] [blame] | 1177 | /* Avoid dumping huge zero page */ |
| 1178 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) |
| 1179 | return ERR_PTR(-EFAULT); |
| 1180 | |
Mel Gorman | 2b4847e | 2013-12-18 17:08:32 -0800 | [diff] [blame] | 1181 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
| 1182 | if ((flags & FOLL_NUMA) && pmd_numa(*pmd)) |
| 1183 | goto out; |
| 1184 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1185 | page = pmd_page(*pmd); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1186 | VM_BUG_ON_PAGE(!PageHead(page), page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1187 | if (flags & FOLL_TOUCH) { |
| 1188 | pmd_t _pmd; |
| 1189 | /* |
| 1190 | * We should set the dirty bit only for FOLL_WRITE but |
| 1191 | * for now the dirty bit in the pmd is meaningless. |
| 1192 | * And if the dirty bit will become meaningful and |
| 1193 | * we'll only set it with FOLL_WRITE, an atomic |
| 1194 | * set_bit will be required on the pmd to set the |
| 1195 | * young bit, instead of the current set_pmd_at. |
| 1196 | */ |
| 1197 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); |
Aneesh Kumar K.V | 8663890a | 2013-06-06 00:20:34 -0700 | [diff] [blame] | 1198 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
| 1199 | pmd, _pmd, 1)) |
| 1200 | update_mmu_cache_pmd(vma, addr, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1201 | } |
David Rientjes | b676b29 | 2012-10-08 16:34:03 -0700 | [diff] [blame] | 1202 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
| 1203 | if (page->mapping && trylock_page(page)) { |
| 1204 | lru_add_drain(); |
| 1205 | if (page->mapping) |
| 1206 | mlock_vma_page(page); |
| 1207 | unlock_page(page); |
| 1208 | } |
| 1209 | } |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1210 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1211 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1212 | if (flags & FOLL_GET) |
Andrea Arcangeli | 70b50f9 | 2011-11-02 13:36:59 -0700 | [diff] [blame] | 1213 | get_page_foll(page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1214 | |
| 1215 | out: |
| 1216 | return page; |
| 1217 | } |
| 1218 | |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1219 | /* NUMA hinting page fault entry point for trans huge pmds */ |
Mel Gorman | 4daae3b | 2012-11-02 11:33:45 +0000 | [diff] [blame] | 1220 | int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, |
| 1221 | unsigned long addr, pmd_t pmd, pmd_t *pmdp) |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1222 | { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1223 | spinlock_t *ptl; |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1224 | struct anon_vma *anon_vma = NULL; |
Mel Gorman | b32967f | 2012-11-19 12:35:47 +0000 | [diff] [blame] | 1225 | struct page *page; |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1226 | unsigned long haddr = addr & HPAGE_PMD_MASK; |
Mel Gorman | 8191acb | 2013-10-07 11:28:45 +0100 | [diff] [blame] | 1227 | int page_nid = -1, this_nid = numa_node_id(); |
Peter Zijlstra | 9057289 | 2013-10-07 11:29:20 +0100 | [diff] [blame] | 1228 | int target_nid, last_cpupid = -1; |
Mel Gorman | 8191acb | 2013-10-07 11:28:45 +0100 | [diff] [blame] | 1229 | bool page_locked; |
| 1230 | bool migrated = false; |
Peter Zijlstra | 6688cc0 | 2013-10-07 11:29:24 +0100 | [diff] [blame] | 1231 | int flags = 0; |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1232 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1233 | ptl = pmd_lock(mm, pmdp); |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1234 | if (unlikely(!pmd_same(pmd, *pmdp))) |
| 1235 | goto out_unlock; |
| 1236 | |
Mel Gorman | de466bd | 2013-12-18 17:08:42 -0800 | [diff] [blame] | 1237 | /* |
| 1238 | * If there are potential migrations, wait for completion and retry |
| 1239 | * without disrupting NUMA hinting information. Do not relock and |
| 1240 | * check_same as the page may no longer be mapped. |
| 1241 | */ |
| 1242 | if (unlikely(pmd_trans_migrating(*pmdp))) { |
| 1243 | spin_unlock(ptl); |
| 1244 | wait_migrate_huge_page(vma->anon_vma, pmdp); |
| 1245 | goto out; |
| 1246 | } |
| 1247 | |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1248 | page = pmd_page(pmd); |
Mel Gorman | a1a4618 | 2013-10-07 11:28:50 +0100 | [diff] [blame] | 1249 | BUG_ON(is_huge_zero_page(page)); |
Mel Gorman | 8191acb | 2013-10-07 11:28:45 +0100 | [diff] [blame] | 1250 | page_nid = page_to_nid(page); |
Peter Zijlstra | 9057289 | 2013-10-07 11:29:20 +0100 | [diff] [blame] | 1251 | last_cpupid = page_cpupid_last(page); |
Mel Gorman | 03c5a6e | 2012-11-02 14:52:48 +0000 | [diff] [blame] | 1252 | count_vm_numa_event(NUMA_HINT_FAULTS); |
Rik van Riel | 04bb2f9 | 2013-10-07 11:29:36 +0100 | [diff] [blame] | 1253 | if (page_nid == this_nid) { |
Mel Gorman | 03c5a6e | 2012-11-02 14:52:48 +0000 | [diff] [blame] | 1254 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
Rik van Riel | 04bb2f9 | 2013-10-07 11:29:36 +0100 | [diff] [blame] | 1255 | flags |= TNF_FAULT_LOCAL; |
| 1256 | } |
Mel Gorman | 4daae3b | 2012-11-02 11:33:45 +0000 | [diff] [blame] | 1257 | |
Mel Gorman | ff9042b | 2013-10-07 11:28:43 +0100 | [diff] [blame] | 1258 | /* |
Peter Zijlstra | 6688cc0 | 2013-10-07 11:29:24 +0100 | [diff] [blame] | 1259 | * Avoid grouping on DSO/COW pages in specific and RO pages |
| 1260 | * in general, RO pages shouldn't hurt as much anyway since |
| 1261 | * they can be in shared cache state. |
| 1262 | */ |
| 1263 | if (!pmd_write(pmd)) |
| 1264 | flags |= TNF_NO_GROUP; |
| 1265 | |
| 1266 | /* |
Mel Gorman | ff9042b | 2013-10-07 11:28:43 +0100 | [diff] [blame] | 1267 | * Acquire the page lock to serialise THP migrations but avoid dropping |
| 1268 | * page_table_lock if at all possible |
| 1269 | */ |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1270 | page_locked = trylock_page(page); |
| 1271 | target_nid = mpol_misplaced(page, vma, haddr); |
| 1272 | if (target_nid == -1) { |
| 1273 | /* If the page was locked, there are no parallel migrations */ |
Mel Gorman | a54a407 | 2013-10-07 11:28:46 +0100 | [diff] [blame] | 1274 | if (page_locked) |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1275 | goto clear_pmdnuma; |
Mel Gorman | 2b4847e | 2013-12-18 17:08:32 -0800 | [diff] [blame] | 1276 | } |
Mel Gorman | 4daae3b | 2012-11-02 11:33:45 +0000 | [diff] [blame] | 1277 | |
Mel Gorman | de466bd | 2013-12-18 17:08:42 -0800 | [diff] [blame] | 1278 | /* Migration could have started since the pmd_trans_migrating check */ |
Mel Gorman | 2b4847e | 2013-12-18 17:08:32 -0800 | [diff] [blame] | 1279 | if (!page_locked) { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1280 | spin_unlock(ptl); |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1281 | wait_on_page_locked(page); |
Mel Gorman | a54a407 | 2013-10-07 11:28:46 +0100 | [diff] [blame] | 1282 | page_nid = -1; |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1283 | goto out; |
| 1284 | } |
| 1285 | |
Mel Gorman | 2b4847e | 2013-12-18 17:08:32 -0800 | [diff] [blame] | 1286 | /* |
| 1287 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma |
| 1288 | * to serialises splits |
| 1289 | */ |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1290 | get_page(page); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1291 | spin_unlock(ptl); |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1292 | anon_vma = page_lock_anon_vma_read(page); |
Peter Zijlstra | cbee9f8 | 2012-10-25 14:16:43 +0200 | [diff] [blame] | 1293 | |
Peter Zijlstra | c69307d | 2013-10-07 11:28:41 +0100 | [diff] [blame] | 1294 | /* Confirm the PMD did not change while page_table_lock was released */ |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1295 | spin_lock(ptl); |
Mel Gorman | b32967f | 2012-11-19 12:35:47 +0000 | [diff] [blame] | 1296 | if (unlikely(!pmd_same(pmd, *pmdp))) { |
| 1297 | unlock_page(page); |
| 1298 | put_page(page); |
Mel Gorman | a54a407 | 2013-10-07 11:28:46 +0100 | [diff] [blame] | 1299 | page_nid = -1; |
Mel Gorman | b32967f | 2012-11-19 12:35:47 +0000 | [diff] [blame] | 1300 | goto out_unlock; |
| 1301 | } |
Mel Gorman | ff9042b | 2013-10-07 11:28:43 +0100 | [diff] [blame] | 1302 | |
Mel Gorman | c3a489c | 2013-12-18 17:08:38 -0800 | [diff] [blame] | 1303 | /* Bail if we fail to protect against THP splits for any reason */ |
| 1304 | if (unlikely(!anon_vma)) { |
| 1305 | put_page(page); |
| 1306 | page_nid = -1; |
| 1307 | goto clear_pmdnuma; |
| 1308 | } |
| 1309 | |
Mel Gorman | a54a407 | 2013-10-07 11:28:46 +0100 | [diff] [blame] | 1310 | /* |
| 1311 | * Migrate the THP to the requested node, returns with page unlocked |
| 1312 | * and pmd_numa cleared. |
| 1313 | */ |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1314 | spin_unlock(ptl); |
Mel Gorman | b32967f | 2012-11-19 12:35:47 +0000 | [diff] [blame] | 1315 | migrated = migrate_misplaced_transhuge_page(mm, vma, |
Hugh Dickins | 340ef39 | 2013-02-22 16:34:33 -0800 | [diff] [blame] | 1316 | pmdp, pmd, addr, page, target_nid); |
Peter Zijlstra | 6688cc0 | 2013-10-07 11:29:24 +0100 | [diff] [blame] | 1317 | if (migrated) { |
| 1318 | flags |= TNF_MIGRATED; |
Mel Gorman | 8191acb | 2013-10-07 11:28:45 +0100 | [diff] [blame] | 1319 | page_nid = target_nid; |
Peter Zijlstra | 6688cc0 | 2013-10-07 11:29:24 +0100 | [diff] [blame] | 1320 | } |
Mel Gorman | b32967f | 2012-11-19 12:35:47 +0000 | [diff] [blame] | 1321 | |
Mel Gorman | 8191acb | 2013-10-07 11:28:45 +0100 | [diff] [blame] | 1322 | goto out; |
Mel Gorman | 4daae3b | 2012-11-02 11:33:45 +0000 | [diff] [blame] | 1323 | clear_pmdnuma: |
Mel Gorman | a54a407 | 2013-10-07 11:28:46 +0100 | [diff] [blame] | 1324 | BUG_ON(!PageLocked(page)); |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1325 | pmd = pmd_mknonnuma(pmd); |
| 1326 | set_pmd_at(mm, haddr, pmdp, pmd); |
| 1327 | VM_BUG_ON(pmd_numa(*pmdp)); |
| 1328 | update_mmu_cache_pmd(vma, addr, pmdp); |
Mel Gorman | a54a407 | 2013-10-07 11:28:46 +0100 | [diff] [blame] | 1329 | unlock_page(page); |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1330 | out_unlock: |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 1331 | spin_unlock(ptl); |
Mel Gorman | b891663 | 2013-10-07 11:28:44 +0100 | [diff] [blame] | 1332 | |
| 1333 | out: |
| 1334 | if (anon_vma) |
| 1335 | page_unlock_anon_vma_read(anon_vma); |
| 1336 | |
Mel Gorman | 8191acb | 2013-10-07 11:28:45 +0100 | [diff] [blame] | 1337 | if (page_nid != -1) |
Peter Zijlstra | 6688cc0 | 2013-10-07 11:29:24 +0100 | [diff] [blame] | 1338 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); |
Mel Gorman | 8191acb | 2013-10-07 11:28:45 +0100 | [diff] [blame] | 1339 | |
Mel Gorman | d10e63f | 2012-10-25 14:16:31 +0200 | [diff] [blame] | 1340 | return 0; |
| 1341 | } |
| 1342 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1343 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
Shaohua Li | f21760b | 2012-01-12 17:19:16 -0800 | [diff] [blame] | 1344 | pmd_t *pmd, unsigned long addr) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1345 | { |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1346 | spinlock_t *ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1347 | int ret = 0; |
| 1348 | |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1349 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1350 | struct page *page; |
| 1351 | pgtable_t pgtable; |
David Miller | f5c8ad4 | 2012-10-08 16:34:26 -0700 | [diff] [blame] | 1352 | pmd_t orig_pmd; |
