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         int current_nid = -1;
1030
1031         spin_lock(&mm->page_table_lock);
1032         if (unlikely(!pmd_same(pmd, *pmdp)))
1033                 goto out_unlock;
1034
1035         page = pmd_page(pmd);
1036         get_page(page);
1037         spin_unlock(&mm->page_table_lock);
1038         current_nid = page_to_nid(page);
1039         count_vm_numa_event(NUMA_HINT_FAULTS);
1040         if (current_nid == numa_node_id())
1041                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1042
1043         target_nid = mpol_misplaced(page, vma, haddr);
1044         if (target_nid == -1)
1045                 goto clear_pmdnuma;
1046
1047         /*
1048          * Due to lacking code to migrate thp pages, we'll split
1049          * (which preserves the special PROT_NONE) and re-take the
1050          * fault on the normal pages.
1051          */
1052         split_huge_page(page);
1053         put_page(page);
1054
1055         return 0;
1056
1057 clear_pmdnuma:
1058         spin_lock(&mm->page_table_lock);
1059         if (unlikely(!pmd_same(pmd, *pmdp)))
1060                 goto out_unlock;
1061
1062         pmd = pmd_mknonnuma(pmd);
1063         set_pmd_at(mm, haddr, pmdp, pmd);
1064         VM_BUG_ON(pmd_numa(*pmdp));
1065         update_mmu_cache_pmd(vma, addr, pmdp);
1066
1067 out_unlock:
1068         spin_unlock(&mm->page_table_lock);
1069         if (page) {
1070                 put_page(page);
1071                 task_numa_fault(numa_node_id(), HPAGE_PMD_NR, false);
1072         }
1073         return 0;
1074 }
1075
1076 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1077                  pmd_t *pmd, unsigned long addr)
1078 {
1079         int ret = 0;
1080
1081         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1082                 struct page *page;
1083                 pgtable_t pgtable;
1084                 pmd_t orig_pmd;
1085                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1086                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1087                 page = pmd_page(orig_pmd);
1088                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1089                 page_remove_rmap(page);
1090                 VM_BUG_ON(page_mapcount(page) < 0);
1091                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1092                 VM_BUG_ON(!PageHead(page));
1093                 tlb->mm->nr_ptes--;
1094                 spin_unlock(&tlb->mm->page_table_lock);
1095                 tlb_remove_page(tlb, page);
1096                 pte_free(tlb->mm, pgtable);
1097                 ret = 1;
1098         }
1099         return ret;
1100 }
1101
1102 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1103                 unsigned long addr, unsigned long end,
1104                 unsigned char *vec)
1105 {
1106         int ret = 0;
1107
1108         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1109                 /*
1110                  * All logical pages in the range are present
1111                  * if backed by a huge page.
1112                  */
1113                 spin_unlock(&vma->vm_mm->page_table_lock);
1114                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1115                 ret = 1;
1116         }
1117
1118         return ret;
1119 }
1120
1121 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1122                   unsigned long old_addr,
1123                   unsigned long new_addr, unsigned long old_end,
1124                   pmd_t *old_pmd, pmd_t *new_pmd)
1125 {
1126         int ret = 0;
1127         pmd_t pmd;
1128
1129         struct mm_struct *mm = vma->vm_mm;
1130
1131         if ((old_addr & ~HPAGE_PMD_MASK) ||
1132             (new_addr & ~HPAGE_PMD_MASK) ||
1133             old_end - old_addr < HPAGE_PMD_SIZE ||
1134             (new_vma->vm_flags & VM_NOHUGEPAGE))
1135                 goto out;
1136
1137         /*
1138          * The destination pmd shouldn't be established, free_pgtables()
1139          * should have release it.
1140          */
1141         if (WARN_ON(!pmd_none(*new_pmd))) {
1142                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1143                 goto out;
1144         }
1145
1146         ret = __pmd_trans_huge_lock(old_pmd, vma);
1147         if (ret == 1) {
1148                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1149                 VM_BUG_ON(!pmd_none(*new_pmd));
1150                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1151                 spin_unlock(&mm->page_table_lock);
1152         }
1153 out:
1154         return ret;
1155 }
1156
1157 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1158                 unsigned long addr, pgprot_t newprot, int prot_numa)
1159 {
1160         struct mm_struct *mm = vma->vm_mm;
1161         int ret = 0;
1162
1163         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1164                 pmd_t entry;
1165                 entry = pmdp_get_and_clear(mm, addr, pmd);
1166                 if (!prot_numa)
1167                         entry = pmd_modify(entry, newprot);
1168                 else {
1169                         struct page *page = pmd_page(*pmd);
1170
1171                         /* only check non-shared pages */
1172                         if (page_mapcount(page) == 1 &&
1173                             !pmd_numa(*pmd)) {
1174                                 entry = pmd_mknuma(entry);
1175                         }
1176                 }
1177                 set_pmd_at(mm, addr, pmd, entry);
1178                 spin_unlock(&vma->vm_mm->page_table_lock);
1179                 ret = 1;
1180         }
1181
1182         return ret;
1183 }
1184
1185 /*
1186  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1187  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1188  *
1189  * Note that if it returns 1, this routine returns without unlocking page
1190  * table locks. So callers must unlock them.
1191  */
1192 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1193 {
1194         spin_lock(&vma->vm_mm->page_table_lock);
1195         if (likely(pmd_trans_huge(*pmd))) {
1196                 if (unlikely(pmd_trans_splitting(*pmd))) {
1197                         spin_unlock(&vma->vm_mm->page_table_lock);
1198                         wait_split_huge_page(vma->anon_vma, pmd);
1199                         return -1;
1200                 } else {
1201                         /* Thp mapped by 'pmd' is stable, so we can
1202                          * handle it as it is. */
1203                         return 1;
1204                 }
1205         }
1206         spin_unlock(&vma->vm_mm->page_table_lock);
1207         return 0;
1208 }
1209
1210 pmd_t *page_check_address_pmd(struct page *page,
1211                               struct mm_struct *mm,
1212                               unsigned long address,
1213                               enum page_check_address_pmd_flag flag)
1214 {
1215         pgd_t *pgd;
1216         pud_t *pud;
1217         pmd_t *pmd, *ret = NULL;
1218
1219         if (address & ~HPAGE_PMD_MASK)
1220                 goto out;
1221
1222         pgd = pgd_offset(mm, address);
1223         if (!pgd_present(*pgd))
1224                 goto out;
1225
1226         pud = pud_offset(pgd, address);
1227         if (!pud_present(*pud))
1228                 goto out;
1229
1230         pmd = pmd_offset(pud, address);
1231         if (pmd_none(*pmd))
1232                 goto out;
1233         if (pmd_page(*pmd) != page)
1234                 goto out;
1235         /*
1236          * split_vma() may create temporary aliased mappings. There is
1237          * no risk as long as all huge pmd are found and have their
1238          * splitting bit set before __split_huge_page_refcount
1239          * runs. Finding the same huge pmd more than once during the
1240          * same rmap walk is not a problem.
