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)
 651 {
 652         return GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT);
 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 {
 659         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag),
 660                                HPAGE_PMD_ORDER, vma, haddr, nd);
 661 }
 662
 663 #ifndef CONFIG_NUMA
 664 static inline struct page *alloc_hugepage(int defrag)
 665 {
 666         return alloc_pages(alloc_hugepage_gfpmask(defrag),
 667                            HPAGE_PMD_ORDER);
 668 }
 669 #endif
 670
 671 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 672                                unsigned long address, pmd_t *pmd,
 673                                unsigned int flags)
 674 {
 675         struct page *page;
 676         unsigned long haddr = address & HPAGE_PMD_MASK;
 677         pte_t *pte;
 678
 679         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 680                 if (unlikely(anon_vma_prepare(vma)))
 681                         return VM_FAULT_OOM;
 682                 if (unlikely(khugepaged_enter(vma)))
 683                         return VM_FAULT_OOM;
 684                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 685                                           vma, haddr, numa_node_id());
 686                 if (unlikely(!page))
 687                         goto out;
 688                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 689                         put_page(page);
 690                         goto out;
 691                 }
 692
 693                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
 694         }
 695 out:
 696         /*
 697          * Use __pte_alloc instead of pte_alloc_map, because we can't
 698          * run pte_offset_map on the pmd, if an huge pmd could
 699          * materialize from under us from a different thread.
 700          */
 701         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 702                 return VM_FAULT_OOM;
 703         /* if an huge pmd materialized from under us just retry later */
 704         if (unlikely(pmd_trans_huge(*pmd)))
 705                 return 0;
 706         /*
 707          * A regular pmd is established and it can't morph into a huge pmd
 708          * from under us anymore at this point because we hold the mmap_sem
 709          * read mode and khugepaged takes it in write mode. So now it's
 710          * safe to run pte_offset_map().
 711          */
 712         pte = pte_offset_map(pmd, address);
 713         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 714 }
 715
 716 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 717                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 718                   struct vm_area_struct *vma)
 719 {
 720         struct page *src_page;
 721         pmd_t pmd;
 722         pgtable_t pgtable;
 723         int ret;
 724
 725         ret = -ENOMEM;
 726         pgtable = pte_alloc_one(dst_mm, addr);
 727         if (unlikely(!pgtable))
 728                 goto out;
 729
 730         spin_lock(&dst_mm->page_table_lock);
 731         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 732
 733         ret = -EAGAIN;
 734         pmd = *src_pmd;
 735         if (unlikely(!pmd_trans_huge(pmd))) {
 736                 pte_free(dst_mm, pgtable);
 737                 goto out_unlock;
 738         }
 739         if (unlikely(pmd_trans_splitting(pmd))) {
 740                 /* split huge page running from under us */
 741                 spin_unlock(&src_mm->page_table_lock);
 742                 spin_unlock(&dst_mm->page_table_lock);
 743                 pte_free(dst_mm, pgtable);
 744
 745                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 746                 goto out;
 747         }
 748         src_page = pmd_page(pmd);
 749         VM_BUG_ON(!PageHead(src_page));
 750         get_page(src_page);
 751         page_dup_rmap(src_page);
 752         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 753
 754         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 755         pmd = pmd_mkold(pmd_wrprotect(pmd));
 756         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 757         prepare_pmd_huge_pte(pgtable, dst_mm);
 758
 759         ret = 0;
 760 out_unlock:
 761         spin_unlock(&src_mm->page_table_lock);
 762         spin_unlock(&dst_mm->page_table_lock);
 763 out:
 764         return ret;
 765 }
 766
 767 /* no "address" argument so destroys page coloring of some arch */
 768 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 769 {
 770         pgtable_t pgtable;
 771
 772         assert_spin_locked(&mm->page_table_lock);
 773
 774         /* FIFO */
 775         pgtable = mm->pmd_huge_pte;
 776         if (list_empty(&pgtable->lru))
 777                 mm->pmd_huge_pte = NULL;
 778         else {
 779                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 780                                               struct page, lru);
 781                 list_del(&pgtable->lru);
 782         }
 783         return pgtable;
 784 }
 785
 786 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 787                                         struct vm_area_struct *vma,
 788                                         unsigned long address,
 789                                         pmd_t *pmd, pmd_t orig_pmd,
 790                                         struct page *page,
 791                                         unsigned long haddr)
 792 {
 793         pgtable_t pgtable;
 794         pmd_t _pmd;
 795         int ret = 0, i;
 796         struct page **pages;
 797
 798         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 799                         GFP_KERNEL);
 800         if (unlikely(!pages)) {
 801                 ret |= VM_FAULT_OOM;
 802                 goto out;
 803         }
 804
 805         for (i = 0; i < HPAGE_PMD_NR; i++) {
 806                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE,
 807                                                vma, address, page_to_nid(page));
 808                 if (unlikely(!pages[i] ||
 809                              mem_cgroup_newpage_charge(pages[i], mm,
 810                                                        GFP_KERNEL))) {
 811                         if (pages[i])
 812                                 put_page(pages[i]);
 813                         mem_cgroup_uncharge_start();
 814                         while (--i >= 0) {
 815                                 mem_cgroup_uncharge_page(pages[i]);
 816                                 put_page(pages[i]);
 817                         }
 818                         mem_cgroup_uncharge_end();
 819                         kfree(pages);
 820                         ret |= VM_FAULT_OOM;
 821                         goto out;
 822                 }
 823         }
 824
 825         for (i = 0; i < HPAGE_PMD_NR; i++) {
 826                 copy_user_highpage(pages[i], page + i,
 827                                    haddr + PAGE_SHIFT*i, vma);
 828                 __SetPageUptodate(pages[i]);
 829                 cond_resched();
 830         }
 831
 832         spin_lock(&mm->page_table_lock);
 833         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 834                 goto out_free_pages;
 835         VM_BUG_ON(!PageHead(page));
 836
 837         pmdp_clear_flush_notify(vma, haddr, pmd);
 838         /* leave pmd empty until pte is filled */
 839
 840         pgtable = get_pmd_huge_pte(mm);
 841         pmd_populate(mm, &_pmd, pgtable);
 842
 843         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 844                 pte_t *pte, entry;
 845                 entry = mk_pte(pages[i], vma->vm_page_prot);
 846                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 847                 page_add_new_anon_rmap(pages[i], vma, haddr);
 848                 pte = pte_offset_map(&_pmd, haddr);
 849                 VM_BUG_ON(!pte_none(*pte));
 850                 set_pte_at(mm, haddr, pte, entry);
 851                 pte_unmap(pte);
 852         }
 853         kfree(pages);
 854
 855         mm->nr_ptes++;
 856         smp_wmb(); /* make pte visible before pmd */
 857         pmd_populate(mm, pmd, pgtable);
 858         page_remove_rmap(page);
 859         spin_unlock(&mm->page_table_lock);
 860
 861         ret |= VM_FAULT_WRITE;
 862         put_page(page);
 863
 864 out:
 865         return ret;
 866
 867 out_free_pages:
 868         spin_unlock(&mm->page_table_lock);
 869         mem_cgroup_uncharge_start();
 870         for (i = 0; i < HPAGE_PMD_NR; i++) {
 871                 mem_cgroup_uncharge_page(pages[i]);
 872                 put_page(pages[i]);
 873         }
 874         mem_cgroup_uncharge_end();
 875         kfree(pages);
 876         goto out;
 877 }
 878
 879 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 880                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 881 {
 882         int ret = 0;
