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_numa(struct vm_area_struct *vma, pmd_t pmd)
 731 {
 732         /*
 733          * See pte_numa().
 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_numa_page(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         spin_unlock(&mm->page_table_lock);
 786
 787         lock_page(page);
 788         spin_lock(&mm->page_table_lock);
 789         if (unlikely(!pmd_same(*pmd, entry))) {
 790                 spin_unlock(&mm->page_table_lock);
 791                 unlock_page(page);
 792                 put_page(page);
 793                 return;
 794         }
 795         spin_unlock(&mm->page_table_lock);
 796
 797         new_page = alloc_pages_node(node,
 798             (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT,
 799             HPAGE_PMD_ORDER);
 800
 801         if (!new_page)
 802                 goto alloc_fail;
 803
 804         lru = PageLRU(page);
 805
 806         if (lru && isolate_lru_page(page)) /* does an implicit get_page() */
 807                 goto alloc_fail;
 808
 809         if (!trylock_page(new_page))
 810                 BUG();
 811
 812         /* anon mapping, we can simply copy page->mapping to the new page: */
 813         new_page->mapping = page->mapping;
 814         new_page->index = page->index;
 815
 816         migrate_page_copy(new_page, page);
 817
 818         WARN_ON(PageLRU(new_page));
 819
 820         spin_lock(&mm->page_table_lock);
 821         if (unlikely(!pmd_same(*pmd, entry))) {
 822                 spin_unlock(&mm->page_table_lock);
 823                 if (lru)
 824                         putback_lru_page(page);
 825
 826                 unlock_page(new_page);
 827                 ClearPageActive(new_page);      /* Set by migrate_page_copy() */
 828                 new_page->mapping = NULL;
 829                 put_page(new_page);             /* Free it */
 830
 831                 unlock_page(page);
 832                 put_page(page);                 /* Drop the local reference */
 833
 834                 return;
 835         }
 836
 837         entry = mk_pmd(new_page, vma->vm_page_prot);
 838         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 839         entry = pmd_mkhuge(entry);
 840
 841         page_add_new_anon_rmap(new_page, vma, haddr);
 842
 843         set_pmd_at(mm, haddr, pmd, entry);
 844         update_mmu_cache_pmd(vma, address, entry);
 845         page_remove_rmap(page);
 846         spin_unlock(&mm->page_table_lock);
 847
 848         put_page(page);                 /* Drop the rmap reference */
 849
 850         task_numa_fault(node, HPAGE_PMD_NR);
 851
 852         if (lru)
 853                 put_page(page);         /* drop the LRU isolation reference */
 854
 855         unlock_page(new_page);
 856         unlock_page(page);
 857         put_page(page);                 /* Drop the local reference */
 858
 859         return;
 860
 861 alloc_fail:
 862         if (new_page)
 863                 put_page(new_page);
 864
 865         unlock_page(page);
 866
 867         spin_lock(&mm->page_table_lock);
 868         if (unlikely(!pmd_same(*pmd, entry))) {
 869                 put_page(page);
 870                 page = NULL;
 871                 goto unlock;
 872         }
 873         goto fixup;
 874 }
 875
 876 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 877                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 878                   struct vm_area_struct *vma)
 879 {
 880         struct page *src_page;
 881         pmd_t pmd;
 882         pgtable_t pgtable;
 883         int ret;
 884
 885         ret = -ENOMEM;
 886         pgtable = pte_alloc_one(dst_mm, addr);
 887         if (unlikely(!pgtable))
 888                 goto out;
 889
 890         spin_lock(&dst_mm->page_table_lock);
 891         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 892
 893         ret = -EAGAIN;
 894         pmd = *src_pmd;
 895         if (unlikely(!pmd_trans_huge(pmd))) {
 896                 pte_free(dst_mm, pgtable);
 897                 goto out_unlock;
 898         }
 899         if (unlikely(pmd_trans_splitting(pmd))) {
 900                 /* split huge page running from under us */
 901                 spin_unlock(&src_mm->page_table_lock);
 902                 spin_unlock(&dst_mm->page_table_lock);
 903                 pte_free(dst_mm, pgtable);
 904
 905                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 906                 goto out;
 907         }
 908         src_page = pmd_page(pmd);
 909         VM_BUG_ON(!PageHead(src_page));
 910         get_page(src_page);
 911         page_dup_rmap(src_page);
 912         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 913
 914         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 915         pmd = pmd_mkold(pmd_wrprotect(pmd));
 916         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 917         pgtable_trans_huge_deposit(dst_mm, pgtable);
 918         dst_mm->nr_ptes++;
 919
 920         ret = 0;
 921 out_unlock:
 922         spin_unlock(&src_mm->page_table_lock);
 923         spin_unlock(&dst_mm->page_table_lock);
 924 out:
 925         return ret;
 926 }
 927
 928 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 929                                         struct vm_area_struct *vma,
 930                                         unsigned long address,
 931                                         pmd_t *pmd, pmd_t orig_pmd,
 932                                         struct page *page,
 933                                         unsigned long haddr)
 934 {
 935         pgtable_t pgtable;
 936         pmd_t _pmd;
 937         int ret = 0, i;
 938         struct page **pages;
 939         unsigned long mmun_start;       /* For mmu_notifiers */
 940         unsigned long mmun_end;         /* For mmu_notifiers */
 941
 942         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 943                         GFP_KERNEL);
 944         if (unlikely(!pages)) {
 945                 ret |= VM_FAULT_OOM;
 946                 goto out;
 947         }
 948
 949         for (i = 0; i < HPAGE_PMD_NR; i++) {
 950                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 951                                                __GFP_OTHER_NODE,
 952                                                vma, address, page_to_nid(page));
 953                 if (unlikely(!pages[i] ||
 954                              mem_cgroup_newpage_charge(pages[i], mm,
 955                                                        GFP_KERNEL))) {
 956                         if (pages[i])
 957                                 put_page(pages[i]);
 958                         mem_cgroup_uncharge_start();
 959                         while (--i >= 0) {
 960                                 mem_cgroup_uncharge_page(pages[i]);
 961                                 put_page(pages[i]);
 962                         }
 963                         mem_cgroup_uncharge_end();
 964                         kfree(pages);
 965                         ret |= VM_FAULT_OOM;
 966                         goto out;
 967                 }
 968         }
 969
 970         for (i = 0; i < HPAGE_PMD_NR; i++) {
 971                 copy_user_highpage(pages[i], page + i,
 972                                    haddr + PAGE_SIZE * i, vma);
 973                 __SetPageUptodate(pages[i]);
 974                 cond_resched();
 975         }
 976
 977         mmun_start = haddr;
 978         mmun_end   = haddr + HPAGE_PMD_SIZE;
 979         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 980
 981         spin_lock(&mm->page_table_lock);
 982         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 983                 goto out_free_pages;
 984         VM_BUG_ON(!PageHead(page));
 985
 986         pmdp_clear_flush(vma, haddr, pmd);
 987         /* leave pmd empty until pte is filled */
 988
 989         pgtable = pgtable_trans_huge_withdraw(mm);
 990         pmd_populate(mm, &_pmd, pgtable);
 991
 992         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 993                 pte_t *pte, entry;
 994                 entry = mk_pte(pages[i], vma->vm_page_prot);
 995                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 996                 page_add_new_anon_rmap(pages[i], vma, haddr);
 997                 pte = pte_offset_map(&_pmd, haddr);
 998                 VM_BUG_ON(!pte_none(*pte));
 999                 set_pte_at(mm, haddr, pte, entry);
1000                 pte_unmap(pte);
1001         }
1002         kfree(pages);
1003
1004         smp_wmb(); /* make pte visible before pmd */
1005         pmd_populate(mm, pmd, pgtable);
1006         page_remove_rmap(page);
1007         spin_unlock(&mm->page_table_lock);
1008