1241          */
1242         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1243             pmd_trans_splitting(*pmd))
1244                 goto out;
1245         if (pmd_trans_huge(*pmd)) {
1246                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1247                           !pmd_trans_splitting(*pmd));
1248                 ret = pmd;
1249         }
1250 out:
1251         return ret;
1252 }
1253
1254 static int __split_huge_page_splitting(struct page *page,
1255                                        struct vm_area_struct *vma,
1256                                        unsigned long address)
1257 {
1258         struct mm_struct *mm = vma->vm_mm;
1259         pmd_t *pmd;
1260         int ret = 0;
1261         /* For mmu_notifiers */
1262         const unsigned long mmun_start = address;
1263         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1264
1265         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1266         spin_lock(&mm->page_table_lock);
1267         pmd = page_check_address_pmd(page, mm, address,
1268                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1269         if (pmd) {
1270                 /*
1271                  * We can't temporarily set the pmd to null in order
1272                  * to split it, the pmd must remain marked huge at all
1273                  * times or the VM won't take the pmd_trans_huge paths
1274                  * and it won't wait on the anon_vma->root->mutex to
1275                  * serialize against split_huge_page*.
1276                  */
1277                 pmdp_splitting_flush(vma, address, pmd);
1278                 ret = 1;
1279         }
1280         spin_unlock(&mm->page_table_lock);
1281         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1282
1283         return ret;
1284 }
1285
1286 static void __split_huge_page_refcount(struct page *page)
1287 {
1288         int i;
1289         struct zone *zone = page_zone(page);
1290         struct lruvec *lruvec;
1291         int tail_count = 0;
1292
1293         /* prevent PageLRU to go away from under us, and freeze lru stats */
1294         spin_lock_irq(&zone->lru_lock);
1295         lruvec = mem_cgroup_page_lruvec(page, zone);
1296
1297         compound_lock(page);
1298         /* complete memcg works before add pages to LRU */
1299         mem_cgroup_split_huge_fixup(page);
1300
1301         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1302                 struct page *page_tail = page + i;
1303
1304                 /* tail_page->_mapcount cannot change */
1305                 BUG_ON(page_mapcount(page_tail) < 0);
1306                 tail_count += page_mapcount(page_tail);
1307                 /* check for overflow */
1308                 BUG_ON(tail_count < 0);
1309                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1310                 /*
1311                  * tail_page->_count is zero and not changing from
1312                  * under us. But get_page_unless_zero() may be running
1313                  * from under us on the tail_page. If we used
1314                  * atomic_set() below instead of atomic_add(), we
1315                  * would then run atomic_set() concurrently with
1316                  * get_page_unless_zero(), and atomic_set() is
1317                  * implemented in C not using locked ops. spin_unlock
1318                  * on x86 sometime uses locked ops because of PPro
1319                  * errata 66, 92, so unless somebody can guarantee
1320                  * atomic_set() here would be safe on all archs (and
1321                  * not only on x86), it's safer to use atomic_add().
1322                  */
1323                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1324                            &page_tail->_count);
1325
1326                 /* after clearing PageTail the gup refcount can be released */
1327                 smp_mb();
1328
1329                 /*
1330                  * retain hwpoison flag of the poisoned tail page:
1331                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1332                  *   by the memory-failure.
1333                  */
1334                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1335                 page_tail->flags |= (page->flags &
1336                                      ((1L << PG_referenced) |
1337                                       (1L << PG_swapbacked) |
1338                                       (1L << PG_mlocked) |
1339                                       (1L << PG_uptodate)));
1340                 page_tail->flags |= (1L << PG_dirty);
1341
1342                 /* clear PageTail before overwriting first_page */
1343                 smp_wmb();
1344
1345                 /*
1346                  * __split_huge_page_splitting() already set the
1347                  * splitting bit in all pmd that could map this
1348                  * hugepage, that will ensure no CPU can alter the
1349                  * mapcount on the head page. The mapcount is only
1350                  * accounted in the head page and it has to be
1351                  * transferred to all tail pages in the below code. So
1352                  * for this code to be safe, the split the mapcount
1353                  * can't change. But that doesn't mean userland can't
1354                  * keep changing and reading the page contents while
1355                  * we transfer the mapcount, so the pmd splitting
1356                  * status is achieved setting a reserved bit in the
1357                  * pmd, not by clearing the present bit.
1358                 */
1359                 page_tail->_mapcount = page->_mapcount;
1360
1361                 BUG_ON(page_tail->mapping);
1362                 page_tail->mapping = page->mapping;
1363
1364                 page_tail->index = page->index + i;
1365                 page_xchg_last_nid(page_tail, page_last_nid(page));
1366
1367                 BUG_ON(!PageAnon(page_tail));
1368                 BUG_ON(!PageUptodate(page_tail));
1369                 BUG_ON(!PageDirty(page_tail));
1370                 BUG_ON(!PageSwapBacked(page_tail));
1371
1372                 lru_add_page_tail(page, page_tail, lruvec);
1373         }
1374         atomic_sub(tail_count, &page->_count);
1375         BUG_ON(atomic_read(&page->_count) <= 0);
1376
1377         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1378         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1379
1380         ClearPageCompound(page);
1381         compound_unlock(page);
1382         spin_unlock_irq(&zone->lru_lock);
1383
1384         for (i = 1; i < HPAGE_PMD_NR; i++) {
1385                 struct page *page_tail = page + i;
1386                 BUG_ON(page_count(page_tail) <= 0);
1387                 /*
1388                  * Tail pages may be freed if there wasn't any mapping
1389                  * like if add_to_swap() is running on a lru page that
1390                  * had its mapping zapped. And freeing these pages
1391                  * requires taking the lru_lock so we do the put_page
1392                  * of the tail pages after the split is complete.