 883         struct page *page, *new_page;
 884         unsigned long haddr;
 885
 886         VM_BUG_ON(!vma->anon_vma);
 887         spin_lock(&mm->page_table_lock);
 888         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 889                 goto out_unlock;
 890
 891         page = pmd_page(orig_pmd);
 892         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 893         haddr = address & HPAGE_PMD_MASK;
 894         if (page_mapcount(page) == 1) {
 895                 pmd_t entry;
 896                 entry = pmd_mkyoung(orig_pmd);
 897                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 898                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 899                         update_mmu_cache(vma, address, entry);
 900                 ret |= VM_FAULT_WRITE;
 901                 goto out_unlock;
 902         }
 903         get_page(page);
 904         spin_unlock(&mm->page_table_lock);
 905
 906         if (transparent_hugepage_enabled(vma) &&
 907             !transparent_hugepage_debug_cow())
 908                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 909                                               vma, haddr, numa_node_id());
 910         else
 911                 new_page = NULL;
 912
 913         if (unlikely(!new_page)) {
 914                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 915                                                    pmd, orig_pmd, page, haddr);
 916                 put_page(page);
 917                 goto out;
 918         }
 919
 920         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 921                 put_page(new_page);
 922                 put_page(page);
 923                 ret |= VM_FAULT_OOM;
 924                 goto out;
 925         }
 926
 927         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 928         __SetPageUptodate(new_page);
 929
 930         spin_lock(&mm->page_table_lock);
 931         put_page(page);
 932         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 933                 mem_cgroup_uncharge_page(new_page);
 934                 put_page(new_page);
 935         } else {
 936                 pmd_t entry;
 937                 VM_BUG_ON(!PageHead(page));
 938                 entry = mk_pmd(new_page, vma->vm_page_prot);
 939                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 940                 entry = pmd_mkhuge(entry);
 941                 pmdp_clear_flush_notify(vma, haddr, pmd);
 942                 page_add_new_anon_rmap(new_page, vma, haddr);
 943                 set_pmd_at(mm, haddr, pmd, entry);
 944                 update_mmu_cache(vma, address, entry);
 945                 page_remove_rmap(page);
 946                 put_page(page);
 947                 ret |= VM_FAULT_WRITE;
 948         }
 949 out_unlock:
 950         spin_unlock(&mm->page_table_lock);
 951 out:
 952         return ret;
 953 }
 954
 955 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 956                                    unsigned long addr,
 957                                    pmd_t *pmd,
 958                                    unsigned int flags)
 959 {
 960         struct page *page = NULL;
 961
 962         assert_spin_locked(&mm->page_table_lock);
 963
 964         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 965                 goto out;
 966
 967         page = pmd_page(*pmd);
 968         VM_BUG_ON(!PageHead(page));
 969         if (flags & FOLL_TOUCH) {
 970                 pmd_t _pmd;
 971                 /*
 972                  * We should set the dirty bit only for FOLL_WRITE but
 973                  * for now the dirty bit in the pmd is meaningless.
 974                  * And if the dirty bit will become meaningful and
 975                  * we'll only set it with FOLL_WRITE, an atomic
 976                  * set_bit will be required on the pmd to set the
 977                  * young bit, instead of the current set_pmd_at.
 978                  */
 979                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 980                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
 981         }
 982         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
 983         VM_BUG_ON(!PageCompound(page));
 984         if (flags & FOLL_GET)
 985                 get_page(page);
 986
 987 out:
 988         return page;
 989 }
 990
 991 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 992                  pmd_t *pmd)
 993 {
 994         int ret = 0;
 995
 996         spin_lock(&tlb->mm->page_table_lock);
 997         if (likely(pmd_trans_huge(*pmd))) {
 998                 if (unlikely(pmd_trans_splitting(*pmd))) {
 999                         spin_unlock(&tlb->mm->page_table_lock);
1000                         wait_split_huge_page(vma->anon_vma,
1001                                              pmd);
1002                 } else {
1003                         struct page *page;
1004                         pgtable_t pgtable;
1005                         pgtable = get_pmd_huge_pte(tlb->mm);
1006                         page = pmd_page(*pmd);
1007                         pmd_clear(pmd);
1008                         page_remove_rmap(page);
1009                         VM_BUG_ON(page_mapcount(page) < 0);
1010                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1011                         VM_BUG_ON(!PageHead(page));
1012                         spin_unlock(&tlb->mm->page_table_lock);
1013                         tlb_remove_page(tlb, page);
1014                         pte_free(tlb->mm, pgtable);
1015                         ret = 1;
1016                 }
1017         } else
1018                 spin_unlock(&tlb->mm->page_table_lock);
1019
1020         return ret;
1021 }
1022
1023 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1024                 unsigned long addr, unsigned long end,
1025                 unsigned char *vec)
1026 {
1027         int ret = 0;
1028
1029         spin_lock(&vma->vm_mm->page_table_lock);
1030         if (likely(pmd_trans_huge(*pmd))) {
1031                 ret = !pmd_trans_splitting(*pmd);
1032                 spin_unlock(&vma->vm_mm->page_table_lock);
1033                 if (unlikely(!ret))
1034                         wait_split_huge_page(vma->anon_vma, pmd);
1035                 else {
1036                         /*
1037                          * All logical pages in the range are present
1038                          * if backed by a huge page.
1039                          */
1040                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1041                 }
1042         } else
1043                 spin_unlock(&vma->vm_mm->page_table_lock);
1044
1045         return ret;
1046 }
1047
1048 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1049                 unsigned long addr, pgprot_t newprot)
1050 {
1051         struct mm_struct *mm = vma->vm_mm;
1052         int ret = 0;
1053
1054         spin_lock(&mm->page_table_lock);
1055         if (likely(pmd_trans_huge(*pmd))) {
1056                 if (unlikely(pmd_trans_splitting(*pmd))) {
1057                         spin_unlock(&mm->page_table_lock);
1058                         wait_split_huge_page(vma->anon_vma, pmd);
1059                 } else {
1060                         pmd_t entry;
1061
1062                         entry = pmdp_get_and_clear(mm, addr, pmd);
1063                         entry = pmd_modify(entry, newprot);
1064                         set_pmd_at(mm, addr, pmd, entry);
1065                         spin_unlock(&vma->vm_mm->page_table_lock);
1066                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1067                         ret = 1;
1068                 }
1069         } else
1070                 spin_unlock(&vma->vm_mm->page_table_lock);
1071
1072         return ret;
1073 }
1074
1075 pmd_t *page_check_address_pmd(struct page *page,
1076                               struct mm_struct *mm,
1077                               unsigned long address,
1078                               enum page_check_address_pmd_flag flag)
1079 {
1080         pgd_t *pgd;
1081         pud_t *pud;
1082         pmd_t *pmd, *ret = NULL;
1083
1084         if (address & ~HPAGE_PMD_MASK)
1085                 goto out;
1086
1087         pgd = pgd_offset(mm, address);
1088         if (!pgd_present(*pgd))
1089                 goto out;
1090
1091         pud = pud_offset(pgd, address);
1092         if (!pud_present(*pud))
1093                 goto out;
1094
1095         pmd = pmd_offset(pud, address);
1096         if (pmd_none(*pmd))
1097                 goto out;
1098         if (pmd_page(*pmd) != page)
1099                 goto out;
1100         /*
1101          * split_vma() may create temporary aliased mappings. There is
1102          * no risk as long as all huge pmd are found and have their
1103          * splitting bit set before __split_huge_page_refcount
1104          * runs. Finding the same huge pmd more than once during the
1105          * same rmap walk is not a problem.