1009         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1010
1011         ret |= VM_FAULT_WRITE;
1012         put_page(page);
1013
1014 out:
1015         return ret;
1016
1017 out_free_pages:
1018         spin_unlock(&mm->page_table_lock);
1019         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1020         mem_cgroup_uncharge_start();
1021         for (i = 0; i < HPAGE_PMD_NR; i++) {
1022                 mem_cgroup_uncharge_page(pages[i]);
1023                 put_page(pages[i]);
1024         }
1025         mem_cgroup_uncharge_end();
1026         kfree(pages);
1027         goto out;
1028 }
1029
1030 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1031                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1032 {
1033         int ret = 0;
1034         struct page *page, *new_page;
1035         unsigned long haddr;
1036         unsigned long mmun_start;       /* For mmu_notifiers */
1037         unsigned long mmun_end;         /* For mmu_notifiers */
1038
1039         VM_BUG_ON(!vma->anon_vma);
1040         spin_lock(&mm->page_table_lock);
1041         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1042                 goto out_unlock;
1043
1044         page = pmd_page(orig_pmd);
1045         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1046         haddr = address & HPAGE_PMD_MASK;
1047         if (page_mapcount(page) == 1) {
1048                 pmd_t entry;
1049                 entry = pmd_mkyoung(orig_pmd);
1050                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1051                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1052                         update_mmu_cache_pmd(vma, address, pmd);
1053                 ret |= VM_FAULT_WRITE;
1054                 goto out_unlock;
1055         }
1056         get_page(page);
1057         spin_unlock(&mm->page_table_lock);
1058
1059         if (transparent_hugepage_enabled(vma) &&
1060             !transparent_hugepage_debug_cow())
1061                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1062                                               vma, haddr, numa_node_id(), 0);
1063         else
1064                 new_page = NULL;
1065
1066         if (unlikely(!new_page)) {
1067                 count_vm_event(THP_FAULT_FALLBACK);
1068                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1069                                                    pmd, orig_pmd, page, haddr);
1070                 if (ret & VM_FAULT_OOM)
1071                         split_huge_page(page);
1072                 put_page(page);
1073                 goto out;
1074         }
1075         count_vm_event(THP_FAULT_ALLOC);
1076
1077         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1078                 put_page(new_page);
1079                 split_huge_page(page);
1080                 put_page(page);
1081                 ret |= VM_FAULT_OOM;
1082                 goto out;
1083         }
1084
1085         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1086         __SetPageUptodate(new_page);
1087
1088         mmun_start = haddr;
1089         mmun_end   = haddr + HPAGE_PMD_SIZE;
1090         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1091
1092         spin_lock(&mm->page_table_lock);
1093         put_page(page);
1094         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1095                 spin_unlock(&mm->page_table_lock);
1096                 mem_cgroup_uncharge_page(new_page);
1097                 put_page(new_page);
1098                 goto out_mn;
1099         } else {
1100                 pmd_t entry;
1101                 VM_BUG_ON(!PageHead(page));
1102                 entry = mk_pmd(new_page, vma->vm_page_prot);
1103                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1104                 entry = pmd_mkhuge(entry);
1105                 pmdp_clear_flush(vma, haddr, pmd);
1106                 page_add_new_anon_rmap(new_page, vma, haddr);
1107                 set_pmd_at(mm, haddr, pmd, entry);
1108                 update_mmu_cache_pmd(vma, address, pmd);
1109                 page_remove_rmap(page);
1110                 put_page(page);
1111                 ret |= VM_FAULT_WRITE;
1112         }
1113         spin_unlock(&mm->page_table_lock);
1114 out_mn:
1115         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1116 out:
1117         return ret;
1118 out_unlock:
1119         spin_unlock(&mm->page_table_lock);
1120         return ret;
1121 }
1122
1123 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1124                                    unsigned long addr,
1125                                    pmd_t *pmd,
1126                                    unsigned int flags)
1127 {
1128         struct mm_struct *mm = vma->vm_mm;
1129         struct page *page = NULL;
1130
1131         assert_spin_locked(&mm->page_table_lock);
1132
1133         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1134                 goto out;
1135
1136         page = pmd_page(*pmd);
1137         VM_BUG_ON(!PageHead(page));
1138         if (flags & FOLL_TOUCH) {
1139                 pmd_t _pmd;
1140                 /*
1141                  * We should set the dirty bit only for FOLL_WRITE but
1142                  * for now the dirty bit in the pmd is meaningless.
1143                  * And if the dirty bit will become meaningful and
1144                  * we'll only set it with FOLL_WRITE, an atomic
1145                  * set_bit will be required on the pmd to set the
1146                  * young bit, instead of the current set_pmd_at.
1147                  */
1148                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1149                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1150         }
1151         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1152                 if (page->mapping && trylock_page(page)) {
1153                         lru_add_drain();
1154                         if (page->mapping)
1155                                 mlock_vma_page(page);
1156                         unlock_page(page);
1157                 }
1158         }
1159         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1160         VM_BUG_ON(!PageCompound(page));
1161         if (flags & FOLL_GET)
1162                 get_page_foll(page);
1163
1164 out:
1165         return page;
1166 }
1167
1168 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1169                  pmd_t *pmd, unsigned long addr)
1170 {
1171         int ret = 0;
1172
1173         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1174                 struct page *page;
1175                 pgtable_t pgtable;
1176                 pmd_t orig_pmd;
1177                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1178                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1179                 page = pmd_page(orig_pmd);
1180                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1181                 page_remove_rmap(page);
1182                 VM_BUG_ON(page_mapcount(page) < 0);
1183                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1184                 VM_BUG_ON(!PageHead(page));
1185                 tlb->mm->nr_ptes--;
1186                 spin_unlock(&tlb->mm->page_table_lock);
1187                 tlb_remove_page(tlb, page);
1188                 pte_free(tlb->mm, pgtable);
1189                 ret = 1;
1190         }
1191         return ret;
1192 }
1193
1194 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1195                 unsigned long addr, unsigned long end,
1196                 unsigned char *vec)
1197 {
1198         int ret = 0;
1199
1200         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1201                 /*
1202                  * All logical pages in the range are present
1203                  * if backed by a huge page.
1204                  */
1205                 spin_unlock(&vma->vm_mm->page_table_lock);
1206                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1207                 ret = 1;
1208         }
1209
1210         return ret;
1211 }
1212
1213 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1214                   unsigned long old_addr,
1215                   unsigned long new_addr, unsigned long old_end,
1216                   pmd_t *old_pmd, pmd_t *new_pmd)
1217 {
1218         int ret = 0;
1219         pmd_t pmd;
1220
1221         struct mm_struct *mm = vma->vm_mm;
1222
1223         if ((old_addr & ~HPAGE_PMD_MASK) ||
1224             (new_addr & ~HPAGE_PMD_MASK) ||
1225             old_end - old_addr < HPAGE_PMD_SIZE ||
1226             (new_vma->vm_flags & VM_NOHUGEPAGE))
1227                 goto out;
1228
1229         /*
1230          * The destination pmd shouldn't be established, free_pgtables()
1231          * should have release it.