Aneesh Kumar K.V | a6bf2bb | 2013-06-05 17:14:04 -0700 | [diff] [blame] | 1353 | /* |
| 1354 | * For architectures like ppc64 we look at deposited pgtable |
| 1355 | * when calling pmdp_get_and_clear. So do the |
| 1356 | * pgtable_trans_huge_withdraw after finishing pmdp related |
| 1357 | * operations. |
| 1358 | */ |
David Miller | f5c8ad4 | 2012-10-08 16:34:26 -0700 | [diff] [blame] | 1359 | orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1360 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); |
Aneesh Kumar K.V | a6bf2bb | 2013-06-05 17:14:04 -0700 | [diff] [blame] | 1361 | pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd); |
Kirill A. Shutemov | 479f0ab | 2012-12-12 13:50:50 -0800 | [diff] [blame] | 1362 | if (is_huge_zero_pmd(orig_pmd)) { |
Kirill A. Shutemov | e1f56c8 | 2013-11-14 14:30:48 -0800 | [diff] [blame] | 1363 | atomic_long_dec(&tlb->mm->nr_ptes); |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1364 | spin_unlock(ptl); |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 1365 | put_huge_zero_page(); |
Kirill A. Shutemov | 479f0ab | 2012-12-12 13:50:50 -0800 | [diff] [blame] | 1366 | } else { |
| 1367 | page = pmd_page(orig_pmd); |
| 1368 | page_remove_rmap(page); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1369 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); |
Kirill A. Shutemov | 479f0ab | 2012-12-12 13:50:50 -0800 | [diff] [blame] | 1370 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1371 | VM_BUG_ON_PAGE(!PageHead(page), page); |
Kirill A. Shutemov | e1f56c8 | 2013-11-14 14:30:48 -0800 | [diff] [blame] | 1372 | atomic_long_dec(&tlb->mm->nr_ptes); |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1373 | spin_unlock(ptl); |
Kirill A. Shutemov | 479f0ab | 2012-12-12 13:50:50 -0800 | [diff] [blame] | 1374 | tlb_remove_page(tlb, page); |
| 1375 | } |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1376 | pte_free(tlb->mm, pgtable); |
| 1377 | ret = 1; |
| 1378 | } |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1379 | return ret; |
| 1380 | } |
| 1381 | |
Johannes Weiner | 0ca1634 | 2011-01-13 15:47:02 -0800 | [diff] [blame] | 1382 | int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
| 1383 | unsigned long addr, unsigned long end, |
| 1384 | unsigned char *vec) |
| 1385 | { |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1386 | spinlock_t *ptl; |
Johannes Weiner | 0ca1634 | 2011-01-13 15:47:02 -0800 | [diff] [blame] | 1387 | int ret = 0; |
| 1388 | |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1389 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1390 | /* |
| 1391 | * All logical pages in the range are present |
| 1392 | * if backed by a huge page. |
| 1393 | */ |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1394 | spin_unlock(ptl); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1395 | memset(vec, 1, (end - addr) >> PAGE_SHIFT); |
| 1396 | ret = 1; |
| 1397 | } |
Johannes Weiner | 0ca1634 | 2011-01-13 15:47:02 -0800 | [diff] [blame] | 1398 | |
| 1399 | return ret; |
| 1400 | } |
| 1401 | |
Andrea Arcangeli | 37a1c49 | 2011-10-31 17:08:30 -0700 | [diff] [blame] | 1402 | int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, |
| 1403 | unsigned long old_addr, |
| 1404 | unsigned long new_addr, unsigned long old_end, |
| 1405 | pmd_t *old_pmd, pmd_t *new_pmd) |
| 1406 | { |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1407 | spinlock_t *old_ptl, *new_ptl; |
Andrea Arcangeli | 37a1c49 | 2011-10-31 17:08:30 -0700 | [diff] [blame] | 1408 | int ret = 0; |
| 1409 | pmd_t pmd; |
| 1410 | |
| 1411 | struct mm_struct *mm = vma->vm_mm; |
| 1412 | |
| 1413 | if ((old_addr & ~HPAGE_PMD_MASK) || |
| 1414 | (new_addr & ~HPAGE_PMD_MASK) || |
| 1415 | old_end - old_addr < HPAGE_PMD_SIZE || |
| 1416 | (new_vma->vm_flags & VM_NOHUGEPAGE)) |
| 1417 | goto out; |
| 1418 | |
| 1419 | /* |
| 1420 | * The destination pmd shouldn't be established, free_pgtables() |
| 1421 | * should have release it. |
| 1422 | */ |
| 1423 | if (WARN_ON(!pmd_none(*new_pmd))) { |
| 1424 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); |
| 1425 | goto out; |
| 1426 | } |
| 1427 | |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1428 | /* |
| 1429 | * We don't have to worry about the ordering of src and dst |
| 1430 | * ptlocks because exclusive mmap_sem prevents deadlock. |
| 1431 | */ |
| 1432 | ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1433 | if (ret == 1) { |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1434 | new_ptl = pmd_lockptr(mm, new_pmd); |
| 1435 | if (new_ptl != old_ptl) |
| 1436 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1437 | pmd = pmdp_get_and_clear(mm, old_addr, old_pmd); |
| 1438 | VM_BUG_ON(!pmd_none(*new_pmd)); |
Kirill A. Shutemov | 3592806 | 2013-12-12 17:12:33 -0800 | [diff] [blame] | 1439 | |
Aneesh Kumar K.V | b3084f4 | 2014-01-13 11:34:24 +0530 | [diff] [blame] | 1440 | if (pmd_move_must_withdraw(new_ptl, old_ptl)) { |
| 1441 | pgtable_t pgtable; |
Kirill A. Shutemov | 3592806 | 2013-12-12 17:12:33 -0800 | [diff] [blame] | 1442 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
| 1443 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); |
Kirill A. Shutemov | 3592806 | 2013-12-12 17:12:33 -0800 | [diff] [blame] | 1444 | } |
Aneesh Kumar K.V | b3084f4 | 2014-01-13 11:34:24 +0530 | [diff] [blame] | 1445 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); |
| 1446 | if (new_ptl != old_ptl) |
| 1447 | spin_unlock(new_ptl); |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1448 | spin_unlock(old_ptl); |
Andrea Arcangeli | 37a1c49 | 2011-10-31 17:08:30 -0700 | [diff] [blame] | 1449 | } |
| 1450 | out: |
| 1451 | return ret; |
| 1452 | } |
| 1453 | |
Mel Gorman | f123d74 | 2013-10-07 11:28:49 +0100 | [diff] [blame] | 1454 | /* |
| 1455 | * Returns |
| 1456 | * - 0 if PMD could not be locked |
| 1457 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary |
| 1458 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary |
| 1459 | */ |
Johannes Weiner | cd7548a | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 1460 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
Mel Gorman | 4b10e7d | 2012-10-25 14:16:32 +0200 | [diff] [blame] | 1461 | unsigned long addr, pgprot_t newprot, int prot_numa) |
Johannes Weiner | cd7548a | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 1462 | { |
| 1463 | struct mm_struct *mm = vma->vm_mm; |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1464 | spinlock_t *ptl; |
Johannes Weiner | cd7548a | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 1465 | int ret = 0; |
| 1466 | |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1467 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1468 | pmd_t entry; |
Mel Gorman | f123d74 | 2013-10-07 11:28:49 +0100 | [diff] [blame] | 1469 | ret = 1; |
Hugh Dickins | a4f1de1 | 2012-12-16 18:56:58 -0800 | [diff] [blame] | 1470 | if (!prot_numa) { |
Mel Gorman | f123d74 | 2013-10-07 11:28:49 +0100 | [diff] [blame] | 1471 | entry = pmdp_get_and_clear(mm, addr, pmd); |
Mel Gorman | 1667918 | 2013-12-18 17:08:41 -0800 | [diff] [blame] | 1472 | if (pmd_numa(entry)) |
| 1473 | entry = pmd_mknonnuma(entry); |
Mel Gorman | 4b10e7d | 2012-10-25 14:16:32 +0200 | [diff] [blame] | 1474 | entry = pmd_modify(entry, newprot); |
Mel Gorman | f123d74 | 2013-10-07 11:28:49 +0100 | [diff] [blame] | 1475 | ret = HPAGE_PMD_NR; |
Aneesh Kumar K.V | 56eecdb | 2014-02-12 09:13:38 +0530 | [diff] [blame] | 1476 | set_pmd_at(mm, addr, pmd, entry); |
Hugh Dickins | a4f1de1 | 2012-12-16 18:56:58 -0800 | [diff] [blame] | 1477 | BUG_ON(pmd_write(entry)); |
| 1478 | } else { |
Mel Gorman | 4b10e7d | 2012-10-25 14:16:32 +0200 | [diff] [blame] | 1479 | struct page *page = pmd_page(*pmd); |
| 1480 | |
Mel Gorman | a1a4618 | 2013-10-07 11:28:50 +0100 | [diff] [blame] | 1481 | /* |
Mel Gorman | 1bc115d | 2013-10-07 11:29:05 +0100 | [diff] [blame] | 1482 | * Do not trap faults against the zero page. The |
| 1483 | * read-only data is likely to be read-cached on the |
| 1484 | * local CPU cache and it is less useful to know about |
| 1485 | * local vs remote hits on the zero page. |
Mel Gorman | a1a4618 | 2013-10-07 11:28:50 +0100 | [diff] [blame] | 1486 | */ |
Mel Gorman | 1bc115d | 2013-10-07 11:29:05 +0100 | [diff] [blame] | 1487 | if (!is_huge_zero_page(page) && |
Mel Gorman | 4b10e7d | 2012-10-25 14:16:32 +0200 | [diff] [blame] | 1488 | !pmd_numa(*pmd)) { |
Aneesh Kumar K.V | 56eecdb | 2014-02-12 09:13:38 +0530 | [diff] [blame] | 1489 | pmdp_set_numa(mm, addr, pmd); |
Mel Gorman | f123d74 | 2013-10-07 11:28:49 +0100 | [diff] [blame] | 1490 | ret = HPAGE_PMD_NR; |
Mel Gorman | 4b10e7d | 2012-10-25 14:16:32 +0200 | [diff] [blame] | 1491 | } |
| 1492 | } |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1493 | spin_unlock(ptl); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1494 | } |
Johannes Weiner | cd7548a | 2011-01-13 15:47:04 -0800 | [diff] [blame] | 1495 | |
| 1496 | return ret; |
| 1497 | } |
| 1498 | |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1499 | /* |
| 1500 | * Returns 1 if a given pmd maps a stable (not under splitting) thp. |
| 1501 | * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. |
| 1502 | * |
| 1503 | * Note that if it returns 1, this routine returns without unlocking page |
| 1504 | * table locks. So callers must unlock them. |
| 1505 | */ |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1506 | int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, |
| 1507 | spinlock_t **ptl) |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1508 | { |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1509 | *ptl = pmd_lock(vma->vm_mm, pmd); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1510 | if (likely(pmd_trans_huge(*pmd))) { |
| 1511 | if (unlikely(pmd_trans_splitting(*pmd))) { |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1512 | spin_unlock(*ptl); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1513 | wait_split_huge_page(vma->anon_vma, pmd); |
| 1514 | return -1; |
| 1515 | } else { |
| 1516 | /* Thp mapped by 'pmd' is stable, so we can |
| 1517 | * handle it as it is. */ |
| 1518 | return 1; |
| 1519 | } |
| 1520 | } |
Kirill A. Shutemov | bf92915 | 2013-11-14 14:30:54 -0800 | [diff] [blame] | 1521 | spin_unlock(*ptl); |
Naoya Horiguchi | 025c5b2 | 2012-03-21 16:33:57 -0700 | [diff] [blame] | 1522 | return 0; |
| 1523 | } |
| 1524 | |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1525 | /* |
| 1526 | * This function returns whether a given @page is mapped onto the @address |
| 1527 | * in the virtual space of @mm. |
| 1528 | * |
| 1529 | * When it's true, this function returns *pmd with holding the page table lock |
| 1530 | * and passing it back to the caller via @ptl. |
| 1531 | * If it's false, returns NULL without holding the page table lock. |
| 1532 | */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1533 | pmd_t *page_check_address_pmd(struct page *page, |
| 1534 | struct mm_struct *mm, |
| 1535 | unsigned long address, |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1536 | enum page_check_address_pmd_flag flag, |
| 1537 | spinlock_t **ptl) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1538 | { |
Kirill A. Shutemov | b5a8cad | 2014-04-18 15:07:25 -0700 | [diff] [blame] | 1539 | pgd_t *pgd; |
