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                 mm->nr_ptes++;
 645                 spin_unlock(&mm->page_table_lock);
 646         }
 647
 648         return ret;
 649 }
 650
 651 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 652 {
 653         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 654 }
 655
 656 static inline struct page *alloc_hugepage_vma(int defrag,
 657                                               struct vm_area_struct *vma,
 658                                               unsigned long haddr, int nd,
 659                                               gfp_t extra_gfp)
 660 {
 661         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 662                                HPAGE_PMD_ORDER, vma, haddr, nd);
 663 }
 664
 665 #ifndef CONFIG_NUMA
 666 static inline struct page *alloc_hugepage(int defrag)
 667 {
 668         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 669                            HPAGE_PMD_ORDER);
 670 }
 671 #endif
 672
 673 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 674                                unsigned long address, pmd_t *pmd,
 675                                unsigned int flags)
 676 {
 677         struct page *page;
 678         unsigned long haddr = address & HPAGE_PMD_MASK;
 679         pte_t *pte;
 680
 681         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 682                 if (unlikely(anon_vma_prepare(vma)))
 683                         return VM_FAULT_OOM;
 684                 if (unlikely(khugepaged_enter(vma)))
 685                         return VM_FAULT_OOM;
 686                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 687                                           vma, haddr, numa_node_id(), 0);
 688                 if (unlikely(!page)) {
 689                         count_vm_event(THP_FAULT_FALLBACK);
 690                         goto out;
 691                 }
 692                 count_vm_event(THP_FAULT_ALLOC);
 693                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 694                         put_page(page);
 695                         goto out;
 696                 }
 697
 698                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
 699         }
 700 out:
 701         /*
 702          * Use __pte_alloc instead of pte_alloc_map, because we can't
 703          * run pte_offset_map on the pmd, if an huge pmd could
 704          * materialize from under us from a different thread.
 705          */
 706         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 707                 return VM_FAULT_OOM;
 708         /* if an huge pmd materialized from under us just retry later */
 709         if (unlikely(pmd_trans_huge(*pmd)))
 710                 return 0;
 711         /*
 712          * A regular pmd is established and it can't morph into a huge pmd
 713          * from under us anymore at this point because we hold the mmap_sem
 714          * read mode and khugepaged takes it in write mode. So now it's
 715          * safe to run pte_offset_map().
 716          */
 717         pte = pte_offset_map(pmd, address);
 718         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 719 }
 720
 721 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 722                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 723                   struct vm_area_struct *vma)
 724 {
 725         struct page *src_page;
 726         pmd_t pmd;
 727         pgtable_t pgtable;
 728         int ret;
 729
 730         ret = -ENOMEM;
 731         pgtable = pte_alloc_one(dst_mm, addr);
 732         if (unlikely(!pgtable))
 733                 goto out;
 734
 735         spin_lock(&dst_mm->page_table_lock);
 736         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 737
 738         ret = -EAGAIN;
 739         pmd = *src_pmd;
 740         if (unlikely(!pmd_trans_huge(pmd))) {
 741                 pte_free(dst_mm, pgtable);
 742                 goto out_unlock;
 743         }
 744         if (unlikely(pmd_trans_splitting(pmd))) {
 745                 /* split huge page running from under us */
 746                 spin_unlock(&src_mm->page_table_lock);
 747                 spin_unlock(&dst_mm->page_table_lock);
 748                 pte_free(dst_mm, pgtable);
 749
 750                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 751                 goto out;
 752         }
 753         src_page = pmd_page(pmd);
 754         VM_BUG_ON(!PageHead(src_page));
 755         get_page(src_page);
 756         page_dup_rmap(src_page);
 757         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 758
 759         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 760         pmd = pmd_mkold(pmd_wrprotect(pmd));
 761         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 762         prepare_pmd_huge_pte(pgtable, dst_mm);
 763         dst_mm->nr_ptes++;
 764
 765         ret = 0;
 766 out_unlock:
 767         spin_unlock(&src_mm->page_table_lock);
 768         spin_unlock(&dst_mm->page_table_lock);
 769 out:
 770         return ret;
 771 }
 772
 773 /* no "address" argument so destroys page coloring of some arch */
 774 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 775 {
 776         pgtable_t pgtable;
 777
 778         assert_spin_locked(&mm->page_table_lock);
 779
 780         /* FIFO */
 781         pgtable = mm->pmd_huge_pte;
 782         if (list_empty(&pgtable->lru))
 783                 mm->pmd_huge_pte = NULL;
 784         else {
 785                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 786                                               struct page, lru);
 787                 list_del(&pgtable->lru);
 788         }
 789         return pgtable;
 790 }
 791
 792 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 793                                         struct vm_area_struct *vma,
 794                                         unsigned long address,
 795                                         pmd_t *pmd, pmd_t orig_pmd,
 796                                         struct page *page,
 797                                         unsigned long haddr)
 798 {
 799         pgtable_t pgtable;
 800         pmd_t _pmd;
 801         int ret = 0, i;
 802         struct page **pages;
 803
 804         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 805                         GFP_KERNEL);
 806         if (unlikely(!pages)) {
 807                 ret |= VM_FAULT_OOM;
 808                 goto out;
 809         }
 810
 811         for (i = 0; i < HPAGE_PMD_NR; i++) {
 812                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 813                                                __GFP_OTHER_NODE,
 814                                                vma, address, page_to_nid(page));
 815                 if (unlikely(!pages[i] ||
 816                              mem_cgroup_newpage_charge(pages[i], mm,
 817                                                        GFP_KERNEL))) {
 818                         if (pages[i])
 819                                 put_page(pages[i]);
 820                         mem_cgroup_uncharge_start();
 821                         while (--i >= 0) {
 822                                 mem_cgroup_uncharge_page(pages[i]);
 823                                 put_page(pages[i]);
 824                         }
 825                         mem_cgroup_uncharge_end();
 826                         kfree(pages);
 827                         ret |= VM_FAULT_OOM;
 828                         goto out;
 829                 }
 830         }
 831
 832         for (i = 0; i < HPAGE_PMD_NR; i++) {
 833                 copy_user_highpage(pages[i], page + i,
 834                                    haddr + PAGE_SHIFT*i, vma);
 835                 __SetPageUptodate(pages[i]);
 836                 cond_resched();
 837         }
 838
 839         spin_lock(&mm->page_table_lock);
 840         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 841                 goto out_free_pages;
 842         VM_BUG_ON(!PageHead(page));
 843
 844         pmdp_clear_flush_notify(vma, haddr, pmd);
 845         /* leave pmd empty until pte is filled */
 846
 847         pgtable = get_pmd_huge_pte(mm);
 848         pmd_populate(mm, &_pmd, pgtable);
 849
 850         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 851                 pte_t *pte, entry;
 852                 entry = mk_pte(pages[i], vma->vm_page_prot);
 853                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 854                 page_add_new_anon_rmap(pages[i], vma, haddr);
 855                 pte = pte_offset_map(&_pmd, haddr);
 856                 VM_BUG_ON(!pte_none(*pte));
 857                 set_pte_at(mm, haddr, pte, entry);
 858                 pte_unmap(pte);
 859         }
 860         kfree(pages);
 861
 862         smp_wmb(); /* make pte visible before pmd */
 863         pmd_populate(mm, pmd, pgtable);
 864         page_remove_rmap(page);
 865         spin_unlock(&mm->page_table_lock);
 866
 867         ret |= VM_FAULT_WRITE;
 868         put_page(page);
 869
 870 out:
 871         return ret;
 872
 873 out_free_pages:
 874         spin_unlock(&mm->page_table_lock);
 875         mem_cgroup_uncharge_start();
 876         for (i = 0; i < HPAGE_PMD_NR; i++) {
 877                 mem_cgroup_uncharge_page(pages[i]);
 878                 put_page(pages[i]);
 879         }
 880         mem_cgroup_uncharge_end();
