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/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is disabled in order that avoid
  31  * to risk increase the memory footprint of applications without a guaranteed
  32  * benefit. When transparent hugepage support is enabled, is for all mappings,
  33  * and khugepaged scans all mappings.
  34  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  35  * for all hugepage allocations.
  36  */
  37 unsigned long transparent_hugepage_flags __read_mostly =
  38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  39         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  40 #endif
  41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  42         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  43 #endif
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  46         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  47
  48 /* default scan 8*512 pte (or vmas) every 30 second */
  49 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  50 static unsigned int khugepaged_pages_collapsed;
  51 static unsigned int khugepaged_full_scans;
  52 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  53 /* during fragmentation poll the hugepage allocator once every minute */
  54 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  55 static struct task_struct *khugepaged_thread __read_mostly;
  56 static DEFINE_MUTEX(khugepaged_mutex);
  57 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  58 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  59 /*
  60  * default collapse hugepages if there is at least one pte mapped like
  61  * it would have happened if the vma was large enough during page
  62  * fault.
  63  */
  64 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  65
  66 static int khugepaged(void *none);
  67 static int khugepaged_slab_init(void);
  68
  69 #define MM_SLOTS_HASH_BITS 10
  70 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  71
  72 static struct kmem_cache *mm_slot_cache __read_mostly;
  73
  74 /**
  75  * struct mm_slot - hash lookup from mm to mm_slot
  76  * @hash: hash collision list
  77  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  78  * @mm: the mm that this information is valid for
  79  */
  80 struct mm_slot {
  81         struct hlist_node hash;
  82         struct list_head mm_node;
  83         struct mm_struct *mm;
  84 };
  85
  86 /**
  87  * struct khugepaged_scan - cursor for scanning
  88  * @mm_head: the head of the mm list to scan
  89  * @mm_slot: the current mm_slot we are scanning
  90  * @address: the next address inside that to be scanned
  91  *
  92  * There is only the one khugepaged_scan instance of this cursor structure.
  93  */
  94 struct khugepaged_scan {
  95         struct list_head mm_head;
  96         struct mm_slot *mm_slot;
  97         unsigned long address;
  98 };
  99 static struct khugepaged_scan khugepaged_scan = {
 100         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 101 };
 102
 103
 104 static int set_recommended_min_free_kbytes(void)
 105 {
 106         struct zone *zone;
 107         int nr_zones = 0;
 108         unsigned long recommended_min;
 109
 110         if (!khugepaged_enabled())
 111                 return 0;
 112
 113         for_each_populated_zone(zone)
 114                 nr_zones++;
 115
 116         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 117         recommended_min = pageblock_nr_pages * nr_zones * 2;
 118
 119         /*
 120          * Make sure that on average at least two pageblocks are almost free
 121          * of another type, one for a migratetype to fall back to and a
 122          * second to avoid subsequent fallbacks of other types There are 3
 123          * MIGRATE_TYPES we care about.
 124          */
 125         recommended_min += pageblock_nr_pages * nr_zones *
 126                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 127
 128         /* don't ever allow to reserve more than 5% of the lowmem */
 129         recommended_min = min(recommended_min,
 130                               (unsigned long) nr_free_buffer_pages() / 20);
 131         recommended_min <<= (PAGE_SHIFT-10);
 132
 133         if (recommended_min > min_free_kbytes) {
 134                 if (user_min_free_kbytes >= 0)
 135                         pr_info("raising min_free_kbytes from %d to %lu "
 136                                 "to help transparent hugepage allocations\n",
 137                                 min_free_kbytes, recommended_min);
 138
 139                 min_free_kbytes = recommended_min;
 140         }
 141         setup_per_zone_wmarks();
 142         return 0;
 143 }
 144 late_initcall(set_recommended_min_free_kbytes);
 145
 146 static int start_khugepaged(void)
 147 {
 148         int err = 0;
 149         if (khugepaged_enabled()) {
 150                 if (!khugepaged_thread)
 151                         khugepaged_thread = kthread_run(khugepaged, NULL,
 152                                                         "khugepaged");
 153                 if (unlikely(IS_ERR(khugepaged_thread))) {
 154                         printk(KERN_ERR
 155                                "khugepaged: kthread_run(khugepaged) failed\n");
 156                         err = PTR_ERR(khugepaged_thread);
 157                         khugepaged_thread = NULL;
 158                 }
 159
 160                 if (!list_empty(&khugepaged_scan.mm_head))
 161                         wake_up_interruptible(&khugepaged_wait);
 162
 163                 set_recommended_min_free_kbytes();
 164         } else if (khugepaged_thread) {
 165                 kthread_stop(khugepaged_thread);
 166                 khugepaged_thread = NULL;
 167         }
 168
 169         return err;
 170 }
 171
 172 static atomic_t huge_zero_refcount;
 173 static struct page *huge_zero_page __read_mostly;
 174
 175 static inline bool is_huge_zero_page(struct page *page)
 176 {
 177         return ACCESS_ONCE(huge_zero_page) == page;
 178 }
 179
 180 static inline bool is_huge_zero_pmd(pmd_t pmd)
 181 {
 182         return is_huge_zero_page(pmd_page(pmd));
 183 }
 184
 185 static struct page *get_huge_zero_page(void)
 186 {
 187         struct page *zero_page;
 188 retry:
 189         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 190                 return ACCESS_ONCE(huge_zero_page);
 191
 192         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 193                         HPAGE_PMD_ORDER);
 194         if (!zero_page) {
 195                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 196                 return NULL;
 197         }
 198         count_vm_event(THP_ZERO_PAGE_ALLOC);
 199         preempt_disable();
 200         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 201                 preempt_enable();
 202                 __free_page(zero_page);
 203                 goto retry;
 204         }
 205
 206         /* We take additional reference here. It will be put back by shrinker */
 207         atomic_set(&huge_zero_refcount, 2);
 208         preempt_enable();
 209         return ACCESS_ONCE(huge_zero_page);
 210 }
 211
 212 static void put_huge_zero_page(void)
 213 {
 214         /*
 215          * Counter should never go to zero here. Only shrinker can put
 216          * last reference.
