mm/huge_memory.c

   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>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/migrate.h>
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 /*
  27  * By default transparent hugepage support is enabled for all mappings
  28  * and khugepaged scans all mappings. Defrag is only invoked by
  29  * khugepaged hugepage allocations and by page faults inside
  30  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  31  * allocations.
  32  */
  33 unsigned long transparent_hugepage_flags __read_mostly =
  34 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  35         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  36 #endif
  37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  38         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  39 #endif
  40         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  41         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  42
  43 /* default scan 8*512 pte (or vmas) every 30 second */
  44 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  45 static unsigned int khugepaged_pages_collapsed;
  46 static unsigned int khugepaged_full_scans;
  47 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  48 /* during fragmentation poll the hugepage allocator once every minute */
  49 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  50 static struct task_struct *khugepaged_thread __read_mostly;
  51 static DEFINE_MUTEX(khugepaged_mutex);
  52 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  53 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  54 /*
  55  * default collapse hugepages if there is at least one pte mapped like
  56  * it would have happened if the vma was large enough during page
  57  * fault.
  58  */
  59 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  60
  61 static int khugepaged(void *none);
  62 static int mm_slots_hash_init(void);
  63 static int khugepaged_slab_init(void);
  64 static void khugepaged_slab_free(void);
  65
  66 #define MM_SLOTS_HASH_HEADS 1024
  67 static struct hlist_head *mm_slots_hash __read_mostly;
  68 static struct kmem_cache *mm_slot_cache __read_mostly;
  69
  70 /**
  71  * struct mm_slot - hash lookup from mm to mm_slot
  72  * @hash: hash collision list
  73  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  74  * @mm: the mm that this information is valid for
  75  */
  76 struct mm_slot {
  77         struct hlist_node hash;
  78         struct list_head mm_node;
  79         struct mm_struct *mm;
  80 };
  81
  82 /**
  83  * struct khugepaged_scan - cursor for scanning
  84  * @mm_head: the head of the mm list to scan
  85  * @mm_slot: the current mm_slot we are scanning
  86  * @address: the next address inside that to be scanned
  87  *
  88  * There is only the one khugepaged_scan instance of this cursor structure.
  89  */
  90 struct khugepaged_scan {
  91         struct list_head mm_head;
  92         struct mm_slot *mm_slot;
  93         unsigned long address;
  94 };
  95 static struct khugepaged_scan khugepaged_scan = {
  96         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  97 };
  98
  99
 100 static int set_recommended_min_free_kbytes(void)
 101 {
 102         struct zone *zone;
 103         int nr_zones = 0;
 104         unsigned long recommended_min;
 105         extern int min_free_kbytes;
 106
 107         if (!khugepaged_enabled())
 108                 return 0;
 109
 110         for_each_populated_zone(zone)
 111                 nr_zones++;
 112
 113         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 114         recommended_min = pageblock_nr_pages * nr_zones * 2;
 115
 116         /*
 117          * Make sure that on average at least two pageblocks are almost free
 118          * of another type, one for a migratetype to fall back to and a
 119          * second to avoid subsequent fallbacks of other types There are 3
 120          * MIGRATE_TYPES we care about.
 121          */
 122         recommended_min += pageblock_nr_pages * nr_zones *
 123                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 124
 125         /* don't ever allow to reserve more than 5% of the lowmem */
 126         recommended_min = min(recommended_min,
 127                               (unsigned long) nr_free_buffer_pages() / 20);
 128         recommended_min <<= (PAGE_SHIFT-10);
 129
 130         if (recommended_min > min_free_kbytes)
 131                 min_free_kbytes = recommended_min;
 132         setup_per_zone_wmarks();
 133         return 0;
 134 }
 135 late_initcall(set_recommended_min_free_kbytes);
 136
 137 static int start_khugepaged(void)
 138 {
 139         int err = 0;
 140         if (khugepaged_enabled()) {
 141                 if (!khugepaged_thread)
 142                         khugepaged_thread = kthread_run(khugepaged, NULL,
 143                                                         "khugepaged");
 144                 if (unlikely(IS_ERR(khugepaged_thread))) {
 145                         printk(KERN_ERR
 146                                "khugepaged: kthread_run(khugepaged) failed\n");
 147                         err = PTR_ERR(khugepaged_thread);
 148                         khugepaged_thread = NULL;
 149                 }
 150
 151                 if (!list_empty(&khugepaged_scan.mm_head))
 152                         wake_up_interruptible(&khugepaged_wait);
 153
 154                 set_recommended_min_free_kbytes();
 155         } else if (khugepaged_thread) {
 156                 kthread_stop(khugepaged_thread);
 157                 khugepaged_thread = NULL;
 158         }
 159
 160         return err;
 161 }
 162
 163 #ifdef CONFIG_SYSFS
 164
 165 static ssize_t double_flag_show(struct kobject *kobj,
 166                                 struct kobj_attribute *attr, char *buf,
 167                                 enum transparent_hugepage_flag enabled,
 168                                 enum transparent_hugepage_flag req_madv)
 169 {
 170         if (test_bit(enabled, &transparent_hugepage_flags)) {
 171                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 172                 return sprintf(buf, "[always] madvise never\n");
 173         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 174                 return sprintf(buf, "always [madvise] never\n");
 175         else
 176                 return sprintf(buf, "always madvise [never]\n");
 177 }
 178 static ssize_t double_flag_store(struct kobject *kobj,
 179                                  struct kobj_attribute *attr,
 180                                  const char *buf, size_t count,
 181                                  enum transparent_hugepage_flag enabled,
 182                                  enum transparent_hugepage_flag req_madv)
 183 {
 184         if (!memcmp("always", buf,
 185                     min(sizeof("always")-1, count))) {
 186                 set_bit(enabled, &transparent_hugepage_flags);
 187                 clear_bit(req_madv, &transparent_hugepage_flags);
 188         } else if (!memcmp("madvise", buf,
 189                            min(sizeof("madvise")-1, count))) {
 190                 clear_bit(enabled, &transparent_hugepage_flags);
 191                 set_bit(req_madv, &transparent_hugepage_flags);
 192         } else if (!memcmp("never", buf,
 193                            min(sizeof("never")-1, count))) {
 194                 clear_bit(enabled, &transparent_hugepage_flags);
 195                 clear_bit(req_madv, &transparent_hugepage_flags);
 196         } else
 197                 return -EINVAL;
 198
 199         return count;
 200 }
 201
 202 static ssize_t enabled_show(struct kobject *kobj,
 203                             struct kobj_attribute *attr, char *buf)
 204 {
 205         return double_flag_show(kobj, attr, buf,
 206                                 TRANSPARENT_HUGEPAGE_FLAG,
 207                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 208 }
 209 static ssize_t enabled_store(struct kobject *kobj,
 210                              struct kobj_attribute *attr,
 211                              const char *buf, size_t count)
 212 {
 213         ssize_t ret;
 214
 215         ret = double_flag_store(kobj, attr, buf, count,
 216                                 TRANSPARENT_HUGEPAGE_FLAG,
 217                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 218
 219         if (ret > 0) {
 220                 int err;
 221
 222                 mutex_lock(&khugepaged_mutex);
 223                 err = start_khugepaged();
 224                 mutex_unlock(&khugepaged_mutex);
 225
 226                 if (err)
 227                         ret = err;
 228         }
 229
 230         return ret;
 231 }
 232 static struct kobj_attribute enabled_attr =
 233         __ATTR(enabled, 0644, enabled_show, enabled_store);
 234
 235 static ssize_t single_flag_show(struct kobject *kobj,
 236                                 struct kobj_attribute *attr, char *buf,
 237                                 enum transparent_hugepage_flag flag)
 238 {
 239         return sprintf(buf, "%d\n",
 240                        !!test_bit(flag, &transparent_hugepage_flags));
 241 }
 242
 243 static ssize_t single_flag_store(struct kobject *kobj,
 244                                  struct kobj_attribute *attr,
 245                                  const char *buf, size_t count,
 246                                  enum transparent_hugepage_flag flag)
 247 {
 248         unsigned long value;
 249         int ret;
 250
 251         ret = kstrtoul(buf, 10, &value);
 252         if (ret < 0)
 253                 return ret;
 254         if (value > 1)
 255                 return -EINVAL;
 256
 257         if (value)
 258                 set_bit(flag, &transparent_hugepage_flags);
 259         else
 260                 clear_bit(flag, &transparent_hugepage_flags);
 261
 262         return count;
 263 }
 264
 265 /*
 266  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 267  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 268  * memory just to allocate one more hugepage.