| 1540 | pud_t *pud; |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1541 | pmd_t *pmd; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1542 | |
| 1543 | if (address & ~HPAGE_PMD_MASK) |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1544 | return NULL; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1545 | |
Kirill A. Shutemov | b5a8cad | 2014-04-18 15:07:25 -0700 | [diff] [blame] | 1546 | pgd = pgd_offset(mm, address); |
| 1547 | if (!pgd_present(*pgd)) |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1548 | return NULL; |
Kirill A. Shutemov | b5a8cad | 2014-04-18 15:07:25 -0700 | [diff] [blame] | 1549 | pud = pud_offset(pgd, address); |
| 1550 | if (!pud_present(*pud)) |
| 1551 | return NULL; |
| 1552 | pmd = pmd_offset(pud, address); |
| 1553 | |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1554 | *ptl = pmd_lock(mm, pmd); |
Kirill A. Shutemov | b5a8cad | 2014-04-18 15:07:25 -0700 | [diff] [blame] | 1555 | if (!pmd_present(*pmd)) |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1556 | goto unlock; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1557 | if (pmd_page(*pmd) != page) |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1558 | goto unlock; |
Andrea Arcangeli | 94fcc58 | 2011-01-13 15:47:08 -0800 | [diff] [blame] | 1559 | /* |
| 1560 | * split_vma() may create temporary aliased mappings. There is |
| 1561 | * no risk as long as all huge pmd are found and have their |
| 1562 | * splitting bit set before __split_huge_page_refcount |
| 1563 | * runs. Finding the same huge pmd more than once during the |
| 1564 | * same rmap walk is not a problem. |
| 1565 | */ |
| 1566 | if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && |
| 1567 | pmd_trans_splitting(*pmd)) |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1568 | goto unlock; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1569 | if (pmd_trans_huge(*pmd)) { |
| 1570 | VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && |
| 1571 | !pmd_trans_splitting(*pmd)); |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1572 | return pmd; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1573 | } |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1574 | unlock: |
| 1575 | spin_unlock(*ptl); |
| 1576 | return NULL; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1577 | } |
| 1578 | |
| 1579 | static int __split_huge_page_splitting(struct page *page, |
| 1580 | struct vm_area_struct *vma, |
| 1581 | unsigned long address) |
| 1582 | { |
| 1583 | struct mm_struct *mm = vma->vm_mm; |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1584 | spinlock_t *ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1585 | pmd_t *pmd; |
| 1586 | int ret = 0; |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1587 | /* For mmu_notifiers */ |
| 1588 | const unsigned long mmun_start = address; |
| 1589 | const unsigned long mmun_end = address + HPAGE_PMD_SIZE; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1590 | |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1591 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1592 | pmd = page_check_address_pmd(page, mm, address, |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1593 | PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1594 | if (pmd) { |
| 1595 | /* |
| 1596 | * We can't temporarily set the pmd to null in order |
| 1597 | * to split it, the pmd must remain marked huge at all |
| 1598 | * times or the VM won't take the pmd_trans_huge paths |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 1599 | * and it won't wait on the anon_vma->root->rwsem to |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1600 | * serialize against split_huge_page*. |
| 1601 | */ |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1602 | pmdp_splitting_flush(vma, address, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1603 | ret = 1; |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1604 | spin_unlock(ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1605 | } |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 1606 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1607 | |
| 1608 | return ret; |
| 1609 | } |
| 1610 | |
Shaohua Li | 5bc7b8a | 2013-04-29 15:08:36 -0700 | [diff] [blame] | 1611 | static void __split_huge_page_refcount(struct page *page, |
| 1612 | struct list_head *list) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1613 | { |
| 1614 | int i; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1615 | struct zone *zone = page_zone(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1616 | struct lruvec *lruvec; |
Andrea Arcangeli | 70b50f9 | 2011-11-02 13:36:59 -0700 | [diff] [blame] | 1617 | int tail_count = 0; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1618 | |
| 1619 | /* prevent PageLRU to go away from under us, and freeze lru stats */ |
| 1620 | spin_lock_irq(&zone->lru_lock); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1621 | lruvec = mem_cgroup_page_lruvec(page, zone); |
| 1622 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1623 | compound_lock(page); |
KAMEZAWA Hiroyuki | e94c8a9 | 2012-01-12 17:18:20 -0800 | [diff] [blame] | 1624 | /* complete memcg works before add pages to LRU */ |
| 1625 | mem_cgroup_split_huge_fixup(page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1626 | |
Shaohua Li | 4567688 | 2012-01-12 17:19:18 -0800 | [diff] [blame] | 1627 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1628 | struct page *page_tail = page + i; |
| 1629 | |
Andrea Arcangeli | 70b50f9 | 2011-11-02 13:36:59 -0700 | [diff] [blame] | 1630 | /* tail_page->_mapcount cannot change */ |
| 1631 | BUG_ON(page_mapcount(page_tail) < 0); |
| 1632 | tail_count += page_mapcount(page_tail); |
| 1633 | /* check for overflow */ |
| 1634 | BUG_ON(tail_count < 0); |
| 1635 | BUG_ON(atomic_read(&page_tail->_count) != 0); |
| 1636 | /* |
| 1637 | * tail_page->_count is zero and not changing from |
| 1638 | * under us. But get_page_unless_zero() may be running |
| 1639 | * from under us on the tail_page. If we used |
| 1640 | * atomic_set() below instead of atomic_add(), we |
| 1641 | * would then run atomic_set() concurrently with |
| 1642 | * get_page_unless_zero(), and atomic_set() is |
| 1643 | * implemented in C not using locked ops. spin_unlock |
| 1644 | * on x86 sometime uses locked ops because of PPro |
| 1645 | * errata 66, 92, so unless somebody can guarantee |
| 1646 | * atomic_set() here would be safe on all archs (and |
| 1647 | * not only on x86), it's safer to use atomic_add(). |
| 1648 | */ |
| 1649 | atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, |
| 1650 | &page_tail->_count); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1651 | |
| 1652 | /* after clearing PageTail the gup refcount can be released */ |
| 1653 | smp_mb(); |
| 1654 | |
Jin Dongming | a6d30dd | 2011-02-01 15:52:40 -0800 | [diff] [blame] | 1655 | /* |
| 1656 | * retain hwpoison flag of the poisoned tail page: |
| 1657 | * fix for the unsuitable process killed on Guest Machine(KVM) |
| 1658 | * by the memory-failure. |
| 1659 | */ |
| 1660 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1661 | page_tail->flags |= (page->flags & |
| 1662 | ((1L << PG_referenced) | |
| 1663 | (1L << PG_swapbacked) | |
| 1664 | (1L << PG_mlocked) | |
Kirill A. Shutemov | e180cf8 | 2013-07-31 13:53:39 -0700 | [diff] [blame] | 1665 | (1L << PG_uptodate) | |
| 1666 | (1L << PG_active) | |
| 1667 | (1L << PG_unevictable))); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1668 | page_tail->flags |= (1L << PG_dirty); |
| 1669 | |
Andrea Arcangeli | 70b50f9 | 2011-11-02 13:36:59 -0700 | [diff] [blame] | 1670 | /* clear PageTail before overwriting first_page */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1671 | smp_wmb(); |
| 1672 | |
| 1673 | /* |
| 1674 | * __split_huge_page_splitting() already set the |
| 1675 | * splitting bit in all pmd that could map this |
| 1676 | * hugepage, that will ensure no CPU can alter the |
| 1677 | * mapcount on the head page. The mapcount is only |
| 1678 | * accounted in the head page and it has to be |
| 1679 | * transferred to all tail pages in the below code. So |
| 1680 | * for this code to be safe, the split the mapcount |
| 1681 | * can't change. But that doesn't mean userland can't |
| 1682 | * keep changing and reading the page contents while |
| 1683 | * we transfer the mapcount, so the pmd splitting |
| 1684 | * status is achieved setting a reserved bit in the |
| 1685 | * pmd, not by clearing the present bit. |
| 1686 | */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1687 | page_tail->_mapcount = page->_mapcount; |
| 1688 | |
| 1689 | BUG_ON(page_tail->mapping); |
| 1690 | page_tail->mapping = page->mapping; |
| 1691 | |
Shaohua Li | 4567688 | 2012-01-12 17:19:18 -0800 | [diff] [blame] | 1692 | page_tail->index = page->index + i; |
Peter Zijlstra | 9057289 | 2013-10-07 11:29:20 +0100 | [diff] [blame] | 1693 | page_cpupid_xchg_last(page_tail, page_cpupid_last(page)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1694 | |
| 1695 | BUG_ON(!PageAnon(page_tail)); |
| 1696 | BUG_ON(!PageUptodate(page_tail)); |
| 1697 | BUG_ON(!PageDirty(page_tail)); |
| 1698 | BUG_ON(!PageSwapBacked(page_tail)); |
| 1699 | |
Shaohua Li | 5bc7b8a | 2013-04-29 15:08:36 -0700 | [diff] [blame] | 1700 | lru_add_page_tail(page, page_tail, lruvec, list); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1701 | } |
Andrea Arcangeli | 70b50f9 | 2011-11-02 13:36:59 -0700 | [diff] [blame] | 1702 | atomic_sub(tail_count, &page->_count); |
| 1703 | BUG_ON(atomic_read(&page->_count) <= 0); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1704 | |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1705 | __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); |
Andrea Arcangeli | 7913417 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1706 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1707 | ClearPageCompound(page); |
| 1708 | compound_unlock(page); |
| 1709 | spin_unlock_irq(&zone->lru_lock); |
| 1710 | |
| 1711 | for (i = 1; i < HPAGE_PMD_NR; i++) { |
| 1712 | struct page *page_tail = page + i; |
| 1713 | BUG_ON(page_count(page_tail) <= 0); |
| 1714 | /* |
| 1715 | * Tail pages may be freed if there wasn't any mapping |
| 1716 | * like if add_to_swap() is running on a lru page that |
| 1717 | * had its mapping zapped. And freeing these pages |
| 1718 | * requires taking the lru_lock so we do the put_page |
| 1719 | * of the tail pages after the split is complete. |
| 1720 | */ |
| 1721 | put_page(page_tail); |
| 1722 | } |
| 1723 | |
| 1724 | /* |
| 1725 | * Only the head page (now become a regular page) is required |
| 1726 | * to be pinned by the caller. |
| 1727 | */ |
| 1728 | BUG_ON(page_count(page) <= 0); |
| 1729 | } |
| 1730 | |
| 1731 | static int __split_huge_page_map(struct page *page, |
| 1732 | struct vm_area_struct *vma, |
| 1733 | unsigned long address) |
| 1734 | { |
| 1735 | struct mm_struct *mm = vma->vm_mm; |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1736 | spinlock_t *ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1737 | pmd_t *pmd, _pmd; |
| 1738 | int ret = 0, i; |
| 1739 | pgtable_t pgtable; |
| 1740 | unsigned long haddr; |
| 1741 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1742 | pmd = page_check_address_pmd(page, mm, address, |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1743 | PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1744 | if (pmd) { |