1106          */
1107         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1108             pmd_trans_splitting(*pmd))
1109                 goto out;
1110         if (pmd_trans_huge(*pmd)) {
1111                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1112                           !pmd_trans_splitting(*pmd));
1113                 ret = pmd;
1114         }
1115 out:
1116         return ret;
1117 }
1118
1119 static int __split_huge_page_splitting(struct page *page,
1120                                        struct vm_area_struct *vma,
1121                                        unsigned long address)
1122 {
1123         struct mm_struct *mm = vma->vm_mm;
1124         pmd_t *pmd;
1125         int ret = 0;
1126
1127         spin_lock(&mm->page_table_lock);
1128         pmd = page_check_address_pmd(page, mm, address,
1129                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1130         if (pmd) {
1131                 /*
1132                  * We can't temporarily set the pmd to null in order
1133                  * to split it, the pmd must remain marked huge at all
1134                  * times or the VM won't take the pmd_trans_huge paths
1135                  * and it won't wait on the anon_vma->root->lock to
1136                  * serialize against split_huge_page*.
1137                  */
1138                 pmdp_splitting_flush_notify(vma, address, pmd);
1139                 ret = 1;
1140         }
1141         spin_unlock(&mm->page_table_lock);
1142
1143         return ret;
1144 }
1145
1146 static void __split_huge_page_refcount(struct page *page)
1147 {
1148         int i;
1149         unsigned long head_index = page->index;
1150         struct zone *zone = page_zone(page);
1151         int zonestat;
1152
1153         /* prevent PageLRU to go away from under us, and freeze lru stats */
1154         spin_lock_irq(&zone->lru_lock);
1155         compound_lock(page);
1156
1157         for (i = 1; i < HPAGE_PMD_NR; i++) {
1158                 struct page *page_tail = page + i;
1159
1160                 /* tail_page->_count cannot change */
1161                 atomic_sub(atomic_read(&page_tail->_count), &page->_count);
1162                 BUG_ON(page_count(page) <= 0);
1163                 atomic_add(page_mapcount(page) + 1, &page_tail->_count);
1164                 BUG_ON(atomic_read(&page_tail->_count) <= 0);
1165
1166                 /* after clearing PageTail the gup refcount can be released */
1167                 smp_mb();
1168
1169                 /*
1170                  * retain hwpoison flag of the poisoned tail page:
1171                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1172                  *   by the memory-failure.
1173                  */
1174                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1175                 page_tail->flags |= (page->flags &
1176                                      ((1L << PG_referenced) |
1177                                       (1L << PG_swapbacked) |
1178                                       (1L << PG_mlocked) |
1179                                       (1L << PG_uptodate)));
1180                 page_tail->flags |= (1L << PG_dirty);
1181
1182                 /*
1183                  * 1) clear PageTail before overwriting first_page
1184                  * 2) clear PageTail before clearing PageHead for VM_BUG_ON
1185                  */
1186                 smp_wmb();
1187
1188                 /*
1189                  * __split_huge_page_splitting() already set the
1190                  * splitting bit in all pmd that could map this
1191                  * hugepage, that will ensure no CPU can alter the
1192                  * mapcount on the head page. The mapcount is only
1193                  * accounted in the head page and it has to be
1194                  * transferred to all tail pages in the below code. So
1195                  * for this code to be safe, the split the mapcount
1196                  * can't change. But that doesn't mean userland can't
1197                  * keep changing and reading the page contents while
1198                  * we transfer the mapcount, so the pmd splitting
1199                  * status is achieved setting a reserved bit in the
1200                  * pmd, not by clearing the present bit.
1201                 */
1202                 BUG_ON(page_mapcount(page_tail));
1203                 page_tail->_mapcount = page->_mapcount;
1204
1205                 BUG_ON(page_tail->mapping);
1206                 page_tail->mapping = page->mapping;
1207
1208                 page_tail->index = ++head_index;
1209
1210                 BUG_ON(!PageAnon(page_tail));
1211                 BUG_ON(!PageUptodate(page_tail));
1212                 BUG_ON(!PageDirty(page_tail));
1213                 BUG_ON(!PageSwapBacked(page_tail));
1214
1215                 mem_cgroup_split_huge_fixup(page, page_tail);
1216
1217                 lru_add_page_tail(zone, page, page_tail);
1218         }
1219
1220         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1221         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1222
1223         /*
1224          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1225          * so adjust those appropriately if this page is on the LRU.
1226          */
1227         if (PageLRU(page)) {
1228                 zonestat = NR_LRU_BASE + page_lru(page);
1229                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1230         }
1231
1232         ClearPageCompound(page);
1233         compound_unlock(page);
1234         spin_unlock_irq(&zone->lru_lock);
1235
1236         for (i = 1; i < HPAGE_PMD_NR; i++) {
1237                 struct page *page_tail = page + i;
1238                 BUG_ON(page_count(page_tail) <= 0);
1239                 /*
1240                  * Tail pages may be freed if there wasn't any mapping
1241                  * like if add_to_swap() is running on a lru page that
1242                  * had its mapping zapped. And freeing these pages
1243                  * requires taking the lru_lock so we do the put_page
1244                  * of the tail pages after the split is complete.
1245                  */
1246                 put_page(page_tail);
1247         }
1248
1249         /*
1250          * Only the head page (now become a regular page) is required
1251          * to be pinned by the caller.