 269  */
 270 static ssize_t defrag_show(struct kobject *kobj,
 271                            struct kobj_attribute *attr, char *buf)
 272 {
 273         return double_flag_show(kobj, attr, buf,
 274                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 275                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 276 }
 277 static ssize_t defrag_store(struct kobject *kobj,
 278                             struct kobj_attribute *attr,
 279                             const char *buf, size_t count)
 280 {
 281         return double_flag_store(kobj, attr, buf, count,
 282                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 283                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 284 }
 285 static struct kobj_attribute defrag_attr =
 286         __ATTR(defrag, 0644, defrag_show, defrag_store);
 287
 288 #ifdef CONFIG_DEBUG_VM
 289 static ssize_t debug_cow_show(struct kobject *kobj,
 290                                 struct kobj_attribute *attr, char *buf)
 291 {
 292         return single_flag_show(kobj, attr, buf,
 293                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 294 }
 295 static ssize_t debug_cow_store(struct kobject *kobj,
 296                                struct kobj_attribute *attr,
 297                                const char *buf, size_t count)
 298 {
 299         return single_flag_store(kobj, attr, buf, count,
 300                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 301 }
 302 static struct kobj_attribute debug_cow_attr =
 303         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 304 #endif /* CONFIG_DEBUG_VM */
 305
 306 static struct attribute *hugepage_attr[] = {
 307         &enabled_attr.attr,
 308         &defrag_attr.attr,
 309 #ifdef CONFIG_DEBUG_VM
 310         &debug_cow_attr.attr,
 311 #endif
 312         NULL,
 313 };
 314
 315 static struct attribute_group hugepage_attr_group = {
 316         .attrs = hugepage_attr,
 317 };
 318
 319 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 320                                          struct kobj_attribute *attr,
 321                                          char *buf)
 322 {
 323         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 324 }
 325
 326 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 327                                           struct kobj_attribute *attr,
 328                                           const char *buf, size_t count)
 329 {
 330         unsigned long msecs;
 331         int err;
 332
 333         err = strict_strtoul(buf, 10, &msecs);
 334         if (err || msecs > UINT_MAX)
 335                 return -EINVAL;
 336
 337         khugepaged_scan_sleep_millisecs = msecs;
 338         wake_up_interruptible(&khugepaged_wait);
 339
 340         return count;
 341 }
 342 static struct kobj_attribute scan_sleep_millisecs_attr =
 343         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 344                scan_sleep_millisecs_store);
 345
 346 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 347                                           struct kobj_attribute *attr,
 348                                           char *buf)
 349 {
 350         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 351 }
 352
 353 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 354                                            struct kobj_attribute *attr,
 355                                            const char *buf, size_t count)
 356 {
 357         unsigned long msecs;
 358         int err;
 359
 360         err = strict_strtoul(buf, 10, &msecs);
 361         if (err || msecs > UINT_MAX)
 362                 return -EINVAL;
 363
 364         khugepaged_alloc_sleep_millisecs = msecs;
 365         wake_up_interruptible(&khugepaged_wait);
 366
 367         return count;
 368 }
 369 static struct kobj_attribute alloc_sleep_millisecs_attr =
 370         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 371                alloc_sleep_millisecs_store);
 372
 373 static ssize_t pages_to_scan_show(struct kobject *kobj,
 374                                   struct kobj_attribute *attr,
 375                                   char *buf)
 376 {
 377         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 378 }
 379 static ssize_t pages_to_scan_store(struct kobject *kobj,
 380                                    struct kobj_attribute *attr,
 381                                    const char *buf, size_t count)
 382 {
 383         int err;
 384         unsigned long pages;
 385
 386         err = strict_strtoul(buf, 10, &pages);
 387         if (err || !pages || pages > UINT_MAX)
 388                 return -EINVAL;
 389
 390         khugepaged_pages_to_scan = pages;
 391
 392         return count;
 393 }
 394 static struct kobj_attribute pages_to_scan_attr =
 395         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 396                pages_to_scan_store);
 397
 398 static ssize_t pages_collapsed_show(struct kobject *kobj,
 399                                     struct kobj_attribute *attr,
 400                                     char *buf)
 401 {
 402         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 403 }
 404 static struct kobj_attribute pages_collapsed_attr =
 405         __ATTR_RO(pages_collapsed);
 406
 407 static ssize_t full_scans_show(struct kobject *kobj,
 408                                struct kobj_attribute *attr,
 409                                char *buf)
 410 {
 411         return sprintf(buf, "%u\n", khugepaged_full_scans);
 412 }
 413 static struct kobj_attribute full_scans_attr =
 414         __ATTR_RO(full_scans);
 415
 416 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 417                                       struct kobj_attribute *attr, char *buf)
 418 {
 419         return single_flag_show(kobj, attr, buf,
 420                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 421 }
 422 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 423                                        struct kobj_attribute *attr,
 424                                        const char *buf, size_t count)
 425 {
 426         return single_flag_store(kobj, attr, buf, count,
 427                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 428 }
 429 static struct kobj_attribute khugepaged_defrag_attr =
 430         __ATTR(defrag, 0644, khugepaged_defrag_show,
 431                khugepaged_defrag_store);
 432
 433 /*
 434  * max_ptes_none controls if khugepaged should collapse hugepages over
 435  * any unmapped ptes in turn potentially increasing the memory
 436  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 437  * reduce the available free memory in the system as it
 438  * runs. Increasing max_ptes_none will instead potentially reduce the
 439  * free memory in the system during the khugepaged scan.
 440  */
 441 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 442                                              struct kobj_attribute *attr,
 443                                              char *buf)
 444 {
 445         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 446 }
 447 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 448                                               struct kobj_attribute *attr,
 449                                               const char *buf, size_t count)
 450 {
 451         int err;
 452         unsigned long max_ptes_none;
 453
 454         err = strict_strtoul(buf, 10, &max_ptes_none);
 455         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 456                 return -EINVAL;
 457
 458         khugepaged_max_ptes_none = max_ptes_none;
 459
 460         return count;
 461 }
 462 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 463         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 464                khugepaged_max_ptes_none_store);
 465
 466 static struct attribute *khugepaged_attr[] = {
 467         &khugepaged_defrag_attr.attr,
 468         &khugepaged_max_ptes_none_attr.attr,
 469         &pages_to_scan_attr.attr,
 470         &pages_collapsed_attr.attr,
 471         &full_scans_attr.attr,
 472         &scan_sleep_millisecs_attr.attr,
 473         &alloc_sleep_millisecs_attr.attr,
 474         NULL,
 475 };
 476
 477 static struct attribute_group khugepaged_attr_group = {
 478         .attrs = khugepaged_attr,
 479         .name = "khugepaged",
 480 };
 481
 482 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 483 {
 484         int err;
 485
 486         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 487         if (unlikely(!*hugepage_kobj)) {
 488                 printk(KERN_ERR "hugepage: failed kobject create\n");
 489                 return -ENOMEM;
 490         }
 491
 492         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 493         if (err) {
 494                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 495                 goto delete_obj;
 496         }
 497
 498         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 499         if (err) {
 500                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 501                 goto remove_hp_group;
 502         }
 503
 504         return 0;
 505
 506 remove_hp_group:
 507         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 508 delete_obj:
 509         kobject_put(*hugepage_kobj);
 510         return err;
 511 }
 512
 513 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 514 {
 515         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 516         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 517         kobject_put(hugepage_kobj);
 518 }
 519 #else
 520 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 521 {
 522         return 0;
 523 }
 524
 525 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 526 {
 527 }
 528 #endif /* CONFIG_SYSFS */
 529
 530 static int __init hugepage_init(void)
 531 {
 532         int err;
 533         struct kobject *hugepage_kobj;
 534
 535         if (!has_transparent_hugepage()) {
 536                 transparent_hugepage_flags = 0;
 537                 return -EINVAL;
 538         }
 539
 540         err = hugepage_init_sysfs(&hugepage_kobj);
 541         if (err)
 542                 return err;
 543
 544         err = khugepaged_slab_init();
 545         if (err)
 546                 goto out;
 547
 548         err = mm_slots_hash_init();
 549         if (err) {
 550                 khugepaged_slab_free();
 551                 goto out;
 552         }
 553
 554         /*
 555          * By default disable transparent hugepages on smaller systems,
 556          * where the extra memory used could hurt more than TLB overhead
 557          * is likely to save.  The admin can still enable it through /sys.
 558          */
 559         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 560                 transparent_hugepage_flags = 0;
 561
 562         start_khugepaged();
 563
 564         return 0;
 565 out:
 566         hugepage_exit_sysfs(hugepage_kobj);
 567         return err;
 568 }
 569 module_init(hugepage_init)
 570
 571 static int __init setup_transparent_hugepage(char *str)
 572 {
 573         int ret = 0;
 574         if (!str)
 575                 goto out;
 576         if (!strcmp(str, "always")) {
 577                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 578                         &transparent_hugepage_flags);
 579                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 580                           &transparent_hugepage_flags);
 581                 ret = 1;
 582         } else if (!strcmp(str, "madvise")) {
 583                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 584                           &transparent_hugepage_flags);
 585                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 586                         &transparent_hugepage_flags);
 587                 ret = 1;
 588         } else if (!strcmp(str, "never")) {
 589                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 590                           &transparent_hugepage_flags);
 591                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 592                           &transparent_hugepage_flags);
 593                 ret = 1;
 594         }
 595 out:
 596         if (!ret)
 597                 printk(KERN_WARNING
 598                        "transparent_hugepage= cannot parse, ignored\n");
 599         return ret;
 600 }
 601 __setup("transparent_hugepage=", setup_transparent_hugepage);
 602
 603 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 604 {
 605         if (likely(vma->vm_flags & VM_WRITE))
 606                 pmd = pmd_mkwrite(pmd);
 607         return pmd;
 608 }
 609
 610 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 611                                         struct vm_area_struct *vma,
 612                                         unsigned long haddr, pmd_t *pmd,
 613                                         struct page *page)
 614 {
 615         pgtable_t pgtable;
 616
 617         VM_BUG_ON(!PageCompound(page));
 618         pgtable = pte_alloc_one(mm, haddr);
 619         if (unlikely(!pgtable))
 620                 return VM_FAULT_OOM;
 621
 622         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 623         __SetPageUptodate(page);
 624
 625         spin_lock(&mm->page_table_lock);
 626         if (unlikely(!pmd_none(*pmd))) {
 627                 spin_unlock(&mm->page_table_lock);
 628                 mem_cgroup_uncharge_page(page);
 629                 put_page(page);
 630                 pte_free(mm, pgtable);
 631         } else {
 632                 pmd_t entry;
 633                 entry = mk_pmd(page, vma->vm_page_prot);
 634                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 635                 entry = pmd_mkhuge(entry);
 636                 /*
 637                  * The spinlocking to take the lru_lock inside
 638                  * page_add_new_anon_rmap() acts as a full memory
 639                  * barrier to be sure clear_huge_page writes become
 640                  * visible after the set_pmd_at() write.