1393                  */
1394                 put_page(page_tail);
1395         }
1396
1397         /*
1398          * Only the head page (now become a regular page) is required
1399          * to be pinned by the caller.
1400          */
1401         BUG_ON(page_count(page) <= 0);
1402 }
1403
1404 static int __split_huge_page_map(struct page *page,
1405                                  struct vm_area_struct *vma,
1406                                  unsigned long address)
1407 {
1408         struct mm_struct *mm = vma->vm_mm;
1409         pmd_t *pmd, _pmd;
1410         int ret = 0, i;
1411         pgtable_t pgtable;
1412         unsigned long haddr;
1413
1414         spin_lock(&mm->page_table_lock);
1415         pmd = page_check_address_pmd(page, mm, address,
1416                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1417         if (pmd) {
1418                 pgtable = pgtable_trans_huge_withdraw(mm);
1419                 pmd_populate(mm, &_pmd, pgtable);
1420
1421                 haddr = address;
1422                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1423                         pte_t *pte, entry;
1424                         BUG_ON(PageCompound(page+i));
1425                         entry = mk_pte(page + i, vma->vm_page_prot);
1426                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1427                         if (!pmd_write(*pmd))
1428                                 entry = pte_wrprotect(entry);
1429                         else
1430                                 BUG_ON(page_mapcount(page) != 1);
1431                         if (!pmd_young(*pmd))
1432                                 entry = pte_mkold(entry);
1433                         if (pmd_numa(*pmd))
1434                                 entry = pte_mknuma(entry);
1435                         pte = pte_offset_map(&_pmd, haddr);
1436                         BUG_ON(!pte_none(*pte));
1437                         set_pte_at(mm, haddr, pte, entry);
1438                         pte_unmap(pte);
1439                 }
1440
1441                 smp_wmb(); /* make pte visible before pmd */
1442                 /*
1443                  * Up to this point the pmd is present and huge and
1444                  * userland has the whole access to the hugepage
1445                  * during the split (which happens in place). If we
1446                  * overwrite the pmd with the not-huge version
1447                  * pointing to the pte here (which of course we could
1448                  * if all CPUs were bug free), userland could trigger
1449                  * a small page size TLB miss on the small sized TLB
1450                  * while the hugepage TLB entry is still established
1451                  * in the huge TLB. Some CPU doesn't like that. See
1452                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1453                  * Erratum 383 on page 93. Intel should be safe but is
1454                  * also warns that it's only safe if the permission
1455                  * and cache attributes of the two entries loaded in
1456                  * the two TLB is identical (which should be the case
1457                  * here). But it is generally safer to never allow
1458                  * small and huge TLB entries for the same virtual
1459                  * address to be loaded simultaneously. So instead of
1460                  * doing "pmd_populate(); flush_tlb_range();" we first
1461                  * mark the current pmd notpresent (atomically because
1462                  * here the pmd_trans_huge and pmd_trans_splitting
1463                  * must remain set at all times on the pmd until the
1464                  * split is complete for this pmd), then we flush the
1465                  * SMP TLB and finally we write the non-huge version
1466                  * of the pmd entry with pmd_populate.
1467                  */
1468                 pmdp_invalidate(vma, address, pmd);
1469                 pmd_populate(mm, pmd, pgtable);
1470                 ret = 1;
1471         }
1472         spin_unlock(&mm->page_table_lock);
1473
1474         return ret;
1475 }
1476
1477 /* must be called with anon_vma->root->mutex hold */
1478 static void __split_huge_page(struct page *page,
1479                               struct anon_vma *anon_vma)
1480 {
1481         int mapcount, mapcount2;
1482         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1483         struct anon_vma_chain *avc;
1484
1485         BUG_ON(!PageHead(page));
1486         BUG_ON(PageTail(page));
1487
1488         mapcount = 0;
1489         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1490                 struct vm_area_struct *vma = avc->vma;
1491                 unsigned long addr = vma_address(page, vma);
1492                 BUG_ON(is_vma_temporary_stack(vma));
1493                 mapcount += __split_huge_page_splitting(page, vma, addr);
1494         }
1495         /*
1496          * It is critical that new vmas are added to the tail of the
1497          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1498          * and establishes a child pmd before
1499          * __split_huge_page_splitting() freezes the parent pmd (so if
1500          * we fail to prevent copy_huge_pmd() from running until the
1501          * whole __split_huge_page() is complete), we will still see
1502          * the newly established pmd of the child later during the
1503          * walk, to be able to set it as pmd_trans_splitting too.
1504          */
1505         if (mapcount != page_mapcount(page))
1506                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1507                        mapcount, page_mapcount(page));
1508         BUG_ON(mapcount != page_mapcount(page));
1509
1510         __split_huge_page_refcount(page);
1511
1512         mapcount2 = 0;
1513         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1514                 struct vm_area_struct *vma = avc->vma;
1515                 unsigned long addr = vma_address(page, vma);
1516                 BUG_ON(is_vma_temporary_stack(vma));
1517                 mapcount2 += __split_huge_page_map(page, vma, addr);
1518         }
1519         if (mapcount != mapcount2)
1520                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1521                        mapcount, mapcount2, page_mapcount(page));
1522         BUG_ON(mapcount != mapcount2);
1523 }
1524
1525 int split_huge_page(struct page *page)
1526 {
1527         struct anon_vma *anon_vma;
1528         int ret = 1;
1529
1530         BUG_ON(!PageAnon(page));
1531         anon_vma = page_lock_anon_vma(page);
1532         if (!anon_vma)
1533                 goto out;
1534         ret = 0;
1535         if (!PageCompound(page))
1536                 goto out_unlock;
1537
1538         BUG_ON(!PageSwapBacked(page));
1539         __split_huge_page(page, anon_vma);
1540         count_vm_event(THP_SPLIT);
1541
1542         BUG_ON(PageCompound(page));
1543 out_unlock:
1544         page_unlock_anon_vma(anon_vma);
1545 out:
1546         return ret;
1547 }
1548
1549 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1550
1551 int hugepage_madvise(struct vm_area_struct *vma,
1552                      unsigned long *vm_flags, int advice)
1553 {
1554         struct mm_struct *mm = vma->vm_mm;
1555
1556         switch (advice) {
1557         case MADV_HUGEPAGE:
1558                 /*
1559                  * Be somewhat over-protective like KSM for now!