1252          */
1253         BUG_ON(page_count(page) <= 0);
1254 }
1255
1256 static int __split_huge_page_map(struct page *page,
1257                                  struct vm_area_struct *vma,
1258                                  unsigned long address)
1259 {
1260         struct mm_struct *mm = vma->vm_mm;
1261         pmd_t *pmd, _pmd;
1262         int ret = 0, i;
1263         pgtable_t pgtable;
1264         unsigned long haddr;
1265
1266         spin_lock(&mm->page_table_lock);
1267         pmd = page_check_address_pmd(page, mm, address,
1268                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1269         if (pmd) {
1270                 pgtable = get_pmd_huge_pte(mm);
1271                 pmd_populate(mm, &_pmd, pgtable);
1272
1273                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1274                      i++, haddr += PAGE_SIZE) {
1275                         pte_t *pte, entry;
1276                         BUG_ON(PageCompound(page+i));
1277                         entry = mk_pte(page + i, vma->vm_page_prot);
1278                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1279                         if (!pmd_write(*pmd))
1280                                 entry = pte_wrprotect(entry);
1281                         else
1282                                 BUG_ON(page_mapcount(page) != 1);
1283                         if (!pmd_young(*pmd))
1284                                 entry = pte_mkold(entry);
1285                         pte = pte_offset_map(&_pmd, haddr);
1286                         BUG_ON(!pte_none(*pte));
1287                         set_pte_at(mm, haddr, pte, entry);
1288                         pte_unmap(pte);
1289                 }
1290
1291                 mm->nr_ptes++;
1292                 smp_wmb(); /* make pte visible before pmd */
1293                 /*
1294                  * Up to this point the pmd is present and huge and
1295                  * userland has the whole access to the hugepage
1296                  * during the split (which happens in place). If we
1297                  * overwrite the pmd with the not-huge version
1298                  * pointing to the pte here (which of course we could
1299                  * if all CPUs were bug free), userland could trigger
1300                  * a small page size TLB miss on the small sized TLB
1301                  * while the hugepage TLB entry is still established
1302                  * in the huge TLB. Some CPU doesn't like that. See
1303                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1304                  * Erratum 383 on page 93. Intel should be safe but is
1305                  * also warns that it's only safe if the permission
1306                  * and cache attributes of the two entries loaded in
1307                  * the two TLB is identical (which should be the case
1308                  * here). But it is generally safer to never allow
1309                  * small and huge TLB entries for the same virtual
1310                  * address to be loaded simultaneously. So instead of
1311                  * doing "pmd_populate(); flush_tlb_range();" we first
1312                  * mark the current pmd notpresent (atomically because
1313                  * here the pmd_trans_huge and pmd_trans_splitting
1314                  * must remain set at all times on the pmd until the
1315                  * split is complete for this pmd), then we flush the
1316                  * SMP TLB and finally we write the non-huge version
1317                  * of the pmd entry with pmd_populate.
1318                  */
1319                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1320                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1321                 pmd_populate(mm, pmd, pgtable);
1322                 ret = 1;
1323         }
1324         spin_unlock(&mm->page_table_lock);
1325
1326         return ret;
1327 }
1328
1329 /* must be called with anon_vma->root->lock hold */
1330 static void __split_huge_page(struct page *page,
1331                               struct anon_vma *anon_vma)
1332 {
1333         int mapcount, mapcount2;
1334         struct anon_vma_chain *avc;
1335
1336         BUG_ON(!PageHead(page));
1337         BUG_ON(PageTail(page));
1338
1339         mapcount = 0;
1340         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1341                 struct vm_area_struct *vma = avc->vma;
1342                 unsigned long addr = vma_address(page, vma);
1343                 BUG_ON(is_vma_temporary_stack(vma));
1344                 if (addr == -EFAULT)
1345                         continue;
1346                 mapcount += __split_huge_page_splitting(page, vma, addr);
1347         }
1348         /*
1349          * It is critical that new vmas are added to the tail of the
1350          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1351          * and establishes a child pmd before
1352          * __split_huge_page_splitting() freezes the parent pmd (so if
1353          * we fail to prevent copy_huge_pmd() from running until the
1354          * whole __split_huge_page() is complete), we will still see
1355          * the newly established pmd of the child later during the
1356          * walk, to be able to set it as pmd_trans_splitting too.
1357          */
1358         if (mapcount != page_mapcount(page))
1359                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1360                        mapcount, page_mapcount(page));
1361         BUG_ON(mapcount != page_mapcount(page));
1362
1363         __split_huge_page_refcount(page);
1364
1365         mapcount2 = 0;
1366         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1367                 struct vm_area_struct *vma = avc->vma;
1368                 unsigned long addr = vma_address(page, vma);
1369                 BUG_ON(is_vma_temporary_stack(vma));
1370                 if (addr == -EFAULT)
1371                         continue;
1372                 mapcount2 += __split_huge_page_map(page, vma, addr);
1373         }
1374         if (mapcount != mapcount2)
1375                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1376                        mapcount, mapcount2, page_mapcount(page));
1377         BUG_ON(mapcount != mapcount2);
1378 }
1379
1380 int split_huge_page(struct page *page)
1381 {
1382         struct anon_vma *anon_vma;
1383         int ret = 1;
1384
1385         BUG_ON(!PageAnon(page));
1386         anon_vma = page_lock_anon_vma(page);
1387         if (!anon_vma)
1388                 goto out;
1389         ret = 0;
1390         if (!PageCompound(page))
1391                 goto out_unlock;
1392
1393         BUG_ON(!PageSwapBacked(page));
1394         __split_huge_page(page, anon_vma);
1395
1396         BUG_ON(PageCompound(page));
1397 out_unlock:
1398         page_unlock_anon_vma(anon_vma);
1399 out:
1400         return ret;
1401 }
1402
1403 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1404                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1405
1406 int hugepage_madvise(struct vm_area_struct *vma,
1407                      unsigned long *vm_flags, int advice)
1408 {
1409         switch (advice) {
1410         case MADV_HUGEPAGE:
1411                 /*
1412                  * Be somewhat over-protective like KSM for now!
1413                  */
1414                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1415                         return -EINVAL;
1416                 *vm_flags &= ~VM_NOHUGEPAGE;
1417                 *vm_flags |= VM_HUGEPAGE;
1418                 /*
1419                  * If the vma become good for khugepaged to scan,
1420                  * register it here without waiting a page fault that
1421                  * may not happen any time soon.
1422                  */
1423                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1424                         return -ENOMEM;
1425                 break;
1426         case MADV_NOHUGEPAGE:
1427                 /*
1428                  * Be somewhat over-protective like KSM for now!
1429                  */
1430                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1431                         return -EINVAL;
1432                 *vm_flags &= ~VM_HUGEPAGE;
1433                 *vm_flags |= VM_NOHUGEPAGE;
1434                 /*
1435                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1436                  * this vma even if we leave the mm registered in khugepaged if
1437                  * it got registered before VM_NOHUGEPAGE was set.