 641                  */
 642                 page_add_new_anon_rmap(page, vma, haddr);
 643                 set_pmd_at(mm, haddr, pmd, entry);
 644                 pgtable_trans_huge_deposit(mm, pgtable);
 645                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 646                 mm->nr_ptes++;
 647                 spin_unlock(&mm->page_table_lock);
 648         }
 649
 650         return 0;
 651 }
 652
 653 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 654 {
 655         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 656 }
 657
 658 static inline struct page *alloc_hugepage_vma(int defrag,
 659                                               struct vm_area_struct *vma,
 660                                               unsigned long haddr, int nd,
 661                                               gfp_t extra_gfp)
 662 {
 663         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 664                                HPAGE_PMD_ORDER, vma, haddr, nd);
 665 }
 666
 667 #ifndef CONFIG_NUMA
 668 static inline struct page *alloc_hugepage(int defrag)
 669 {
 670         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 671                            HPAGE_PMD_ORDER);
 672 }
 673 #endif
 674
 675 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 676                                unsigned long address, pmd_t *pmd,
 677                                unsigned int flags)
 678 {
 679         struct page *page;
 680         unsigned long haddr = address & HPAGE_PMD_MASK;
 681         pte_t *pte;
 682
 683         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 684                 if (unlikely(anon_vma_prepare(vma)))
 685                         return VM_FAULT_OOM;
 686                 if (unlikely(khugepaged_enter(vma)))
 687                         return VM_FAULT_OOM;
 688                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 689                                           vma, haddr, numa_node_id(), 0);
 690                 if (unlikely(!page)) {
 691                         count_vm_event(THP_FAULT_FALLBACK);
 692                         goto out;
 693                 }
 694                 count_vm_event(THP_FAULT_ALLOC);
 695                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 696                         put_page(page);
 697                         goto out;
 698                 }
 699                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 700                                                           page))) {
 701                         mem_cgroup_uncharge_page(page);
 702                         put_page(page);
 703                         goto out;
 704                 }
 705
 706                 return 0;
 707         }
 708 out:
 709         /*
 710          * Use __pte_alloc instead of pte_alloc_map, because we can't
 711          * run pte_offset_map on the pmd, if an huge pmd could
 712          * materialize from under us from a different thread.
 713          */
 714         if (unlikely(pmd_none(*pmd)) &&
 715             unlikely(__pte_alloc(mm, vma, pmd, address)))
 716                 return VM_FAULT_OOM;
 717         /* if an huge pmd materialized from under us just retry later */
 718         if (unlikely(pmd_trans_huge(*pmd)))
 719                 return 0;
 720         /*
 721          * A regular pmd is established and it can't morph into a huge pmd
 722          * from under us anymore at this point because we hold the mmap_sem
 723          * read mode and khugepaged takes it in write mode. So now it's
 724          * safe to run pte_offset_map().
 725          */
 726         pte = pte_offset_map(pmd, address);
 727         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 728 }
 729
 730 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 731                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 732                   struct vm_area_struct *vma)
 733 {
 734         struct page *src_page;
 735         pmd_t pmd;
 736         pgtable_t pgtable;
 737         int ret;
 738
 739         ret = -ENOMEM;
 740         pgtable = pte_alloc_one(dst_mm, addr);
 741         if (unlikely(!pgtable))
 742                 goto out;
 743
 744         spin_lock(&dst_mm->page_table_lock);
 745         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 746
 747         ret = -EAGAIN;
 748         pmd = *src_pmd;
 749         if (unlikely(!pmd_trans_huge(pmd))) {
 750                 pte_free(dst_mm, pgtable);
 751                 goto out_unlock;
 752         }
 753         if (unlikely(pmd_trans_splitting(pmd))) {
 754                 /* split huge page running from under us */
 755                 spin_unlock(&src_mm->page_table_lock);
 756                 spin_unlock(&dst_mm->page_table_lock);
 757                 pte_free(dst_mm, pgtable);
 758
 759                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 760                 goto out;
 761         }
 762         src_page = pmd_page(pmd);
 763         VM_BUG_ON(!PageHead(src_page));
 764         get_page(src_page);
 765         page_dup_rmap(src_page);
 766         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 767
 768         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 769         pmd = pmd_mkold(pmd_wrprotect(pmd));
 770         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 771         pgtable_trans_huge_deposit(dst_mm, pgtable);
 772         dst_mm->nr_ptes++;
 773
 774         ret = 0;
 775 out_unlock:
 776         spin_unlock(&src_mm->page_table_lock);
 777         spin_unlock(&dst_mm->page_table_lock);
 778 out:
 779         return ret;
 780 }
 781
 782 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 783                                         struct vm_area_struct *vma,
 784                                         unsigned long address,
 785                                         pmd_t *pmd, pmd_t orig_pmd,
 786                                         struct page *page,
 787                                         unsigned long haddr)
 788 {
 789         pgtable_t pgtable;
 790         pmd_t _pmd;
 791         int ret = 0, i;
 792         struct page **pages;
 793         unsigned long mmun_start;       /* For mmu_notifiers */
 794         unsigned long mmun_end;         /* For mmu_notifiers */
 795
 796         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 797                         GFP_KERNEL);
 798         if (unlikely(!pages)) {
 799                 ret |= VM_FAULT_OOM;
 800                 goto out;
 801         }
 802
 803         for (i = 0; i < HPAGE_PMD_NR; i++) {
 804                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 805                                                __GFP_OTHER_NODE,
 806                                                vma, address, page_to_nid(page));
 807                 if (unlikely(!pages[i] ||
 808                              mem_cgroup_newpage_charge(pages[i], mm,
 809                                                        GFP_KERNEL))) {
 810                         if (pages[i])
 811                                 put_page(pages[i]);
 812                         mem_cgroup_uncharge_start();
 813                         while (--i >= 0) {
 814                                 mem_cgroup_uncharge_page(pages[i]);
 815                                 put_page(pages[i]);
 816                         }
 817                         mem_cgroup_uncharge_end();
 818                         kfree(pages);
 819                         ret |= VM_FAULT_OOM;
 820                         goto out;
 821                 }
 822         }
 823
 824         for (i = 0; i < HPAGE_PMD_NR; i++) {
 825                 copy_user_highpage(pages[i], page + i,
 826                                    haddr + PAGE_SIZE * i, vma);
 827                 __SetPageUptodate(pages[i]);
 828                 cond_resched();
 829         }
 830
 831         mmun_start = haddr;
 832         mmun_end   = haddr + HPAGE_PMD_SIZE;
 833         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 834
 835         spin_lock(&mm->page_table_lock);
 836         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 837                 goto out_free_pages;
 838         VM_BUG_ON(!PageHead(page));
 839
 840         pmdp_clear_flush(vma, haddr, pmd);
 841         /* leave pmd empty until pte is filled */
 842
 843         pgtable = pgtable_trans_huge_withdraw(mm);
 844         pmd_populate(mm, &_pmd, pgtable);
 845
 846         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 847                 pte_t *pte, entry;
 848                 entry = mk_pte(pages[i], vma->vm_page_prot);
 849                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 850                 page_add_new_anon_rmap(pages[i], vma, haddr);
 851                 pte = pte_offset_map(&_pmd, haddr);
 852                 VM_BUG_ON(!pte_none(*pte));
 853                 set_pte_at(mm, haddr, pte, entry);
 854                 pte_unmap(pte);
 855         }
 856         kfree(pages);
 857
 858         smp_wmb(); /* make pte visible before pmd */
 859         pmd_populate(mm, pmd, pgtable);
 860         page_remove_rmap(page);
 861         spin_unlock(&mm->page_table_lock);
 862
 863         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 864
 865         ret |= VM_FAULT_WRITE;
 866         put_page(page);
 867
 868 out:
 869         return ret;
 870
 871 out_free_pages:
 872         spin_unlock(&mm->page_table_lock);
 873         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 874         mem_cgroup_uncharge_start();
 875         for (i = 0; i < HPAGE_PMD_NR; i++) {
 876                 mem_cgroup_uncharge_page(pages[i]);
 877                 put_page(pages[i]);
 878         }
 879         mem_cgroup_uncharge_end();
 880         kfree(pages);
 881         goto out;
 882 }
 883
 884 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 885                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 886 {
 887         int ret = 0;
 888         struct page *page, *new_page;
 889         unsigned long haddr;
 890         unsigned long mmun_start;       /* For mmu_notifiers */
 891         unsigned long mmun_end;         /* For mmu_notifiers */
 892
 893         VM_BUG_ON(!vma->anon_vma);
 894         spin_lock(&mm->page_table_lock);
 895         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 896                 goto out_unlock;
 897
 898         page = pmd_page(orig_pmd);
 899         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 900         haddr = address & HPAGE_PMD_MASK;
 901         if (page_mapcount(page) == 1) {
 902                 pmd_t entry;
 903                 entry = pmd_mkyoung(orig_pmd);
 904                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 905                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 906                         update_mmu_cache_pmd(vma, address, pmd);
 907                 ret |= VM_FAULT_WRITE;
 908                 goto out_unlock;
 909         }
 910         get_page(page);
 911         spin_unlock(&mm->page_table_lock);
 912
 913         if (transparent_hugepage_enabled(vma) &&
 914             !transparent_hugepage_debug_cow())
 915                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 916                                               vma, haddr, numa_node_id(), 0);
 917         else
 918                 new_page = NULL;
 919
 920         if (unlikely(!new_page)) {
 921                 count_vm_event(THP_FAULT_FALLBACK);
 922                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 923                                                    pmd, orig_pmd, page, haddr);
 924                 if (ret & VM_FAULT_OOM)
 925                         split_huge_page(page);
 926                 put_page(page);
 927                 goto out;
 928         }
 929         count_vm_event(THP_FAULT_ALLOC);
 930
 931         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 932                 put_page(new_page);
 933                 split_huge_page(page);
 934                 put_page(page);
 935                 ret |= VM_FAULT_OOM;
 936                 goto out;
 937         }
 938
 939         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 940         __SetPageUptodate(new_page);
 941
 942         mmun_start = haddr;
 943         mmun_end   = haddr + HPAGE_PMD_SIZE;
 944         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 945
 946         spin_lock(&mm->page_table_lock);
 947         put_page(page);
 948         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 949                 spin_unlock(&mm->page_table_lock);
 950                 mem_cgroup_uncharge_page(new_page);
 951                 put_page(new_page);
 952                 goto out_mn;
 953         } else {
 954                 pmd_t entry;
 955                 VM_BUG_ON(!PageHead(page));
 956                 entry = mk_pmd(new_page, vma->vm_page_prot);
 957                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 958                 entry = pmd_mkhuge(entry);
 959                 pmdp_clear_flush(vma, haddr, pmd);
 960                 page_add_new_anon_rmap(new_page, vma, haddr);
 961                 set_pmd_at(mm, haddr, pmd, entry);
 962                 update_mmu_cache_pmd(vma, address, pmd);
 963                 page_remove_rmap(page);
 964                 put_page(page);
 965                 ret |= VM_FAULT_WRITE;
 966         }
 967         spin_unlock(&mm->page_table_lock);
 968 out_mn:
 969         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 970 out:
 971         return ret;
 972 out_unlock:
 973         spin_unlock(&mm->page_table_lock);
 974         return ret;
 975 }
 976
 977 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
 978                                    unsigned long addr,
 979                                    pmd_t *pmd,
 980                                    unsigned int flags)
 981 {
 982         struct mm_struct *mm = vma->vm_mm;
 983         struct page *page = NULL;
 984
 985         assert_spin_locked(&mm->page_table_lock);
 986
 987         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 988                 goto out;
 989
 990         page = pmd_page(*pmd);
 991         VM_BUG_ON(!PageHead(page));
 992         if (flags & FOLL_TOUCH) {
 993                 pmd_t _pmd;
 994                 /*
 995                  * We should set the dirty bit only for FOLL_WRITE but
 996                  * for now the dirty bit in the pmd is meaningless.