1560                  */
1561                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1562                         return -EINVAL;
1563                 if (mm->def_flags & VM_NOHUGEPAGE)
1564                         return -EINVAL;
1565                 *vm_flags &= ~VM_NOHUGEPAGE;
1566                 *vm_flags |= VM_HUGEPAGE;
1567                 /*
1568                  * If the vma become good for khugepaged to scan,
1569                  * register it here without waiting a page fault that
1570                  * may not happen any time soon.
1571                  */
1572                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1573                         return -ENOMEM;
1574                 break;
1575         case MADV_NOHUGEPAGE:
1576                 /*
1577                  * Be somewhat over-protective like KSM for now!
1578                  */
1579                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1580                         return -EINVAL;
1581                 *vm_flags &= ~VM_HUGEPAGE;
1582                 *vm_flags |= VM_NOHUGEPAGE;
1583                 /*
1584                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1585                  * this vma even if we leave the mm registered in khugepaged if
1586                  * it got registered before VM_NOHUGEPAGE was set.
1587                  */
1588                 break;
1589         }
1590
1591         return 0;
1592 }
1593
1594 static int __init khugepaged_slab_init(void)
1595 {
1596         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1597                                           sizeof(struct mm_slot),
1598                                           __alignof__(struct mm_slot), 0, NULL);
1599         if (!mm_slot_cache)
1600                 return -ENOMEM;
1601
1602         return 0;
1603 }
1604
1605 static void __init khugepaged_slab_free(void)
1606 {
1607         kmem_cache_destroy(mm_slot_cache);
1608         mm_slot_cache = NULL;
1609 }
1610
1611 static inline struct mm_slot *alloc_mm_slot(void)
1612 {
1613         if (!mm_slot_cache)     /* initialization failed */
1614                 return NULL;
1615         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1616 }
1617
1618 static inline void free_mm_slot(struct mm_slot *mm_slot)
1619 {
1620         kmem_cache_free(mm_slot_cache, mm_slot);
1621 }
1622
1623 static int __init mm_slots_hash_init(void)
1624 {
1625         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1626                                 GFP_KERNEL);
1627         if (!mm_slots_hash)
1628                 return -ENOMEM;
1629         return 0;
1630 }
1631
1632 #if 0
1633 static void __init mm_slots_hash_free(void)
1634 {
1635         kfree(mm_slots_hash);
1636         mm_slots_hash = NULL;
1637 }
1638 #endif
1639
1640 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1641 {
1642         struct mm_slot *mm_slot;
1643         struct hlist_head *bucket;
1644         struct hlist_node *node;
1645
1646         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1647                                 % MM_SLOTS_HASH_HEADS];
1648         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1649                 if (mm == mm_slot->mm)
1650                         return mm_slot;
1651         }
1652         return NULL;
1653 }
1654
1655 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1656                                     struct mm_slot *mm_slot)
1657 {
1658         struct hlist_head *bucket;
1659
1660         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1661                                 % MM_SLOTS_HASH_HEADS];
1662         mm_slot->mm = mm;
1663         hlist_add_head(&mm_slot->hash, bucket);
1664 }
1665
1666 static inline int khugepaged_test_exit(struct mm_struct *mm)
1667 {
1668         return atomic_read(&mm->mm_users) == 0;
1669 }
1670
1671 int __khugepaged_enter(struct mm_struct *mm)
1672 {
1673         struct mm_slot *mm_slot;
1674         int wakeup;
1675
1676         mm_slot = alloc_mm_slot();
1677         if (!mm_slot)
1678                 return -ENOMEM;
1679
1680         /* __khugepaged_exit() must not run from under us */
1681         VM_BUG_ON(khugepaged_test_exit(mm));
1682         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1683                 free_mm_slot(mm_slot);
1684                 return 0;
1685         }
1686
1687         spin_lock(&khugepaged_mm_lock);
1688         insert_to_mm_slots_hash(mm, mm_slot);
1689         /*
1690          * Insert just behind the scanning cursor, to let the area settle
1691          * down a little.
1692          */
1693         wakeup = list_empty(&khugepaged_scan.mm_head);
1694         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1695         spin_unlock(&khugepaged_mm_lock);
1696
1697         atomic_inc(&mm->mm_count);
1698         if (wakeup)
1699                 wake_up_interruptible(&khugepaged_wait);
1700
1701         return 0;
1702 }
1703
1704 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1705 {
1706         unsigned long hstart, hend;
1707         if (!vma->anon_vma)
1708                 /*
1709                  * Not yet faulted in so we will register later in the
1710                  * page fault if needed.
1711                  */
1712                 return 0;
1713         if (vma->vm_ops)
1714                 /* khugepaged not yet working on file or special mappings */
1715                 return 0;
1716         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1717         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1718         hend = vma->vm_end & HPAGE_PMD_MASK;
1719         if (hstart < hend)
1720                 return khugepaged_enter(vma);
1721         return 0;
1722 }
1723
1724 void __khugepaged_exit(struct mm_struct *mm)
1725 {
1726         struct mm_slot *mm_slot;
1727         int free = 0;
1728
1729         spin_lock(&khugepaged_mm_lock);
1730         mm_slot = get_mm_slot(mm);
1731         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1732                 hlist_del(&mm_slot->hash);
1733                 list_del(&mm_slot->mm_node);
1734                 free = 1;
1735         }
1736         spin_unlock(&khugepaged_mm_lock);
1737
1738         if (free) {
1739                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1740                 free_mm_slot(mm_slot);
1741                 mmdrop(mm);
1742         } else if (mm_slot) {
1743                 /*
1744                  * This is required to serialize against
1745                  * khugepaged_test_exit() (which is guaranteed to run
1746                  * under mmap sem read mode). Stop here (after we
1747                  * return all pagetables will be destroyed) until
1748                  * khugepaged has finished working on the pagetables
1749                  * under the mmap_sem.