 881         kfree(pages);
 882         goto out;
 883 }
 884
 885 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 886                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 887 {
 888         int ret = 0;
 889         struct page *page, *new_page;
 890         unsigned long haddr;
 891
 892         VM_BUG_ON(!vma->anon_vma);
 893         spin_lock(&mm->page_table_lock);
 894         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 895                 goto out_unlock;
 896
 897         page = pmd_page(orig_pmd);
 898         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 899         haddr = address & HPAGE_PMD_MASK;
 900         if (page_mapcount(page) == 1) {
 901                 pmd_t entry;
 902                 entry = pmd_mkyoung(orig_pmd);
 903                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 904                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 905                         update_mmu_cache(vma, address, entry);
 906                 ret |= VM_FAULT_WRITE;
 907                 goto out_unlock;
 908         }
 909         get_page(page);
 910         spin_unlock(&mm->page_table_lock);
 911
 912         if (transparent_hugepage_enabled(vma) &&
 913             !transparent_hugepage_debug_cow())
 914                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 915                                               vma, haddr, numa_node_id(), 0);
 916         else
 917                 new_page = NULL;
 918
 919         if (unlikely(!new_page)) {
 920                 count_vm_event(THP_FAULT_FALLBACK);
 921                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 922                                                    pmd, orig_pmd, page, haddr);
 923                 put_page(page);
 924                 goto out;
 925         }
 926         count_vm_event(THP_FAULT_ALLOC);
 927
 928         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 929                 put_page(new_page);
 930                 put_page(page);
 931                 ret |= VM_FAULT_OOM;
 932                 goto out;
 933         }
 934
 935         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 936         __SetPageUptodate(new_page);
 937
 938         spin_lock(&mm->page_table_lock);
 939         put_page(page);
 940         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 941                 mem_cgroup_uncharge_page(new_page);
 942                 put_page(new_page);
 943         } else {
 944                 pmd_t entry;
 945                 VM_BUG_ON(!PageHead(page));
 946                 entry = mk_pmd(new_page, vma->vm_page_prot);
 947                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 948                 entry = pmd_mkhuge(entry);
 949                 pmdp_clear_flush_notify(vma, haddr, pmd);
 950                 page_add_new_anon_rmap(new_page, vma, haddr);
 951                 set_pmd_at(mm, haddr, pmd, entry);
 952                 update_mmu_cache(vma, address, entry);
 953                 page_remove_rmap(page);
 954                 put_page(page);
 955                 ret |= VM_FAULT_WRITE;
 956         }
 957 out_unlock:
 958         spin_unlock(&mm->page_table_lock);
 959 out:
 960         return ret;
 961 }
 962
 963 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 964                                    unsigned long addr,
 965                                    pmd_t *pmd,
 966                                    unsigned int flags)
 967 {
 968         struct page *page = NULL;
 969
 970         assert_spin_locked(&mm->page_table_lock);
 971
 972         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 973                 goto out;
 974
 975         page = pmd_page(*pmd);
 976         VM_BUG_ON(!PageHead(page));
 977         if (flags & FOLL_TOUCH) {
 978                 pmd_t _pmd;
 979                 /*
 980                  * We should set the dirty bit only for FOLL_WRITE but
 981                  * for now the dirty bit in the pmd is meaningless.
 982                  * And if the dirty bit will become meaningful and
 983                  * we'll only set it with FOLL_WRITE, an atomic
 984                  * set_bit will be required on the pmd to set the
 985                  * young bit, instead of the current set_pmd_at.
 986                  */
 987                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 988                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
 989         }
 990         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
 991         VM_BUG_ON(!PageCompound(page));
 992         if (flags & FOLL_GET)
 993                 get_page_foll(page);
 994
 995 out:
 996         return page;
 997 }
 998
 999 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1000                  pmd_t *pmd)
1001 {
1002         int ret = 0;
1003
1004         spin_lock(&tlb->mm->page_table_lock);
1005         if (likely(pmd_trans_huge(*pmd))) {
1006                 if (unlikely(pmd_trans_splitting(*pmd))) {
1007                         spin_unlock(&tlb->mm->page_table_lock);
1008                         wait_split_huge_page(vma->anon_vma,
1009                                              pmd);
1010                 } else {
1011                         struct page *page;
1012                         pgtable_t pgtable;
1013                         pgtable = get_pmd_huge_pte(tlb->mm);
1014                         page = pmd_page(*pmd);
1015                         pmd_clear(pmd);
1016                         page_remove_rmap(page);
1017                         VM_BUG_ON(page_mapcount(page) < 0);
1018                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1019                         VM_BUG_ON(!PageHead(page));
1020                         tlb->mm->nr_ptes--;
1021                         spin_unlock(&tlb->mm->page_table_lock);
1022                         tlb_remove_page(tlb, page);
1023                         pte_free(tlb->mm, pgtable);
1024                         ret = 1;
1025                 }
1026         } else
1027                 spin_unlock(&tlb->mm->page_table_lock);
1028
1029         return ret;
1030 }
1031
1032 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1033                 unsigned long addr, unsigned long end,
1034                 unsigned char *vec)
1035 {
1036         int ret = 0;
1037
1038         spin_lock(&vma->vm_mm->page_table_lock);
1039         if (likely(pmd_trans_huge(*pmd))) {
1040                 ret = !pmd_trans_splitting(*pmd);
1041                 spin_unlock(&vma->vm_mm->page_table_lock);
1042                 if (unlikely(!ret))
1043                         wait_split_huge_page(vma->anon_vma, pmd);
1044                 else {
1045                         /*
1046                          * All logical pages in the range are present
1047                          * if backed by a huge page.
1048                          */
1049                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1050                 }
1051         } else
1052                 spin_unlock(&vma->vm_mm->page_table_lock);
1053
1054         return ret;
1055 }
1056
1057 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1058                 unsigned long addr, pgprot_t newprot)
1059 {
1060         struct mm_struct *mm = vma->vm_mm;
1061         int ret = 0;
1062
1063         spin_lock(&mm->page_table_lock);
1064         if (likely(pmd_trans_huge(*pmd))) {
1065                 if (unlikely(pmd_trans_splitting(*pmd))) {
1066                         spin_unlock(&mm->page_table_lock);
1067                         wait_split_huge_page(vma->anon_vma, pmd);
1068                 } else {
1069                         pmd_t entry;
1070
1071                         entry = pmdp_get_and_clear(mm, addr, pmd);
1072                         entry = pmd_modify(entry, newprot);
1073                         set_pmd_at(mm, addr, pmd, entry);
1074                         spin_unlock(&vma->vm_mm->page_table_lock);
1075                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1076                         ret = 1;
1077                 }
1078         } else
1079                 spin_unlock(&vma->vm_mm->page_table_lock);
1080
1081         return ret;
1082 }
1083
1084 pmd_t *page_check_address_pmd(struct page *page,
1085                               struct mm_struct *mm,
1086                               unsigned long address,
1087                               enum page_check_address_pmd_flag flag)
1088 {
1089         pgd_t *pgd;
1090         pud_t *pud;
1091         pmd_t *pmd, *ret = NULL;
1092
1093         if (address & ~HPAGE_PMD_MASK)
1094                 goto out;
1095
1096         pgd = pgd_offset(mm, address);
1097         if (!pgd_present(*pgd))
1098                 goto out;
1099
1100         pud = pud_offset(pgd, address);
1101         if (!pud_present(*pud))
1102                 goto out;
1103
1104         pmd = pmd_offset(pud, address);
1105         if (pmd_none(*pmd))
1106                 goto out;
1107         if (pmd_page(*pmd) != page)
1108                 goto out;
1109         /*
1110          * split_vma() may create temporary aliased mappings. There is
1111          * no risk as long as all huge pmd are found and have their
1112          * splitting bit set before __split_huge_page_refcount
1113          * runs. Finding the same huge pmd more than once during the
1114          * same rmap walk is not a problem.