 997                  * And if the dirty bit will become meaningful and
 998                  * we'll only set it with FOLL_WRITE, an atomic
 999                  * set_bit will be required on the pmd to set the
1000                  * young bit, instead of the current set_pmd_at.
1001                  */
1002                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1003                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1004         }
1005         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1006                 if (page->mapping && trylock_page(page)) {
1007                         lru_add_drain();
1008                         if (page->mapping)
1009                                 mlock_vma_page(page);
1010                         unlock_page(page);
1011                 }
1012         }
1013         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1014         VM_BUG_ON(!PageCompound(page));
1015         if (flags & FOLL_GET)
1016                 get_page_foll(page);
1017
1018 out:
1019         return page;
1020 }
1021
1022 /* NUMA hinting page fault entry point for trans huge pmds */
1023 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1024                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1025 {
1026         struct page *page = NULL;
1027         unsigned long haddr = addr & HPAGE_PMD_MASK;
1028         int target_nid;
1029
1030         spin_lock(&mm->page_table_lock);
1031         if (unlikely(!pmd_same(pmd, *pmdp)))
1032                 goto out_unlock;
1033
1034         page = pmd_page(pmd);
1035         get_page(page);
1036         spin_unlock(&mm->page_table_lock);
1037
1038         target_nid = mpol_misplaced(page, vma, haddr);
1039         if (target_nid == -1)
1040                 goto clear_pmdnuma;
1041
1042         /*
1043          * Due to lacking code to migrate thp pages, we'll split
1044          * (which preserves the special PROT_NONE) and re-take the
1045          * fault on the normal pages.
1046          */
1047         split_huge_page(page);
1048         put_page(page);
1049         return 0;
1050
1051 clear_pmdnuma:
1052         spin_lock(&mm->page_table_lock);
1053         if (unlikely(!pmd_same(pmd, *pmdp)))
1054                 goto out_unlock;
1055
1056         pmd = pmd_mknonnuma(pmd);
1057         set_pmd_at(mm, haddr, pmdp, pmd);
1058         VM_BUG_ON(pmd_numa(*pmdp));
1059         update_mmu_cache_pmd(vma, addr, pmdp);
1060
1061 out_unlock:
1062         spin_unlock(&mm->page_table_lock);
1063         if (page)
1064                 put_page(page);
1065         return 0;
1066 }
1067
1068 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1069                  pmd_t *pmd, unsigned long addr)
1070 {
1071         int ret = 0;
1072
1073         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1074                 struct page *page;
1075                 pgtable_t pgtable;
1076                 pmd_t orig_pmd;
1077                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1078                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1079                 page = pmd_page(orig_pmd);
1080                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1081                 page_remove_rmap(page);
1082                 VM_BUG_ON(page_mapcount(page) < 0);
1083                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1084                 VM_BUG_ON(!PageHead(page));
1085                 tlb->mm->nr_ptes--;
1086                 spin_unlock(&tlb->mm->page_table_lock);
1087                 tlb_remove_page(tlb, page);
1088                 pte_free(tlb->mm, pgtable);
1089                 ret = 1;
1090         }
1091         return ret;
1092 }
1093
1094 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1095                 unsigned long addr, unsigned long end,
1096                 unsigned char *vec)
1097 {
1098         int ret = 0;
1099
1100         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1101                 /*
1102                  * All logical pages in the range are present
1103                  * if backed by a huge page.
1104                  */
1105                 spin_unlock(&vma->vm_mm->page_table_lock);
1106                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1107                 ret = 1;
1108         }
1109
1110         return ret;
1111 }
1112
1113 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1114                   unsigned long old_addr,
1115                   unsigned long new_addr, unsigned long old_end,
1116                   pmd_t *old_pmd, pmd_t *new_pmd)
1117 {
1118         int ret = 0;
1119         pmd_t pmd;
1120
1121         struct mm_struct *mm = vma->vm_mm;
1122
1123         if ((old_addr & ~HPAGE_PMD_MASK) ||
1124             (new_addr & ~HPAGE_PMD_MASK) ||
1125             old_end - old_addr < HPAGE_PMD_SIZE ||
1126             (new_vma->vm_flags & VM_NOHUGEPAGE))
1127                 goto out;
1128
1129         /*
1130          * The destination pmd shouldn't be established, free_pgtables()
1131          * should have release it.
1132          */
1133         if (WARN_ON(!pmd_none(*new_pmd))) {
1134                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1135                 goto out;
1136         }
1137
1138         ret = __pmd_trans_huge_lock(old_pmd, vma);
1139         if (ret == 1) {
1140                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1141                 VM_BUG_ON(!pmd_none(*new_pmd));
1142                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1143                 spin_unlock(&mm->page_table_lock);
1144         }
1145 out:
1146         return ret;
1147 }
1148
1149 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1150                 unsigned long addr, pgprot_t newprot)
1151 {
1152         struct mm_struct *mm = vma->vm_mm;
1153         int ret = 0;
1154
1155         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1156                 pmd_t entry;
1157                 entry = pmdp_get_and_clear(mm, addr, pmd);
1158                 entry = pmd_modify(entry, newprot);
1159                 set_pmd_at(mm, addr, pmd, entry);
1160                 spin_unlock(&vma->vm_mm->page_table_lock);
1161                 ret = 1;
1162         }
1163
1164         return ret;
1165 }
1166
1167 /*
1168  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1169  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1170  *
1171  * Note that if it returns 1, this routine returns without unlocking page
1172  * table locks. So callers must unlock them.
1173  */
1174 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1175 {
1176         spin_lock(&vma->vm_mm->page_table_lock);
1177         if (likely(pmd_trans_huge(*pmd))) {
1178                 if (unlikely(pmd_trans_splitting(*pmd))) {
1179                         spin_unlock(&vma->vm_mm->page_table_lock);
1180                         wait_split_huge_page(vma->anon_vma, pmd);
1181                         return -1;
1182                 } else {
1183                         /* Thp mapped by 'pmd' is stable, so we can
1184                          * handle it as it is. */
1185                         return 1;
1186                 }
1187         }
1188         spin_unlock(&vma->vm_mm->page_table_lock);
1189         return 0;
1190 }
1191
1192 pmd_t *page_check_address_pmd(struct page *page,
1193                               struct mm_struct *mm,
1194                               unsigned long address,
1195                               enum page_check_address_pmd_flag flag)
1196 {
1197         pgd_t *pgd;
1198         pud_t *pud;
1199         pmd_t *pmd, *ret = NULL;
1200
1201         if (address & ~HPAGE_PMD_MASK)
1202                 goto out;
1203
1204         pgd = pgd_offset(mm, address);
1205         if (!pgd_present(*pgd))
1206                 goto out;
1207
1208         pud = pud_offset(pgd, address);
1209         if (!pud_present(*pud))
1210                 goto out;
1211
1212         pmd = pmd_offset(pud, address);
1213         if (pmd_none(*pmd))
1214                 goto out;
1215         if (pmd_page(*pmd) != page)
1216                 goto out;
1217         /*
1218          * split_vma() may create temporary aliased mappings. There is
1219          * no risk as long as all huge pmd are found and have their
1220          * splitting bit set before __split_huge_page_refcount
1221          * runs. Finding the same huge pmd more than once during the
1222          * same rmap walk is not a problem.
1223          */
1224         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1225             pmd_trans_splitting(*pmd))
1226                 goto out;
1227         if (pmd_trans_huge(*pmd)) {
1228                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1229                           !pmd_trans_splitting(*pmd));
1230                 ret = pmd;
1231         }
1232 out:
1233         return ret;
1234 }
1235
1236 static int __split_huge_page_splitting(struct page *page,
1237                                        struct vm_area_struct *vma,
1238                                        unsigned long address)
1239 {
1240         struct mm_struct *mm = vma->vm_mm;
1241         pmd_t *pmd;
1242         int ret = 0;
1243         /* For mmu_notifiers */
1244         const unsigned long mmun_start = address;
1245         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1246
1247         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1248         spin_lock(&mm->page_table_lock);
1249         pmd = page_check_address_pmd(page, mm, address,
1250                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1251         if (pmd) {
1252                 /*
1253                  * We can't temporarily set the pmd to null in order
1254                  * to split it, the pmd must remain marked huge at all
1255                  * times or the VM won't take the pmd_trans_huge paths
1256                  * and it won't wait on the anon_vma->root->mutex to
1257                  * serialize against split_huge_page*.