1750                  */
1751                 down_write(&mm->mmap_sem);
1752                 up_write(&mm->mmap_sem);
1753         }
1754 }
1755
1756 static void release_pte_page(struct page *page)
1757 {
1758         /* 0 stands for page_is_file_cache(page) == false */
1759         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1760         unlock_page(page);
1761         putback_lru_page(page);
1762 }
1763
1764 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1765 {
1766         while (--_pte >= pte) {
1767                 pte_t pteval = *_pte;
1768                 if (!pte_none(pteval))
1769                         release_pte_page(pte_page(pteval));
1770         }
1771 }
1772
1773 static void release_all_pte_pages(pte_t *pte)
1774 {
1775         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1776 }
1777
1778 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1779                                         unsigned long address,
1780                                         pte_t *pte)
1781 {
1782         struct page *page;
1783         pte_t *_pte;
1784         int referenced = 0, isolated = 0, none = 0;
1785         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1786              _pte++, address += PAGE_SIZE) {
1787                 pte_t pteval = *_pte;
1788                 if (pte_none(pteval)) {
1789                         if (++none <= khugepaged_max_ptes_none)
1790                                 continue;
1791                         else {
1792                                 release_pte_pages(pte, _pte);
1793                                 goto out;
1794                         }
1795                 }
1796                 if (!pte_present(pteval) || !pte_write(pteval)) {
1797                         release_pte_pages(pte, _pte);
1798                         goto out;
1799                 }
1800                 page = vm_normal_page(vma, address, pteval);
1801                 if (unlikely(!page)) {
1802                         release_pte_pages(pte, _pte);
1803                         goto out;
1804                 }
1805                 VM_BUG_ON(PageCompound(page));
1806                 BUG_ON(!PageAnon(page));
1807                 VM_BUG_ON(!PageSwapBacked(page));
1808
1809                 /* cannot use mapcount: can't collapse if there's a gup pin */
1810                 if (page_count(page) != 1) {
1811                         release_pte_pages(pte, _pte);
1812                         goto out;
1813                 }
1814                 /*
1815                  * We can do it before isolate_lru_page because the
1816                  * page can't be freed from under us. NOTE: PG_lock
1817                  * is needed to serialize against split_huge_page
1818                  * when invoked from the VM.
1819                  */
1820                 if (!trylock_page(page)) {
1821                         release_pte_pages(pte, _pte);
1822                         goto out;
1823                 }
1824                 /*
1825                  * Isolate the page to avoid collapsing an hugepage
1826                  * currently in use by the VM.
1827                  */
1828                 if (isolate_lru_page(page)) {
1829                         unlock_page(page);
1830                         release_pte_pages(pte, _pte);
1831                         goto out;
1832                 }
1833                 /* 0 stands for page_is_file_cache(page) == false */
1834                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1835                 VM_BUG_ON(!PageLocked(page));
1836                 VM_BUG_ON(PageLRU(page));
1837
1838                 /* If there is no mapped pte young don't collapse the page */
1839                 if (pte_young(pteval) || PageReferenced(page) ||
1840                     mmu_notifier_test_young(vma->vm_mm, address))
1841                         referenced = 1;
1842         }
1843         if (unlikely(!referenced))
1844                 release_all_pte_pages(pte);
1845         else
1846                 isolated = 1;
1847 out:
1848         return isolated;
1849 }
1850
1851 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1852                                       struct vm_area_struct *vma,
1853                                       unsigned long address,
1854                                       spinlock_t *ptl)
1855 {
1856         pte_t *_pte;
1857         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1858                 pte_t pteval = *_pte;
1859                 struct page *src_page;
1860
1861                 if (pte_none(pteval)) {
1862                         clear_user_highpage(page, address);
1863                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1864                 } else {
1865                         src_page = pte_page(pteval);
1866                         copy_user_highpage(page, src_page, address, vma);
1867                         VM_BUG_ON(page_mapcount(src_page) != 1);
1868                         release_pte_page(src_page);
1869                         /*
1870                          * ptl mostly unnecessary, but preempt has to
1871                          * be disabled to update the per-cpu stats
1872                          * inside page_remove_rmap().
1873                          */
1874                         spin_lock(ptl);
1875                         /*
1876                          * paravirt calls inside pte_clear here are
1877                          * superfluous.
1878                          */
1879                         pte_clear(vma->vm_mm, address, _pte);
1880                         page_remove_rmap(src_page);
1881                         spin_unlock(ptl);
1882                         free_page_and_swap_cache(src_page);
1883                 }
1884
1885                 address += PAGE_SIZE;
1886                 page++;
1887         }
1888 }
1889
1890 static void khugepaged_alloc_sleep(void)
1891 {
1892         wait_event_freezable_timeout(khugepaged_wait, false,
1893                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1894 }
1895
1896 #ifdef CONFIG_NUMA
1897 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1898 {
1899         if (IS_ERR(*hpage)) {
1900                 if (!*wait)
1901                         return false;
1902
1903                 *wait = false;
1904                 *hpage = NULL;
1905                 khugepaged_alloc_sleep();
1906         } else if (*hpage) {
1907                 put_page(*hpage);
1908                 *hpage = NULL;
1909         }
1910
1911         return true;
1912 }
1913
1914 static struct page
1915 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1916                        struct vm_area_struct *vma, unsigned long address,
1917                        int node)
1918 {
1919         VM_BUG_ON(*hpage);
1920         /*
1921          * Allocate the page while the vma is still valid and under
1922          * the mmap_sem read mode so there is no memory allocation
1923          * later when we take the mmap_sem in write mode. This is more
1924          * friendly behavior (OTOH it may actually hide bugs) to
1925          * filesystems in userland with daemons allocating memory in
1926          * the userland I/O paths.  Allocating memory with the
1927          * mmap_sem in read mode is good idea also to allow greater
1928          * scalability.