1232          */
1233         if (WARN_ON(!pmd_none(*new_pmd))) {
1234                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1235                 goto out;
1236         }
1237
1238         ret = __pmd_trans_huge_lock(old_pmd, vma);
1239         if (ret == 1) {
1240                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1241                 VM_BUG_ON(!pmd_none(*new_pmd));
1242                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1243                 spin_unlock(&mm->page_table_lock);
1244         }
1245 out:
1246         return ret;
1247 }
1248
1249 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1250                 unsigned long addr, pgprot_t newprot)
1251 {
1252         struct mm_struct *mm = vma->vm_mm;
1253         int ret = 0;
1254
1255         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1256                 pmd_t entry;
1257                 entry = pmdp_get_and_clear(mm, addr, pmd);
1258                 entry = pmd_modify(entry, newprot);
1259                 set_pmd_at(mm, addr, pmd, entry);
1260                 spin_unlock(&vma->vm_mm->page_table_lock);
1261                 ret = 1;
1262         }
1263
1264         return ret;
1265 }
1266
1267 /*
1268  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1269  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1270  *
1271  * Note that if it returns 1, this routine returns without unlocking page
1272  * table locks. So callers must unlock them.
1273  */
1274 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1275 {
1276         spin_lock(&vma->vm_mm->page_table_lock);
1277         if (likely(pmd_trans_huge(*pmd))) {
1278                 if (unlikely(pmd_trans_splitting(*pmd))) {
1279                         spin_unlock(&vma->vm_mm->page_table_lock);
1280                         wait_split_huge_page(vma->anon_vma, pmd);
1281                         return -1;
1282                 } else {
1283                         /* Thp mapped by 'pmd' is stable, so we can
1284                          * handle it as it is. */
1285                         return 1;
1286                 }
1287         }
1288         spin_unlock(&vma->vm_mm->page_table_lock);
1289         return 0;
1290 }
1291
1292 pmd_t *page_check_address_pmd(struct page *page,
1293                               struct mm_struct *mm,
1294                               unsigned long address,
1295                               enum page_check_address_pmd_flag flag)
1296 {
1297         pgd_t *pgd;
1298         pud_t *pud;
1299         pmd_t *pmd, *ret = NULL;
1300
1301         if (address & ~HPAGE_PMD_MASK)
1302                 goto out;
1303
1304         pgd = pgd_offset(mm, address);
1305         if (!pgd_present(*pgd))
1306                 goto out;
1307
1308         pud = pud_offset(pgd, address);
1309         if (!pud_present(*pud))
1310                 goto out;
1311
1312         pmd = pmd_offset(pud, address);
1313         if (pmd_none(*pmd))
1314                 goto out;
1315         if (pmd_page(*pmd) != page)
1316                 goto out;
1317         /*
1318          * split_vma() may create temporary aliased mappings. There is
1319          * no risk as long as all huge pmd are found and have their
1320          * splitting bit set before __split_huge_page_refcount
1321          * runs. Finding the same huge pmd more than once during the
1322          * same rmap walk is not a problem.
1323          */
1324         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1325             pmd_trans_splitting(*pmd))
1326                 goto out;
1327         if (pmd_trans_huge(*pmd)) {
1328                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1329                           !pmd_trans_splitting(*pmd));
1330                 ret = pmd;
1331         }
1332 out:
1333         return ret;
1334 }
1335
1336 static int __split_huge_page_splitting(struct page *page,
1337                                        struct vm_area_struct *vma,
1338                                        unsigned long address)
1339 {
1340         struct mm_struct *mm = vma->vm_mm;
1341         pmd_t *pmd;
1342         int ret = 0;
1343         /* For mmu_notifiers */
1344         const unsigned long mmun_start = address;
1345         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1346
1347         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1348         spin_lock(&mm->page_table_lock);
1349         pmd = page_check_address_pmd(page, mm, address,
1350                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1351         if (pmd) {
1352                 /*
1353                  * We can't temporarily set the pmd to null in order
1354                  * to split it, the pmd must remain marked huge at all
1355                  * times or the VM won't take the pmd_trans_huge paths
1356                  * and it won't wait on the anon_vma->root->mutex to
1357                  * serialize against split_huge_page*.
1358                  */
1359                 pmdp_splitting_flush(vma, address, pmd);
1360                 ret = 1;
1361         }
1362         spin_unlock(&mm->page_table_lock);
1363         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1364
1365         return ret;
1366 }
1367
1368 static void __split_huge_page_refcount(struct page *page)
1369 {
1370         int i;
1371         struct zone *zone = page_zone(page);
1372         struct lruvec *lruvec;
1373         int tail_count = 0;
1374
1375         /* prevent PageLRU to go away from under us, and freeze lru stats */
1376         spin_lock_irq(&zone->lru_lock);
1377         lruvec = mem_cgroup_page_lruvec(page, zone);
1378
1379         compound_lock(page);
1380         /* complete memcg works before add pages to LRU */
1381         mem_cgroup_split_huge_fixup(page);
1382
1383         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1384                 struct page *page_tail = page + i;
1385
1386                 /* tail_page->_mapcount cannot change */
1387                 BUG_ON(page_mapcount(page_tail) < 0);
1388                 tail_count += page_mapcount(page_tail);
1389                 /* check for overflow */
1390                 BUG_ON(tail_count < 0);
1391                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1392                 /*
1393                  * tail_page->_count is zero and not changing from
1394                  * under us. But get_page_unless_zero() may be running
1395                  * from under us on the tail_page. If we used
1396                  * atomic_set() below instead of atomic_add(), we
1397                  * would then run atomic_set() concurrently with
1398                  * get_page_unless_zero(), and atomic_set() is
1399                  * implemented in C not using locked ops. spin_unlock
1400                  * on x86 sometime uses locked ops because of PPro
1401                  * errata 66, 92, so unless somebody can guarantee
1402                  * atomic_set() here would be safe on all archs (and
1403                  * not only on x86), it's safer to use atomic_add().
1404                  */
1405                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1406                            &page_tail->_count);
1407
1408                 /* after clearing PageTail the gup refcount can be released */
1409                 smp_mb();
1410
1411                 /*
1412                  * retain hwpoison flag of the poisoned tail page:
1413                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1414                  *   by the memory-failure.
1415                  */
1416                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1417                 page_tail->flags |= (page->flags &
1418                                      ((1L << PG_referenced) |
1419                                       (1L << PG_swapbacked) |
1420                                       (1L << PG_mlocked) |
1421                                       (1L << PG_uptodate)));
1422                 page_tail->flags |= (1L << PG_dirty);
1423
1424                 /* clear PageTail before overwriting first_page */
1425                 smp_wmb();
1426
1427                 /*
1428                  * __split_huge_page_splitting() already set the
1429                  * splitting bit in all pmd that could map this
1430                  * hugepage, that will ensure no CPU can alter the
1431                  * mapcount on the head page. The mapcount is only
1432                  * accounted in the head page and it has to be
1433                  * transferred to all tail pages in the below code. So
1434                  * for this code to be safe, the split the mapcount
1435                  * can't change. But that doesn't mean userland can't
1436                  * keep changing and reading the page contents while
1437                  * we transfer the mapcount, so the pmd splitting
1438                  * status is achieved setting a reserved bit in the
1439                  * pmd, not by clearing the present bit.