1438                  */
1439                 break;
1440         }
1441
1442         return 0;
1443 }
1444
1445 static int __init khugepaged_slab_init(void)
1446 {
1447         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1448                                           sizeof(struct mm_slot),
1449                                           __alignof__(struct mm_slot), 0, NULL);
1450         if (!mm_slot_cache)
1451                 return -ENOMEM;
1452
1453         return 0;
1454 }
1455
1456 static void __init khugepaged_slab_free(void)
1457 {
1458         kmem_cache_destroy(mm_slot_cache);
1459         mm_slot_cache = NULL;
1460 }
1461
1462 static inline struct mm_slot *alloc_mm_slot(void)
1463 {
1464         if (!mm_slot_cache)     /* initialization failed */
1465                 return NULL;
1466         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1467 }
1468
1469 static inline void free_mm_slot(struct mm_slot *mm_slot)
1470 {
1471         kmem_cache_free(mm_slot_cache, mm_slot);
1472 }
1473
1474 static int __init mm_slots_hash_init(void)
1475 {
1476         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1477                                 GFP_KERNEL);
1478         if (!mm_slots_hash)
1479                 return -ENOMEM;
1480         return 0;
1481 }
1482
1483 #if 0
1484 static void __init mm_slots_hash_free(void)
1485 {
1486         kfree(mm_slots_hash);
1487         mm_slots_hash = NULL;
1488 }
1489 #endif
1490
1491 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1492 {
1493         struct mm_slot *mm_slot;
1494         struct hlist_head *bucket;
1495         struct hlist_node *node;
1496
1497         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1498                                 % MM_SLOTS_HASH_HEADS];
1499         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1500                 if (mm == mm_slot->mm)
1501                         return mm_slot;
1502         }
1503         return NULL;
1504 }
1505
1506 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1507                                     struct mm_slot *mm_slot)
1508 {
1509         struct hlist_head *bucket;
1510
1511         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1512                                 % MM_SLOTS_HASH_HEADS];
1513         mm_slot->mm = mm;
1514         hlist_add_head(&mm_slot->hash, bucket);
1515 }
1516
1517 static inline int khugepaged_test_exit(struct mm_struct *mm)
1518 {
1519         return atomic_read(&mm->mm_users) == 0;
1520 }
1521
1522 int __khugepaged_enter(struct mm_struct *mm)
1523 {
1524         struct mm_slot *mm_slot;
1525         int wakeup;
1526
1527         mm_slot = alloc_mm_slot();
1528         if (!mm_slot)
1529                 return -ENOMEM;
1530
1531         /* __khugepaged_exit() must not run from under us */
1532         VM_BUG_ON(khugepaged_test_exit(mm));
1533         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1534                 free_mm_slot(mm_slot);
1535                 return 0;
1536         }
1537
1538         spin_lock(&khugepaged_mm_lock);
1539         insert_to_mm_slots_hash(mm, mm_slot);
1540         /*
1541          * Insert just behind the scanning cursor, to let the area settle
1542          * down a little.
1543          */
1544         wakeup = list_empty(&khugepaged_scan.mm_head);
1545         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1546         spin_unlock(&khugepaged_mm_lock);
1547
1548         atomic_inc(&mm->mm_count);
1549         if (wakeup)
1550                 wake_up_interruptible(&khugepaged_wait);
1551
1552         return 0;
1553 }
1554
1555 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1556 {
1557         unsigned long hstart, hend;
1558         if (!vma->anon_vma)
1559                 /*
1560                  * Not yet faulted in so we will register later in the
1561                  * page fault if needed.
1562                  */
1563                 return 0;
1564         if (vma->vm_ops)
1565                 /* khugepaged not yet working on file or special mappings */
1566                 return 0;
1567         /*
1568          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1569          * true too, verify it here.
1570          */
1571         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1572         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1573         hend = vma->vm_end & HPAGE_PMD_MASK;
1574         if (hstart < hend)
1575                 return khugepaged_enter(vma);
1576         return 0;
1577 }
1578
1579 void __khugepaged_exit(struct mm_struct *mm)
1580 {
1581         struct mm_slot *mm_slot;
1582         int free = 0;
1583
1584         spin_lock(&khugepaged_mm_lock);
1585         mm_slot = get_mm_slot(mm);
1586         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1587                 hlist_del(&mm_slot->hash);
1588                 list_del(&mm_slot->mm_node);
1589                 free = 1;
1590         }
1591
1592         if (free) {
1593                 spin_unlock(&khugepaged_mm_lock);
1594                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1595                 free_mm_slot(mm_slot);
1596                 mmdrop(mm);
1597         } else if (mm_slot) {
1598                 spin_unlock(&khugepaged_mm_lock);
1599                 /*
1600                  * This is required to serialize against
1601                  * khugepaged_test_exit() (which is guaranteed to run
1602                  * under mmap sem read mode). Stop here (after we
1603                  * return all pagetables will be destroyed) until
1604                  * khugepaged has finished working on the pagetables
1605                  * under the mmap_sem.
1606                  */
1607                 down_write(&mm->mmap_sem);
1608                 up_write(&mm->mmap_sem);
1609         } else
1610                 spin_unlock(&khugepaged_mm_lock);
1611 }
1612
1613 static void release_pte_page(struct page *page)
1614 {
1615         /* 0 stands for page_is_file_cache(page) == false */
1616         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1617         unlock_page(page);
1618         putback_lru_page(page);
1619 }
1620
1621 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1622 {
1623         while (--_pte >= pte) {
1624                 pte_t pteval = *_pte;
1625                 if (!pte_none(pteval))
1626                         release_pte_page(pte_page(pteval));
1627         }
1628 }
1629
1630 static void release_all_pte_pages(pte_t *pte)
1631 {
1632         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1633 }
1634
1635 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1636                                         unsigned long address,
1637                                         pte_t *pte)
1638 {
1639         struct page *page;
1640         pte_t *_pte;
1641         int referenced = 0, isolated = 0, none = 0;
1642         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1643              _pte++, address += PAGE_SIZE) {
1644                 pte_t pteval = *_pte;
1645                 if (pte_none(pteval)) {
1646                         if (++none <= khugepaged_max_ptes_none)
1647                                 continue;
1648                         else {
1649                                 release_pte_pages(pte, _pte);
1650                                 goto out;
1651                         }
1652                 }
1653                 if (!pte_present(pteval) || !pte_write(pteval)) {
1654                         release_pte_pages(pte, _pte);
1655                         goto out;
1656                 }
1657                 page = vm_normal_page(vma, address, pteval);
1658                 if (unlikely(!page)) {
1659                         release_pte_pages(pte, _pte);
1660                         goto out;
1661                 }
1662                 VM_BUG_ON(PageCompound(page));
1663                 BUG_ON(!PageAnon(page));
1664                 VM_BUG_ON(!PageSwapBacked(page));
1665
1666                 /* cannot use mapcount: can't collapse if there's a gup pin */
1667                 if (page_count(page) != 1) {
1668                         release_pte_pages(pte, _pte);
1669                         goto out;
1670                 }
1671                 /*
1672                  * We can do it before isolate_lru_page because the
1673                  * page can't be freed from under us. NOTE: PG_lock
1674                  * is needed to serialize against split_huge_page
1675                  * when invoked from the VM.
1676                  */
1677                 if (!trylock_page(page)) {
1678                         release_pte_pages(pte, _pte);
1679                         goto out;
1680                 }
1681                 /*
1682                  * Isolate the page to avoid collapsing an hugepage
1683                  * currently in use by the VM.
1684                  */
1685                 if (isolate_lru_page(page)) {
1686                         unlock_page(page);
1687                         release_pte_pages(pte, _pte);
1688                         goto out;
1689                 }
1690                 /* 0 stands for page_is_file_cache(page) == false */
1691                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1692                 VM_BUG_ON(!PageLocked(page));
1693                 VM_BUG_ON(PageLRU(page));
1694
1695                 /* If there is no mapped pte young don't collapse the page */
1696                 if (pte_young(pteval) || PageReferenced(page) ||
1697                     mmu_notifier_test_young(vma->vm_mm, address))
1698                         referenced = 1;
1699         }
1700         if (unlikely(!referenced))
1701                 release_all_pte_pages(pte);
1702         else
1703                 isolated = 1;
1704 out:
1705         return isolated;
1706 }
1707
1708 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1709                                       struct vm_area_struct *vma,
1710                                       unsigned long address,
1711                                       spinlock_t *ptl)
1712 {
1713         pte_t *_pte;
1714         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1715                 pte_t pteval = *_pte;
1716                 struct page *src_page;
1717
1718                 if (pte_none(pteval)) {
1719                         clear_user_highpage(page, address);
1720                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1721                 } else {
1722                         src_page = pte_page(pteval);
1723                         copy_user_highpage(page, src_page, address, vma);
1724                         VM_BUG_ON(page_mapcount(src_page) != 1);
1725                         VM_BUG_ON(page_count(src_page) != 2);
1726                         release_pte_page(src_page);
1727                         /*
1728                          * ptl mostly unnecessary, but preempt has to
1729                          * be disabled to update the per-cpu stats
1730                          * inside page_remove_rmap().