1258                  */
1259                 pmdp_splitting_flush(vma, address, pmd);
1260                 ret = 1;
1261         }
1262         spin_unlock(&mm->page_table_lock);
1263         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1264
1265         return ret;
1266 }
1267
1268 static void __split_huge_page_refcount(struct page *page)
1269 {
1270         int i;
1271         struct zone *zone = page_zone(page);
1272         struct lruvec *lruvec;
1273         int tail_count = 0;
1274
1275         /* prevent PageLRU to go away from under us, and freeze lru stats */
1276         spin_lock_irq(&zone->lru_lock);
1277         lruvec = mem_cgroup_page_lruvec(page, zone);
1278
1279         compound_lock(page);
1280         /* complete memcg works before add pages to LRU */
1281         mem_cgroup_split_huge_fixup(page);
1282
1283         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1284                 struct page *page_tail = page + i;
1285
1286                 /* tail_page->_mapcount cannot change */
1287                 BUG_ON(page_mapcount(page_tail) < 0);
1288                 tail_count += page_mapcount(page_tail);
1289                 /* check for overflow */
1290                 BUG_ON(tail_count < 0);
1291                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1292                 /*
1293                  * tail_page->_count is zero and not changing from
1294                  * under us. But get_page_unless_zero() may be running
1295                  * from under us on the tail_page. If we used
1296                  * atomic_set() below instead of atomic_add(), we
1297                  * would then run atomic_set() concurrently with
1298                  * get_page_unless_zero(), and atomic_set() is
1299                  * implemented in C not using locked ops. spin_unlock
1300                  * on x86 sometime uses locked ops because of PPro
1301                  * errata 66, 92, so unless somebody can guarantee
1302                  * atomic_set() here would be safe on all archs (and
1303                  * not only on x86), it's safer to use atomic_add().
1304                  */
1305                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1306                            &page_tail->_count);
1307
1308                 /* after clearing PageTail the gup refcount can be released */
1309                 smp_mb();
1310
1311                 /*
1312                  * retain hwpoison flag of the poisoned tail page:
1313                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1314                  *   by the memory-failure.
1315                  */
1316                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1317                 page_tail->flags |= (page->flags &
1318                                      ((1L << PG_referenced) |
1319                                       (1L << PG_swapbacked) |
1320                                       (1L << PG_mlocked) |
1321                                       (1L << PG_uptodate)));
1322                 page_tail->flags |= (1L << PG_dirty);
1323
1324                 /* clear PageTail before overwriting first_page */
1325                 smp_wmb();
1326
1327                 /*
1328                  * __split_huge_page_splitting() already set the
1329                  * splitting bit in all pmd that could map this
1330                  * hugepage, that will ensure no CPU can alter the
1331                  * mapcount on the head page. The mapcount is only
1332                  * accounted in the head page and it has to be
1333                  * transferred to all tail pages in the below code. So
1334                  * for this code to be safe, the split the mapcount
1335                  * can't change. But that doesn't mean userland can't
1336                  * keep changing and reading the page contents while
1337                  * we transfer the mapcount, so the pmd splitting
1338                  * status is achieved setting a reserved bit in the
1339                  * pmd, not by clearing the present bit.
1340                 */
1341                 page_tail->_mapcount = page->_mapcount;
1342
1343                 BUG_ON(page_tail->mapping);
1344                 page_tail->mapping = page->mapping;
1345
1346                 page_tail->index = page->index + i;
1347
1348                 BUG_ON(!PageAnon(page_tail));
1349                 BUG_ON(!PageUptodate(page_tail));
1350                 BUG_ON(!PageDirty(page_tail));
1351                 BUG_ON(!PageSwapBacked(page_tail));
1352
1353                 lru_add_page_tail(page, page_tail, lruvec);
1354         }
1355         atomic_sub(tail_count, &page->_count);
1356         BUG_ON(atomic_read(&page->_count) <= 0);
1357
1358         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1359         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1360
1361         ClearPageCompound(page);
1362         compound_unlock(page);
1363         spin_unlock_irq(&zone->lru_lock);
1364
1365         for (i = 1; i < HPAGE_PMD_NR; i++) {
1366                 struct page *page_tail = page + i;
1367                 BUG_ON(page_count(page_tail) <= 0);
1368                 /*
1369                  * Tail pages may be freed if there wasn't any mapping
1370                  * like if add_to_swap() is running on a lru page that
1371                  * had its mapping zapped. And freeing these pages
1372                  * requires taking the lru_lock so we do the put_page
1373                  * of the tail pages after the split is complete.
1374                  */
1375                 put_page(page_tail);
1376         }
1377
1378         /*
1379          * Only the head page (now become a regular page) is required
1380          * to be pinned by the caller.
1381          */
1382         BUG_ON(page_count(page) <= 0);
1383 }
1384
1385 static int __split_huge_page_map(struct page *page,
1386                                  struct vm_area_struct *vma,
1387                                  unsigned long address)
1388 {
1389         struct mm_struct *mm = vma->vm_mm;
1390         pmd_t *pmd, _pmd;
1391         int ret = 0, i;
1392         pgtable_t pgtable;
1393         unsigned long haddr;
1394
1395         spin_lock(&mm->page_table_lock);
1396         pmd = page_check_address_pmd(page, mm, address,
1397                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1398         if (pmd) {
1399                 pgtable = pgtable_trans_huge_withdraw(mm);
1400                 pmd_populate(mm, &_pmd, pgtable);
1401
1402                 haddr = address;
1403                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1404                         pte_t *pte, entry;
1405                         BUG_ON(PageCompound(page+i));
1406                         entry = mk_pte(page + i, vma->vm_page_prot);
1407                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1408                         if (!pmd_write(*pmd))
1409                                 entry = pte_wrprotect(entry);
1410                         else
1411                                 BUG_ON(page_mapcount(page) != 1);
1412                         if (!pmd_young(*pmd))
1413                                 entry = pte_mkold(entry);
1414                         if (pmd_numa(*pmd))
1415                                 entry = pte_mknuma(entry);
1416                         pte = pte_offset_map(&_pmd, haddr);
1417                         BUG_ON(!pte_none(*pte));
1418                         set_pte_at(mm, haddr, pte, entry);
1419                         pte_unmap(pte);
1420                 }
1421
1422                 smp_wmb(); /* make pte visible before pmd */
1423                 /*
1424                  * Up to this point the pmd is present and huge and
1425                  * userland has the whole access to the hugepage
1426                  * during the split (which happens in place). If we
1427                  * overwrite the pmd with the not-huge version
1428                  * pointing to the pte here (which of course we could
1429                  * if all CPUs were bug free), userland could trigger
1430                  * a small page size TLB miss on the small sized TLB
1431                  * while the hugepage TLB entry is still established
1432                  * in the huge TLB. Some CPU doesn't like that. See
1433                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1434                  * Erratum 383 on page 93. Intel should be safe but is
1435                  * also warns that it's only safe if the permission
1436                  * and cache attributes of the two entries loaded in
1437                  * the two TLB is identical (which should be the case
1438                  * here). But it is generally safer to never allow
1439                  * small and huge TLB entries for the same virtual
1440                  * address to be loaded simultaneously. So instead of
1441                  * doing "pmd_populate(); flush_tlb_range();" we first
1442                  * mark the current pmd notpresent (atomically because
1443                  * here the pmd_trans_huge and pmd_trans_splitting
1444                  * must remain set at all times on the pmd until the
1445                  * split is complete for this pmd), then we flush the
1446                  * SMP TLB and finally we write the non-huge version
1447                  * of the pmd entry with pmd_populate.
1448                  */
1449                 pmdp_invalidate(vma, address, pmd);
1450                 pmd_populate(mm, pmd, pgtable);
1451                 ret = 1;
1452         }
1453         spin_unlock(&mm->page_table_lock);
1454
1455         return ret;
1456 }
1457
1458 /* must be called with anon_vma->root->mutex hold */
1459 static void __split_huge_page(struct page *page,
1460                               struct anon_vma *anon_vma)
1461 {
1462         int mapcount, mapcount2;
1463         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1464         struct anon_vma_chain *avc;
1465
1466         BUG_ON(!PageHead(page));
1467         BUG_ON(PageTail(page));
1468
1469         mapcount = 0;
1470         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1471                 struct vm_area_struct *vma = avc->vma;
1472                 unsigned long addr = vma_address(page, vma);
1473                 BUG_ON(is_vma_temporary_stack(vma));
1474                 mapcount += __split_huge_page_splitting(page, vma, addr);
1475         }
1476         /*
1477          * It is critical that new vmas are added to the tail of the
1478          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1479          * and establishes a child pmd before
1480          * __split_huge_page_splitting() freezes the parent pmd (so if
1481          * we fail to prevent copy_huge_pmd() from running until the
1482          * whole __split_huge_page() is complete), we will still see
1483          * the newly established pmd of the child later during the
1484          * walk, to be able to set it as pmd_trans_splitting too.
1485          */
1486         if (mapcount != page_mapcount(page))
1487                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1488                        mapcount, page_mapcount(page));
1489         BUG_ON(mapcount != page_mapcount(page));
1490
1491         __split_huge_page_refcount(page);
1492
1493         mapcount2 = 0;
1494         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1495                 struct vm_area_struct *vma = avc->vma;
1496                 unsigned long addr = vma_address(page, vma);
1497                 BUG_ON(is_vma_temporary_stack(vma));
1498                 mapcount2 += __split_huge_page_map(page, vma, addr);
1499         }
1500         if (mapcount != mapcount2)
1501                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1502                        mapcount, mapcount2, page_mapcount(page));
1503         BUG_ON(mapcount != mapcount2);
1504 }
1505
1506 int split_huge_page(struct page *page)
1507 {
1508         struct anon_vma *anon_vma;
1509         int ret = 1;
1510
1511         BUG_ON(!PageAnon(page));
1512         anon_vma = page_lock_anon_vma(page);
1513         if (!anon_vma)
1514                 goto out;
1515         ret = 0;
1516         if (!PageCompound(page))
1517                 goto out_unlock;
1518
1519         BUG_ON(!PageSwapBacked(page));
1520         __split_huge_page(page, anon_vma);
1521         count_vm_event(THP_SPLIT);
1522
1523         BUG_ON(PageCompound(page));
1524 out_unlock:
1525         page_unlock_anon_vma(anon_vma);
1526 out:
1527         return ret;
1528 }
1529
1530 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1531
1532 int hugepage_madvise(struct vm_area_struct *vma,
1533                      unsigned long *vm_flags, int advice)
1534 {
1535         struct mm_struct *mm = vma->vm_mm;
1536
1537         switch (advice) {
1538         case MADV_HUGEPAGE:
1539                 /*
1540                  * Be somewhat over-protective like KSM for now!