1440                 */
1441                 page_tail->_mapcount = page->_mapcount;
1442
1443                 BUG_ON(page_tail->mapping);
1444                 page_tail->mapping = page->mapping;
1445
1446                 page_tail->index = page->index + i;
1447                 page_xchg_last_nid(page, page_last_nid(page_tail));
1448
1449                 BUG_ON(!PageAnon(page_tail));
1450                 BUG_ON(!PageUptodate(page_tail));
1451                 BUG_ON(!PageDirty(page_tail));
1452                 BUG_ON(!PageSwapBacked(page_tail));
1453
1454                 lru_add_page_tail(page, page_tail, lruvec);
1455         }
1456         atomic_sub(tail_count, &page->_count);
1457         BUG_ON(atomic_read(&page->_count) <= 0);
1458
1459         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1460         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1461
1462         ClearPageCompound(page);
1463         compound_unlock(page);
1464         spin_unlock_irq(&zone->lru_lock);
1465
1466         for (i = 1; i < HPAGE_PMD_NR; i++) {
1467                 struct page *page_tail = page + i;
1468                 BUG_ON(page_count(page_tail) <= 0);
1469                 /*
1470                  * Tail pages may be freed if there wasn't any mapping
1471                  * like if add_to_swap() is running on a lru page that
1472                  * had its mapping zapped. And freeing these pages
1473                  * requires taking the lru_lock so we do the put_page
1474                  * of the tail pages after the split is complete.
1475                  */
1476                 put_page(page_tail);
1477         }
1478
1479         /*
1480          * Only the head page (now become a regular page) is required
1481          * to be pinned by the caller.
1482          */
1483         BUG_ON(page_count(page) <= 0);
1484 }
1485
1486 static int __split_huge_page_map(struct page *page,
1487                                  struct vm_area_struct *vma,
1488                                  unsigned long address)
1489 {
1490         struct mm_struct *mm = vma->vm_mm;
1491         pmd_t *pmd, _pmd;
1492         int ret = 0, i;
1493         pgtable_t pgtable;
1494         unsigned long haddr;
1495         pgprot_t prot;
1496
1497         spin_lock(&mm->page_table_lock);
1498         pmd = page_check_address_pmd(page, mm, address,
1499                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1500         if (!pmd)
1501                 goto unlock;
1502
1503         prot = pmd_pgprot(*pmd);
1504         pgtable = pgtable_trans_huge_withdraw(mm);
1505         pmd_populate(mm, &_pmd, pgtable);
1506
1507         for (i = 0, haddr = address; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1508                 pte_t *pte, entry;
1509
1510                 BUG_ON(PageCompound(page+i));
1511                 entry = mk_pte(page + i, prot);
1512                 entry = pte_mkdirty(entry);
1513                 if (!pmd_young(*pmd))
1514                         entry = pte_mkold(entry);
1515                 pte = pte_offset_map(&_pmd, haddr);
1516                 BUG_ON(!pte_none(*pte));
1517                 set_pte_at(mm, haddr, pte, entry);
1518                 pte_unmap(pte);
1519         }
1520
1521         smp_wmb(); /* make ptes visible before pmd, see __pte_alloc */
1522         /*
1523          * Up to this point the pmd is present and huge.
1524          *
1525          * If we overwrite the pmd with the not-huge version, we could trigger
1526          * a small page size TLB miss on the small sized TLB while the hugepage
1527          * TLB entry is still established in the huge TLB.
1528          *
1529          * Some CPUs don't like that. See
1530          * http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 383
1531          * on page 93.
1532          *
1533          * Thus it is generally safer to never allow small and huge TLB entries
1534          * for overlapping virtual addresses to be loaded. So we first mark the
1535          * current pmd not present, then we flush the TLB and finally we write
1536          * the non-huge version of the pmd entry with pmd_populate.
1537          *
1538          * The above needs to be done under the ptl because pmd_trans_huge and
1539          * pmd_trans_splitting must remain set on the pmd until the split is
1540          * complete. The ptl also protects against concurrent faults due to
1541          * making the pmd not-present.
1542          */
1543         set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1544         flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1545         pmd_populate(mm, pmd, pgtable);
1546         ret = 1;
1547
1548 unlock:
1549         spin_unlock(&mm->page_table_lock);
1550
1551         return ret;
1552 }
1553
1554 /* must be called with anon_vma->root->mutex hold */
1555 static void __split_huge_page(struct page *page,
1556                               struct anon_vma *anon_vma)
1557 {
1558         int mapcount, mapcount2;
1559         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1560         struct anon_vma_chain *avc;
1561
1562         BUG_ON(!PageHead(page));
1563         BUG_ON(PageTail(page));
1564
1565         mapcount = 0;
1566         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1567                 struct vm_area_struct *vma = avc->vma;
1568                 unsigned long addr = vma_address(page, vma);
1569                 BUG_ON(is_vma_temporary_stack(vma));
1570                 mapcount += __split_huge_page_splitting(page, vma, addr);
1571         }
1572         /*
1573          * It is critical that new vmas are added to the tail of the
1574          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1575          * and establishes a child pmd before
1576          * __split_huge_page_splitting() freezes the parent pmd (so if
1577          * we fail to prevent copy_huge_pmd() from running until the
1578          * whole __split_huge_page() is complete), we will still see
1579          * the newly established pmd of the child later during the
1580          * walk, to be able to set it as pmd_trans_splitting too.
1581          */
1582         if (mapcount != page_mapcount(page))
1583                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1584                        mapcount, page_mapcount(page));
1585         BUG_ON(mapcount != page_mapcount(page));
1586
1587         __split_huge_page_refcount(page);
1588
1589         mapcount2 = 0;
1590         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1591                 struct vm_area_struct *vma = avc->vma;
1592                 unsigned long addr = vma_address(page, vma);
1593                 BUG_ON(is_vma_temporary_stack(vma));
1594                 mapcount2 += __split_huge_page_map(page, vma, addr);
1595         }
1596         if (mapcount != mapcount2)
1597                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1598                        mapcount, mapcount2, page_mapcount(page));
1599         BUG_ON(mapcount != mapcount2);
1600 }
1601
1602 int split_huge_page(struct page *page)
1603 {
1604         struct anon_vma *anon_vma;
1605         int ret = 1;
1606
1607         BUG_ON(!PageAnon(page));
1608         anon_vma = page_lock_anon_vma(page);
1609         if (!anon_vma)
1610                 goto out;
1611         ret = 0;
1612         if (!PageCompound(page))
1613                 goto out_unlock;
1614
1615         BUG_ON(!PageSwapBacked(page));
1616         __split_huge_page(page, anon_vma);
1617         count_vm_event(THP_SPLIT);
1618
1619         BUG_ON(PageCompound(page));
1620 out_unlock:
1621         page_unlock_anon_vma(anon_vma);
1622 out:
1623         return ret;
1624 }
1625
1626 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1627
1628 int hugepage_madvise(struct vm_area_struct *vma,
1629                      unsigned long *vm_flags, int advice)
1630 {
1631         struct mm_struct *mm = vma->vm_mm;
1632
1633         switch (advice) {
1634         case MADV_HUGEPAGE:
1635                 /*
1636                  * Be somewhat over-protective like KSM for now!