1115          */
1116         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1117             pmd_trans_splitting(*pmd))
1118                 goto out;
1119         if (pmd_trans_huge(*pmd)) {
1120                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1121                           !pmd_trans_splitting(*pmd));
1122                 ret = pmd;
1123         }
1124 out:
1125         return ret;
1126 }
1127
1128 static int __split_huge_page_splitting(struct page *page,
1129                                        struct vm_area_struct *vma,
1130                                        unsigned long address)
1131 {
1132         struct mm_struct *mm = vma->vm_mm;
1133         pmd_t *pmd;
1134         int ret = 0;
1135
1136         spin_lock(&mm->page_table_lock);
1137         pmd = page_check_address_pmd(page, mm, address,
1138                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1139         if (pmd) {
1140                 /*
1141                  * We can't temporarily set the pmd to null in order
1142                  * to split it, the pmd must remain marked huge at all
1143                  * times or the VM won't take the pmd_trans_huge paths
1144                  * and it won't wait on the anon_vma->root->mutex to
1145                  * serialize against split_huge_page*.
1146                  */
1147                 pmdp_splitting_flush_notify(vma, address, pmd);
1148                 ret = 1;
1149         }
1150         spin_unlock(&mm->page_table_lock);
1151
1152         return ret;
1153 }
1154
1155 static void __split_huge_page_refcount(struct page *page)
1156 {
1157         int i;
1158         unsigned long head_index = page->index;
1159         struct zone *zone = page_zone(page);
1160         int zonestat;
1161         int tail_count = 0;
1162
1163         /* prevent PageLRU to go away from under us, and freeze lru stats */
1164         spin_lock_irq(&zone->lru_lock);
1165         compound_lock(page);
1166
1167         for (i = 1; i < HPAGE_PMD_NR; i++) {
1168                 struct page *page_tail = page + i;
1169
1170                 /* tail_page->_mapcount cannot change */
1171                 BUG_ON(page_mapcount(page_tail) < 0);
1172                 tail_count += page_mapcount(page_tail);
1173                 /* check for overflow */
1174                 BUG_ON(tail_count < 0);
1175                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1176                 /*
1177                  * tail_page->_count is zero and not changing from
1178                  * under us. But get_page_unless_zero() may be running
1179                  * from under us on the tail_page. If we used
1180                  * atomic_set() below instead of atomic_add(), we
1181                  * would then run atomic_set() concurrently with
1182                  * get_page_unless_zero(), and atomic_set() is
1183                  * implemented in C not using locked ops. spin_unlock
1184                  * on x86 sometime uses locked ops because of PPro
1185                  * errata 66, 92, so unless somebody can guarantee
1186                  * atomic_set() here would be safe on all archs (and
1187                  * not only on x86), it's safer to use atomic_add().
1188                  */
1189                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1190                            &page_tail->_count);
1191
1192                 /* after clearing PageTail the gup refcount can be released */
1193                 smp_mb();
1194
1195                 /*
1196                  * retain hwpoison flag of the poisoned tail page:
1197                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1198                  *   by the memory-failure.
1199                  */
1200                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1201                 page_tail->flags |= (page->flags &
1202                                      ((1L << PG_referenced) |
1203                                       (1L << PG_swapbacked) |
1204                                       (1L << PG_mlocked) |
1205                                       (1L << PG_uptodate)));
1206                 page_tail->flags |= (1L << PG_dirty);
1207
1208                 /* clear PageTail before overwriting first_page */
1209                 smp_wmb();
1210
1211                 /*
1212                  * __split_huge_page_splitting() already set the
1213                  * splitting bit in all pmd that could map this
1214                  * hugepage, that will ensure no CPU can alter the
1215                  * mapcount on the head page. The mapcount is only
1216                  * accounted in the head page and it has to be
1217                  * transferred to all tail pages in the below code. So
1218                  * for this code to be safe, the split the mapcount
1219                  * can't change. But that doesn't mean userland can't
1220                  * keep changing and reading the page contents while
1221                  * we transfer the mapcount, so the pmd splitting
1222                  * status is achieved setting a reserved bit in the
1223                  * pmd, not by clearing the present bit.
1224                 */
1225                 page_tail->_mapcount = page->_mapcount;
1226
1227                 BUG_ON(page_tail->mapping);
1228                 page_tail->mapping = page->mapping;
1229
1230                 page_tail->index = ++head_index;
1231
1232                 BUG_ON(!PageAnon(page_tail));
1233                 BUG_ON(!PageUptodate(page_tail));
1234                 BUG_ON(!PageDirty(page_tail));
1235                 BUG_ON(!PageSwapBacked(page_tail));
1236
1237                 mem_cgroup_split_huge_fixup(page, page_tail);
1238
1239                 lru_add_page_tail(zone, page, page_tail);
1240         }
1241         atomic_sub(tail_count, &page->_count);
1242         BUG_ON(atomic_read(&page->_count) <= 0);
1243
1244         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1245         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1246
1247         /*
1248          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1249          * so adjust those appropriately if this page is on the LRU.
1250          */
1251         if (PageLRU(page)) {
1252                 zonestat = NR_LRU_BASE + page_lru(page);
1253                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1254         }
1255
1256         ClearPageCompound(page);
1257         compound_unlock(page);
1258         spin_unlock_irq(&zone->lru_lock);
1259
1260         for (i = 1; i < HPAGE_PMD_NR; i++) {
1261                 struct page *page_tail = page + i;
1262                 BUG_ON(page_count(page_tail) <= 0);
1263                 /*
1264                  * Tail pages may be freed if there wasn't any mapping
1265                  * like if add_to_swap() is running on a lru page that
1266                  * had its mapping zapped. And freeing these pages
1267                  * requires taking the lru_lock so we do the put_page
1268                  * of the tail pages after the split is complete.
1269                  */
1270                 put_page(page_tail);
1271         }
1272
1273         /*
1274          * Only the head page (now become a regular page) is required
1275          * to be pinned by the caller.