1731                          */
1732                         spin_lock(ptl);
1733                         /*
1734                          * paravirt calls inside pte_clear here are
1735                          * superfluous.
1736                          */
1737                         pte_clear(vma->vm_mm, address, _pte);
1738                         page_remove_rmap(src_page);
1739                         spin_unlock(ptl);
1740                         free_page_and_swap_cache(src_page);
1741                 }
1742
1743                 address += PAGE_SIZE;
1744                 page++;
1745         }
1746 }
1747
1748 static void collapse_huge_page(struct mm_struct *mm,
1749                                unsigned long address,
1750                                struct page **hpage,
1751                                struct vm_area_struct *vma,
1752                                int node)
1753 {
1754         pgd_t *pgd;
1755         pud_t *pud;
1756         pmd_t *pmd, _pmd;
1757         pte_t *pte;
1758         pgtable_t pgtable;
1759         struct page *new_page;
1760         spinlock_t *ptl;
1761         int isolated;
1762         unsigned long hstart, hend;
1763
1764         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1765 #ifndef CONFIG_NUMA
1766         VM_BUG_ON(!*hpage);
1767         new_page = *hpage;
1768         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1769                 up_read(&mm->mmap_sem);
1770                 return;
1771         }
1772 #else
1773         VM_BUG_ON(*hpage);
1774         /*
1775          * Allocate the page while the vma is still valid and under
1776          * the mmap_sem read mode so there is no memory allocation
1777          * later when we take the mmap_sem in write mode. This is more
1778          * friendly behavior (OTOH it may actually hide bugs) to
1779          * filesystems in userland with daemons allocating memory in
1780          * the userland I/O paths.  Allocating memory with the
1781          * mmap_sem in read mode is good idea also to allow greater
1782          * scalability.
1783          */
1784         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1785                                       node);
1786         if (unlikely(!new_page)) {
1787                 up_read(&mm->mmap_sem);
1788                 *hpage = ERR_PTR(-ENOMEM);
1789                 return;
1790         }
1791         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1792                 up_read(&mm->mmap_sem);
1793                 put_page(new_page);
1794                 return;
1795         }
1796 #endif
1797
1798         /* after allocating the hugepage upgrade to mmap_sem write mode */
1799         up_read(&mm->mmap_sem);
1800
1801         /*
1802          * Prevent all access to pagetables with the exception of
1803          * gup_fast later hanlded by the ptep_clear_flush and the VM
1804          * handled by the anon_vma lock + PG_lock.
1805          */
1806         down_write(&mm->mmap_sem);
1807         if (unlikely(khugepaged_test_exit(mm)))
1808                 goto out;
1809
1810         vma = find_vma(mm, address);
1811         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1812         hend = vma->vm_end & HPAGE_PMD_MASK;
1813         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1814                 goto out;
1815
1816         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1817             (vma->vm_flags & VM_NOHUGEPAGE))
1818                 goto out;
1819
1820         if (!vma->anon_vma || vma->vm_ops)
1821                 goto out;
1822         if (is_vma_temporary_stack(vma))
1823                 goto out;
1824         /*
1825          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1826          * true too, verify it here.
1827          */
1828         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1829
1830         pgd = pgd_offset(mm, address);
1831         if (!pgd_present(*pgd))
1832                 goto out;
1833
1834         pud = pud_offset(pgd, address);
1835         if (!pud_present(*pud))
1836                 goto out;
1837
1838         pmd = pmd_offset(pud, address);
1839         /* pmd can't go away or become huge under us */
1840         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1841                 goto out;
1842
1843         anon_vma_lock(vma->anon_vma);
1844
1845         pte = pte_offset_map(pmd, address);
1846         ptl = pte_lockptr(mm, pmd);
1847
1848         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1849         /*
1850          * After this gup_fast can't run anymore. This also removes
1851          * any huge TLB entry from the CPU so we won't allow
1852          * huge and small TLB entries for the same virtual address
1853          * to avoid the risk of CPU bugs in that area.
1854          */
1855         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1856         spin_unlock(&mm->page_table_lock);
1857
1858         spin_lock(ptl);
1859         isolated = __collapse_huge_page_isolate(vma, address, pte);
1860         spin_unlock(ptl);
1861
1862         if (unlikely(!isolated)) {
1863                 pte_unmap(pte);
1864                 spin_lock(&mm->page_table_lock);
1865                 BUG_ON(!pmd_none(*pmd));
1866                 set_pmd_at(mm, address, pmd, _pmd);
1867                 spin_unlock(&mm->page_table_lock);
1868                 anon_vma_unlock(vma->anon_vma);
1869                 goto out;
1870         }
1871
1872         /*
1873          * All pages are isolated and locked so anon_vma rmap
1874          * can't run anymore.
1875          */
1876         anon_vma_unlock(vma->anon_vma);
1877
1878         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1879         pte_unmap(pte);
1880         __SetPageUptodate(new_page);
1881         pgtable = pmd_pgtable(_pmd);
1882         VM_BUG_ON(page_count(pgtable) != 1);
1883         VM_BUG_ON(page_mapcount(pgtable) != 0);
1884
1885         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1886         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1887         _pmd = pmd_mkhuge(_pmd);
1888
1889         /*
1890          * spin_lock() below is not the equivalent of smp_wmb(), so
1891          * this is needed to avoid the copy_huge_page writes to become
1892          * visible after the set_pmd_at() write.
1893          */
1894         smp_wmb();
1895
1896         spin_lock(&mm->page_table_lock);
1897         BUG_ON(!pmd_none(*pmd));
1898         page_add_new_anon_rmap(new_page, vma, address);
1899         set_pmd_at(mm, address, pmd, _pmd);
1900         update_mmu_cache(vma, address, entry);
1901         prepare_pmd_huge_pte(pgtable, mm);
1902         mm->nr_ptes--;
1903         spin_unlock(&mm->page_table_lock);
1904
1905 #ifndef CONFIG_NUMA
1906         *hpage = NULL;
1907 #endif
1908         khugepaged_pages_collapsed++;
1909 out_up_write:
1910         up_write(&mm->mmap_sem);
1911         return;
1912
1913 out:
1914         mem_cgroup_uncharge_page(new_page);
1915 #ifdef CONFIG_NUMA
1916         put_page(new_page);
1917 #endif
1918         goto out_up_write;
1919 }
1920
1921 static int khugepaged_scan_pmd(struct mm_struct *mm,
1922                                struct vm_area_struct *vma,
1923                                unsigned long address,
1924                                struct page **hpage)
1925 {
1926         pgd_t *pgd;
1927         pud_t *pud;
1928         pmd_t *pmd;
1929         pte_t *pte, *_pte;
1930         int ret = 0, referenced = 0, none = 0;
1931         struct page *page;
1932         unsigned long _address;
1933         spinlock_t *ptl;
1934         int node = -1;
1935
1936         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1937
1938         pgd = pgd_offset(mm, address);
1939         if (!pgd_present(*pgd))
1940                 goto out;
1941
1942         pud = pud_offset(pgd, address);
1943         if (!pud_present(*pud))
1944                 goto out;
1945
1946         pmd = pmd_offset(pud, address);
1947         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1948                 goto out;
1949
1950         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1951         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1952              _pte++, _address += PAGE_SIZE) {
1953                 pte_t pteval = *_pte;
1954                 if (pte_none(pteval)) {
1955                         if (++none <= khugepaged_max_ptes_none)
1956                                 continue;
1957                         else
1958                                 goto out_unmap;
1959                 }
1960                 if (!pte_present(pteval) || !pte_write(pteval))
1961                         goto out_unmap;
1962                 page = vm_normal_page(vma, _address, pteval);
1963                 if (unlikely(!page))
1964                         goto out_unmap;
1965                 /*
1966                  * Chose the node of the first page. This could
1967                  * be more sophisticated and look at more pages,
1968                  * but isn't for now.