Aneesh Kumar K.V | 6b0b50b | 2013-06-05 17:14:02 -0700 | [diff] [blame] | 1745 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1746 | pmd_populate(mm, &_pmd, pgtable); |
| 1747 | |
Gerald Schaefer | e3ebcf6 | 2012-10-08 16:30:07 -0700 | [diff] [blame] | 1748 | haddr = address; |
| 1749 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1750 | pte_t *pte, entry; |
| 1751 | BUG_ON(PageCompound(page+i)); |
| 1752 | entry = mk_pte(page + i, vma->vm_page_prot); |
| 1753 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
| 1754 | if (!pmd_write(*pmd)) |
| 1755 | entry = pte_wrprotect(entry); |
| 1756 | else |
| 1757 | BUG_ON(page_mapcount(page) != 1); |
| 1758 | if (!pmd_young(*pmd)) |
| 1759 | entry = pte_mkold(entry); |
Andrea Arcangeli | 1ba6e0b | 2012-10-04 01:51:06 +0200 | [diff] [blame] | 1760 | if (pmd_numa(*pmd)) |
| 1761 | entry = pte_mknuma(entry); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1762 | pte = pte_offset_map(&_pmd, haddr); |
| 1763 | BUG_ON(!pte_none(*pte)); |
| 1764 | set_pte_at(mm, haddr, pte, entry); |
| 1765 | pte_unmap(pte); |
| 1766 | } |
| 1767 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1768 | smp_wmb(); /* make pte visible before pmd */ |
| 1769 | /* |
| 1770 | * Up to this point the pmd is present and huge and |
| 1771 | * userland has the whole access to the hugepage |
| 1772 | * during the split (which happens in place). If we |
| 1773 | * overwrite the pmd with the not-huge version |
| 1774 | * pointing to the pte here (which of course we could |
| 1775 | * if all CPUs were bug free), userland could trigger |
| 1776 | * a small page size TLB miss on the small sized TLB |
| 1777 | * while the hugepage TLB entry is still established |
| 1778 | * in the huge TLB. Some CPU doesn't like that. See |
| 1779 | * http://support.amd.com/us/Processor_TechDocs/41322.pdf, |
| 1780 | * Erratum 383 on page 93. Intel should be safe but is |
| 1781 | * also warns that it's only safe if the permission |
| 1782 | * and cache attributes of the two entries loaded in |
| 1783 | * the two TLB is identical (which should be the case |
| 1784 | * here). But it is generally safer to never allow |
| 1785 | * small and huge TLB entries for the same virtual |
| 1786 | * address to be loaded simultaneously. So instead of |
| 1787 | * doing "pmd_populate(); flush_tlb_range();" we first |
| 1788 | * mark the current pmd notpresent (atomically because |
| 1789 | * here the pmd_trans_huge and pmd_trans_splitting |
| 1790 | * must remain set at all times on the pmd until the |
| 1791 | * split is complete for this pmd), then we flush the |
| 1792 | * SMP TLB and finally we write the non-huge version |
| 1793 | * of the pmd entry with pmd_populate. |
| 1794 | */ |
Gerald Schaefer | 46dcde7 | 2012-10-08 16:30:09 -0700 | [diff] [blame] | 1795 | pmdp_invalidate(vma, address, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1796 | pmd_populate(mm, pmd, pgtable); |
| 1797 | ret = 1; |
Kirill A. Shutemov | 117b079 | 2013-11-14 14:30:56 -0800 | [diff] [blame] | 1798 | spin_unlock(ptl); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1799 | } |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1800 | |
| 1801 | return ret; |
| 1802 | } |
| 1803 | |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 1804 | /* must be called with anon_vma->root->rwsem held */ |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1805 | static void __split_huge_page(struct page *page, |
Shaohua Li | 5bc7b8a | 2013-04-29 15:08:36 -0700 | [diff] [blame] | 1806 | struct anon_vma *anon_vma, |
| 1807 | struct list_head *list) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1808 | { |
| 1809 | int mapcount, mapcount2; |
Michel Lespinasse | bf181b9 | 2012-10-08 16:31:39 -0700 | [diff] [blame] | 1810 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1811 | struct anon_vma_chain *avc; |
| 1812 | |
| 1813 | BUG_ON(!PageHead(page)); |
| 1814 | BUG_ON(PageTail(page)); |
| 1815 | |
| 1816 | mapcount = 0; |
Michel Lespinasse | bf181b9 | 2012-10-08 16:31:39 -0700 | [diff] [blame] | 1817 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1818 | struct vm_area_struct *vma = avc->vma; |
| 1819 | unsigned long addr = vma_address(page, vma); |
| 1820 | BUG_ON(is_vma_temporary_stack(vma)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1821 | mapcount += __split_huge_page_splitting(page, vma, addr); |
| 1822 | } |
Andrea Arcangeli | 05759d3 | 2011-01-13 15:46:53 -0800 | [diff] [blame] | 1823 | /* |
| 1824 | * It is critical that new vmas are added to the tail of the |
| 1825 | * anon_vma list. This guarantes that if copy_huge_pmd() runs |
| 1826 | * and establishes a child pmd before |
| 1827 | * __split_huge_page_splitting() freezes the parent pmd (so if |
| 1828 | * we fail to prevent copy_huge_pmd() from running until the |
| 1829 | * whole __split_huge_page() is complete), we will still see |
| 1830 | * the newly established pmd of the child later during the |
| 1831 | * walk, to be able to set it as pmd_trans_splitting too. |
| 1832 | */ |
| 1833 | if (mapcount != page_mapcount(page)) |
| 1834 | printk(KERN_ERR "mapcount %d page_mapcount %d\n", |
| 1835 | mapcount, page_mapcount(page)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1836 | BUG_ON(mapcount != page_mapcount(page)); |
| 1837 | |
Shaohua Li | 5bc7b8a | 2013-04-29 15:08:36 -0700 | [diff] [blame] | 1838 | __split_huge_page_refcount(page, list); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1839 | |
| 1840 | mapcount2 = 0; |
Michel Lespinasse | bf181b9 | 2012-10-08 16:31:39 -0700 | [diff] [blame] | 1841 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1842 | struct vm_area_struct *vma = avc->vma; |
| 1843 | unsigned long addr = vma_address(page, vma); |
| 1844 | BUG_ON(is_vma_temporary_stack(vma)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1845 | mapcount2 += __split_huge_page_map(page, vma, addr); |
| 1846 | } |
Andrea Arcangeli | 05759d3 | 2011-01-13 15:46:53 -0800 | [diff] [blame] | 1847 | if (mapcount != mapcount2) |
| 1848 | printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n", |
| 1849 | mapcount, mapcount2, page_mapcount(page)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1850 | BUG_ON(mapcount != mapcount2); |
| 1851 | } |
| 1852 | |
Shaohua Li | 5bc7b8a | 2013-04-29 15:08:36 -0700 | [diff] [blame] | 1853 | /* |
| 1854 | * Split a hugepage into normal pages. This doesn't change the position of head |
| 1855 | * page. If @list is null, tail pages will be added to LRU list, otherwise, to |
| 1856 | * @list. Both head page and tail pages will inherit mapping, flags, and so on |
| 1857 | * from the hugepage. |
| 1858 | * Return 0 if the hugepage is split successfully otherwise return 1. |
| 1859 | */ |
| 1860 | int split_huge_page_to_list(struct page *page, struct list_head *list) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1861 | { |
| 1862 | struct anon_vma *anon_vma; |
| 1863 | int ret = 1; |
| 1864 | |
Kirill A. Shutemov | 5918d10 | 2013-04-29 15:08:44 -0700 | [diff] [blame] | 1865 | BUG_ON(is_huge_zero_page(page)); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1866 | BUG_ON(!PageAnon(page)); |
Mel Gorman | 062f1af | 2013-01-11 14:32:02 -0800 | [diff] [blame] | 1867 | |
| 1868 | /* |
| 1869 | * The caller does not necessarily hold an mmap_sem that would prevent |
| 1870 | * the anon_vma disappearing so we first we take a reference to it |
| 1871 | * and then lock the anon_vma for write. This is similar to |
| 1872 | * page_lock_anon_vma_read except the write lock is taken to serialise |
| 1873 | * against parallel split or collapse operations. |
| 1874 | */ |
| 1875 | anon_vma = page_get_anon_vma(page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1876 | if (!anon_vma) |
| 1877 | goto out; |
Mel Gorman | 062f1af | 2013-01-11 14:32:02 -0800 | [diff] [blame] | 1878 | anon_vma_lock_write(anon_vma); |
| 1879 | |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1880 | ret = 0; |
| 1881 | if (!PageCompound(page)) |
| 1882 | goto out_unlock; |
| 1883 | |
| 1884 | BUG_ON(!PageSwapBacked(page)); |
Shaohua Li | 5bc7b8a | 2013-04-29 15:08:36 -0700 | [diff] [blame] | 1885 | __split_huge_page(page, anon_vma, list); |
Andi Kleen | 81ab420 | 2011-04-14 15:22:06 -0700 | [diff] [blame] | 1886 | count_vm_event(THP_SPLIT); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1887 | |
| 1888 | BUG_ON(PageCompound(page)); |
| 1889 | out_unlock: |
Konstantin Khlebnikov | 08b5270 | 2013-02-22 16:34:40 -0800 | [diff] [blame] | 1890 | anon_vma_unlock_write(anon_vma); |
Mel Gorman | 062f1af | 2013-01-11 14:32:02 -0800 | [diff] [blame] | 1891 | put_anon_vma(anon_vma); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 1892 | out: |
| 1893 | return ret; |
| 1894 | } |
| 1895 | |
Vlastimil Babka | 9050d7e | 2014-03-03 15:38:27 -0800 | [diff] [blame] | 1896 | #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) |
Andrea Arcangeli | 78f11a2 | 2011-04-27 15:26:45 -0700 | [diff] [blame] | 1897 | |
Andrea Arcangeli | 60ab324 | 2011-01-13 15:47:18 -0800 | [diff] [blame] | 1898 | int hugepage_madvise(struct vm_area_struct *vma, |
| 1899 | unsigned long *vm_flags, int advice) |
Andrea Arcangeli | 0af4e98 | 2011-01-13 15:46:55 -0800 | [diff] [blame] | 1900 | { |
Andrea Arcangeli | a664b2d | 2011-01-13 15:47:17 -0800 | [diff] [blame] | 1901 | switch (advice) { |
| 1902 | case MADV_HUGEPAGE: |
Alex Thorlton | 1e1836e | 2014-04-07 15:37:09 -0700 | [diff] [blame] | 1903 | #ifdef CONFIG_S390 |
| 1904 | /* |
| 1905 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 |
| 1906 | * can't handle this properly after s390_enable_sie, so we simply |
| 1907 | * ignore the madvise to prevent qemu from causing a SIGSEGV. |
| 1908 | */ |
| 1909 | if (mm_has_pgste(vma->vm_mm)) |
| 1910 | return 0; |
| 1911 | #endif |
Andrea Arcangeli | a664b2d | 2011-01-13 15:47:17 -0800 | [diff] [blame] | 1912 | /* |
| 1913 | * Be somewhat over-protective like KSM for now! |
| 1914 | */ |
Andrea Arcangeli | 78f11a2 | 2011-04-27 15:26:45 -0700 | [diff] [blame] | 1915 | if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP)) |
Andrea Arcangeli | a664b2d | 2011-01-13 15:47:17 -0800 | [diff] [blame] | 1916 | return -EINVAL; |
| 1917 | *vm_flags &= ~VM_NOHUGEPAGE; |
| 1918 | *vm_flags |= VM_HUGEPAGE; |
Andrea Arcangeli | 60ab324 | 2011-01-13 15:47:18 -0800 | [diff] [blame] | 1919 | /* |
| 1920 | * If the vma become good for khugepaged to scan, |
| 1921 | * register it here without waiting a page fault that |
| 1922 | * may not happen any time soon. |
| 1923 | */ |
| 1924 | if (unlikely(khugepaged_enter_vma_merge(vma))) |
| 1925 | return -ENOMEM; |
Andrea Arcangeli | a664b2d | 2011-01-13 15:47:17 -0800 | [diff] [blame] | 1926 | break; |
| 1927 | case MADV_NOHUGEPAGE: |
| 1928 | /* |
| 1929 | * Be somewhat over-protective like KSM for now! |
| 1930 | */ |
Andrea Arcangeli | 78f11a2 | 2011-04-27 15:26:45 -0700 | [diff] [blame] | 1931 | if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP)) |
Andrea Arcangeli | a664b2d | 2011-01-13 15:47:17 -0800 | [diff] [blame] | 1932 | return -EINVAL; |
| 1933 | *vm_flags &= ~VM_HUGEPAGE; |
| 1934 | *vm_flags |= VM_NOHUGEPAGE; |
Andrea Arcangeli | 60ab324 | 2011-01-13 15:47:18 -0800 | [diff] [blame] | 1935 | /* |
| 1936 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning |
| 1937 | * this vma even if we leave the mm registered in khugepaged if |
| 1938 | * it got registered before VM_NOHUGEPAGE was set. |
| 1939 | */ |
Andrea Arcangeli | a664b2d | 2011-01-13 15:47:17 -0800 | [diff] [blame] | 1940 | break; |
| 1941 | } |
Andrea Arcangeli | 0af4e98 | 2011-01-13 15:46:55 -0800 | [diff] [blame] | 1942 | |
| 1943 | return 0; |
| 1944 | } |
| 1945 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1946 | static int __init khugepaged_slab_init(void) |
| 1947 | { |
| 1948 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", |
| 1949 | sizeof(struct mm_slot), |
| 1950 | __alignof__(struct mm_slot), 0, NULL); |
| 1951 | if (!mm_slot_cache) |
| 1952 | return -ENOMEM; |
| 1953 | |
| 1954 | return 0; |
| 1955 | } |
| 1956 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1957 | static inline struct mm_slot *alloc_mm_slot(void) |
| 1958 | { |
| 1959 | if (!mm_slot_cache) /* initialization failed */ |
| 1960 | return NULL; |
| 1961 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); |
| 1962 | } |
| 1963 | |
| 1964 | static inline void free_mm_slot(struct mm_slot *mm_slot) |
| 1965 | { |
| 1966 | kmem_cache_free(mm_slot_cache, mm_slot); |
| 1967 | } |
| 1968 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1969 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
| 1970 | { |
| 1971 | struct mm_slot *mm_slot; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1972 | |
Sasha Levin | b67bfe0 | 2013-02-27 17:06:00 -0800 | [diff] [blame] | 1973 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1974 | if (mm == mm_slot->mm) |
| 1975 | return mm_slot; |
Sasha Levin | 43b5fbb | 2013-02-22 16:32:27 -0800 | [diff] [blame] | 1976 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1977 | return NULL; |
| 1978 | } |
| 1979 | |
| 1980 | static void insert_to_mm_slots_hash(struct mm_struct *mm, |
| 1981 | struct mm_slot *mm_slot) |
| 1982 | { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1983 | mm_slot->mm = mm; |
Sasha Levin | 43b5fbb | 2013-02-22 16:32:27 -0800 | [diff] [blame] | 1984 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 1985 | } |
| 1986 | |
| 1987 | static inline int khugepaged_test_exit(struct mm_struct *mm) |
| 1988 | { |
| 1989 | return atomic_read(&mm->mm_users) == 0; |
| 1990 | } |
| 1991 | |
| 1992 | int __khugepaged_enter(struct mm_struct *mm) |
| 1993 | { |
| 1994 | struct mm_slot *mm_slot; |
| 1995 | int wakeup; |
| 1996 | |
| 1997 | mm_slot = alloc_mm_slot(); |
| 1998 | if (!mm_slot) |
| 1999 | return -ENOMEM; |
| 2000 | |
| 2001 | /* __khugepaged_exit() must not run from under us */ |
| 2002 | VM_BUG_ON(khugepaged_test_exit(mm)); |
| 2003 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
| 2004 | free_mm_slot(mm_slot); |
| 2005 | return 0; |
| 2006 | } |
| 2007 | |
| 2008 | spin_lock(&khugepaged_mm_lock); |
| 2009 | insert_to_mm_slots_hash(mm, mm_slot); |
| 2010 | /* |
| 2011 | * Insert just behind the scanning cursor, to let the area settle |
| 2012 | * down a little. |
| 2013 | */ |
| 2014 | wakeup = list_empty(&khugepaged_scan.mm_head); |
| 2015 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); |
| 2016 | spin_unlock(&khugepaged_mm_lock); |
| 2017 | |
| 2018 | atomic_inc(&mm->mm_count); |
| 2019 | if (wakeup) |
| 2020 | wake_up_interruptible(&khugepaged_wait); |
| 2021 | |
| 2022 | return 0; |
| 2023 | } |
| 2024 | |
| 2025 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma) |
| 2026 | { |
| 2027 | unsigned long hstart, hend; |
| 2028 | if (!vma->anon_vma) |
| 2029 | /* |
| 2030 | * Not yet faulted in so we will register later in the |
| 2031 | * page fault if needed. |
| 2032 | */ |
| 2033 | return 0; |
Andrea Arcangeli | 78f11a2 | 2011-04-27 15:26:45 -0700 | [diff] [blame] | 2034 | if (vma->vm_ops) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2035 | /* khugepaged not yet working on file or special mappings */ |
| 2036 | return 0; |
Konstantin Khlebnikov | b3b9c29 | 2012-10-08 16:28:34 -0700 | [diff] [blame] | 2037 | VM_BUG_ON(vma->vm_flags & VM_NO_THP); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2038 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
| 2039 | hend = vma->vm_end & HPAGE_PMD_MASK; |
| 2040 | if (hstart < hend) |
| 2041 | return khugepaged_enter(vma); |
| 2042 | return 0; |
| 2043 | } |
| 2044 | |
| 2045 | void __khugepaged_exit(struct mm_struct *mm) |
| 2046 | { |
| 2047 | struct mm_slot *mm_slot; |
| 2048 | int free = 0; |
| 2049 | |
| 2050 | spin_lock(&khugepaged_mm_lock); |
| 2051 | mm_slot = get_mm_slot(mm); |
| 2052 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { |
Sasha Levin | 43b5fbb | 2013-02-22 16:32:27 -0800 | [diff] [blame] | 2053 | hash_del(&mm_slot->hash); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2054 | list_del(&mm_slot->mm_node); |
| 2055 | free = 1; |
| 2056 | } |
Chris Wright | d788e80 | 2011-07-25 17:12:14 -0700 | [diff] [blame] | 2057 | spin_unlock(&khugepaged_mm_lock); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2058 | |
| 2059 | if (free) { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2060 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); |
| 2061 | free_mm_slot(mm_slot); |
| 2062 | mmdrop(mm); |
| 2063 | } else if (mm_slot) { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2064 | /* |
| 2065 | * This is required to serialize against |
| 2066 | * khugepaged_test_exit() (which is guaranteed to run |
| 2067 | * under mmap sem read mode). Stop here (after we |
| 2068 | * return all pagetables will be destroyed) until |
| 2069 | * khugepaged has finished working on the pagetables |
| 2070 | * under the mmap_sem. |
| 2071 | */ |
| 2072 | down_write(&mm->mmap_sem); |
| 2073 | up_write(&mm->mmap_sem); |
Chris Wright | d788e80 | 2011-07-25 17:12:14 -0700 | [diff] [blame] | 2074 | } |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2075 | } |
| 2076 | |
| 2077 | static void release_pte_page(struct page *page) |
| 2078 | { |
| 2079 | /* 0 stands for page_is_file_cache(page) == false */ |
| 2080 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); |
| 2081 | unlock_page(page); |
| 2082 | putback_lru_page(page); |
| 2083 | } |
| 2084 | |
| 2085 | static void release_pte_pages(pte_t *pte, pte_t *_pte) |
| 2086 | { |
| 2087 | while (--_pte >= pte) { |
| 2088 | pte_t pteval = *_pte; |
| 2089 | if (!pte_none(pteval)) |
| 2090 | release_pte_page(pte_page(pteval)); |
| 2091 | } |
| 2092 | } |
| 2093 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2094 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
| 2095 | unsigned long address, |
| 2096 | pte_t *pte) |
| 2097 | { |
| 2098 | struct page *page; |
| 2099 | pte_t *_pte; |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2100 | int referenced = 0, none = 0; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2101 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; |
| 2102 | _pte++, address += PAGE_SIZE) { |
| 2103 | pte_t pteval = *_pte; |
| 2104 | if (pte_none(pteval)) { |
| 2105 | if (++none <= khugepaged_max_ptes_none) |
| 2106 | continue; |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2107 | else |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2108 | goto out; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2109 | } |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2110 | if (!pte_present(pteval) || !pte_write(pteval)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2111 | goto out; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2112 | page = vm_normal_page(vma, address, pteval); |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2113 | if (unlikely(!page)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2114 | goto out; |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2115 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 2116 | VM_BUG_ON_PAGE(PageCompound(page), page); |
| 2117 | VM_BUG_ON_PAGE(!PageAnon(page), page); |
| 2118 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2119 | |
| 2120 | /* cannot use mapcount: can't collapse if there's a gup pin */ |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2121 | if (page_count(page) != 1) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2122 | goto out; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2123 | /* |
| 2124 | * We can do it before isolate_lru_page because the |
| 2125 | * page can't be freed from under us. NOTE: PG_lock |
| 2126 | * is needed to serialize against split_huge_page |
| 2127 | * when invoked from the VM. |
| 2128 | */ |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2129 | if (!trylock_page(page)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2130 | goto out; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2131 | /* |
| 2132 | * Isolate the page to avoid collapsing an hugepage |
| 2133 | * currently in use by the VM. |
| 2134 | */ |
| 2135 | if (isolate_lru_page(page)) { |
| 2136 | unlock_page(page); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2137 | goto out; |
| 2138 | } |
| 2139 | /* 0 stands for page_is_file_cache(page) == false */ |
| 2140 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 2141 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
| 2142 | VM_BUG_ON_PAGE(PageLRU(page), page); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2143 | |
| 2144 | /* If there is no mapped pte young don't collapse the page */ |
Andrea Arcangeli | 8ee5382 | 2011-01-13 15:47:10 -0800 | [diff] [blame] | 2145 | if (pte_young(pteval) || PageReferenced(page) || |
| 2146 | mmu_notifier_test_young(vma->vm_mm, address)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2147 | referenced = 1; |
| 2148 | } |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2149 | if (likely(referenced)) |
| 2150 | return 1; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2151 | out: |
Bob Liu | 344aa35 | 2012-12-11 16:00:34 -0800 | [diff] [blame] | 2152 | release_pte_pages(pte, _pte); |
| 2153 | return 0; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2154 | } |
| 2155 | |
| 2156 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, |
| 2157 | struct vm_area_struct *vma, |
| 2158 | unsigned long address, |
| 2159 | spinlock_t *ptl) |
| 2160 | { |
| 2161 | pte_t *_pte; |
| 2162 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { |
| 2163 | pte_t pteval = *_pte; |
| 2164 | struct page *src_page; |
| 2165 | |
| 2166 | if (pte_none(pteval)) { |
| 2167 | clear_user_highpage(page, address); |
| 2168 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); |
| 2169 | } else { |
| 2170 | src_page = pte_page(pteval); |
| 2171 | copy_user_highpage(page, src_page, address, vma); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 2172 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2173 | release_pte_page(src_page); |
| 2174 | /* |
| 2175 | * ptl mostly unnecessary, but preempt has to |
| 2176 | * be disabled to update the per-cpu stats |
| 2177 | * inside page_remove_rmap(). |
| 2178 | */ |
| 2179 | spin_lock(ptl); |
| 2180 | /* |
| 2181 | * paravirt calls inside pte_clear here are |
| 2182 | * superfluous. |
| 2183 | */ |
| 2184 | pte_clear(vma->vm_mm, address, _pte); |
| 2185 | page_remove_rmap(src_page); |
| 2186 | spin_unlock(ptl); |
| 2187 | free_page_and_swap_cache(src_page); |
| 2188 | } |
| 2189 | |
| 2190 | address += PAGE_SIZE; |
| 2191 | page++; |
| 2192 | } |
| 2193 | } |
| 2194 | |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2195 | static void khugepaged_alloc_sleep(void) |