1276          */
1277         BUG_ON(page_count(page) <= 0);
1278 }
1279
1280 static int __split_huge_page_map(struct page *page,
1281                                  struct vm_area_struct *vma,
1282                                  unsigned long address)
1283 {
1284         struct mm_struct *mm = vma->vm_mm;
1285         pmd_t *pmd, _pmd;
1286         int ret = 0, i;
1287         pgtable_t pgtable;
1288         unsigned long haddr;
1289
1290         spin_lock(&mm->page_table_lock);
1291         pmd = page_check_address_pmd(page, mm, address,
1292                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1293         if (pmd) {
1294                 pgtable = get_pmd_huge_pte(mm);
1295                 pmd_populate(mm, &_pmd, pgtable);
1296
1297                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1298                      i++, haddr += PAGE_SIZE) {
1299                         pte_t *pte, entry;
1300                         BUG_ON(PageCompound(page+i));
1301                         entry = mk_pte(page + i, vma->vm_page_prot);
1302                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1303                         if (!pmd_write(*pmd))
1304                                 entry = pte_wrprotect(entry);
1305                         else
1306                                 BUG_ON(page_mapcount(page) != 1);
1307                         if (!pmd_young(*pmd))
1308                                 entry = pte_mkold(entry);
1309                         pte = pte_offset_map(&_pmd, haddr);
1310                         BUG_ON(!pte_none(*pte));
1311                         set_pte_at(mm, haddr, pte, entry);
1312                         pte_unmap(pte);
1313                 }
1314
1315                 smp_wmb(); /* make pte visible before pmd */
1316                 /*
1317                  * Up to this point the pmd is present and huge and
1318                  * userland has the whole access to the hugepage
1319                  * during the split (which happens in place). If we
1320                  * overwrite the pmd with the not-huge version
1321                  * pointing to the pte here (which of course we could
1322                  * if all CPUs were bug free), userland could trigger
1323                  * a small page size TLB miss on the small sized TLB
1324                  * while the hugepage TLB entry is still established
1325                  * in the huge TLB. Some CPU doesn't like that. See
1326                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1327                  * Erratum 383 on page 93. Intel should be safe but is
1328                  * also warns that it's only safe if the permission
1329                  * and cache attributes of the two entries loaded in
1330                  * the two TLB is identical (which should be the case
1331                  * here). But it is generally safer to never allow
1332                  * small and huge TLB entries for the same virtual
1333                  * address to be loaded simultaneously. So instead of
1334                  * doing "pmd_populate(); flush_tlb_range();" we first
1335                  * mark the current pmd notpresent (atomically because
1336                  * here the pmd_trans_huge and pmd_trans_splitting
1337                  * must remain set at all times on the pmd until the
1338                  * split is complete for this pmd), then we flush the
1339                  * SMP TLB and finally we write the non-huge version
1340                  * of the pmd entry with pmd_populate.
1341                  */
1342                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1343                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1344                 pmd_populate(mm, pmd, pgtable);
1345                 ret = 1;
1346         }
1347         spin_unlock(&mm->page_table_lock);
1348
1349         return ret;
1350 }
1351
1352 /* must be called with anon_vma->root->mutex hold */
1353 static void __split_huge_page(struct page *page,
1354                               struct anon_vma *anon_vma)
1355 {
1356         int mapcount, mapcount2;
1357         struct anon_vma_chain *avc;
1358
1359         BUG_ON(!PageHead(page));
1360         BUG_ON(PageTail(page));
1361
1362         mapcount = 0;
1363         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1364                 struct vm_area_struct *vma = avc->vma;
1365                 unsigned long addr = vma_address(page, vma);
1366                 BUG_ON(is_vma_temporary_stack(vma));
1367                 if (addr == -EFAULT)
1368                         continue;
1369                 mapcount += __split_huge_page_splitting(page, vma, addr);
1370         }
1371         /*
1372          * It is critical that new vmas are added to the tail of the
1373          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1374          * and establishes a child pmd before
1375          * __split_huge_page_splitting() freezes the parent pmd (so if
1376          * we fail to prevent copy_huge_pmd() from running until the
1377          * whole __split_huge_page() is complete), we will still see
1378          * the newly established pmd of the child later during the
1379          * walk, to be able to set it as pmd_trans_splitting too.
1380          */
1381         if (mapcount != page_mapcount(page))
1382                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1383                        mapcount, page_mapcount(page));
1384         BUG_ON(mapcount != page_mapcount(page));
1385
1386         __split_huge_page_refcount(page);
1387
1388         mapcount2 = 0;
1389         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1390                 struct vm_area_struct *vma = avc->vma;
1391                 unsigned long addr = vma_address(page, vma);
1392                 BUG_ON(is_vma_temporary_stack(vma));
1393                 if (addr == -EFAULT)
1394                         continue;
1395                 mapcount2 += __split_huge_page_map(page, vma, addr);
1396         }
1397         if (mapcount != mapcount2)
1398                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1399                        mapcount, mapcount2, page_mapcount(page));
1400         BUG_ON(mapcount != mapcount2);
1401 }
1402
1403 int split_huge_page(struct page *page)
1404 {
1405         struct anon_vma *anon_vma;
1406         int ret = 1;
1407
1408         BUG_ON(!PageAnon(page));
1409         anon_vma = page_lock_anon_vma(page);
1410         if (!anon_vma)
1411                 goto out;
1412         ret = 0;
1413         if (!PageCompound(page))
1414                 goto out_unlock;
1415
1416         BUG_ON(!PageSwapBacked(page));
1417         __split_huge_page(page, anon_vma);
1418         count_vm_event(THP_SPLIT);
1419
1420         BUG_ON(PageCompound(page));
1421 out_unlock:
1422         page_unlock_anon_vma(anon_vma);
1423 out:
1424         return ret;
1425 }
1426
1427 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1428                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1429
1430 int hugepage_madvise(struct vm_area_struct *vma,
1431                      unsigned long *vm_flags, int advice)
1432 {
1433         switch (advice) {
1434         case MADV_HUGEPAGE:
1435                 /*
1436                  * Be somewhat over-protective like KSM for now!
1437                  */
1438                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1439                         return -EINVAL;
1440                 *vm_flags &= ~VM_NOHUGEPAGE;
1441                 *vm_flags |= VM_HUGEPAGE;
1442                 /*
1443                  * If the vma become good for khugepaged to scan,
1444                  * register it here without waiting a page fault that
1445                  * may not happen any time soon.
1446                  */
1447                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1448                         return -ENOMEM;
1449                 break;
1450         case MADV_NOHUGEPAGE:
1451                 /*
1452                  * Be somewhat over-protective like KSM for now!
1453                  */
1454                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1455                         return -EINVAL;
1456                 *vm_flags &= ~VM_HUGEPAGE;
1457                 *vm_flags |= VM_NOHUGEPAGE;
1458                 /*
1459                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1460                  * this vma even if we leave the mm registered in khugepaged if
1461                  * it got registered before VM_NOHUGEPAGE was set.