1541                  */
1542                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1543                         return -EINVAL;
1544                 if (mm->def_flags & VM_NOHUGEPAGE)
1545                         return -EINVAL;
1546                 *vm_flags &= ~VM_NOHUGEPAGE;
1547                 *vm_flags |= VM_HUGEPAGE;
1548                 /*
1549                  * If the vma become good for khugepaged to scan,
1550                  * register it here without waiting a page fault that
1551                  * may not happen any time soon.
1552                  */
1553                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1554                         return -ENOMEM;
1555                 break;
1556         case MADV_NOHUGEPAGE:
1557                 /*
1558                  * Be somewhat over-protective like KSM for now!
1559                  */
1560                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1561                         return -EINVAL;
1562                 *vm_flags &= ~VM_HUGEPAGE;
1563                 *vm_flags |= VM_NOHUGEPAGE;
1564                 /*
1565                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1566                  * this vma even if we leave the mm registered in khugepaged if
1567                  * it got registered before VM_NOHUGEPAGE was set.
1568                  */
1569                 break;
1570         }
1571
1572         return 0;
1573 }
1574
1575 static int __init khugepaged_slab_init(void)
1576 {
1577         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1578                                           sizeof(struct mm_slot),
1579                                           __alignof__(struct mm_slot), 0, NULL);
1580         if (!mm_slot_cache)
1581                 return -ENOMEM;
1582
1583         return 0;
1584 }
1585
1586 static void __init khugepaged_slab_free(void)
1587 {
1588         kmem_cache_destroy(mm_slot_cache);
1589         mm_slot_cache = NULL;
1590 }
1591
1592 static inline struct mm_slot *alloc_mm_slot(void)
1593 {
1594         if (!mm_slot_cache)     /* initialization failed */
1595                 return NULL;
1596         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1597 }
1598
1599 static inline void free_mm_slot(struct mm_slot *mm_slot)
1600 {
1601         kmem_cache_free(mm_slot_cache, mm_slot);
1602 }
1603
1604 static int __init mm_slots_hash_init(void)
1605 {
1606         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1607                                 GFP_KERNEL);
1608         if (!mm_slots_hash)
1609                 return -ENOMEM;
1610         return 0;
1611 }
1612
1613 #if 0
1614 static void __init mm_slots_hash_free(void)
1615 {
1616         kfree(mm_slots_hash);
1617         mm_slots_hash = NULL;
1618 }
1619 #endif
1620
1621 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1622 {
1623         struct mm_slot *mm_slot;
1624         struct hlist_head *bucket;
1625         struct hlist_node *node;
1626
1627         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1628                                 % MM_SLOTS_HASH_HEADS];
1629         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1630                 if (mm == mm_slot->mm)
1631                         return mm_slot;
1632         }
1633         return NULL;
1634 }
1635
1636 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1637                                     struct mm_slot *mm_slot)
1638 {
1639         struct hlist_head *bucket;
1640
1641         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1642                                 % MM_SLOTS_HASH_HEADS];
1643         mm_slot->mm = mm;
1644         hlist_add_head(&mm_slot->hash, bucket);
1645 }
1646
1647 static inline int khugepaged_test_exit(struct mm_struct *mm)
1648 {
1649         return atomic_read(&mm->mm_users) == 0;
1650 }
1651
1652 int __khugepaged_enter(struct mm_struct *mm)
1653 {
1654         struct mm_slot *mm_slot;
1655         int wakeup;
1656
1657         mm_slot = alloc_mm_slot();
1658         if (!mm_slot)
1659                 return -ENOMEM;
1660
1661         /* __khugepaged_exit() must not run from under us */
1662         VM_BUG_ON(khugepaged_test_exit(mm));
1663         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1664                 free_mm_slot(mm_slot);
1665                 return 0;
1666         }
1667
1668         spin_lock(&khugepaged_mm_lock);
1669         insert_to_mm_slots_hash(mm, mm_slot);
1670         /*
1671          * Insert just behind the scanning cursor, to let the area settle
1672          * down a little.
1673          */
1674         wakeup = list_empty(&khugepaged_scan.mm_head);
1675         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1676         spin_unlock(&khugepaged_mm_lock);
1677
1678         atomic_inc(&mm->mm_count);
1679         if (wakeup)
1680                 wake_up_interruptible(&khugepaged_wait);
1681
1682         return 0;
1683 }
1684
1685 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1686 {
1687         unsigned long hstart, hend;
1688         if (!vma->anon_vma)
1689                 /*
1690                  * Not yet faulted in so we will register later in the
1691                  * page fault if needed.
1692                  */
1693                 return 0;
1694         if (vma->vm_ops)
1695                 /* khugepaged not yet working on file or special mappings */
1696                 return 0;
1697         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1698         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1699         hend = vma->vm_end & HPAGE_PMD_MASK;
1700         if (hstart < hend)
1701                 return khugepaged_enter(vma);
1702         return 0;
1703 }
1704
1705 void __khugepaged_exit(struct mm_struct *mm)
1706 {
1707         struct mm_slot *mm_slot;
1708         int free = 0;
1709
1710         spin_lock(&khugepaged_mm_lock);
1711         mm_slot = get_mm_slot(mm);
1712         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1713                 hlist_del(&mm_slot->hash);
1714                 list_del(&mm_slot->mm_node);
1715                 free = 1;
1716         }
1717         spin_unlock(&khugepaged_mm_lock);
1718
1719         if (free) {
1720                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1721                 free_mm_slot(mm_slot);
1722                 mmdrop(mm);
1723         } else if (mm_slot) {
1724                 /*
1725                  * This is required to serialize against
1726                  * khugepaged_test_exit() (which is guaranteed to run
1727                  * under mmap sem read mode). Stop here (after we
1728                  * return all pagetables will be destroyed) until
1729                  * khugepaged has finished working on the pagetables
1730                  * under the mmap_sem.
1731                  */
1732                 down_write(&mm->mmap_sem);
1733                 up_write(&mm->mmap_sem);
1734         }
1735 }
1736
1737 static void release_pte_page(struct page *page)
1738 {
1739         /* 0 stands for page_is_file_cache(page) == false */
1740         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1741         unlock_page(page);
1742         putback_lru_page(page);
1743 }
1744
1745 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1746 {
1747         while (--_pte >= pte) {
1748                 pte_t pteval = *_pte;
1749                 if (!pte_none(pteval))
1750                         release_pte_page(pte_page(pteval));
1751         }
1752 }
1753
1754 static void release_all_pte_pages(pte_t *pte)
1755 {
1756         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1757 }
1758
1759 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1760                                         unsigned long address,
1761                                         pte_t *pte)
1762 {
1763         struct page *page;
1764         pte_t *_pte;
1765         int referenced = 0, isolated = 0, none = 0;
1766         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1767              _pte++, address += PAGE_SIZE) {
1768                 pte_t pteval = *_pte;
1769                 if (pte_none(pteval)) {
1770                         if (++none <= khugepaged_max_ptes_none)
1771                                 continue;
1772                         else {
1773                                 release_pte_pages(pte, _pte);
1774                                 goto out;
1775                         }
1776                 }
1777                 if (!pte_present(pteval) || !pte_write(pteval)) {
1778                         release_pte_pages(pte, _pte);
1779                         goto out;
1780                 }
1781                 page = vm_normal_page(vma, address, pteval);
1782                 if (unlikely(!page)) {
1783                         release_pte_pages(pte, _pte);
1784                         goto out;
1785                 }
1786                 VM_BUG_ON(PageCompound(page));
1787                 BUG_ON(!PageAnon(page));
1788                 VM_BUG_ON(!PageSwapBacked(page));
1789
1790                 /* cannot use mapcount: can't collapse if there's a gup pin */
1791                 if (page_count(page) != 1) {
1792                         release_pte_pages(pte, _pte);
1793                         goto out;
1794                 }
1795                 /*
1796                  * We can do it before isolate_lru_page because the
1797                  * page can't be freed from under us. NOTE: PG_lock
1798                  * is needed to serialize against split_huge_page
1799                  * when invoked from the VM.
1800                  */
1801                 if (!trylock_page(page)) {
1802                         release_pte_pages(pte, _pte);
1803                         goto out;
1804                 }
1805                 /*
1806                  * Isolate the page to avoid collapsing an hugepage
1807                  * currently in use by the VM.
1808                  */
1809                 if (isolate_lru_page(page)) {
1810                         unlock_page(page);
1811                         release_pte_pages(pte, _pte);
1812                         goto out;
1813                 }
1814                 /* 0 stands for page_is_file_cache(page) == false */
1815                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1816                 VM_BUG_ON(!PageLocked(page));
1817                 VM_BUG_ON(PageLRU(page));
1818
1819                 /* If there is no mapped pte young don't collapse the page */
1820                 if (pte_young(pteval) || PageReferenced(page) ||
1821                     mmu_notifier_test_young(vma->vm_mm, address))
1822                         referenced = 1;
1823         }
1824         if (unlikely(!referenced))
1825                 release_all_pte_pages(pte);
1826         else
1827                 isolated = 1;
1828 out:
1829         return isolated;
1830 }
1831
1832 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1833                                       struct vm_area_struct *vma,
1834                                       unsigned long address,
1835                                       spinlock_t *ptl)
1836 {
1837         pte_t *_pte;
1838         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1839                 pte_t pteval = *_pte;
1840                 struct page *src_page;
1841
1842                 if (pte_none(pteval)) {
1843                         clear_user_highpage(page, address);
1844                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1845                 } else {
1846                         src_page = pte_page(pteval);
1847                         copy_user_highpage(page, src_page, address, vma);
1848                         VM_BUG_ON(page_mapcount(src_page) != 1);
1849                         release_pte_page(src_page);
1850                         /*
1851                          * ptl mostly unnecessary, but preempt has to
1852                          * be disabled to update the per-cpu stats
1853                          * inside page_remove_rmap().