1637                  */
1638                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1639                         return -EINVAL;
1640                 if (mm->def_flags & VM_NOHUGEPAGE)
1641                         return -EINVAL;
1642                 *vm_flags &= ~VM_NOHUGEPAGE;
1643                 *vm_flags |= VM_HUGEPAGE;
1644                 /*
1645                  * If the vma become good for khugepaged to scan,
1646                  * register it here without waiting a page fault that
1647                  * may not happen any time soon.
1648                  */
1649                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1650                         return -ENOMEM;
1651                 break;
1652         case MADV_NOHUGEPAGE:
1653                 /*
1654                  * Be somewhat over-protective like KSM for now!
1655                  */
1656                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1657                         return -EINVAL;
1658                 *vm_flags &= ~VM_HUGEPAGE;
1659                 *vm_flags |= VM_NOHUGEPAGE;
1660                 /*
1661                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1662                  * this vma even if we leave the mm registered in khugepaged if
1663                  * it got registered before VM_NOHUGEPAGE was set.
1664                  */
1665                 break;
1666         }
1667
1668         return 0;
1669 }
1670
1671 static int __init khugepaged_slab_init(void)
1672 {
1673         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1674                                           sizeof(struct mm_slot),
1675                                           __alignof__(struct mm_slot), 0, NULL);
1676         if (!mm_slot_cache)
1677                 return -ENOMEM;
1678
1679         return 0;
1680 }
1681
1682 static void __init khugepaged_slab_free(void)
1683 {
1684         kmem_cache_destroy(mm_slot_cache);
1685         mm_slot_cache = NULL;
1686 }
1687
1688 static inline struct mm_slot *alloc_mm_slot(void)
1689 {
1690         if (!mm_slot_cache)     /* initialization failed */
1691                 return NULL;
1692         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1693 }
1694
1695 static inline void free_mm_slot(struct mm_slot *mm_slot)
1696 {
1697         kmem_cache_free(mm_slot_cache, mm_slot);
1698 }
1699
1700 static int __init mm_slots_hash_init(void)
1701 {
1702         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1703                                 GFP_KERNEL);
1704         if (!mm_slots_hash)
1705                 return -ENOMEM;
1706         return 0;
1707 }
1708
1709 #if 0
1710 static void __init mm_slots_hash_free(void)
1711 {
1712         kfree(mm_slots_hash);
1713         mm_slots_hash = NULL;
1714 }
1715 #endif
1716
1717 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1718 {
1719         struct mm_slot *mm_slot;
1720         struct hlist_head *bucket;
1721         struct hlist_node *node;
1722
1723         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1724                                 % MM_SLOTS_HASH_HEADS];
1725         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1726                 if (mm == mm_slot->mm)
1727                         return mm_slot;
1728         }
1729         return NULL;
1730 }
1731
1732 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1733                                     struct mm_slot *mm_slot)
1734 {
1735         struct hlist_head *bucket;
1736
1737         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1738                                 % MM_SLOTS_HASH_HEADS];
1739         mm_slot->mm = mm;
1740         hlist_add_head(&mm_slot->hash, bucket);
1741 }
1742
1743 static inline int khugepaged_test_exit(struct mm_struct *mm)
1744 {
1745         return atomic_read(&mm->mm_users) == 0;
1746 }
1747
1748 int __khugepaged_enter(struct mm_struct *mm)
1749 {
1750         struct mm_slot *mm_slot;
1751         int wakeup;
1752
1753         mm_slot = alloc_mm_slot();
1754         if (!mm_slot)
1755                 return -ENOMEM;
1756
1757         /* __khugepaged_exit() must not run from under us */
1758         VM_BUG_ON(khugepaged_test_exit(mm));
1759         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1760                 free_mm_slot(mm_slot);
1761                 return 0;
1762         }
1763
1764         spin_lock(&khugepaged_mm_lock);
1765         insert_to_mm_slots_hash(mm, mm_slot);
1766         /*
1767          * Insert just behind the scanning cursor, to let the area settle
1768          * down a little.
1769          */
1770         wakeup = list_empty(&khugepaged_scan.mm_head);
1771         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1772         spin_unlock(&khugepaged_mm_lock);
1773
1774         atomic_inc(&mm->mm_count);
1775         if (wakeup)
1776                 wake_up_interruptible(&khugepaged_wait);
1777
1778         return 0;
1779 }
1780
1781 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1782 {
1783         unsigned long hstart, hend;
1784         if (!vma->anon_vma)
1785                 /*
1786                  * Not yet faulted in so we will register later in the
1787                  * page fault if needed.
1788                  */
1789                 return 0;
1790         if (vma->vm_ops)
1791                 /* khugepaged not yet working on file or special mappings */
1792                 return 0;
1793         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1794         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1795         hend = vma->vm_end & HPAGE_PMD_MASK;
1796         if (hstart < hend)
1797                 return khugepaged_enter(vma);
1798         return 0;
1799 }
1800
1801 void __khugepaged_exit(struct mm_struct *mm)
1802 {
1803         struct mm_slot *mm_slot;
1804         int free = 0;
1805
1806         spin_lock(&khugepaged_mm_lock);
1807         mm_slot = get_mm_slot(mm);
1808         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1809                 hlist_del(&mm_slot->hash);
1810                 list_del(&mm_slot->mm_node);
1811                 free = 1;
1812         }
1813         spin_unlock(&khugepaged_mm_lock);
1814
1815         if (free) {
1816                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1817                 free_mm_slot(mm_slot);
1818                 mmdrop(mm);
1819         } else if (mm_slot) {
1820                 /*
1821                  * This is required to serialize against
1822                  * khugepaged_test_exit() (which is guaranteed to run
1823                  * under mmap sem read mode). Stop here (after we
1824                  * return all pagetables will be destroyed) until
1825                  * khugepaged has finished working on the pagetables
1826                  * under the mmap_sem.
1827                  */
1828                 down_write(&mm->mmap_sem);
1829                 up_write(&mm->mmap_sem);
1830         }
1831 }
1832
1833 static void release_pte_page(struct page *page)
1834 {
1835         /* 0 stands for page_is_file_cache(page) == false */
1836         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1837         unlock_page(page);
1838         putback_lru_page(page);
1839 }
1840
1841 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1842 {
1843         while (--_pte >= pte) {
1844                 pte_t pteval = *_pte;
1845                 if (!pte_none(pteval))
1846                         release_pte_page(pte_page(pteval));
1847         }
1848 }
1849
1850 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1851                                         unsigned long address,
1852                                         pte_t *pte)
1853 {
1854         struct page *page;
1855         pte_t *_pte;
1856         int referenced = 0, isolated = 1, none = 0;
1857         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1858              _pte++, address += PAGE_SIZE) {
1859                 pte_t pteval = *_pte;
1860                 if (pte_none(pteval)) {
1861                         if (++none <= khugepaged_max_ptes_none)
1862                                 continue;
1863                         else
1864                                 goto out;
1865                 }
1866                 if (!pte_present(pteval) || !pte_write(pteval))
1867                         goto out;
1868                 page = vm_normal_page(vma, address, pteval);
1869                 if (unlikely(!page))
1870                         goto out;
1871
1872                 VM_BUG_ON(PageCompound(page));
1873                 BUG_ON(!PageAnon(page));
1874                 VM_BUG_ON(!PageSwapBacked(page));
1875
1876                 /* cannot use mapcount: can't collapse if there's a gup pin */
1877                 if (page_count(page) != 1)
1878                         goto out;
1879                 /*
1880                  * We can do it before isolate_lru_page because the
1881                  * page can't be freed from under us. NOTE: PG_lock
1882                  * is needed to serialize against split_huge_page
1883                  * when invoked from the VM.