1969                  */
1970                 if (node == -1)
1971                         node = page_to_nid(page);
1972                 VM_BUG_ON(PageCompound(page));
1973                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
1974                         goto out_unmap;
1975                 /* cannot use mapcount: can't collapse if there's a gup pin */
1976                 if (page_count(page) != 1)
1977                         goto out_unmap;
1978                 if (pte_young(pteval) || PageReferenced(page) ||
1979                     mmu_notifier_test_young(vma->vm_mm, address))
1980                         referenced = 1;
1981         }
1982         if (referenced)
1983                 ret = 1;
1984 out_unmap:
1985         pte_unmap_unlock(pte, ptl);
1986         if (ret)
1987                 /* collapse_huge_page will return with the mmap_sem released */
1988                 collapse_huge_page(mm, address, hpage, vma, node);
1989 out:
1990         return ret;
1991 }
1992
1993 static void collect_mm_slot(struct mm_slot *mm_slot)
1994 {
1995         struct mm_struct *mm = mm_slot->mm;
1996
1997         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1998
1999         if (khugepaged_test_exit(mm)) {
2000                 /* free mm_slot */
2001                 hlist_del(&mm_slot->hash);
2002                 list_del(&mm_slot->mm_node);
2003
2004                 /*
2005                  * Not strictly needed because the mm exited already.
2006                  *
2007                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2008                  */
2009
2010                 /* khugepaged_mm_lock actually not necessary for the below */
2011                 free_mm_slot(mm_slot);
2012                 mmdrop(mm);
2013         }
2014 }
2015
2016 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2017                                             struct page **hpage)
2018 {
2019         struct mm_slot *mm_slot;
2020         struct mm_struct *mm;
2021         struct vm_area_struct *vma;
2022         int progress = 0;
2023
2024         VM_BUG_ON(!pages);
2025         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2026
2027         if (khugepaged_scan.mm_slot)
2028                 mm_slot = khugepaged_scan.mm_slot;
2029         else {
2030                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2031                                      struct mm_slot, mm_node);
2032                 khugepaged_scan.address = 0;
2033                 khugepaged_scan.mm_slot = mm_slot;
2034         }
2035         spin_unlock(&khugepaged_mm_lock);
2036
2037         mm = mm_slot->mm;
2038         down_read(&mm->mmap_sem);
2039         if (unlikely(khugepaged_test_exit(mm)))
2040                 vma = NULL;
2041         else
2042                 vma = find_vma(mm, khugepaged_scan.address);
2043
2044         progress++;
2045         for (; vma; vma = vma->vm_next) {
2046                 unsigned long hstart, hend;
2047
2048                 cond_resched();
2049                 if (unlikely(khugepaged_test_exit(mm))) {
2050                         progress++;
2051                         break;
2052                 }
2053
2054                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2055                      !khugepaged_always()) ||
2056                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2057                 skip:
2058                         progress++;
2059                         continue;
2060                 }
2061                 if (!vma->anon_vma || vma->vm_ops)
2062                         goto skip;
2063                 if (is_vma_temporary_stack(vma))
2064                         goto skip;
2065                 /*
2066                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2067                  * must be true too, verify it here.
2068                  */
2069                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2070                           vma->vm_flags & VM_NO_THP);
2071
2072                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2073                 hend = vma->vm_end & HPAGE_PMD_MASK;
2074                 if (hstart >= hend)
2075                         goto skip;
2076                 if (khugepaged_scan.address > hend)
2077                         goto skip;
2078                 if (khugepaged_scan.address < hstart)
2079                         khugepaged_scan.address = hstart;
2080                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2081
2082                 while (khugepaged_scan.address < hend) {
2083                         int ret;
2084                         cond_resched();
2085                         if (unlikely(khugepaged_test_exit(mm)))
2086                                 goto breakouterloop;
2087
2088                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2089                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2090                                   hend);
2091                         ret = khugepaged_scan_pmd(mm, vma,
2092                                                   khugepaged_scan.address,
2093                                                   hpage);
2094                         /* move to next address */
2095                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2096                         progress += HPAGE_PMD_NR;
2097                         if (ret)
2098                                 /* we released mmap_sem so break loop */
2099                                 goto breakouterloop_mmap_sem;
2100                         if (progress >= pages)
2101                                 goto breakouterloop;
2102                 }
2103         }
2104 breakouterloop:
2105         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2106 breakouterloop_mmap_sem:
2107
2108         spin_lock(&khugepaged_mm_lock);
2109         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2110         /*
2111          * Release the current mm_slot if this mm is about to die, or
2112          * if we scanned all vmas of this mm.
2113          */
2114         if (khugepaged_test_exit(mm) || !vma) {
2115                 /*
2116                  * Make sure that if mm_users is reaching zero while
2117                  * khugepaged runs here, khugepaged_exit will find
2118                  * mm_slot not pointing to the exiting mm.