| 2196 | { |
| 2197 | wait_event_freezable_timeout(khugepaged_wait, false, |
| 2198 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); |
| 2199 | } |
| 2200 | |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2201 | static int khugepaged_node_load[MAX_NUMNODES]; |
| 2202 | |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2203 | #ifdef CONFIG_NUMA |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2204 | static int khugepaged_find_target_node(void) |
| 2205 | { |
| 2206 | static int last_khugepaged_target_node = NUMA_NO_NODE; |
| 2207 | int nid, target_node = 0, max_value = 0; |
| 2208 | |
| 2209 | /* find first node with max normal pages hit */ |
| 2210 | for (nid = 0; nid < MAX_NUMNODES; nid++) |
| 2211 | if (khugepaged_node_load[nid] > max_value) { |
| 2212 | max_value = khugepaged_node_load[nid]; |
| 2213 | target_node = nid; |
| 2214 | } |
| 2215 | |
| 2216 | /* do some balance if several nodes have the same hit record */ |
| 2217 | if (target_node <= last_khugepaged_target_node) |
| 2218 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; |
| 2219 | nid++) |
| 2220 | if (max_value == khugepaged_node_load[nid]) { |
| 2221 | target_node = nid; |
| 2222 | break; |
| 2223 | } |
| 2224 | |
| 2225 | last_khugepaged_target_node = target_node; |
| 2226 | return target_node; |
| 2227 | } |
| 2228 | |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2229 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) |
| 2230 | { |
| 2231 | if (IS_ERR(*hpage)) { |
| 2232 | if (!*wait) |
| 2233 | return false; |
| 2234 | |
| 2235 | *wait = false; |
Xiao Guangrong | e3b4126 | 2012-10-08 16:32:57 -0700 | [diff] [blame] | 2236 | *hpage = NULL; |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2237 | khugepaged_alloc_sleep(); |
| 2238 | } else if (*hpage) { |
| 2239 | put_page(*hpage); |
| 2240 | *hpage = NULL; |
| 2241 | } |
| 2242 | |
| 2243 | return true; |
| 2244 | } |
| 2245 | |
| 2246 | static struct page |
| 2247 | *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, |
| 2248 | struct vm_area_struct *vma, unsigned long address, |
| 2249 | int node) |
| 2250 | { |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 2251 | VM_BUG_ON_PAGE(*hpage, *hpage); |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2252 | /* |
| 2253 | * Allocate the page while the vma is still valid and under |
| 2254 | * the mmap_sem read mode so there is no memory allocation |
| 2255 | * later when we take the mmap_sem in write mode. This is more |
| 2256 | * friendly behavior (OTOH it may actually hide bugs) to |
| 2257 | * filesystems in userland with daemons allocating memory in |
| 2258 | * the userland I/O paths. Allocating memory with the |
| 2259 | * mmap_sem in read mode is good idea also to allow greater |
| 2260 | * scalability. |
| 2261 | */ |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2262 | *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask( |
| 2263 | khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER); |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2264 | /* |
| 2265 | * After allocating the hugepage, release the mmap_sem read lock in |
| 2266 | * preparation for taking it in write mode. |
| 2267 | */ |
| 2268 | up_read(&mm->mmap_sem); |
| 2269 | if (unlikely(!*hpage)) { |
| 2270 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); |
| 2271 | *hpage = ERR_PTR(-ENOMEM); |
| 2272 | return NULL; |
| 2273 | } |
| 2274 | |
| 2275 | count_vm_event(THP_COLLAPSE_ALLOC); |
| 2276 | return *hpage; |
| 2277 | } |
| 2278 | #else |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2279 | static int khugepaged_find_target_node(void) |
| 2280 | { |
| 2281 | return 0; |
| 2282 | } |
| 2283 | |
Bob Liu | 10dc415 | 2013-11-12 15:07:35 -0800 | [diff] [blame] | 2284 | static inline struct page *alloc_hugepage(int defrag) |
| 2285 | { |
| 2286 | return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), |
| 2287 | HPAGE_PMD_ORDER); |
| 2288 | } |
| 2289 | |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2290 | static struct page *khugepaged_alloc_hugepage(bool *wait) |
| 2291 | { |
| 2292 | struct page *hpage; |
| 2293 | |
| 2294 | do { |
| 2295 | hpage = alloc_hugepage(khugepaged_defrag()); |
| 2296 | if (!hpage) { |
| 2297 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); |
| 2298 | if (!*wait) |
| 2299 | return NULL; |
| 2300 | |
| 2301 | *wait = false; |
| 2302 | khugepaged_alloc_sleep(); |
| 2303 | } else |
| 2304 | count_vm_event(THP_COLLAPSE_ALLOC); |
| 2305 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); |
| 2306 | |
| 2307 | return hpage; |
| 2308 | } |
| 2309 | |
| 2310 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) |
| 2311 | { |
| 2312 | if (!*hpage) |
| 2313 | *hpage = khugepaged_alloc_hugepage(wait); |
| 2314 | |
| 2315 | if (unlikely(!*hpage)) |
| 2316 | return false; |
| 2317 | |
| 2318 | return true; |
| 2319 | } |
| 2320 | |
| 2321 | static struct page |
| 2322 | *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, |
| 2323 | struct vm_area_struct *vma, unsigned long address, |
| 2324 | int node) |
| 2325 | { |
| 2326 | up_read(&mm->mmap_sem); |
| 2327 | VM_BUG_ON(!*hpage); |
| 2328 | return *hpage; |
| 2329 | } |
| 2330 | #endif |
| 2331 | |
Bob Liu | fa475e5 | 2012-12-11 16:00:39 -0800 | [diff] [blame] | 2332 | static bool hugepage_vma_check(struct vm_area_struct *vma) |
| 2333 | { |
| 2334 | if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || |
| 2335 | (vma->vm_flags & VM_NOHUGEPAGE)) |
| 2336 | return false; |
| 2337 | |
| 2338 | if (!vma->anon_vma || vma->vm_ops) |
| 2339 | return false; |
| 2340 | if (is_vma_temporary_stack(vma)) |
| 2341 | return false; |
| 2342 | VM_BUG_ON(vma->vm_flags & VM_NO_THP); |
| 2343 | return true; |
| 2344 | } |
| 2345 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2346 | static void collapse_huge_page(struct mm_struct *mm, |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2347 | unsigned long address, |
| 2348 | struct page **hpage, |
| 2349 | struct vm_area_struct *vma, |
| 2350 | int node) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2351 | { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2352 | pmd_t *pmd, _pmd; |
| 2353 | pte_t *pte; |
| 2354 | pgtable_t pgtable; |
| 2355 | struct page *new_page; |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2356 | spinlock_t *pmd_ptl, *pte_ptl; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2357 | int isolated; |
| 2358 | unsigned long hstart, hend; |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 2359 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 2360 | unsigned long mmun_end; /* For mmu_notifiers */ |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2361 | |
| 2362 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); |
Andrea Arcangeli | 692e0b3 | 2011-05-24 17:12:14 -0700 | [diff] [blame] | 2363 | |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2364 | /* release the mmap_sem read lock. */ |
| 2365 | new_page = khugepaged_alloc_page(hpage, mm, vma, address, node); |
| 2366 | if (!new_page) |
Andrea Arcangeli | ce83d21 | 2011-01-13 15:47:06 -0800 | [diff] [blame] | 2367 | return; |
Andrea Arcangeli | ce83d21 | 2011-01-13 15:47:06 -0800 | [diff] [blame] | 2368 | |
Michal Hocko | d715ae0 | 2014-04-07 15:37:46 -0700 | [diff] [blame] | 2369 | if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))) |
Andrea Arcangeli | 692e0b3 | 2011-05-24 17:12:14 -0700 | [diff] [blame] | 2370 | return; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2371 | |
| 2372 | /* |
| 2373 | * Prevent all access to pagetables with the exception of |
| 2374 | * gup_fast later hanlded by the ptep_clear_flush and the VM |
| 2375 | * handled by the anon_vma lock + PG_lock. |
| 2376 | */ |
| 2377 | down_write(&mm->mmap_sem); |
| 2378 | if (unlikely(khugepaged_test_exit(mm))) |
| 2379 | goto out; |
| 2380 | |
| 2381 | vma = find_vma(mm, address); |
Libin | a8f531eb | 2013-09-11 14:20:38 -0700 | [diff] [blame] | 2382 | if (!vma) |
| 2383 | goto out; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2384 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
| 2385 | hend = vma->vm_end & HPAGE_PMD_MASK; |
| 2386 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) |
| 2387 | goto out; |
Bob Liu | fa475e5 | 2012-12-11 16:00:39 -0800 | [diff] [blame] | 2388 | if (!hugepage_vma_check(vma)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2389 | goto out; |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 2390 | pmd = mm_find_pmd(mm, address); |
| 2391 | if (!pmd) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2392 | goto out; |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 2393 | if (pmd_trans_huge(*pmd)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2394 | goto out; |
| 2395 | |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 2396 | anon_vma_lock_write(vma->anon_vma); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2397 | |
| 2398 | pte = pte_offset_map(pmd, address); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2399 | pte_ptl = pte_lockptr(mm, pmd); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2400 | |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 2401 | mmun_start = address; |
| 2402 | mmun_end = address + HPAGE_PMD_SIZE; |
| 2403 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2404 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2405 | /* |
| 2406 | * After this gup_fast can't run anymore. This also removes |
| 2407 | * any huge TLB entry from the CPU so we won't allow |
| 2408 | * huge and small TLB entries for the same virtual address |
| 2409 | * to avoid the risk of CPU bugs in that area. |
| 2410 | */ |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 2411 | _pmd = pmdp_clear_flush(vma, address, pmd); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2412 | spin_unlock(pmd_ptl); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 2413 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2414 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2415 | spin_lock(pte_ptl); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2416 | isolated = __collapse_huge_page_isolate(vma, address, pte); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2417 | spin_unlock(pte_ptl); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2418 | |
| 2419 | if (unlikely(!isolated)) { |
Johannes Weiner | 453c719 | 2011-01-20 14:44:18 -0800 | [diff] [blame] | 2420 | pte_unmap(pte); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2421 | spin_lock(pmd_ptl); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2422 | BUG_ON(!pmd_none(*pmd)); |
Aneesh Kumar K.V | 7c34251 | 2013-05-24 15:55:21 -0700 | [diff] [blame] | 2423 | /* |
| 2424 | * We can only use set_pmd_at when establishing |
| 2425 | * hugepmds and never for establishing regular pmds that |
| 2426 | * points to regular pagetables. Use pmd_populate for that |
| 2427 | */ |
| 2428 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2429 | spin_unlock(pmd_ptl); |
Konstantin Khlebnikov | 08b5270 | 2013-02-22 16:34:40 -0800 | [diff] [blame] | 2430 | anon_vma_unlock_write(vma->anon_vma); |
Andrea Arcangeli | ce83d21 | 2011-01-13 15:47:06 -0800 | [diff] [blame] | 2431 | goto out; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2432 | } |
| 2433 | |
| 2434 | /* |
| 2435 | * All pages are isolated and locked so anon_vma rmap |
