mm/huge_memory.c

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