2119                  */
2120                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2121                         khugepaged_scan.mm_slot = list_entry(
2122                                 mm_slot->mm_node.next,
2123                                 struct mm_slot, mm_node);
2124                         khugepaged_scan.address = 0;
2125                 } else {
2126                         khugepaged_scan.mm_slot = NULL;
2127                         khugepaged_full_scans++;
2128                 }
2129
2130                 collect_mm_slot(mm_slot);
2131         }
2132
2133         return progress;
2134 }
2135
2136 static int khugepaged_has_work(void)
2137 {
2138         return !list_empty(&khugepaged_scan.mm_head) &&
2139                 khugepaged_enabled();
2140 }
2141
2142 static int khugepaged_wait_event(void)
2143 {
2144         return !list_empty(&khugepaged_scan.mm_head) ||
2145                 !khugepaged_enabled();
2146 }
2147
2148 static void khugepaged_do_scan(struct page **hpage)
2149 {
2150         unsigned int progress = 0, pass_through_head = 0;
2151         unsigned int pages = khugepaged_pages_to_scan;
2152
2153         barrier(); /* write khugepaged_pages_to_scan to local stack */
2154
2155         while (progress < pages) {
2156                 cond_resched();
2157
2158 #ifndef CONFIG_NUMA
2159                 if (!*hpage) {
2160                         *hpage = alloc_hugepage(khugepaged_defrag());
2161                         if (unlikely(!*hpage))
2162                                 break;
2163                 }
2164 #else
2165                 if (IS_ERR(*hpage))
2166                         break;
2167 #endif
2168
2169                 if (unlikely(kthread_should_stop() || freezing(current)))
2170                         break;
2171
2172                 spin_lock(&khugepaged_mm_lock);
2173                 if (!khugepaged_scan.mm_slot)
2174                         pass_through_head++;
2175                 if (khugepaged_has_work() &&
2176                     pass_through_head < 2)
2177                         progress += khugepaged_scan_mm_slot(pages - progress,
2178                                                             hpage);
2179                 else
2180                         progress = pages;
2181                 spin_unlock(&khugepaged_mm_lock);
2182         }
2183 }
2184
2185 static void khugepaged_alloc_sleep(void)
2186 {
2187         DEFINE_WAIT(wait);
2188         add_wait_queue(&khugepaged_wait, &wait);
2189         schedule_timeout_interruptible(
2190                 msecs_to_jiffies(
2191                         khugepaged_alloc_sleep_millisecs));
2192         remove_wait_queue(&khugepaged_wait, &wait);
2193 }
2194
2195 #ifndef CONFIG_NUMA
2196 static struct page *khugepaged_alloc_hugepage(void)
2197 {
2198         struct page *hpage;
2199
2200         do {
2201                 hpage = alloc_hugepage(khugepaged_defrag());
2202                 if (!hpage)
2203                         khugepaged_alloc_sleep();
2204         } while (unlikely(!hpage) &&
2205                  likely(khugepaged_enabled()));
2206         return hpage;
2207 }
2208 #endif
2209
2210 static void khugepaged_loop(void)
2211 {
2212         struct page *hpage;
2213
2214 #ifdef CONFIG_NUMA
2215         hpage = NULL;
2216 #endif
2217         while (likely(khugepaged_enabled())) {
2218 #ifndef CONFIG_NUMA
2219                 hpage = khugepaged_alloc_hugepage();
2220                 if (unlikely(!hpage))
2221                         break;
2222 #else
2223                 if (IS_ERR(hpage)) {
2224                         khugepaged_alloc_sleep();
2225                         hpage = NULL;
2226                 }
2227 #endif
2228
2229                 khugepaged_do_scan(&hpage);
2230 #ifndef CONFIG_NUMA
2231                 if (hpage)
2232                         put_page(hpage);
2233 #endif
2234                 try_to_freeze();
2235                 if (unlikely(kthread_should_stop()))
2236                         break;
2237                 if (khugepaged_has_work()) {
2238                         DEFINE_WAIT(wait);
2239                         if (!khugepaged_scan_sleep_millisecs)
2240                                 continue;
2241                         add_wait_queue(&khugepaged_wait, &wait);
2242                         schedule_timeout_interruptible(
2243                                 msecs_to_jiffies(
2244                                         khugepaged_scan_sleep_millisecs));
2245                         remove_wait_queue(&khugepaged_wait, &wait);
2246                 } else if (khugepaged_enabled())
2247                         wait_event_freezable(khugepaged_wait,
2248                                              khugepaged_wait_event());
2249         }
2250 }
2251
2252 static int khugepaged(void *none)
2253 {
2254         struct mm_slot *mm_slot;
2255
2256         set_freezable();
2257         set_user_nice(current, 19);
2258
2259         /* serialize with start_khugepaged() */
2260         mutex_lock(&khugepaged_mutex);
2261
2262         for (;;) {
2263                 mutex_unlock(&khugepaged_mutex);
2264                 VM_BUG_ON(khugepaged_thread != current);
2265                 khugepaged_loop();
2266                 VM_BUG_ON(khugepaged_thread != current);
2267
2268                 mutex_lock(&khugepaged_mutex);
2269                 if (!khugepaged_enabled())
2270                         break;
2271                 if (unlikely(kthread_should_stop()))
2272                         break;
2273         }
2274
2275         spin_lock(&khugepaged_mm_lock);
2276         mm_slot = khugepaged_scan.mm_slot;
2277         khugepaged_scan.mm_slot = NULL;
2278         if (mm_slot)
2279                 collect_mm_slot(mm_slot);
2280         spin_unlock(&khugepaged_mm_lock);
2281
2282         khugepaged_thread = NULL;
2283         mutex_unlock(&khugepaged_mutex);
2284
2285         return 0;
2286 }
2287
2288 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2289 {
2290         struct page *page;
2291
2292         spin_lock(&mm->page_table_lock);
2293         if (unlikely(!pmd_trans_huge(*pmd))) {
2294                 spin_unlock(&mm->page_table_lock);
2295                 return;
2296         }
2297         page = pmd_page(*pmd);
2298         VM_BUG_ON(!page_count(page));
2299         get_page(page);
2300         spin_unlock(&mm->page_table_lock);
2301
2302         split_huge_page(page);
2303
2304         put_page(page);
2305         BUG_ON(pmd_trans_huge(*pmd));
2306 }
2307
2308 static void split_huge_page_address(struct mm_struct *mm,
2309                                     unsigned long address)
2310 {
2311         pgd_t *pgd;
2312         pud_t *pud;
2313         pmd_t *pmd;
2314
2315         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2316
2317         pgd = pgd_offset(mm, address);
2318         if (!pgd_present(*pgd))
2319                 return;
2320
2321         pud = pud_offset(pgd, address);
2322         if (!pud_present(*pud))
2323                 return;
2324
2325         pmd = pmd_offset(pud, address);
2326         if (!pmd_present(*pmd))
2327                 return;
2328         /*
2329          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2330          * materialize from under us.
2331          */
2332         split_huge_page_pmd(mm, pmd);
2333 }
2334
2335 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2336                              unsigned long start,
2337                              unsigned long end,
2338                              long adjust_next)
2339 {
2340         /*
2341          * If the new start address isn't hpage aligned and it could
2342          * previously contain an hugepage: check if we need to split
2343          * an huge pmd.
2344          */
2345         if (start & ~HPAGE_PMD_MASK &&
2346             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2347             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2348                 split_huge_page_address(vma->vm_mm, start);
2349
2350         /*
2351          * If the new end address isn't hpage aligned and it could
2352          * previously contain an hugepage: check if we need to split
2353          * an huge pmd.
2354          */
2355         if (end & ~HPAGE_PMD_MASK &&
2356             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2357             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2358                 split_huge_page_address(vma->vm_mm, end);
2359
2360         /*
2361          * If we're also updating the vma->vm_next->vm_start, if the new
2362          * vm_next->vm_start isn't page aligned and it could previously
2363          * contain an hugepage: check if we need to split an huge pmd.
2364          */
2365         if (adjust_next > 0) {
2366                 struct vm_area_struct *next = vma->vm_next;
2367                 unsigned long nstart = next->vm_start;
2368                 nstart += adjust_next << PAGE_SHIFT;
2369                 if (nstart & ~HPAGE_PMD_MASK &&
2370                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2371                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2372                         split_huge_page_address(next->vm_mm, nstart);
2373         }
2374 }