1929          */
1930         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1931                                       node, __GFP_OTHER_NODE);
1932
1933         /*
1934          * After allocating the hugepage, release the mmap_sem read lock in
1935          * preparation for taking it in write mode.
1936          */
1937         up_read(&mm->mmap_sem);
1938         if (unlikely(!*hpage)) {
1939                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1940                 *hpage = ERR_PTR(-ENOMEM);
1941                 return NULL;
1942         }
1943
1944         count_vm_event(THP_COLLAPSE_ALLOC);
1945         return *hpage;
1946 }
1947 #else
1948 static struct page *khugepaged_alloc_hugepage(bool *wait)
1949 {
1950         struct page *hpage;
1951
1952         do {
1953                 hpage = alloc_hugepage(khugepaged_defrag());
1954                 if (!hpage) {
1955                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1956                         if (!*wait)
1957                                 return NULL;
1958
1959                         *wait = false;
1960                         khugepaged_alloc_sleep();
1961                 } else
1962                         count_vm_event(THP_COLLAPSE_ALLOC);
1963         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1964
1965         return hpage;
1966 }
1967
1968 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1969 {
1970         if (!*hpage)
1971                 *hpage = khugepaged_alloc_hugepage(wait);
1972
1973         if (unlikely(!*hpage))
1974                 return false;
1975
1976         return true;
1977 }
1978
1979 static struct page
1980 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1981                        struct vm_area_struct *vma, unsigned long address,
1982                        int node)
1983 {
1984         up_read(&mm->mmap_sem);
1985         VM_BUG_ON(!*hpage);
1986         return  *hpage;
1987 }
1988 #endif
1989
1990 static void collapse_huge_page(struct mm_struct *mm,
1991                                    unsigned long address,
1992                                    struct page **hpage,
1993                                    struct vm_area_struct *vma,
1994                                    int node)
1995 {
1996         pgd_t *pgd;
1997         pud_t *pud;
1998         pmd_t *pmd, _pmd;
1999         pte_t *pte;
2000         pgtable_t pgtable;
2001         struct page *new_page;
2002         spinlock_t *ptl;
2003         int isolated;
2004         unsigned long hstart, hend;
2005         unsigned long mmun_start;       /* For mmu_notifiers */
2006         unsigned long mmun_end;         /* For mmu_notifiers */
2007
2008         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2009
2010         /* release the mmap_sem read lock. */
2011         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2012         if (!new_page)
2013                 return;
2014
2015         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2016                 return;
2017
2018         /*
2019          * Prevent all access to pagetables with the exception of
2020          * gup_fast later hanlded by the ptep_clear_flush and the VM
2021          * handled by the anon_vma lock + PG_lock.
2022          */
2023         down_write(&mm->mmap_sem);
2024         if (unlikely(khugepaged_test_exit(mm)))
2025                 goto out;
2026
2027         vma = find_vma(mm, address);
2028         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2029         hend = vma->vm_end & HPAGE_PMD_MASK;
2030         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2031                 goto out;
2032
2033         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2034             (vma->vm_flags & VM_NOHUGEPAGE))
2035                 goto out;
2036
2037         if (!vma->anon_vma || vma->vm_ops)
2038                 goto out;
2039         if (is_vma_temporary_stack(vma))
2040                 goto out;
2041         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2042
2043         pgd = pgd_offset(mm, address);
2044         if (!pgd_present(*pgd))
2045                 goto out;
2046
2047         pud = pud_offset(pgd, address);
2048         if (!pud_present(*pud))
2049                 goto out;
2050
2051         pmd = pmd_offset(pud, address);
2052         /* pmd can't go away or become huge under us */
2053         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2054                 goto out;
2055
2056         anon_vma_lock(vma->anon_vma);
2057
2058         pte = pte_offset_map(pmd, address);
2059         ptl = pte_lockptr(mm, pmd);
2060
2061         mmun_start = address;
2062         mmun_end   = address + HPAGE_PMD_SIZE;
2063         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2064         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2065         /*
2066          * After this gup_fast can't run anymore. This also removes
2067          * any huge TLB entry from the CPU so we won't allow
2068          * huge and small TLB entries for the same virtual address
2069          * to avoid the risk of CPU bugs in that area.
2070          */
2071         _pmd = pmdp_clear_flush(vma, address, pmd);
2072         spin_unlock(&mm->page_table_lock);
2073         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2074
2075         spin_lock(ptl);
2076         isolated = __collapse_huge_page_isolate(vma, address, pte);
2077         spin_unlock(ptl);
2078
2079         if (unlikely(!isolated)) {
2080                 pte_unmap(pte);
2081                 spin_lock(&mm->page_table_lock);
2082                 BUG_ON(!pmd_none(*pmd));
2083                 set_pmd_at(mm, address, pmd, _pmd);
2084                 spin_unlock(&mm->page_table_lock);
2085                 anon_vma_unlock(vma->anon_vma);
2086                 goto out;
2087         }
2088
2089         /*
2090          * All pages are isolated and locked so anon_vma rmap
2091          * can't run anymore.
2092          */
2093         anon_vma_unlock(vma->anon_vma);
2094
2095         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2096         pte_unmap(pte);
2097         __SetPageUptodate(new_page);
2098         pgtable = pmd_pgtable(_pmd);
2099
2100         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2101         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2102         _pmd = pmd_mkhuge(_pmd);
2103
2104         /*
2105          * spin_lock() below is not the equivalent of smp_wmb(), so
2106          * this is needed to avoid the copy_huge_page writes to become
2107          * visible after the set_pmd_at() write.