1854                          */
1855                         spin_lock(ptl);
1856                         /*
1857                          * paravirt calls inside pte_clear here are
1858                          * superfluous.
1859                          */
1860                         pte_clear(vma->vm_mm, address, _pte);
1861                         page_remove_rmap(src_page);
1862                         spin_unlock(ptl);
1863                         free_page_and_swap_cache(src_page);
1864                 }
1865
1866                 address += PAGE_SIZE;
1867                 page++;
1868         }
1869 }
1870
1871 static void khugepaged_alloc_sleep(void)
1872 {
1873         wait_event_freezable_timeout(khugepaged_wait, false,
1874                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1875 }
1876
1877 #ifdef CONFIG_NUMA
1878 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1879 {
1880         if (IS_ERR(*hpage)) {
1881                 if (!*wait)
1882                         return false;
1883
1884                 *wait = false;
1885                 *hpage = NULL;
1886                 khugepaged_alloc_sleep();
1887         } else if (*hpage) {
1888                 put_page(*hpage);
1889                 *hpage = NULL;
1890         }
1891
1892         return true;
1893 }
1894
1895 static struct page
1896 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1897                        struct vm_area_struct *vma, unsigned long address,
1898                        int node)
1899 {
1900         VM_BUG_ON(*hpage);
1901         /*
1902          * Allocate the page while the vma is still valid and under
1903          * the mmap_sem read mode so there is no memory allocation
1904          * later when we take the mmap_sem in write mode. This is more
1905          * friendly behavior (OTOH it may actually hide bugs) to
1906          * filesystems in userland with daemons allocating memory in
1907          * the userland I/O paths.  Allocating memory with the
1908          * mmap_sem in read mode is good idea also to allow greater
1909          * scalability.
1910          */
1911         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1912                                       node, __GFP_OTHER_NODE);
1913
1914         /*
1915          * After allocating the hugepage, release the mmap_sem read lock in
1916          * preparation for taking it in write mode.
1917          */
1918         up_read(&mm->mmap_sem);
1919         if (unlikely(!*hpage)) {
1920                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1921                 *hpage = ERR_PTR(-ENOMEM);
1922                 return NULL;
1923         }
1924
1925         count_vm_event(THP_COLLAPSE_ALLOC);
1926         return *hpage;
1927 }
1928 #else
1929 static struct page *khugepaged_alloc_hugepage(bool *wait)
1930 {
1931         struct page *hpage;
1932
1933         do {
1934                 hpage = alloc_hugepage(khugepaged_defrag());
1935                 if (!hpage) {
1936                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1937                         if (!*wait)
1938                                 return NULL;
1939
1940                         *wait = false;
1941                         khugepaged_alloc_sleep();
1942                 } else
1943                         count_vm_event(THP_COLLAPSE_ALLOC);
1944         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1945
1946         return hpage;
1947 }
1948
1949 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1950 {
1951         if (!*hpage)
1952                 *hpage = khugepaged_alloc_hugepage(wait);
1953
1954         if (unlikely(!*hpage))
1955                 return false;
1956
1957         return true;
1958 }
1959
1960 static struct page
1961 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1962                        struct vm_area_struct *vma, unsigned long address,
1963                        int node)
1964 {
1965         up_read(&mm->mmap_sem);
1966         VM_BUG_ON(!*hpage);
1967         return  *hpage;
1968 }
1969 #endif
1970
1971 static void collapse_huge_page(struct mm_struct *mm,
1972                                    unsigned long address,
1973                                    struct page **hpage,
1974                                    struct vm_area_struct *vma,
1975                                    int node)
1976 {
1977         pgd_t *pgd;
1978         pud_t *pud;
1979         pmd_t *pmd, _pmd;
1980         pte_t *pte;
1981         pgtable_t pgtable;
1982         struct page *new_page;
1983         spinlock_t *ptl;
1984         int isolated;
1985         unsigned long hstart, hend;
1986         unsigned long mmun_start;       /* For mmu_notifiers */
1987         unsigned long mmun_end;         /* For mmu_notifiers */
1988
1989         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1990
1991         /* release the mmap_sem read lock. */
1992         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1993         if (!new_page)
1994                 return;
1995
1996         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1997                 return;
1998
1999         /*
2000          * Prevent all access to pagetables with the exception of
2001          * gup_fast later hanlded by the ptep_clear_flush and the VM
2002          * handled by the anon_vma lock + PG_lock.
2003          */
2004         down_write(&mm->mmap_sem);
2005         if (unlikely(khugepaged_test_exit(mm)))
2006                 goto out;
2007
2008         vma = find_vma(mm, address);
2009         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2010         hend = vma->vm_end & HPAGE_PMD_MASK;
2011         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2012                 goto out;
2013
2014         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2015             (vma->vm_flags & VM_NOHUGEPAGE))
2016                 goto out;
2017
2018         if (!vma->anon_vma || vma->vm_ops)
2019                 goto out;
2020         if (is_vma_temporary_stack(vma))
2021                 goto out;
2022         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2023
2024         pgd = pgd_offset(mm, address);
2025         if (!pgd_present(*pgd))
2026                 goto out;
2027
2028         pud = pud_offset(pgd, address);
2029         if (!pud_present(*pud))
2030                 goto out;
2031
2032         pmd = pmd_offset(pud, address);
2033         /* pmd can't go away or become huge under us */
2034         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2035                 goto out;
2036
2037         anon_vma_lock(vma->anon_vma);
2038
2039         pte = pte_offset_map(pmd, address);
2040         ptl = pte_lockptr(mm, pmd);
2041
2042         mmun_start = address;
2043         mmun_end   = address + HPAGE_PMD_SIZE;
2044         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2045         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2046         /*
2047          * After this gup_fast can't run anymore. This also removes
2048          * any huge TLB entry from the CPU so we won't allow
2049          * huge and small TLB entries for the same virtual address
2050          * to avoid the risk of CPU bugs in that area.
2051          */
2052         _pmd = pmdp_clear_flush(vma, address, pmd);
2053         spin_unlock(&mm->page_table_lock);
2054         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2055
2056         spin_lock(ptl);
2057         isolated = __collapse_huge_page_isolate(vma, address, pte);
2058         spin_unlock(ptl);
2059
2060         if (unlikely(!isolated)) {
2061                 pte_unmap(pte);
2062                 spin_lock(&mm->page_table_lock);
2063                 BUG_ON(!pmd_none(*pmd));
2064                 set_pmd_at(mm, address, pmd, _pmd);
2065                 spin_unlock(&mm->page_table_lock);
2066                 anon_vma_unlock(vma->anon_vma);
2067                 goto out;
2068         }
2069
2070         /*
2071          * All pages are isolated and locked so anon_vma rmap
2072          * can't run anymore.
2073          */
2074         anon_vma_unlock(vma->anon_vma);
2075
2076         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2077         pte_unmap(pte);
2078         __SetPageUptodate(new_page);
2079         pgtable = pmd_pgtable(_pmd);
2080
2081         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2082         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2083         _pmd = pmd_mkhuge(_pmd);
2084
2085         /*
2086          * spin_lock() below is not the equivalent of smp_wmb(), so
2087          * this is needed to avoid the copy_huge_page writes to become
2088          * visible after the set_pmd_at() write.
2089          */
2090         smp_wmb();
2091
2092         spin_lock(&mm->page_table_lock);
2093         BUG_ON(!pmd_none(*pmd));
2094         page_add_new_anon_rmap(new_page, vma, address);
2095         set_pmd_at(mm, address, pmd, _pmd);
2096         update_mmu_cache_pmd(vma, address, pmd);
2097         pgtable_trans_huge_deposit(mm, pgtable);
2098         spin_unlock(&mm->page_table_lock);
2099
2100         *hpage = NULL;
2101
2102         khugepaged_pages_collapsed++;
2103 out_up_write:
2104         up_write(&mm->mmap_sem);
2105         return;
2106
2107 out:
2108         mem_cgroup_uncharge_page(new_page);
2109         goto out_up_write;
2110 }
2111
2112 static int khugepaged_scan_pmd(struct mm_struct *mm,
2113                                struct vm_area_struct *vma,
2114                                unsigned long address,
2115                                struct page **hpage)
2116 {
2117         pgd_t *pgd;
2118         pud_t *pud;
2119         pmd_t *pmd;
2120         pte_t *pte, *_pte;
2121         int ret = 0, referenced = 0, none = 0;
2122         struct page *page;
2123         unsigned long _address;
2124         spinlock_t *ptl;
2125         int node = -1;
2126
2127         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2128
2129         pgd = pgd_offset(mm, address);
2130         if (!pgd_present(*pgd))
2131                 goto out;
2132
2133         pud = pud_offset(pgd, address);
2134         if (!pud_present(*pud))
2135                 goto out;
2136
2137         pmd = pmd_offset(pud, address);
2138         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2139                 goto out;
2140
2141         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2142         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2143              _pte++, _address += PAGE_SIZE) {
2144                 pte_t pteval = *_pte;
2145                 if (pte_none(pteval)) {
2146                         if (++none <= khugepaged_max_ptes_none)
2147                                 continue;
2148                         else
2149                                 goto out_unmap;
2150                 }
2151                 if (!pte_present(pteval) || !pte_write(pteval))
2152                         goto out_unmap;
2153                 page = vm_normal_page(vma, _address, pteval);
2154                 if (unlikely(!page))
2155                         goto out_unmap;
2156                 /*
2157                  * Chose the node of the first page. This could
2158                  * be more sophisticated and look at more pages,
2159                  * but isn't for now.