1884                  */
1885                 if (!trylock_page(page))
1886                         goto out;
1887                 /*
1888                  * Isolate the page to avoid collapsing an hugepage
1889                  * currently in use by the VM.
1890                  */
1891                 if (isolate_lru_page(page)) {
1892                         unlock_page(page);
1893                         goto out;
1894                 }
1895                 /* 0 stands for page_is_file_cache(page) == false */
1896                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1897                 VM_BUG_ON(!PageLocked(page));
1898                 VM_BUG_ON(PageLRU(page));
1899
1900                 /* If there is no mapped pte young don't collapse the page */
1901                 if (pte_young(pteval) || PageReferenced(page) ||
1902                     mmu_notifier_test_young(vma->vm_mm, address))
1903                         referenced = 1;
1904         }
1905         if (unlikely(!referenced)) {
1906 out:
1907                 release_pte_pages(pte, _pte);
1908                 isolated = 0;
1909         }
1910         return isolated;
1911 }
1912
1913 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1914                                       struct vm_area_struct *vma,
1915                                       unsigned long address,
1916                                       spinlock_t *ptl)
1917 {
1918         pte_t *_pte;
1919         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1920                 pte_t pteval = *_pte;
1921                 struct page *src_page;
1922
1923                 if (pte_none(pteval)) {
1924                         clear_user_highpage(page, address);
1925                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1926                 } else {
1927                         src_page = pte_page(pteval);
1928                         copy_user_highpage(page, src_page, address, vma);
1929                         VM_BUG_ON(page_mapcount(src_page) != 1);
1930                         release_pte_page(src_page);
1931                         /*
1932                          * ptl mostly unnecessary, but preempt has to
1933                          * be disabled to update the per-cpu stats
1934                          * inside page_remove_rmap().
1935                          */
1936                         spin_lock(ptl);
1937                         /*
1938                          * paravirt calls inside pte_clear here are
1939                          * superfluous.
1940                          */
1941                         pte_clear(vma->vm_mm, address, _pte);
1942                         page_remove_rmap(src_page);
1943                         spin_unlock(ptl);
1944                         free_page_and_swap_cache(src_page);
1945                 }
1946
1947                 address += PAGE_SIZE;
1948                 page++;
1949         }
1950 }
1951
1952 static void khugepaged_alloc_sleep(void)
1953 {
1954         wait_event_freezable_timeout(khugepaged_wait, false,
1955                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1956 }
1957
1958 #ifdef CONFIG_NUMA
1959 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1960 {
1961         if (IS_ERR(*hpage)) {
1962                 if (!*wait)
1963                         return false;
1964
1965                 *wait = false;
1966                 *hpage = NULL;
1967                 khugepaged_alloc_sleep();
1968         } else if (*hpage) {
1969                 put_page(*hpage);
1970                 *hpage = NULL;
1971         }
1972
1973         return true;
1974 }
1975
1976 static struct page
1977 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1978                        struct vm_area_struct *vma, unsigned long address,
1979                        int node)
1980 {
1981         VM_BUG_ON(*hpage);
1982         /*
1983          * Allocate the page while the vma is still valid and under
1984          * the mmap_sem read mode so there is no memory allocation
1985          * later when we take the mmap_sem in write mode. This is more
1986          * friendly behavior (OTOH it may actually hide bugs) to
1987          * filesystems in userland with daemons allocating memory in
1988          * the userland I/O paths.  Allocating memory with the
1989          * mmap_sem in read mode is good idea also to allow greater
1990          * scalability.
1991          */
1992         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1993                                       node, __GFP_OTHER_NODE);
1994
1995         /*
1996          * After allocating the hugepage, release the mmap_sem read lock in
1997          * preparation for taking it in write mode.
1998          */
1999         up_read(&mm->mmap_sem);
2000         if (unlikely(!*hpage)) {
2001                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2002                 *hpage = ERR_PTR(-ENOMEM);
2003                 return NULL;
2004         }
2005
2006         count_vm_event(THP_COLLAPSE_ALLOC);
2007         return *hpage;
2008 }
2009 #else
2010 static struct page *khugepaged_alloc_hugepage(bool *wait)
2011 {
2012         struct page *hpage;
2013
2014         do {
2015                 hpage = alloc_hugepage(khugepaged_defrag());
2016                 if (!hpage) {
2017                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2018                         if (!*wait)
2019                                 return NULL;
2020
2021                         *wait = false;
2022                         khugepaged_alloc_sleep();
2023                 } else
2024                         count_vm_event(THP_COLLAPSE_ALLOC);
2025         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2026
2027         return hpage;
2028 }
2029
2030 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2031 {
2032         if (!*hpage)
2033                 *hpage = khugepaged_alloc_hugepage(wait);
2034
2035         if (unlikely(!*hpage))
2036                 return false;
2037
2038         return true;
2039 }
2040
2041 static struct page
2042 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2043                        struct vm_area_struct *vma, unsigned long address,
2044                        int node)
2045 {
2046         up_read(&mm->mmap_sem);
2047         VM_BUG_ON(!*hpage);
2048         return  *hpage;
2049 }
2050 #endif
2051
2052 static void collapse_huge_page(struct mm_struct *mm,
2053                                    unsigned long address,
2054                                    struct page **hpage,
2055                                    struct vm_area_struct *vma,
2056                                    int node)
2057 {
2058         pgd_t *pgd;
2059         pud_t *pud;
2060         pmd_t *pmd, _pmd;
2061         pte_t *pte;
2062         pgtable_t pgtable;
2063         struct page *new_page;
2064         spinlock_t *ptl;
2065         int isolated;
2066         unsigned long hstart, hend;
2067         unsigned long mmun_start;       /* For mmu_notifiers */
2068         unsigned long mmun_end;         /* For mmu_notifiers */
2069
2070         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2071
2072         /* release the mmap_sem read lock. */
2073         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2074         if (!new_page)
2075                 return;
2076
2077         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2078                 return;
2079
2080         /*
2081          * Prevent all access to pagetables with the exception of
2082          * gup_fast later hanlded by the ptep_clear_flush and the VM
2083          * handled by the anon_vma lock + PG_lock.
2084          */
2085         down_write(&mm->mmap_sem);
2086         if (unlikely(khugepaged_test_exit(mm)))
2087                 goto out;
2088
2089         vma = find_vma(mm, address);
2090         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2091         hend = vma->vm_end & HPAGE_PMD_MASK;
2092         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2093                 goto out;
2094
2095         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2096             (vma->vm_flags & VM_NOHUGEPAGE))
2097                 goto out;
2098
2099         if (!vma->anon_vma || vma->vm_ops)
2100                 goto out;
2101         if (is_vma_temporary_stack(vma))
2102                 goto out;
2103         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2104
2105         pgd = pgd_offset(mm, address);
2106         if (!pgd_present(*pgd))
2107                 goto out;
2108
2109         pud = pud_offset(pgd, address);
2110         if (!pud_present(*pud))
2111                 goto out;
2112
2113         pmd = pmd_offset(pud, address);
2114         /* pmd can't go away or become huge under us */
2115         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2116                 goto out;
2117
2118         anon_vma_lock(vma->anon_vma);
2119
2120         pte = pte_offset_map(pmd, address);
2121         ptl = pte_lockptr(mm, pmd);
2122
2123         mmun_start = address;
2124         mmun_end   = address + HPAGE_PMD_SIZE;
2125         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2126         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2127         /*
2128          * After this gup_fast can't run anymore. This also removes
2129          * any huge TLB entry from the CPU so we won't allow
2130          * huge and small TLB entries for the same virtual address
2131          * to avoid the risk of CPU bugs in that area.
