mm/huge_memory.c

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