2160                  */
2161                 if (node == -1)
2162                         node = page_to_nid(page);
2163                 VM_BUG_ON(PageCompound(page));
2164                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2165                         goto out_unmap;
2166                 /* cannot use mapcount: can't collapse if there's a gup pin */
2167                 if (page_count(page) != 1)
2168                         goto out_unmap;
2169                 if (pte_young(pteval) || PageReferenced(page) ||
2170                     mmu_notifier_test_young(vma->vm_mm, address))
2171                         referenced = 1;
2172         }
2173         if (referenced)
2174                 ret = 1;
2175 out_unmap:
2176         pte_unmap_unlock(pte, ptl);
2177         if (ret)
2178                 /* collapse_huge_page will return with the mmap_sem released */
2179                 collapse_huge_page(mm, address, hpage, vma, node);
2180 out:
2181         return ret;
2182 }
2183
2184 static void collect_mm_slot(struct mm_slot *mm_slot)
2185 {
2186         struct mm_struct *mm = mm_slot->mm;
2187
2188         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2189
2190         if (khugepaged_test_exit(mm)) {
2191                 /* free mm_slot */
2192                 hlist_del(&mm_slot->hash);
2193                 list_del(&mm_slot->mm_node);
2194
2195                 /*
2196                  * Not strictly needed because the mm exited already.
2197                  *
2198                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2199                  */
2200
2201                 /* khugepaged_mm_lock actually not necessary for the below */
2202                 free_mm_slot(mm_slot);
2203                 mmdrop(mm);
2204         }
2205 }
2206
2207 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2208                                             struct page **hpage)
2209         __releases(&khugepaged_mm_lock)
2210         __acquires(&khugepaged_mm_lock)
2211 {
2212         struct mm_slot *mm_slot;
2213         struct mm_struct *mm;
2214         struct vm_area_struct *vma;
2215         int progress = 0;
2216
2217         VM_BUG_ON(!pages);
2218         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2219
2220         if (khugepaged_scan.mm_slot)
2221                 mm_slot = khugepaged_scan.mm_slot;
2222         else {
2223                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2224                                      struct mm_slot, mm_node);
2225                 khugepaged_scan.address = 0;
2226                 khugepaged_scan.mm_slot = mm_slot;
2227         }
2228         spin_unlock(&khugepaged_mm_lock);
2229
2230         mm = mm_slot->mm;
2231         down_read(&mm->mmap_sem);
2232         if (unlikely(khugepaged_test_exit(mm)))
2233                 vma = NULL;
2234         else
2235                 vma = find_vma(mm, khugepaged_scan.address);
2236
2237         progress++;
2238         for (; vma; vma = vma->vm_next) {
2239                 unsigned long hstart, hend;
2240
2241                 cond_resched();
2242                 if (unlikely(khugepaged_test_exit(mm))) {
2243                         progress++;
2244                         break;
2245                 }
2246
2247                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2248                      !khugepaged_always()) ||
2249                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2250                 skip:
2251                         progress++;
2252                         continue;
2253                 }
2254                 if (!vma->anon_vma || vma->vm_ops)
2255                         goto skip;
2256                 if (is_vma_temporary_stack(vma))
2257                         goto skip;
2258                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2259
2260                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2261                 hend = vma->vm_end & HPAGE_PMD_MASK;
2262                 if (hstart >= hend)
2263                         goto skip;
2264                 if (khugepaged_scan.address > hend)
2265                         goto skip;
2266                 if (khugepaged_scan.address < hstart)
2267                         khugepaged_scan.address = hstart;
2268                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2269
2270                 while (khugepaged_scan.address < hend) {
2271                         int ret;
2272                         cond_resched();
2273                         if (unlikely(khugepaged_test_exit(mm)))
2274                                 goto breakouterloop;
2275
2276                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2277                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2278                                   hend);
2279                         ret = khugepaged_scan_pmd(mm, vma,
2280                                                   khugepaged_scan.address,
2281                                                   hpage);
2282                         /* move to next address */
2283                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2284                         progress += HPAGE_PMD_NR;
2285                         if (ret)
2286                                 /* we released mmap_sem so break loop */
2287                                 goto breakouterloop_mmap_sem;
2288                         if (progress >= pages)
2289                                 goto breakouterloop;
2290                 }
2291         }
2292 breakouterloop:
2293         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2294 breakouterloop_mmap_sem:
2295
2296         spin_lock(&khugepaged_mm_lock);
2297         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2298         /*
2299          * Release the current mm_slot if this mm is about to die, or
2300          * if we scanned all vmas of this mm.
2301          */
2302         if (khugepaged_test_exit(mm) || !vma) {
2303                 /*
2304                  * Make sure that if mm_users is reaching zero while
2305                  * khugepaged runs here, khugepaged_exit will find
2306                  * mm_slot not pointing to the exiting mm.
2307                  */
2308                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2309                         khugepaged_scan.mm_slot = list_entry(
2310                                 mm_slot->mm_node.next,
2311                                 struct mm_slot, mm_node);
2312                         khugepaged_scan.address = 0;
2313                 } else {
2314                         khugepaged_scan.mm_slot = NULL;
2315                         khugepaged_full_scans++;
2316                 }
2317
2318                 collect_mm_slot(mm_slot);
2319         }
2320
2321         return progress;
2322 }
2323
2324 static int khugepaged_has_work(void)
2325 {
2326         return !list_empty(&khugepaged_scan.mm_head) &&
2327                 khugepaged_enabled();
2328 }
2329
2330 static int khugepaged_wait_event(void)
2331 {
2332         return !list_empty(&khugepaged_scan.mm_head) ||
2333                 kthread_should_stop();
2334 }
2335
2336 static void khugepaged_do_scan(void)
2337 {
2338         struct page *hpage = NULL;
2339         unsigned int progress = 0, pass_through_head = 0;
2340         unsigned int pages = khugepaged_pages_to_scan;
2341         bool wait = true;
2342
2343         barrier(); /* write khugepaged_pages_to_scan to local stack */
2344
2345         while (progress < pages) {
2346                 if (!khugepaged_prealloc_page(&hpage, &wait))
2347                         break;
2348
2349                 cond_resched();
2350
2351                 if (unlikely(kthread_should_stop() || freezing(current)))
2352                         break;
2353
2354                 spin_lock(&khugepaged_mm_lock);
2355                 if (!khugepaged_scan.mm_slot)
2356                         pass_through_head++;
2357                 if (khugepaged_has_work() &&
2358                     pass_through_head < 2)
2359                         progress += khugepaged_scan_mm_slot(pages - progress,
2360                                                             &hpage);
2361                 else
2362                         progress = pages;
2363                 spin_unlock(&khugepaged_mm_lock);
2364         }
2365
2366         if (!IS_ERR_OR_NULL(hpage))
2367                 put_page(hpage);
2368 }
2369
2370 static void khugepaged_wait_work(void)
2371 {
2372         try_to_freeze();
2373
2374         if (khugepaged_has_work()) {
2375                 if (!khugepaged_scan_sleep_millisecs)
2376                         return;
2377
2378                 wait_event_freezable_timeout(khugepaged_wait,
2379                                              kthread_should_stop(),
2380                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2381                 return;
2382         }
2383
2384         if (khugepaged_enabled())
2385                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2386 }
2387
2388 static int khugepaged(void *none)
2389 {
2390         struct mm_slot *mm_slot;
2391
2392         set_freezable();
2393         set_user_nice(current, 19);
2394
2395         while (!kthread_should_stop()) {
2396                 khugepaged_do_scan();
2397                 khugepaged_wait_work();
2398         }
2399
2400         spin_lock(&khugepaged_mm_lock);
2401         mm_slot = khugepaged_scan.mm_slot;
2402         khugepaged_scan.mm_slot = NULL;
2403         if (mm_slot)
2404                 collect_mm_slot(mm_slot);
2405         spin_unlock(&khugepaged_mm_lock);
2406         return 0;
2407 }
2408
2409 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2410 {
2411         struct page *page;
2412
2413         spin_lock(&mm->page_table_lock);
2414         if (unlikely(!pmd_trans_huge(*pmd))) {
2415                 spin_unlock(&mm->page_table_lock);
2416                 return;
2417         }
2418         page = pmd_page(*pmd);
2419         VM_BUG_ON(!page_count(page));
2420         get_page(page);
2421         spin_unlock(&mm->page_table_lock);
2422
2423         split_huge_page(page);
2424
2425         put_page(page);
2426         BUG_ON(pmd_trans_huge(*pmd));
2427 }
2428
2429 static void split_huge_page_address(struct mm_struct *mm,
2430                                     unsigned long address)
2431 {
2432         pgd_t *pgd;
2433         pud_t *pud;
2434         pmd_t *pmd;
2435
2436         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2437
2438         pgd = pgd_offset(mm, address);
2439         if (!pgd_present(*pgd))
2440                 return;
2441
2442         pud = pud_offset(pgd, address);
2443         if (!pud_present(*pud))
2444                 return;
2445
2446         pmd = pmd_offset(pud, address);
2447         if (!pmd_present(*pmd))
2448                 return;
2449         /*
2450          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2451          * materialize from under us.
2452          */
2453         split_huge_page_pmd(mm, pmd);
2454 }
2455
2456 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2457                              unsigned long start,
2458                              unsigned long end,
2459                              long adjust_next)
2460 {
2461         /*
2462          * If the new start address isn't hpage aligned and it could
2463          * previously contain an hugepage: check if we need to split
2464          * an huge pmd.
2465          */
2466         if (start & ~HPAGE_PMD_MASK &&
2467             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2468             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2469                 split_huge_page_address(vma->vm_mm, start);
2470
2471         /*
2472          * If the new end address isn't hpage aligned and it could
2473          * previously contain an hugepage: check if we need to split
2474          * an huge pmd.
2475          */
2476         if (end & ~HPAGE_PMD_MASK &&
2477             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2478             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2479                 split_huge_page_address(vma->vm_mm, end);
2480
2481         /*
2482          * If we're also updating the vma->vm_next->vm_start, if the new
2483          * vm_next->vm_start isn't page aligned and it could previously
2484          * contain an hugepage: check if we need to split an huge pmd.
2485          */
2486         if (adjust_next > 0) {
2487                 struct vm_area_struct *next = vma->vm_next;
2488                 unsigned long nstart = next->vm_start;
2489                 nstart += adjust_next << PAGE_SHIFT;
2490                 if (nstart & ~HPAGE_PMD_MASK &&
2491                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2492                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2493                         split_huge_page_address(next->vm_mm, nstart);
2494         }
2495 }