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