| 2436 | * can't run anymore. |
| 2437 | */ |
Konstantin Khlebnikov | 08b5270 | 2013-02-22 16:34:40 -0800 | [diff] [blame] | 2438 | anon_vma_unlock_write(vma->anon_vma); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2439 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2440 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); |
Johannes Weiner | 453c719 | 2011-01-20 14:44:18 -0800 | [diff] [blame] | 2441 | pte_unmap(pte); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2442 | __SetPageUptodate(new_page); |
| 2443 | pgtable = pmd_pgtable(_pmd); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2444 | |
Kirill A. Shutemov | 3122359 | 2013-09-12 15:14:01 -0700 | [diff] [blame] | 2445 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); |
| 2446 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2447 | |
| 2448 | /* |
| 2449 | * spin_lock() below is not the equivalent of smp_wmb(), so |
| 2450 | * this is needed to avoid the copy_huge_page writes to become |
| 2451 | * visible after the set_pmd_at() write. |
| 2452 | */ |
| 2453 | smp_wmb(); |
| 2454 | |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2455 | spin_lock(pmd_ptl); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2456 | BUG_ON(!pmd_none(*pmd)); |
| 2457 | page_add_new_anon_rmap(new_page, vma, address); |
Aneesh Kumar K.V | fce144b | 2013-06-05 17:14:06 -0700 | [diff] [blame] | 2458 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2459 | set_pmd_at(mm, address, pmd, _pmd); |
David Miller | b113da6 | 2012-10-08 16:34:25 -0700 | [diff] [blame] | 2460 | update_mmu_cache_pmd(vma, address, pmd); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2461 | spin_unlock(pmd_ptl); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2462 | |
| 2463 | *hpage = NULL; |
Xiao Guangrong | 420256ef | 2012-10-08 16:29:49 -0700 | [diff] [blame] | 2464 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2465 | khugepaged_pages_collapsed++; |
Andrea Arcangeli | ce83d21 | 2011-01-13 15:47:06 -0800 | [diff] [blame] | 2466 | out_up_write: |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2467 | up_write(&mm->mmap_sem); |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 2468 | return; |
| 2469 | |
Andrea Arcangeli | ce83d21 | 2011-01-13 15:47:06 -0800 | [diff] [blame] | 2470 | out: |
KAMEZAWA Hiroyuki | 678ff89 | 2011-02-10 15:01:36 -0800 | [diff] [blame] | 2471 | mem_cgroup_uncharge_page(new_page); |
Andrea Arcangeli | ce83d21 | 2011-01-13 15:47:06 -0800 | [diff] [blame] | 2472 | goto out_up_write; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2473 | } |
| 2474 | |
| 2475 | static int khugepaged_scan_pmd(struct mm_struct *mm, |
| 2476 | struct vm_area_struct *vma, |
| 2477 | unsigned long address, |
| 2478 | struct page **hpage) |
| 2479 | { |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2480 | pmd_t *pmd; |
| 2481 | pte_t *pte, *_pte; |
| 2482 | int ret = 0, referenced = 0, none = 0; |
| 2483 | struct page *page; |
| 2484 | unsigned long _address; |
| 2485 | spinlock_t *ptl; |
David Rientjes | 00ef2d2 | 2013-02-22 16:35:36 -0800 | [diff] [blame] | 2486 | int node = NUMA_NO_NODE; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2487 | |
| 2488 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); |
| 2489 | |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 2490 | pmd = mm_find_pmd(mm, address); |
| 2491 | if (!pmd) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2492 | goto out; |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 2493 | if (pmd_trans_huge(*pmd)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2494 | goto out; |
| 2495 | |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2496 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2497 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
| 2498 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; |
| 2499 | _pte++, _address += PAGE_SIZE) { |
| 2500 | pte_t pteval = *_pte; |
| 2501 | if (pte_none(pteval)) { |
| 2502 | if (++none <= khugepaged_max_ptes_none) |
| 2503 | continue; |
| 2504 | else |
| 2505 | goto out_unmap; |
| 2506 | } |
| 2507 | if (!pte_present(pteval) || !pte_write(pteval)) |
| 2508 | goto out_unmap; |
| 2509 | page = vm_normal_page(vma, _address, pteval); |
| 2510 | if (unlikely(!page)) |
| 2511 | goto out_unmap; |
Andi Kleen | 5c4b4be | 2011-03-04 17:36:32 -0800 | [diff] [blame] | 2512 | /* |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2513 | * Record which node the original page is from and save this |
| 2514 | * information to khugepaged_node_load[]. |
| 2515 | * Khupaged will allocate hugepage from the node has the max |
| 2516 | * hit record. |
Andi Kleen | 5c4b4be | 2011-03-04 17:36:32 -0800 | [diff] [blame] | 2517 | */ |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2518 | node = page_to_nid(page); |
| 2519 | khugepaged_node_load[node]++; |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 2520 | VM_BUG_ON_PAGE(PageCompound(page), page); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2521 | if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) |
| 2522 | goto out_unmap; |
| 2523 | /* cannot use mapcount: can't collapse if there's a gup pin */ |
| 2524 | if (page_count(page) != 1) |
| 2525 | goto out_unmap; |
Andrea Arcangeli | 8ee5382 | 2011-01-13 15:47:10 -0800 | [diff] [blame] | 2526 | if (pte_young(pteval) || PageReferenced(page) || |
| 2527 | mmu_notifier_test_young(vma->vm_mm, address)) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2528 | referenced = 1; |
| 2529 | } |
| 2530 | if (referenced) |
| 2531 | ret = 1; |
| 2532 | out_unmap: |
| 2533 | pte_unmap_unlock(pte, ptl); |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2534 | if (ret) { |
| 2535 | node = khugepaged_find_target_node(); |
Andrea Arcangeli | ce83d21 | 2011-01-13 15:47:06 -0800 | [diff] [blame] | 2536 | /* collapse_huge_page will return with the mmap_sem released */ |
Andi Kleen | 5c4b4be | 2011-03-04 17:36:32 -0800 | [diff] [blame] | 2537 | collapse_huge_page(mm, address, hpage, vma, node); |
Bob Liu | 9f1b868 | 2013-11-12 15:07:37 -0800 | [diff] [blame] | 2538 | } |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2539 | out: |
| 2540 | return ret; |
| 2541 | } |
| 2542 | |
| 2543 | static void collect_mm_slot(struct mm_slot *mm_slot) |
| 2544 | { |
| 2545 | struct mm_struct *mm = mm_slot->mm; |
| 2546 | |
Hugh Dickins | b9980cd | 2012-02-08 17:13:40 -0800 | [diff] [blame] | 2547 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2548 | |
| 2549 | if (khugepaged_test_exit(mm)) { |
| 2550 | /* free mm_slot */ |
Sasha Levin | 43b5fbb | 2013-02-22 16:32:27 -0800 | [diff] [blame] | 2551 | hash_del(&mm_slot->hash); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2552 | list_del(&mm_slot->mm_node); |
| 2553 | |
| 2554 | /* |
| 2555 | * Not strictly needed because the mm exited already. |
| 2556 | * |
| 2557 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); |
| 2558 | */ |
| 2559 | |
| 2560 | /* khugepaged_mm_lock actually not necessary for the below */ |
| 2561 | free_mm_slot(mm_slot); |
| 2562 | mmdrop(mm); |
| 2563 | } |
| 2564 | } |
| 2565 | |
| 2566 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, |
| 2567 | struct page **hpage) |
H Hartley Sweeten | 2f1da64 | 2011-10-31 17:09:25 -0700 | [diff] [blame] | 2568 | __releases(&khugepaged_mm_lock) |
| 2569 | __acquires(&khugepaged_mm_lock) |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2570 | { |
| 2571 | struct mm_slot *mm_slot; |
| 2572 | struct mm_struct *mm; |
| 2573 | struct vm_area_struct *vma; |
| 2574 | int progress = 0; |
| 2575 | |
| 2576 | VM_BUG_ON(!pages); |
Hugh Dickins | b9980cd | 2012-02-08 17:13:40 -0800 | [diff] [blame] | 2577 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2578 | |
| 2579 | if (khugepaged_scan.mm_slot) |
| 2580 | mm_slot = khugepaged_scan.mm_slot; |
| 2581 | else { |
| 2582 | mm_slot = list_entry(khugepaged_scan.mm_head.next, |
| 2583 | struct mm_slot, mm_node); |
| 2584 | khugepaged_scan.address = 0; |
| 2585 | khugepaged_scan.mm_slot = mm_slot; |
| 2586 | } |
| 2587 | spin_unlock(&khugepaged_mm_lock); |
| 2588 | |
| 2589 | mm = mm_slot->mm; |
| 2590 | down_read(&mm->mmap_sem); |
| 2591 | if (unlikely(khugepaged_test_exit(mm))) |
| 2592 | vma = NULL; |
| 2593 | else |
| 2594 | vma = find_vma(mm, khugepaged_scan.address); |
| 2595 | |
| 2596 | progress++; |
| 2597 | for (; vma; vma = vma->vm_next) { |
| 2598 | unsigned long hstart, hend; |
| 2599 | |
| 2600 | cond_resched(); |
| 2601 | if (unlikely(khugepaged_test_exit(mm))) { |
| 2602 | progress++; |
| 2603 | break; |
| 2604 | } |
Bob Liu | fa475e5 | 2012-12-11 16:00:39 -0800 | [diff] [blame] | 2605 | if (!hugepage_vma_check(vma)) { |
| 2606 | skip: |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2607 | progress++; |
| 2608 | continue; |
| 2609 | } |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2610 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
| 2611 | hend = vma->vm_end & HPAGE_PMD_MASK; |
Andrea Arcangeli | a7d6e4e | 2011-02-15 19:02:45 +0100 | [diff] [blame] | 2612 | if (hstart >= hend) |
| 2613 | goto skip; |
| 2614 | if (khugepaged_scan.address > hend) |
| 2615 | goto skip; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2616 | if (khugepaged_scan.address < hstart) |
| 2617 | khugepaged_scan.address = hstart; |
Andrea Arcangeli | a7d6e4e | 2011-02-15 19:02:45 +0100 | [diff] [blame] | 2618 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2619 | |
| 2620 | while (khugepaged_scan.address < hend) { |
| 2621 | int ret; |
| 2622 | cond_resched(); |
| 2623 | if (unlikely(khugepaged_test_exit(mm))) |
| 2624 | goto breakouterloop; |
| 2625 | |
| 2626 | VM_BUG_ON(khugepaged_scan.address < hstart || |
| 2627 | khugepaged_scan.address + HPAGE_PMD_SIZE > |
| 2628 | hend); |
| 2629 | ret = khugepaged_scan_pmd(mm, vma, |
| 2630 | khugepaged_scan.address, |
| 2631 | hpage); |
| 2632 | /* move to next address */ |
| 2633 | khugepaged_scan.address += HPAGE_PMD_SIZE; |
| 2634 | progress += HPAGE_PMD_NR; |
| 2635 | if (ret) |
| 2636 | /* we released mmap_sem so break loop */ |
| 2637 | goto breakouterloop_mmap_sem; |
| 2638 | if (progress >= pages) |
| 2639 | goto breakouterloop; |
| 2640 | } |
| 2641 | } |
| 2642 | breakouterloop: |
| 2643 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ |
| 2644 | breakouterloop_mmap_sem: |
| 2645 | |
| 2646 | spin_lock(&khugepaged_mm_lock); |
Andrea Arcangeli | a7d6e4e | 2011-02-15 19:02:45 +0100 | [diff] [blame] | 2647 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2648 | /* |
| 2649 | * Release the current mm_slot if this mm is about to die, or |
| 2650 | * if we scanned all vmas of this mm. |
| 2651 | */ |
| 2652 | if (khugepaged_test_exit(mm) || !vma) { |
| 2653 | /* |
| 2654 | * Make sure that if mm_users is reaching zero while |
| 2655 | * khugepaged runs here, khugepaged_exit will find |
| 2656 | * mm_slot not pointing to the exiting mm. |
| 2657 | */ |
| 2658 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { |
| 2659 | khugepaged_scan.mm_slot = list_entry( |
| 2660 | mm_slot->mm_node.next, |
| 2661 | struct mm_slot, mm_node); |
| 2662 | khugepaged_scan.address = 0; |
| 2663 | } else { |
| 2664 | khugepaged_scan.mm_slot = NULL; |
| 2665 | khugepaged_full_scans++; |
| 2666 | } |
| 2667 | |
| 2668 | collect_mm_slot(mm_slot); |
| 2669 | } |
| 2670 | |
| 2671 | return progress; |