2108          */
2109         smp_wmb();
2110
2111         spin_lock(&mm->page_table_lock);
2112         BUG_ON(!pmd_none(*pmd));
2113         page_add_new_anon_rmap(new_page, vma, address);
2114         set_pmd_at(mm, address, pmd, _pmd);
2115         update_mmu_cache_pmd(vma, address, pmd);
2116         pgtable_trans_huge_deposit(mm, pgtable);
2117         spin_unlock(&mm->page_table_lock);
2118
2119         *hpage = NULL;
2120
2121         khugepaged_pages_collapsed++;
2122 out_up_write:
2123         up_write(&mm->mmap_sem);
2124         return;
2125
2126 out:
2127         mem_cgroup_uncharge_page(new_page);
2128         goto out_up_write;
2129 }
2130
2131 static int khugepaged_scan_pmd(struct mm_struct *mm,
2132                                struct vm_area_struct *vma,
2133                                unsigned long address,
2134                                struct page **hpage)
2135 {
2136         pgd_t *pgd;
2137         pud_t *pud;
2138         pmd_t *pmd;
2139         pte_t *pte, *_pte;
2140         int ret = 0, referenced = 0, none = 0;
2141         struct page *page;
2142         unsigned long _address;
2143         spinlock_t *ptl;
2144         int node = -1;
2145
2146         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2147
2148         pgd = pgd_offset(mm, address);
2149         if (!pgd_present(*pgd))
2150                 goto out;
2151
2152         pud = pud_offset(pgd, address);
2153         if (!pud_present(*pud))
2154                 goto out;
2155
2156         pmd = pmd_offset(pud, address);
2157         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2158                 goto out;
2159
2160         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2161         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2162              _pte++, _address += PAGE_SIZE) {
2163                 pte_t pteval = *_pte;
2164                 if (pte_none(pteval)) {
2165                         if (++none <= khugepaged_max_ptes_none)
2166                                 continue;
2167                         else
2168                                 goto out_unmap;
2169                 }
2170                 if (!pte_present(pteval) || !pte_write(pteval))
2171                         goto out_unmap;
2172                 page = vm_normal_page(vma, _address, pteval);
2173                 if (unlikely(!page))
2174                         goto out_unmap;
2175                 /*
2176                  * Chose the node of the first page. This could
2177                  * be more sophisticated and look at more pages,
2178                  * but isn't for now.
2179                  */
2180                 if (node == -1)
2181                         node = page_to_nid(page);
2182                 VM_BUG_ON(PageCompound(page));
2183                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2184                         goto out_unmap;
2185                 /* cannot use mapcount: can't collapse if there's a gup pin */
2186                 if (page_count(page) != 1)
2187                         goto out_unmap;
2188                 if (pte_young(pteval) || PageReferenced(page) ||
2189                     mmu_notifier_test_young(vma->vm_mm, address))
2190                         referenced = 1;
2191         }
2192         if (referenced)
2193                 ret = 1;
2194 out_unmap:
2195         pte_unmap_unlock(pte, ptl);
2196         if (ret)
2197                 /* collapse_huge_page will return with the mmap_sem released */
2198                 collapse_huge_page(mm, address, hpage, vma, node);
2199 out:
2200         return ret;
2201 }
2202
2203 static void collect_mm_slot(struct mm_slot *mm_slot)
2204 {
2205         struct mm_struct *mm = mm_slot->mm;
2206
2207         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2208
2209         if (khugepaged_test_exit(mm)) {
2210                 /* free mm_slot */
2211                 hlist_del(&mm_slot->hash);
2212                 list_del(&mm_slot->mm_node);
2213
2214                 /*
2215                  * Not strictly needed because the mm exited already.
2216                  *
2217                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2218                  */
2219
2220                 /* khugepaged_mm_lock actually not necessary for the below */
2221                 free_mm_slot(mm_slot);
2222                 mmdrop(mm);
2223         }
2224 }
2225
2226 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2227                                             struct page **hpage)
2228         __releases(&khugepaged_mm_lock)
2229         __acquires(&khugepaged_mm_lock)
2230 {
2231         struct mm_slot *mm_slot;
2232         struct mm_struct *mm;
2233         struct vm_area_struct *vma;
2234         int progress = 0;
2235
2236         VM_BUG_ON(!pages);
2237         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2238
2239         if (khugepaged_scan.mm_slot)
2240                 mm_slot = khugepaged_scan.mm_slot;
2241         else {
2242                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2243                                      struct mm_slot, mm_node);
2244                 khugepaged_scan.address = 0;
2245                 khugepaged_scan.mm_slot = mm_slot;
2246         }
2247         spin_unlock(&khugepaged_mm_lock);
2248
2249         mm = mm_slot->mm;
2250         down_read(&mm->mmap_sem);
2251         if (unlikely(khugepaged_test_exit(mm)))
2252                 vma = NULL;
2253         else
2254                 vma = find_vma(mm, khugepaged_scan.address);
2255
2256         progress++;
2257         for (; vma; vma = vma->vm_next) {
2258                 unsigned long hstart, hend;
2259
2260                 cond_resched();
2261                 if (unlikely(khugepaged_test_exit(mm))) {
2262                         progress++;
2263                         break;
2264                 }
2265
2266                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2267                      !khugepaged_always()) ||
2268                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2269                 skip:
2270                         progress++;
2271                         continue;
2272                 }
2273                 if (!vma->anon_vma || vma->vm_ops)
2274                         goto skip;
2275                 if (is_vma_temporary_stack(vma))
2276                         goto skip;
2277                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2278
2279                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2280                 hend = vma->vm_end & HPAGE_PMD_MASK;
2281                 if (hstart >= hend)
2282                         goto skip;
2283                 if (khugepaged_scan.address > hend)
2284                         goto skip;
2285                 if (khugepaged_scan.address < hstart)
2286                         khugepaged_scan.address = hstart;
2287                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2288
2289                 while (khugepaged_scan.address < hend) {
2290                         int ret;
2291                         cond_resched();
2292                         if (unlikely(khugepaged_test_exit(mm)))
2293                                 goto breakouterloop;
2294
2295                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2296                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2297                                   hend);
2298                         ret = khugepaged_scan_pmd(mm, vma,
2299                                                   khugepaged_scan.address,
2300                                                   hpage);
2301                         /* move to next address */
2302                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2303                         progress += HPAGE_PMD_NR;
2304                         if (ret)
2305                                 /* we released mmap_sem so break loop */
2306                                 goto breakouterloop_mmap_sem;
2307                         if (progress >= pages)
2308                                 goto breakouterloop;
2309                 }
2310         }
2311 breakouterloop:
2312         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2313 breakouterloop_mmap_sem:
2314
2315         spin_lock(&khugepaged_mm_lock);
2316         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2317         /*
2318          * Release the current mm_slot if this mm is about to die, or
2319          * if we scanned all vmas of this mm.