2132          */
2133         _pmd = pmdp_clear_flush(vma, address, pmd);
2134         spin_unlock(&mm->page_table_lock);
2135         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2136
2137         spin_lock(ptl);
2138         isolated = __collapse_huge_page_isolate(vma, address, pte);
2139         spin_unlock(ptl);
2140
2141         if (unlikely(!isolated)) {
2142                 pte_unmap(pte);
2143                 spin_lock(&mm->page_table_lock);
2144                 BUG_ON(!pmd_none(*pmd));
2145                 set_pmd_at(mm, address, pmd, _pmd);
2146                 spin_unlock(&mm->page_table_lock);
2147                 anon_vma_unlock(vma->anon_vma);
2148                 goto out;
2149         }
2150
2151         /*
2152          * All pages are isolated and locked so anon_vma rmap
2153          * can't run anymore.
2154          */
2155         anon_vma_unlock(vma->anon_vma);
2156
2157         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2158         pte_unmap(pte);
2159         __SetPageUptodate(new_page);
2160         pgtable = pmd_pgtable(_pmd);
2161
2162         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2163         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2164         _pmd = pmd_mkhuge(_pmd);
2165
2166         /*
2167          * spin_lock() below is not the equivalent of smp_wmb(), so
2168          * this is needed to avoid the copy_huge_page writes to become
2169          * visible after the set_pmd_at() write.
2170          */
2171         smp_wmb();
2172
2173         spin_lock(&mm->page_table_lock);
2174         BUG_ON(!pmd_none(*pmd));
2175         page_add_new_anon_rmap(new_page, vma, address);
2176         set_pmd_at(mm, address, pmd, _pmd);
2177         update_mmu_cache_pmd(vma, address, pmd);
2178         pgtable_trans_huge_deposit(mm, pgtable);
2179         spin_unlock(&mm->page_table_lock);
2180
2181         *hpage = NULL;
2182
2183         khugepaged_pages_collapsed++;
2184 out_up_write:
2185         up_write(&mm->mmap_sem);
2186         return;
2187
2188 out:
2189         mem_cgroup_uncharge_page(new_page);
2190         goto out_up_write;
2191 }
2192
2193 static int khugepaged_scan_pmd(struct mm_struct *mm,
2194                                struct vm_area_struct *vma,
2195                                unsigned long address,
2196                                struct page **hpage)
2197 {
2198         pgd_t *pgd;
2199         pud_t *pud;
2200         pmd_t *pmd;
2201         pte_t *pte, *_pte;
2202         int ret = 0, referenced = 0, none = 0;
2203         struct page *page;
2204         unsigned long _address;
2205         spinlock_t *ptl;
2206         int node = -1;
2207
2208         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2209
2210         pgd = pgd_offset(mm, address);
2211         if (!pgd_present(*pgd))
2212                 goto out;
2213
2214         pud = pud_offset(pgd, address);
2215         if (!pud_present(*pud))
2216                 goto out;
2217
2218         pmd = pmd_offset(pud, address);
2219         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2220                 goto out;
2221
2222         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2223         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2224              _pte++, _address += PAGE_SIZE) {
2225                 pte_t pteval = *_pte;
2226                 if (pte_none(pteval)) {
2227                         if (++none <= khugepaged_max_ptes_none)
2228                                 continue;
2229                         else
2230                                 goto out_unmap;
2231                 }
2232                 if (!pte_present(pteval) || !pte_write(pteval))
2233                         goto out_unmap;
2234                 page = vm_normal_page(vma, _address, pteval);
2235                 if (unlikely(!page))
2236                         goto out_unmap;
2237                 /*
2238                  * Chose the node of the first page. This could
2239                  * be more sophisticated and look at more pages,
2240                  * but isn't for now.
2241                  */
2242                 if (node == -1)
2243                         node = page_to_nid(page);
2244                 VM_BUG_ON(PageCompound(page));
2245                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2246                         goto out_unmap;
2247                 /* cannot use mapcount: can't collapse if there's a gup pin */
2248                 if (page_count(page) != 1)
2249                         goto out_unmap;
2250                 if (pte_young(pteval) || PageReferenced(page) ||
2251                     mmu_notifier_test_young(vma->vm_mm, address))
2252                         referenced = 1;
2253         }
2254         if (referenced)
2255                 ret = 1;
2256 out_unmap:
2257         pte_unmap_unlock(pte, ptl);
2258         if (ret)
2259                 /* collapse_huge_page will return with the mmap_sem released */
2260                 collapse_huge_page(mm, address, hpage, vma, node);
2261 out:
2262         return ret;
2263 }
2264
2265 static void collect_mm_slot(struct mm_slot *mm_slot)
2266 {
2267         struct mm_struct *mm = mm_slot->mm;
2268
2269         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2270
2271         if (khugepaged_test_exit(mm)) {
2272                 /* free mm_slot */
2273                 hlist_del(&mm_slot->hash);
2274                 list_del(&mm_slot->mm_node);
2275
2276                 /*
2277                  * Not strictly needed because the mm exited already.
2278                  *
2279                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2280                  */
2281
2282                 /* khugepaged_mm_lock actually not necessary for the below */
2283                 free_mm_slot(mm_slot);
2284                 mmdrop(mm);
2285         }
2286 }
2287
2288 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2289                                             struct page **hpage)
2290         __releases(&khugepaged_mm_lock)
2291         __acquires(&khugepaged_mm_lock)
2292 {
2293         struct mm_slot *mm_slot;
2294         struct mm_struct *mm;
2295         struct vm_area_struct *vma;
2296         int progress = 0;
2297
2298         VM_BUG_ON(!pages);
2299         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2300
2301         if (khugepaged_scan.mm_slot)
2302                 mm_slot = khugepaged_scan.mm_slot;
2303         else {
2304                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2305                                      struct mm_slot, mm_node);
2306                 khugepaged_scan.address = 0;
2307                 khugepaged_scan.mm_slot = mm_slot;
2308         }
2309         spin_unlock(&khugepaged_mm_lock);
2310
2311         mm = mm_slot->mm;
2312         down_read(&mm->mmap_sem);
2313         if (unlikely(khugepaged_test_exit(mm)))
2314                 vma = NULL;
2315         else
2316                 vma = find_vma(mm, khugepaged_scan.address);
2317
2318         progress++;
2319         for (; vma; vma = vma->vm_next) {
2320                 unsigned long hstart, hend;
2321
2322                 cond_resched();
2323                 if (unlikely(khugepaged_test_exit(mm))) {
2324                         progress++;
2325                         break;
2326                 }
2327
2328                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2329                      !khugepaged_always()) ||
2330                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2331                 skip:
2332                         progress++;
2333                         continue;
2334                 }
2335                 if (!vma->anon_vma || vma->vm_ops)
2336                         goto skip;
2337                 if (is_vma_temporary_stack(vma))
2338                         goto skip;
2339                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2340
2341                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2342                 hend = vma->vm_end & HPAGE_PMD_MASK;
2343                 if (hstart >= hend)
2344                         goto skip;
2345                 if (khugepaged_scan.address > hend)
2346                         goto skip;
2347                 if (khugepaged_scan.address < hstart)
2348                         khugepaged_scan.address = hstart;
2349                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2350
2351                 while (khugepaged_scan.address < hend) {
2352                         int ret;
2353                         cond_resched();
2354                         if (unlikely(khugepaged_test_exit(mm)))
2355                                 goto breakouterloop;
2356
2357                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2358                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2359                                   hend);
2360                         ret = khugepaged_scan_pmd(mm, vma,
2361                                                   khugepaged_scan.address,
2362                                                   hpage);
2363                         /* move to next address */
2364                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2365                         progress += HPAGE_PMD_NR;
2366                         if (ret)
2367                                 /* we released mmap_sem so break loop */
2368                                 goto breakouterloop_mmap_sem;
2369                         if (progress >= pages)
2370                                 goto breakouterloop;
2371                 }
2372         }
2373 breakouterloop:
2374         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2375 breakouterloop_mmap_sem:
2376
2377         spin_lock(&khugepaged_mm_lock);
2378         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2379         /*
2380          * Release the current mm_slot if this mm is about to die, or
2381          * if we scanned all vmas of this mm.