| 2672 | } |
| 2673 | |
| 2674 | static int khugepaged_has_work(void) |
| 2675 | { |
| 2676 | return !list_empty(&khugepaged_scan.mm_head) && |
| 2677 | khugepaged_enabled(); |
| 2678 | } |
| 2679 | |
| 2680 | static int khugepaged_wait_event(void) |
| 2681 | { |
| 2682 | return !list_empty(&khugepaged_scan.mm_head) || |
Xiao Guangrong | 2017c0b | 2012-10-08 16:29:44 -0700 | [diff] [blame] | 2683 | kthread_should_stop(); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2684 | } |
| 2685 | |
Xiao Guangrong | d516904 | 2012-10-08 16:29:48 -0700 | [diff] [blame] | 2686 | static void khugepaged_do_scan(void) |
| 2687 | { |
| 2688 | struct page *hpage = NULL; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2689 | unsigned int progress = 0, pass_through_head = 0; |
| 2690 | unsigned int pages = khugepaged_pages_to_scan; |
Xiao Guangrong | d516904 | 2012-10-08 16:29:48 -0700 | [diff] [blame] | 2691 | bool wait = true; |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2692 | |
| 2693 | barrier(); /* write khugepaged_pages_to_scan to local stack */ |
| 2694 | |
| 2695 | while (progress < pages) { |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2696 | if (!khugepaged_prealloc_page(&hpage, &wait)) |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 2697 | break; |
Xiao Guangrong | 26234f3 | 2012-10-08 16:29:51 -0700 | [diff] [blame] | 2698 | |
Xiao Guangrong | 420256ef | 2012-10-08 16:29:49 -0700 | [diff] [blame] | 2699 | cond_resched(); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2700 | |
Andrea Arcangeli | 878aee7 | 2011-01-13 15:47:10 -0800 | [diff] [blame] | 2701 | if (unlikely(kthread_should_stop() || freezing(current))) |
| 2702 | break; |
| 2703 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2704 | spin_lock(&khugepaged_mm_lock); |
| 2705 | if (!khugepaged_scan.mm_slot) |
| 2706 | pass_through_head++; |
| 2707 | if (khugepaged_has_work() && |
| 2708 | pass_through_head < 2) |
| 2709 | progress += khugepaged_scan_mm_slot(pages - progress, |
Xiao Guangrong | d516904 | 2012-10-08 16:29:48 -0700 | [diff] [blame] | 2710 | &hpage); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2711 | else |
| 2712 | progress = pages; |
| 2713 | spin_unlock(&khugepaged_mm_lock); |
| 2714 | } |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2715 | |
Xiao Guangrong | d516904 | 2012-10-08 16:29:48 -0700 | [diff] [blame] | 2716 | if (!IS_ERR_OR_NULL(hpage)) |
| 2717 | put_page(hpage); |
Andrea Arcangeli | 0bbbc0b | 2011-01-13 15:47:05 -0800 | [diff] [blame] | 2718 | } |
| 2719 | |
Xiao Guangrong | 2017c0b | 2012-10-08 16:29:44 -0700 | [diff] [blame] | 2720 | static void khugepaged_wait_work(void) |
| 2721 | { |
| 2722 | try_to_freeze(); |
| 2723 | |
| 2724 | if (khugepaged_has_work()) { |
| 2725 | if (!khugepaged_scan_sleep_millisecs) |
| 2726 | return; |
| 2727 | |
| 2728 | wait_event_freezable_timeout(khugepaged_wait, |
| 2729 | kthread_should_stop(), |
| 2730 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); |
| 2731 | return; |
| 2732 | } |
| 2733 | |
| 2734 | if (khugepaged_enabled()) |
| 2735 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); |
| 2736 | } |
| 2737 | |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2738 | static int khugepaged(void *none) |
| 2739 | { |
| 2740 | struct mm_slot *mm_slot; |
| 2741 | |
Andrea Arcangeli | 878aee7 | 2011-01-13 15:47:10 -0800 | [diff] [blame] | 2742 | set_freezable(); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2743 | set_user_nice(current, 19); |
| 2744 | |
Xiao Guangrong | b723178 | 2012-10-08 16:29:54 -0700 | [diff] [blame] | 2745 | while (!kthread_should_stop()) { |
| 2746 | khugepaged_do_scan(); |
| 2747 | khugepaged_wait_work(); |
| 2748 | } |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2749 | |
| 2750 | spin_lock(&khugepaged_mm_lock); |
| 2751 | mm_slot = khugepaged_scan.mm_slot; |
| 2752 | khugepaged_scan.mm_slot = NULL; |
| 2753 | if (mm_slot) |
| 2754 | collect_mm_slot(mm_slot); |
| 2755 | spin_unlock(&khugepaged_mm_lock); |
Andrea Arcangeli | ba76149 | 2011-01-13 15:46:58 -0800 | [diff] [blame] | 2756 | return 0; |
| 2757 | } |
| 2758 | |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2759 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
| 2760 | unsigned long haddr, pmd_t *pmd) |
| 2761 | { |
| 2762 | struct mm_struct *mm = vma->vm_mm; |
| 2763 | pgtable_t pgtable; |
| 2764 | pmd_t _pmd; |
| 2765 | int i; |
| 2766 | |
| 2767 | pmdp_clear_flush(vma, haddr, pmd); |
| 2768 | /* leave pmd empty until pte is filled */ |
| 2769 | |
Aneesh Kumar K.V | 6b0b50b | 2013-06-05 17:14:02 -0700 | [diff] [blame] | 2770 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2771 | pmd_populate(mm, &_pmd, pgtable); |
| 2772 | |
| 2773 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { |
| 2774 | pte_t *pte, entry; |
| 2775 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); |
| 2776 | entry = pte_mkspecial(entry); |
| 2777 | pte = pte_offset_map(&_pmd, haddr); |
| 2778 | VM_BUG_ON(!pte_none(*pte)); |
| 2779 | set_pte_at(mm, haddr, pte, entry); |
| 2780 | pte_unmap(pte); |
| 2781 | } |
| 2782 | smp_wmb(); /* make pte visible before pmd */ |
| 2783 | pmd_populate(mm, pmd, pgtable); |
Kirill A. Shutemov | 97ae174 | 2012-12-12 13:51:06 -0800 | [diff] [blame] | 2784 | put_huge_zero_page(); |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2785 | } |
| 2786 | |
Kirill A. Shutemov | e180377 | 2012-12-12 13:50:59 -0800 | [diff] [blame] | 2787 | void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, |
| 2788 | pmd_t *pmd) |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2789 | { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2790 | spinlock_t *ptl; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2791 | struct page *page; |
Kirill A. Shutemov | e180377 | 2012-12-12 13:50:59 -0800 | [diff] [blame] | 2792 | struct mm_struct *mm = vma->vm_mm; |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2793 | unsigned long haddr = address & HPAGE_PMD_MASK; |
| 2794 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 2795 | unsigned long mmun_end; /* For mmu_notifiers */ |
Kirill A. Shutemov | e180377 | 2012-12-12 13:50:59 -0800 | [diff] [blame] | 2796 | |
| 2797 | BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2798 | |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2799 | mmun_start = haddr; |
| 2800 | mmun_end = haddr + HPAGE_PMD_SIZE; |
Hugh Dickins | 750e816 | 2013-10-16 13:47:08 -0700 | [diff] [blame] | 2801 | again: |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2802 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2803 | ptl = pmd_lock(mm, pmd); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2804 | if (unlikely(!pmd_trans_huge(*pmd))) { |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2805 | spin_unlock(ptl); |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2806 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
| 2807 | return; |
| 2808 | } |
| 2809 | if (is_huge_zero_pmd(*pmd)) { |
| 2810 | __split_huge_zero_page_pmd(vma, haddr, pmd); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2811 | spin_unlock(ptl); |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2812 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2813 | return; |
| 2814 | } |
| 2815 | page = pmd_page(*pmd); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 2816 | VM_BUG_ON_PAGE(!page_count(page), page); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2817 | get_page(page); |
Kirill A. Shutemov | c4088eb | 2013-11-14 14:31:04 -0800 | [diff] [blame] | 2818 | spin_unlock(ptl); |
Kirill A. Shutemov | c5a647d | 2012-12-12 13:51:00 -0800 | [diff] [blame] | 2819 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2820 | |
| 2821 | split_huge_page(page); |
| 2822 | |
| 2823 | put_page(page); |
Hugh Dickins | 750e816 | 2013-10-16 13:47:08 -0700 | [diff] [blame] | 2824 | |
| 2825 | /* |
| 2826 | * We don't always have down_write of mmap_sem here: a racing |
| 2827 | * do_huge_pmd_wp_page() might have copied-on-write to another |
| 2828 | * huge page before our split_huge_page() got the anon_vma lock. |
| 2829 | */ |
| 2830 | if (unlikely(pmd_trans_huge(*pmd))) |
| 2831 | goto again; |
Andrea Arcangeli | 71e3aac | 2011-01-13 15:46:52 -0800 | [diff] [blame] | 2832 | } |
Andrea Arcangeli | 94fcc58 | 2011-01-13 15:47:08 -0800 | [diff] [blame] | 2833 | |
Kirill A. Shutemov | e180377 | 2012-12-12 13:50:59 -0800 | [diff] [blame] | 2834 | void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address, |
| 2835 | pmd_t *pmd) |
| 2836 | { |
| 2837 | struct vm_area_struct *vma; |
| 2838 | |
| 2839 | vma = find_vma(mm, address); |
| 2840 | BUG_ON(vma == NULL); |
| 2841 | split_huge_page_pmd(vma, address, pmd); |
| 2842 | } |
| 2843 | |
Andrea Arcangeli | 94fcc58 | 2011-01-13 15:47:08 -0800 | [diff] [blame] | 2844 | static void split_huge_page_address(struct mm_struct *mm, |
| 2845 | unsigned long address) |
| 2846 | { |
Andrea Arcangeli | 94fcc58 | 2011-01-13 15:47:08 -0800 | [diff] [blame] | 2847 | pmd_t *pmd; |
| 2848 | |
| 2849 | VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); |
| 2850 | |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 2851 | pmd = mm_find_pmd(mm, address); |
| 2852 | if (!pmd) |
Andrea Arcangeli | 94fcc58 | 2011-01-13 15:47:08 -0800 | [diff] [blame] | 2853 | return; |
| 2854 | /* |
| 2855 | * Caller holds the mmap_sem write mode, so a huge pmd cannot |
| 2856 | * materialize from under us. |
| 2857 | */ |
Kirill A. Shutemov | e180377 | 2012-12-12 13:50:59 -0800 | [diff] [blame] | 2858 | split_huge_page_pmd_mm(mm, address, pmd); |
Andrea Arcangeli | 94fcc58 | 2011-01-13 15:47:08 -0800 | [diff] [blame] | 2859 | } |
| 2860 | |
| 2861 | void __vma_adjust_trans_huge(struct vm_area_struct *vma, |
| 2862 | unsigned long start, |
| 2863 | unsigned long end, |
| 2864 | long adjust_next) |
| 2865 | { |
| 2866 | /* |
| 2867 | * If the new start address isn't hpage aligned and it could |
| 2868 | * previously contain an hugepage: check if we need to split |
| 2869 | * an huge pmd. |
| 2870 | */ |
| 2871 | if (start & ~HPAGE_PMD_MASK && |
| 2872 | (start & HPAGE_PMD_MASK) >= vma->vm_start && |
| 2873 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) |
| 2874 | split_huge_page_address(vma->vm_mm, start); |
| 2875 | |
| 2876 | /* |
| 2877 | * If the new end address isn't hpage aligned and it could |
| 2878 | * previously contain an hugepage: check if we need to split |
| 2879 | * an huge pmd. |
| 2880 | */ |
| 2881 | if (end & ~HPAGE_PMD_MASK && |
| 2882 | (end & HPAGE_PMD_MASK) >= vma->vm_start && |
| 2883 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) |
| 2884 | split_huge_page_address(vma->vm_mm, end); |
| 2885 | |
| 2886 | /* |
| 2887 | * If we're also updating the vma->vm_next->vm_start, if the new |
| 2888 | * vm_next->vm_start isn't page aligned and it could previously |
| 2889 | * contain an hugepage: check if we need to split an huge pmd. |
| 2890 | */ |
| 2891 | if (adjust_next > 0) { |
| 2892 | struct vm_area_struct *next = vma->vm_next; |
| 2893 | unsigned long nstart = next->vm_start; |
| 2894 | nstart += adjust_next << PAGE_SHIFT; |
| 2895 | if (nstart & ~HPAGE_PMD_MASK && |
| 2896 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && |
| 2897 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) |
| 2898 | split_huge_page_address(next->vm_mm, nstart); |
| 2899 | } |
| 2900 | } |