2320          */
2321         if (khugepaged_test_exit(mm) || !vma) {
2322                 /*
2323                  * Make sure that if mm_users is reaching zero while
2324                  * khugepaged runs here, khugepaged_exit will find
2325                  * mm_slot not pointing to the exiting mm.
2326                  */
2327                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2328                         khugepaged_scan.mm_slot = list_entry(
2329                                 mm_slot->mm_node.next,
2330                                 struct mm_slot, mm_node);
2331                         khugepaged_scan.address = 0;
2332                 } else {
2333                         khugepaged_scan.mm_slot = NULL;
2334                         khugepaged_full_scans++;
2335                 }
2336
2337                 collect_mm_slot(mm_slot);
2338         }
2339
2340         return progress;
2341 }
2342
2343 static int khugepaged_has_work(void)
2344 {
2345         return !list_empty(&khugepaged_scan.mm_head) &&
2346                 khugepaged_enabled();
2347 }
2348
2349 static int khugepaged_wait_event(void)
2350 {
2351         return !list_empty(&khugepaged_scan.mm_head) ||
2352                 kthread_should_stop();
2353 }
2354
2355 static void khugepaged_do_scan(void)
2356 {
2357         struct page *hpage = NULL;
2358         unsigned int progress = 0, pass_through_head = 0;
2359         unsigned int pages = khugepaged_pages_to_scan;
2360         bool wait = true;
2361
2362         barrier(); /* write khugepaged_pages_to_scan to local stack */
2363
2364         while (progress < pages) {
2365                 if (!khugepaged_prealloc_page(&hpage, &wait))
2366                         break;
2367
2368                 cond_resched();
2369
2370                 if (unlikely(kthread_should_stop() || freezing(current)))
2371                         break;
2372
2373                 spin_lock(&khugepaged_mm_lock);
2374                 if (!khugepaged_scan.mm_slot)
2375                         pass_through_head++;
2376                 if (khugepaged_has_work() &&
2377                     pass_through_head < 2)
2378                         progress += khugepaged_scan_mm_slot(pages - progress,
2379                                                             &hpage);
2380                 else
2381                         progress = pages;
2382                 spin_unlock(&khugepaged_mm_lock);
2383         }
2384
2385         if (!IS_ERR_OR_NULL(hpage))
2386                 put_page(hpage);
2387 }
2388
2389 static void khugepaged_wait_work(void)
2390 {
2391         try_to_freeze();
2392
2393         if (khugepaged_has_work()) {
2394                 if (!khugepaged_scan_sleep_millisecs)
2395                         return;
2396
2397                 wait_event_freezable_timeout(khugepaged_wait,
2398                                              kthread_should_stop(),
2399                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2400                 return;
2401         }
2402
2403         if (khugepaged_enabled())
2404                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2405 }
2406
2407 static int khugepaged(void *none)
2408 {
2409         struct mm_slot *mm_slot;
2410
2411         set_freezable();
2412         set_user_nice(current, 19);
2413
2414         while (!kthread_should_stop()) {
2415                 khugepaged_do_scan();
2416                 khugepaged_wait_work();
2417         }
2418
2419         spin_lock(&khugepaged_mm_lock);
2420         mm_slot = khugepaged_scan.mm_slot;
2421         khugepaged_scan.mm_slot = NULL;
2422         if (mm_slot)
2423                 collect_mm_slot(mm_slot);
2424         spin_unlock(&khugepaged_mm_lock);
2425         return 0;
2426 }
2427
2428 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2429 {
2430         struct page *page;
2431
2432         spin_lock(&mm->page_table_lock);
2433         if (unlikely(!pmd_trans_huge(*pmd))) {
2434                 spin_unlock(&mm->page_table_lock);
2435                 return;
2436         }
2437         page = pmd_page(*pmd);
2438         VM_BUG_ON(!page_count(page));
2439         get_page(page);
2440         spin_unlock(&mm->page_table_lock);
2441
2442         split_huge_page(page);
2443
2444         put_page(page);
2445         BUG_ON(pmd_trans_huge(*pmd));
2446 }
2447
2448 static void split_huge_page_address(struct mm_struct *mm,
2449                                     unsigned long address)
2450 {
2451         pgd_t *pgd;
2452         pud_t *pud;
2453         pmd_t *pmd;
2454
2455         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2456
2457         pgd = pgd_offset(mm, address);
2458         if (!pgd_present(*pgd))
2459                 return;
2460
2461         pud = pud_offset(pgd, address);
2462         if (!pud_present(*pud))
2463                 return;
2464
2465         pmd = pmd_offset(pud, address);
2466         if (!pmd_present(*pmd))
2467                 return;
2468         /*
2469          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2470          * materialize from under us.
2471          */
2472         split_huge_page_pmd(mm, pmd);
2473 }
2474
2475 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2476                              unsigned long start,
2477                              unsigned long end,
2478                              long adjust_next)
2479 {
2480         /*
2481          * If the new start address isn't hpage aligned and it could
2482          * previously contain an hugepage: check if we need to split
2483          * an huge pmd.
2484          */
2485         if (start & ~HPAGE_PMD_MASK &&
2486             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2487             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2488                 split_huge_page_address(vma->vm_mm, start);
2489
2490         /*
2491          * If the new end address isn't hpage aligned and it could
2492          * previously contain an hugepage: check if we need to split
2493          * an huge pmd.
2494          */
2495         if (end & ~HPAGE_PMD_MASK &&
2496             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2497             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2498                 split_huge_page_address(vma->vm_mm, end);
2499
2500         /*
2501          * If we're also updating the vma->vm_next->vm_start, if the new
2502          * vm_next->vm_start isn't page aligned and it could previously
2503          * contain an hugepage: check if we need to split an huge pmd.
2504          */
2505         if (adjust_next > 0) {
2506                 struct vm_area_struct *next = vma->vm_next;
2507                 unsigned long nstart = next->vm_start;
2508                 nstart += adjust_next << PAGE_SHIFT;
2509                 if (nstart & ~HPAGE_PMD_MASK &&
2510                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2511                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2512                         split_huge_page_address(next->vm_mm, nstart);
2513         }
2514 }