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