1462                  */
1463                 break;
1464         }
1465
1466         return 0;
1467 }
1468
1469 static int __init khugepaged_slab_init(void)
1470 {
1471         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1472                                           sizeof(struct mm_slot),
1473                                           __alignof__(struct mm_slot), 0, NULL);
1474         if (!mm_slot_cache)
1475                 return -ENOMEM;
1476
1477         return 0;
1478 }
1479
1480 static void __init khugepaged_slab_free(void)
1481 {
1482         kmem_cache_destroy(mm_slot_cache);
1483         mm_slot_cache = NULL;
1484 }
1485
1486 static inline struct mm_slot *alloc_mm_slot(void)
1487 {
1488         if (!mm_slot_cache)     /* initialization failed */
1489                 return NULL;
1490         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1491 }
1492
1493 static inline void free_mm_slot(struct mm_slot *mm_slot)
1494 {
1495         kmem_cache_free(mm_slot_cache, mm_slot);
1496 }
1497
1498 static int __init mm_slots_hash_init(void)
1499 {
1500         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1501                                 GFP_KERNEL);
1502         if (!mm_slots_hash)
1503                 return -ENOMEM;
1504         return 0;
1505 }
1506
1507 #if 0
1508 static void __init mm_slots_hash_free(void)
1509 {
1510         kfree(mm_slots_hash);
1511         mm_slots_hash = NULL;
1512 }
1513 #endif
1514
1515 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1516 {
1517         struct mm_slot *mm_slot;
1518         struct hlist_head *bucket;
1519         struct hlist_node *node;
1520
1521         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1522                                 % MM_SLOTS_HASH_HEADS];
1523         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1524                 if (mm == mm_slot->mm)
1525                         return mm_slot;
1526         }
1527         return NULL;
1528 }
1529
1530 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1531                                     struct mm_slot *mm_slot)
1532 {
1533         struct hlist_head *bucket;
1534
1535         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1536                                 % MM_SLOTS_HASH_HEADS];
1537         mm_slot->mm = mm;
1538         hlist_add_head(&mm_slot->hash, bucket);
1539 }
1540
1541 static inline int khugepaged_test_exit(struct mm_struct *mm)
1542 {
1543         return atomic_read(&mm->mm_users) == 0;
1544 }
1545
1546 int __khugepaged_enter(struct mm_struct *mm)
1547 {
1548         struct mm_slot *mm_slot;
1549         int wakeup;
1550
1551         mm_slot = alloc_mm_slot();
1552         if (!mm_slot)
1553                 return -ENOMEM;
1554
1555         /* __khugepaged_exit() must not run from under us */
1556         VM_BUG_ON(khugepaged_test_exit(mm));
1557         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1558                 free_mm_slot(mm_slot);
1559                 return 0;
1560         }
1561
1562         spin_lock(&khugepaged_mm_lock);
1563         insert_to_mm_slots_hash(mm, mm_slot);
1564         /*
1565          * Insert just behind the scanning cursor, to let the area settle
1566          * down a little.
1567          */
1568         wakeup = list_empty(&khugepaged_scan.mm_head);
1569         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1570         spin_unlock(&khugepaged_mm_lock);
1571
1572         atomic_inc(&mm->mm_count);
1573         if (wakeup)
1574                 wake_up_interruptible(&khugepaged_wait);
1575
1576         return 0;
1577 }
1578
1579 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1580 {
1581         unsigned long hstart, hend;
1582         if (!vma->anon_vma)
1583                 /*
1584                  * Not yet faulted in so we will register later in the
1585                  * page fault if needed.
1586                  */
1587                 return 0;
1588         if (vma->vm_ops)
1589                 /* khugepaged not yet working on file or special mappings */
1590                 return 0;
1591         /*
1592          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1593          * true too, verify it here.
1594          */
1595         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1596         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1597         hend = vma->vm_end & HPAGE_PMD_MASK;
1598         if (hstart < hend)
1599                 return khugepaged_enter(vma);
1600         return 0;
1601 }
1602
1603 void __khugepaged_exit(struct mm_struct *mm)
1604 {
1605         struct mm_slot *mm_slot;
1606         int free = 0;
1607
1608         spin_lock(&khugepaged_mm_lock);
1609         mm_slot = get_mm_slot(mm);
1610         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1611                 hlist_del(&mm_slot->hash);
1612                 list_del(&mm_slot->mm_node);
1613                 free = 1;
1614         }
1615
1616         if (free) {
1617                 spin_unlock(&khugepaged_mm_lock);
1618                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1619                 free_mm_slot(mm_slot);
1620                 mmdrop(mm);
1621         } else if (mm_slot) {
1622                 spin_unlock(&khugepaged_mm_lock);
1623                 /*
1624                  * This is required to serialize against
1625                  * khugepaged_test_exit() (which is guaranteed to run
1626                  * under mmap sem read mode). Stop here (after we
1627                  * return all pagetables will be destroyed) until
1628                  * khugepaged has finished working on the pagetables
1629                  * under the mmap_sem.
1630                  */
1631                 down_write(&mm->mmap_sem);
1632                 up_write(&mm->mmap_sem);
1633         } else
1634                 spin_unlock(&khugepaged_mm_lock);
1635 }
1636
1637 static void release_pte_page(struct page *page)
1638 {
1639         /* 0 stands for page_is_file_cache(page) == false */
1640         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1641         unlock_page(page);
1642         putback_lru_page(page);
1643 }
1644
1645 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1646 {
1647         while (--_pte >= pte) {
1648                 pte_t pteval = *_pte;
1649                 if (!pte_none(pteval))
1650                         release_pte_page(pte_page(pteval));
1651         }
1652 }
1653
1654 static void release_all_pte_pages(pte_t *pte)
1655 {
1656         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1657 }
1658
1659 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1660                                         unsigned long address,
1661                                         pte_t *pte)
1662 {
1663         struct page *page;
1664         pte_t *_pte;
1665         int referenced = 0, isolated = 0, none = 0;
1666         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1667              _pte++, address += PAGE_SIZE) {
1668                 pte_t pteval = *_pte;
1669                 if (pte_none(pteval)) {
1670                         if (++none <= khugepaged_max_ptes_none)
1671                                 continue;
1672                         else {
1673                                 release_pte_pages(pte, _pte);
1674                                 goto out;
1675                         }
1676                 }
1677                 if (!pte_present(pteval) || !pte_write(pteval)) {
1678                         release_pte_pages(pte, _pte);
1679                         goto out;
1680                 }
1681                 page = vm_normal_page(vma, address, pteval);
1682                 if (unlikely(!page)) {
1683                         release_pte_pages(pte, _pte);
1684                         goto out;
1685                 }
1686                 VM_BUG_ON(PageCompound(page));
1687                 BUG_ON(!PageAnon(page));
1688                 VM_BUG_ON(!PageSwapBacked(page));
1689
1690                 /* cannot use mapcount: can't collapse if there's a gup pin */
1691                 if (page_count(page) != 1) {
1692                         release_pte_pages(pte, _pte);
1693                         goto out;
1694                 }
1695                 /*
1696                  * We can do it before isolate_lru_page because the
1697                  * page can't be freed from under us. NOTE: PG_lock
1698                  * is needed to serialize against split_huge_page
1699                  * when invoked from the VM.
1700                  */
1701                 if (!trylock_page(page)) {
1702                         release_pte_pages(pte, _pte);
1703                         goto out;
1704                 }
1705                 /*
1706                  * Isolate the page to avoid collapsing an hugepage
1707                  * currently in use by the VM.