2382          */
2383         if (khugepaged_test_exit(mm) || !vma) {
2384                 /*
2385                  * Make sure that if mm_users is reaching zero while
2386                  * khugepaged runs here, khugepaged_exit will find
2387                  * mm_slot not pointing to the exiting mm.
2388                  */
2389                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2390                         khugepaged_scan.mm_slot = list_entry(
2391                                 mm_slot->mm_node.next,
2392                                 struct mm_slot, mm_node);
2393                         khugepaged_scan.address = 0;
2394                 } else {
2395                         khugepaged_scan.mm_slot = NULL;
2396                         khugepaged_full_scans++;
2397                 }
2398
2399                 collect_mm_slot(mm_slot);
2400         }
2401
2402         return progress;
2403 }
2404
2405 static int khugepaged_has_work(void)
2406 {
2407         return !list_empty(&khugepaged_scan.mm_head) &&
2408                 khugepaged_enabled();
2409 }
2410
2411 static int khugepaged_wait_event(void)
2412 {
2413         return !list_empty(&khugepaged_scan.mm_head) ||
2414                 kthread_should_stop();
2415 }
2416
2417 static void khugepaged_do_scan(void)
2418 {
2419         struct page *hpage = NULL;
2420         unsigned int progress = 0, pass_through_head = 0;
2421         bool wait = true;
2422         unsigned int pages = ACCESS_ONCE(khugepaged_pages_to_scan);
2423
2424         while (progress < pages) {
2425                 if (!khugepaged_prealloc_page(&hpage, &wait))
2426                         break;
2427
2428                 cond_resched();
2429
2430                 if (unlikely(kthread_should_stop() || freezing(current)))
2431                         break;
2432
2433                 spin_lock(&khugepaged_mm_lock);
2434                 if (!khugepaged_scan.mm_slot)
2435                         pass_through_head++;
2436                 if (khugepaged_has_work() &&
2437                     pass_through_head < 2)
2438                         progress += khugepaged_scan_mm_slot(pages - progress,
2439                                                             &hpage);
2440                 else
2441                         progress = pages;
2442                 spin_unlock(&khugepaged_mm_lock);
2443         }
2444
2445         if (!IS_ERR_OR_NULL(hpage))
2446                 put_page(hpage);
2447 }
2448
2449 static void khugepaged_wait_work(void)
2450 {
2451         try_to_freeze();
2452
2453         if (khugepaged_has_work()) {
2454                 if (!khugepaged_scan_sleep_millisecs)
2455                         return;
2456
2457                 wait_event_freezable_timeout(khugepaged_wait,
2458                                              kthread_should_stop(),
2459                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2460                 return;
2461         }
2462
2463         if (khugepaged_enabled())
2464                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2465 }
2466
2467 static int khugepaged(void *none)
2468 {
2469         struct mm_slot *mm_slot;
2470
2471         set_freezable();
2472         set_user_nice(current, 19);
2473
2474         while (!kthread_should_stop()) {
2475                 khugepaged_do_scan();
2476                 khugepaged_wait_work();
2477         }
2478
2479         spin_lock(&khugepaged_mm_lock);
2480         mm_slot = khugepaged_scan.mm_slot;
2481         khugepaged_scan.mm_slot = NULL;
2482         if (mm_slot)
2483                 collect_mm_slot(mm_slot);
2484         spin_unlock(&khugepaged_mm_lock);
2485         return 0;
2486 }
2487
2488 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2489 {
2490         struct page *page;
2491
2492         spin_lock(&mm->page_table_lock);
2493         if (unlikely(!pmd_trans_huge(*pmd))) {
2494                 spin_unlock(&mm->page_table_lock);
2495                 return;
2496         }
2497         page = pmd_page(*pmd);
2498         VM_BUG_ON(!page_count(page));
2499         get_page(page);
2500         spin_unlock(&mm->page_table_lock);
2501
2502         split_huge_page(page);
2503
2504         put_page(page);
2505         BUG_ON(pmd_trans_huge(*pmd));
2506 }
2507
2508 static void split_huge_page_address(struct mm_struct *mm,
2509                                     unsigned long address)
2510 {
2511         pgd_t *pgd;
2512         pud_t *pud;
2513         pmd_t *pmd;
2514
2515         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2516
2517         pgd = pgd_offset(mm, address);
2518         if (!pgd_present(*pgd))
2519                 return;
2520
2521         pud = pud_offset(pgd, address);
2522         if (!pud_present(*pud))
2523                 return;
2524
2525         pmd = pmd_offset(pud, address);
2526         if (!pmd_present(*pmd))
2527                 return;
2528         /*
2529          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2530          * materialize from under us.
2531          */
2532         split_huge_page_pmd(mm, pmd);
2533 }
2534
2535 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2536                              unsigned long start,
2537                              unsigned long end,
2538                              long adjust_next)
2539 {
2540         /*
2541          * If the new start address isn't hpage aligned and it could
2542          * previously contain an hugepage: check if we need to split
2543          * an huge pmd.
2544          */
2545         if (start & ~HPAGE_PMD_MASK &&
2546             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2547             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2548                 split_huge_page_address(vma->vm_mm, start);
2549
2550         /*
2551          * If the new end address isn't hpage aligned and it could
2552          * previously contain an hugepage: check if we need to split
2553          * an huge pmd.
2554          */
2555         if (end & ~HPAGE_PMD_MASK &&
2556             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2557             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2558                 split_huge_page_address(vma->vm_mm, end);
2559
2560         /*
2561          * If we're also updating the vma->vm_next->vm_start, if the new
2562          * vm_next->vm_start isn't page aligned and it could previously
2563          * contain an hugepage: check if we need to split an huge pmd.
2564          */
2565         if (adjust_next > 0) {
2566                 struct vm_area_struct *next = vma->vm_next;
2567                 unsigned long nstart = next->vm_start;
2568                 nstart += adjust_next << PAGE_SHIFT;
2569                 if (nstart & ~HPAGE_PMD_MASK &&
2570                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2571                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2572                         split_huge_page_address(next->vm_mm, nstart);
2573         }
2574 }