1708                  */
1709                 if (isolate_lru_page(page)) {
1710                         unlock_page(page);
1711                         release_pte_pages(pte, _pte);
1712                         goto out;
1713                 }
1714                 /* 0 stands for page_is_file_cache(page) == false */
1715                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1716                 VM_BUG_ON(!PageLocked(page));
1717                 VM_BUG_ON(PageLRU(page));
1718
1719                 /* If there is no mapped pte young don't collapse the page */
1720                 if (pte_young(pteval) || PageReferenced(page) ||
1721                     mmu_notifier_test_young(vma->vm_mm, address))
1722                         referenced = 1;
1723         }
1724         if (unlikely(!referenced))
1725                 release_all_pte_pages(pte);
1726         else
1727                 isolated = 1;
1728 out:
1729         return isolated;
1730 }
1731
1732 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1733                                       struct vm_area_struct *vma,
1734                                       unsigned long address,
1735                                       spinlock_t *ptl)
1736 {
1737         pte_t *_pte;
1738         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1739                 pte_t pteval = *_pte;
1740                 struct page *src_page;
1741
1742                 if (pte_none(pteval)) {
1743                         clear_user_highpage(page, address);
1744                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1745                 } else {
1746                         src_page = pte_page(pteval);
1747                         copy_user_highpage(page, src_page, address, vma);
1748                         VM_BUG_ON(page_mapcount(src_page) != 1);
1749                         VM_BUG_ON(page_count(src_page) != 2);
1750                         release_pte_page(src_page);
1751                         /*
1752                          * ptl mostly unnecessary, but preempt has to
1753                          * be disabled to update the per-cpu stats
1754                          * inside page_remove_rmap().
1755                          */
1756                         spin_lock(ptl);
1757                         /*
1758                          * paravirt calls inside pte_clear here are
1759                          * superfluous.
1760                          */
1761                         pte_clear(vma->vm_mm, address, _pte);
1762                         page_remove_rmap(src_page);
1763                         spin_unlock(ptl);
1764                         free_page_and_swap_cache(src_page);
1765                 }
1766
1767                 address += PAGE_SIZE;
1768                 page++;
1769         }
1770 }
1771
1772 static void collapse_huge_page(struct mm_struct *mm,
1773                                unsigned long address,
1774                                struct page **hpage,
1775                                struct vm_area_struct *vma,
1776                                int node)
1777 {
1778         pgd_t *pgd;
1779         pud_t *pud;
1780         pmd_t *pmd, _pmd;
1781         pte_t *pte;
1782         pgtable_t pgtable;
1783         struct page *new_page;
1784         spinlock_t *ptl;
1785         int isolated;
1786         unsigned long hstart, hend;
1787
1788         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1789 #ifndef CONFIG_NUMA
1790         up_read(&mm->mmap_sem);
1791         VM_BUG_ON(!*hpage);
1792         new_page = *hpage;
1793 #else
1794         VM_BUG_ON(*hpage);
1795         /*
1796          * Allocate the page while the vma is still valid and under
1797          * the mmap_sem read mode so there is no memory allocation
1798          * later when we take the mmap_sem in write mode. This is more
1799          * friendly behavior (OTOH it may actually hide bugs) to
1800          * filesystems in userland with daemons allocating memory in
1801          * the userland I/O paths.  Allocating memory with the
1802          * mmap_sem in read mode is good idea also to allow greater
1803          * scalability.
1804          */
1805         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1806                                       node, __GFP_OTHER_NODE);
1807
1808         /*
1809          * After allocating the hugepage, release the mmap_sem read lock in
1810          * preparation for taking it in write mode.
1811          */
1812         up_read(&mm->mmap_sem);
1813         if (unlikely(!new_page)) {
1814                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1815                 *hpage = ERR_PTR(-ENOMEM);
1816                 return;
1817         }
1818 #endif
1819
1820         count_vm_event(THP_COLLAPSE_ALLOC);
1821         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1822 #ifdef CONFIG_NUMA
1823                 put_page(new_page);
1824 #endif
1825                 return;
1826         }
1827
1828         /*
1829          * Prevent all access to pagetables with the exception of
1830          * gup_fast later hanlded by the ptep_clear_flush and the VM
1831          * handled by the anon_vma lock + PG_lock.
1832          */
1833         down_write(&mm->mmap_sem);
1834         if (unlikely(khugepaged_test_exit(mm)))
1835                 goto out;
1836
1837         vma = find_vma(mm, address);
1838         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1839         hend = vma->vm_end & HPAGE_PMD_MASK;
1840         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1841                 goto out;
1842
1843         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1844             (vma->vm_flags & VM_NOHUGEPAGE))
1845                 goto out;
1846
1847         if (!vma->anon_vma || vma->vm_ops)
1848                 goto out;
1849         if (is_vma_temporary_stack(vma))
1850                 goto out;
1851         /*
1852          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1853          * true too, verify it here.
1854          */
1855         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1856
1857         pgd = pgd_offset(mm, address);
1858         if (!pgd_present(*pgd))
1859                 goto out;
1860
1861         pud = pud_offset(pgd, address);
1862         if (!pud_present(*pud))
1863                 goto out;
1864
1865         pmd = pmd_offset(pud, address);
1866         /* pmd can't go away or become huge under us */
1867         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1868                 goto out;
1869
1870         anon_vma_lock(vma->anon_vma);
1871
1872         pte = pte_offset_map(pmd, address);
1873         ptl = pte_lockptr(mm, pmd);
1874
1875         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1876         /*
1877          * After this gup_fast can't run anymore. This also removes
1878          * any huge TLB entry from the CPU so we won't allow
1879          * huge and small TLB entries for the same virtual address
1880          * to avoid the risk of CPU bugs in that area.
1881          */
1882         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1883         spin_unlock(&mm->page_table_lock);
1884
1885         spin_lock(ptl);
1886         isolated = __collapse_huge_page_isolate(vma, address, pte);
1887         spin_unlock(ptl);
1888
1889         if (unlikely(!isolated)) {
1890                 pte_unmap(pte);
1891                 spin_lock(&mm->page_table_lock);
1892                 BUG_ON(!pmd_none(*pmd));
1893                 set_pmd_at(mm, address, pmd, _pmd);
1894                 spin_unlock(&mm->page_table_lock);
1895                 anon_vma_unlock(vma->anon_vma);
1896                 goto out;
1897         }
1898
1899         /*
1900          * All pages are isolated and locked so anon_vma rmap
1901          * can't run anymore.
1902          */
1903         anon_vma_unlock(vma->anon_vma);
1904
1905         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1906         pte_unmap(pte);
1907         __SetPageUptodate(new_page);
1908         pgtable = pmd_pgtable(_pmd);
1909         VM_BUG_ON(page_count(pgtable) != 1);
1910         VM_BUG_ON(page_mapcount(pgtable) != 0);
1911
1912         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1913         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1914         _pmd = pmd_mkhuge(_pmd);
1915
1916         /*
1917          * spin_lock() below is not the equivalent of smp_wmb(), so
1918          * this is needed to avoid the copy_huge_page writes to become
1919          * visible after the set_pmd_at() write.
1920          */
1921         smp_wmb();
1922
1923         spin_lock(&mm->page_table_lock);
1924         BUG_ON(!pmd_none(*pmd));
1925         page_add_new_anon_rmap(new_page, vma, address);
1926         set_pmd_at(mm, address, pmd, _pmd);
1927         update_mmu_cache(vma, address, entry);
1928         prepare_pmd_huge_pte(pgtable, mm);
1929         spin_unlock(&mm->page_table_lock);
1930
1931 #ifndef CONFIG_NUMA
1932         *hpage = NULL;
1933 #endif
1934         khugepaged_pages_collapsed++;
1935 out_up_write:
1936         up_write(&mm->mmap_sem);
1937         return;
1938
1939 out:
1940         mem_cgroup_uncharge_page(new_page);
1941 #ifdef CONFIG_NUMA
1942         put_page(new_page);
1943 #endif
1944         goto out_up_write;
1945 }
1946
1947 static int khugepaged_scan_pmd(struct mm_struct *mm,
1948                                struct vm_area_struct *vma,
1949                                unsigned long address,
1950                                struct page **hpage)
1951 {
1952         pgd_t *pgd;
1953         pud_t *pud;
1954         pmd_t *pmd;
1955         pte_t *pte, *_pte;
1956         int ret = 0, referenced = 0, none = 0;
1957         struct page *page;
1958         unsigned long _address;
1959         spinlock_t *ptl;
1960         int node = -1;
1961
1962         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1963
1964         pgd = pgd_offset(mm, address);
1965         if (!pgd_present(*pgd))
1966                 goto out;
1967
1968         pud = pud_offset(pgd, address);
1969         if (!pud_present(*pud))
1970                 goto out;
1971
1972         pmd = pmd_offset(pud, address);
1973         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1974                 goto out;
1975
1976         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1977         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1978              _pte++, _address += PAGE_SIZE) {
1979                 pte_t pteval = *_pte;
1980                 if (pte_none(pteval)) {
1981                         if (++none <= khugepaged_max_ptes_none)
1982                                 continue;
1983                         else
1984                                 goto out_unmap;
1985                 }
1986                 if (!pte_present(pteval) || !pte_write(pteval))
1987                         goto out_unmap;
1988                 page = vm_normal_page(vma, _address, pteval);
1989                 if (unlikely(!page))
1990                         goto out_unmap;
1991                 /*
1992                  * Chose the node of the first page. This could
1993                  * be more sophisticated and look at more pages,
1994                  * but isn't for now.
1995                  */
1996                 if (node == -1)
1997                         node = page_to_nid(page);
1998                 VM_BUG_ON(PageCompound(page));
1999                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2000                         goto out_unmap;
2001                 /* cannot use mapcount: can't collapse if there's a gup pin */
2002                 if (page_count(page) != 1)
2003                         goto out_unmap;
2004                 if (pte_young(pteval) || PageReferenced(page) ||
2005                     mmu_notifier_test_young(vma->vm_mm, address))
2006                         referenced = 1;
2007         }
2008         if (referenced)
2009                 ret = 1;
2010 out_unmap:
2011         pte_unmap_unlock(pte, ptl);
2012         if (ret)
2013                 /* collapse_huge_page will return with the mmap_sem released */
2014                 collapse_huge_page(mm, address, hpage, vma, node);
2015 out:
2016         return ret;
2017 }
2018
2019 static void collect_mm_slot(struct mm_slot *mm_slot)
2020 {
2021         struct mm_struct *mm = mm_slot->mm;
2022
2023         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2024
2025         if (khugepaged_test_exit(mm)) {
2026                 /* free mm_slot */
2027                 hlist_del(&mm_slot->hash);
2028                 list_del(&mm_slot->mm_node);
2029
2030                 /*
2031                  * Not strictly needed because the mm exited already.
2032                  *
2033                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2034                  */
2035
2036                 /* khugepaged_mm_lock actually not necessary for the below */
2037                 free_mm_slot(mm_slot);
2038                 mmdrop(mm);
2039         }
2040 }
2041
2042 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2043                                             struct page **hpage)
2044 {
2045         struct mm_slot *mm_slot;
2046         struct mm_struct *mm;
2047         struct vm_area_struct *vma;
2048         int progress = 0;
2049
2050         VM_BUG_ON(!pages);
2051         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2052
2053         if (khugepaged_scan.mm_slot)
2054                 mm_slot = khugepaged_scan.mm_slot;
2055         else {
2056                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2057                                      struct mm_slot, mm_node);
2058                 khugepaged_scan.address = 0;
2059                 khugepaged_scan.mm_slot = mm_slot;
2060         }
2061         spin_unlock(&khugepaged_mm_lock);
2062
2063         mm = mm_slot->mm;
2064         down_read(&mm->mmap_sem);
2065         if (unlikely(khugepaged_test_exit(mm)))
2066                 vma = NULL;
2067         else
2068                 vma = find_vma(mm, khugepaged_scan.address);
2069
2070         progress++;
2071         for (; vma; vma = vma->vm_next) {
2072                 unsigned long hstart, hend;
2073
2074                 cond_resched();
2075                 if (unlikely(khugepaged_test_exit(mm))) {
2076                         progress++;
2077                         break;
2078                 }
2079
2080                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2081                      !khugepaged_always()) ||
2082                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2083                 skip:
2084                         progress++;
2085                         continue;
2086                 }
2087                 if (!vma->anon_vma || vma->vm_ops)
2088                         goto skip;
2089                 if (is_vma_temporary_stack(vma))
2090                         goto skip;
2091                 /*
2092                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2093                  * must be true too, verify it here.
2094                  */
2095                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2096                           vma->vm_flags & VM_NO_THP);
2097
2098                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2099                 hend = vma->vm_end & HPAGE_PMD_MASK;
2100                 if (hstart >= hend)
2101                         goto skip;
2102                 if (khugepaged_scan.address > hend)
2103                         goto skip;
2104                 if (khugepaged_scan.address < hstart)
2105                         khugepaged_scan.address = hstart;
2106                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2107
2108                 while (khugepaged_scan.address < hend) {
2109                         int ret;
2110                         cond_resched();
2111                         if (unlikely(khugepaged_test_exit(mm)))
2112                                 goto breakouterloop;
2113
2114                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2115                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2116                                   hend);
2117                         ret = khugepaged_scan_pmd(mm, vma,
2118                                                   khugepaged_scan.address,
2119                                                   hpage);
2120                         /* move to next address */
2121                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2122                         progress += HPAGE_PMD_NR;
2123                         if (ret)
2124                                 /* we released mmap_sem so break loop */
2125                                 goto breakouterloop_mmap_sem;
2126                         if (progress >= pages)
2127                                 goto breakouterloop;
2128                 }
2129         }
2130 breakouterloop:
2131         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2132 breakouterloop_mmap_sem:
2133
2134         spin_lock(&khugepaged_mm_lock);
2135         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2136         /*
2137          * Release the current mm_slot if this mm is about to die, or
2138          * if we scanned all vmas of this mm.
2139          */
2140         if (khugepaged_test_exit(mm) || !vma) {
2141                 /*
2142                  * Make sure that if mm_users is reaching zero while
2143                  * khugepaged runs here, khugepaged_exit will find
2144                  * mm_slot not pointing to the exiting mm.
2145                  */
2146                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2147                         khugepaged_scan.mm_slot = list_entry(
2148                                 mm_slot->mm_node.next,
2149                                 struct mm_slot, mm_node);
2150                         khugepaged_scan.address = 0;
2151                 } else {
2152                         khugepaged_scan.mm_slot = NULL;
2153                         khugepaged_full_scans++;
2154                 }
2155
2156                 collect_mm_slot(mm_slot);
2157         }
2158
2159         return progress;
2160 }
2161
2162 static int khugepaged_has_work(void)
2163 {
2164         return !list_empty(&khugepaged_scan.mm_head) &&
2165                 khugepaged_enabled();
2166 }
2167
2168 static int khugepaged_wait_event(void)
2169 {
2170         return !list_empty(&khugepaged_scan.mm_head) ||
2171                 !khugepaged_enabled();
2172 }
2173
2174 static void khugepaged_do_scan(struct page **hpage)
2175 {
2176         unsigned int progress = 0, pass_through_head = 0;
2177         unsigned int pages = khugepaged_pages_to_scan;
2178
2179         barrier(); /* write khugepaged_pages_to_scan to local stack */
2180
2181         while (progress < pages) {
2182                 cond_resched();
2183
2184 #ifndef CONFIG_NUMA
2185                 if (!*hpage) {
2186                         *hpage = alloc_hugepage(khugepaged_defrag());
2187                         if (unlikely(!*hpage)) {
2188                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2189                                 break;
2190                         }
2191                         count_vm_event(THP_COLLAPSE_ALLOC);
2192                 }
2193 #else
2194                 if (IS_ERR(*hpage))
2195                         break;
2196 #endif
2197
2198                 if (unlikely(kthread_should_stop() || freezing(current)))
2199                         break;
2200
2201                 spin_lock(&khugepaged_mm_lock);
2202                 if (!khugepaged_scan.mm_slot)
2203                         pass_through_head++;
2204                 if (khugepaged_has_work() &&
2205                     pass_through_head < 2)
2206                         progress += khugepaged_scan_mm_slot(pages - progress,
2207                                                             hpage);
2208                 else
2209                         progress = pages;
2210                 spin_unlock(&khugepaged_mm_lock);
2211         }
2212 }
2213
2214 static void khugepaged_alloc_sleep(void)
2215 {
2216         DEFINE_WAIT(wait);
2217         add_wait_queue(&khugepaged_wait, &wait);
2218         schedule_timeout_interruptible(
2219                 msecs_to_jiffies(
2220                         khugepaged_alloc_sleep_millisecs));
2221         remove_wait_queue(&khugepaged_wait, &wait);
2222 }
2223
2224 #ifndef CONFIG_NUMA
2225 static struct page *khugepaged_alloc_hugepage(void)
2226 {
2227         struct page *hpage;
2228
2229         do {
2230                 hpage = alloc_hugepage(khugepaged_defrag());
2231                 if (!hpage) {
2232                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2233                         khugepaged_alloc_sleep();
2234                 } else
2235                         count_vm_event(THP_COLLAPSE_ALLOC);
2236         } while (unlikely(!hpage) &&
2237                  likely(khugepaged_enabled()));
2238         return hpage;
2239 }
2240 #endif
2241
2242 static void khugepaged_loop(void)
2243 {
2244         struct page *hpage;
2245
2246 #ifdef CONFIG_NUMA
2247         hpage = NULL;
2248 #endif
2249         while (likely(khugepaged_enabled())) {
2250 #ifndef CONFIG_NUMA
2251                 hpage = khugepaged_alloc_hugepage();
2252                 if (unlikely(!hpage))
2253                         break;
2254 #else
2255                 if (IS_ERR(hpage)) {
2256                         khugepaged_alloc_sleep();
2257                         hpage = NULL;
2258                 }
2259 #endif
2260
2261                 khugepaged_do_scan(&hpage);
2262 #ifndef CONFIG_NUMA
2263                 if (hpage)
2264                         put_page(hpage);
2265 #endif
2266                 try_to_freeze();
2267                 if (unlikely(kthread_should_stop()))
2268                         break;
2269                 if (khugepaged_has_work()) {
2270                         DEFINE_WAIT(wait);
2271                         if (!khugepaged_scan_sleep_millisecs)
2272                                 continue;
2273                         add_wait_queue(&khugepaged_wait, &wait);
2274                         schedule_timeout_interruptible(
2275                                 msecs_to_jiffies(
2276                                         khugepaged_scan_sleep_millisecs));
2277                         remove_wait_queue(&khugepaged_wait, &wait);
2278                 } else if (khugepaged_enabled())
2279                         wait_event_freezable(khugepaged_wait,
2280                                              khugepaged_wait_event());
2281         }
2282 }
2283
2284 static int khugepaged(void *none)
2285 {
2286         struct mm_slot *mm_slot;
2287
2288         set_freezable();
2289         set_user_nice(current, 19);
2290
2291         /* serialize with start_khugepaged() */
2292         mutex_lock(&khugepaged_mutex);
2293
2294         for (;;) {
2295                 mutex_unlock(&khugepaged_mutex);
2296                 VM_BUG_ON(khugepaged_thread != current);
2297                 khugepaged_loop();
2298                 VM_BUG_ON(khugepaged_thread != current);
2299
2300                 mutex_lock(&khugepaged_mutex);
2301                 if (!khugepaged_enabled())
2302                         break;
2303                 if (unlikely(kthread_should_stop()))
2304                         break;
2305         }
2306
2307         spin_lock(&khugepaged_mm_lock);
2308         mm_slot = khugepaged_scan.mm_slot;
2309         khugepaged_scan.mm_slot = NULL;
2310         if (mm_slot)
2311                 collect_mm_slot(mm_slot);
2312         spin_unlock(&khugepaged_mm_lock);
2313
2314         khugepaged_thread = NULL;
2315         mutex_unlock(&khugepaged_mutex);
2316
2317         return 0;
2318 }
2319
2320 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2321 {
2322         struct page *page;
2323
2324         spin_lock(&mm->page_table_lock);
2325         if (unlikely(!pmd_trans_huge(*pmd))) {
2326                 spin_unlock(&mm->page_table_lock);
2327                 return;
2328         }
2329         page = pmd_page(*pmd);
2330         VM_BUG_ON(!page_count(page));
2331         get_page(page);
2332         spin_unlock(&mm->page_table_lock);
2333
2334         split_huge_page(page);
2335
2336         put_page(page);
2337         BUG_ON(pmd_trans_huge(*pmd));
2338 }
2339
2340 static void split_huge_page_address(struct mm_struct *mm,
2341                                     unsigned long address)
2342 {
2343         pgd_t *pgd;
2344         pud_t *pud;
2345         pmd_t *pmd;
2346
2347         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2348
2349         pgd = pgd_offset(mm, address);
2350         if (!pgd_present(*pgd))
2351                 return;
2352
2353         pud = pud_offset(pgd, address);
2354         if (!pud_present(*pud))
2355                 return;
2356
2357         pmd = pmd_offset(pud, address);
2358         if (!pmd_present(*pmd))
2359                 return;
2360         /*
2361          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2362          * materialize from under us.
2363          */
2364         split_huge_page_pmd(mm, pmd);
2365 }
2366
2367 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2368                              unsigned long start,
2369                              unsigned long end,
2370                              long adjust_next)
2371 {
2372         /*
2373          * If the new start address isn't hpage aligned and it could
2374          * previously contain an hugepage: check if we need to split
2375          * an huge pmd.
2376          */
2377         if (start & ~HPAGE_PMD_MASK &&
2378             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2379             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2380                 split_huge_page_address(vma->vm_mm, start);
2381
2382         /*
2383          * If the new end address isn't hpage aligned and it could
2384          * previously contain an hugepage: check if we need to split
2385          * an huge pmd.
2386          */
2387         if (end & ~HPAGE_PMD_MASK &&
2388             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2389             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2390                 split_huge_page_address(vma->vm_mm, end);
2391
2392         /*
2393          * If we're also updating the vma->vm_next->vm_start, if the new
2394          * vm_next->vm_start isn't page aligned and it could previously
2395          * contain an hugepage: check if we need to split an huge pmd.
2396          */
2397         if (adjust_next > 0) {
2398                 struct vm_area_struct *next = vma->vm_next;
2399                 unsigned long nstart = next->vm_start;
2400                 nstart += adjust_next << PAGE_SHIFT;
2401                 if (nstart & ~HPAGE_PMD_MASK &&
2402                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2403                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2404                         split_huge_page_address(next->vm_mm, nstart);
2405         }
2406 }