 217          */
 218         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 219 }
 220
 221 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 222                                         struct shrink_control *sc)
 223 {
 224         /* we can free zero page only if last reference remains */
 225         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 226 }
 227
 228 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 229                                        struct shrink_control *sc)
 230 {
 231         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 232                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 233                 BUG_ON(zero_page == NULL);
 234                 __free_page(zero_page);
 235                 return HPAGE_PMD_NR;
 236         }
 237
 238         return 0;
 239 }
 240
 241 static struct shrinker huge_zero_page_shrinker = {
 242         .count_objects = shrink_huge_zero_page_count,
 243         .scan_objects = shrink_huge_zero_page_scan,
 244         .seeks = DEFAULT_SEEKS,
 245 };
 246
 247 #ifdef CONFIG_SYSFS
 248
 249 static ssize_t double_flag_show(struct kobject *kobj,
 250                                 struct kobj_attribute *attr, char *buf,
 251                                 enum transparent_hugepage_flag enabled,
 252                                 enum transparent_hugepage_flag req_madv)
 253 {
 254         if (test_bit(enabled, &transparent_hugepage_flags)) {
 255                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 256                 return sprintf(buf, "[always] madvise never\n");
 257         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 258                 return sprintf(buf, "always [madvise] never\n");
 259         else
 260                 return sprintf(buf, "always madvise [never]\n");
 261 }
 262 static ssize_t double_flag_store(struct kobject *kobj,
 263                                  struct kobj_attribute *attr,
 264                                  const char *buf, size_t count,
 265                                  enum transparent_hugepage_flag enabled,
 266                                  enum transparent_hugepage_flag req_madv)
 267 {
 268         if (!memcmp("always", buf,
 269                     min(sizeof("always")-1, count))) {
 270                 set_bit(enabled, &transparent_hugepage_flags);
 271                 clear_bit(req_madv, &transparent_hugepage_flags);
 272         } else if (!memcmp("madvise", buf,
 273                            min(sizeof("madvise")-1, count))) {
 274                 clear_bit(enabled, &transparent_hugepage_flags);
 275                 set_bit(req_madv, &transparent_hugepage_flags);
 276         } else if (!memcmp("never", buf,
 277                            min(sizeof("never")-1, count))) {
 278                 clear_bit(enabled, &transparent_hugepage_flags);
 279                 clear_bit(req_madv, &transparent_hugepage_flags);
 280         } else
 281                 return -EINVAL;
 282
 283         return count;
 284 }
 285
 286 static ssize_t enabled_show(struct kobject *kobj,
 287                             struct kobj_attribute *attr, char *buf)
 288 {
 289         return double_flag_show(kobj, attr, buf,
 290                                 TRANSPARENT_HUGEPAGE_FLAG,
 291                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 292 }
 293 static ssize_t enabled_store(struct kobject *kobj,
 294                              struct kobj_attribute *attr,
 295                              const char *buf, size_t count)
 296 {
 297         ssize_t ret;
 298
 299         ret = double_flag_store(kobj, attr, buf, count,
 300                                 TRANSPARENT_HUGEPAGE_FLAG,
 301                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 302
 303         if (ret > 0) {
 304                 int err;
 305
 306                 mutex_lock(&khugepaged_mutex);
 307                 err = start_khugepaged();
 308                 mutex_unlock(&khugepaged_mutex);
 309
 310                 if (err)
 311                         ret = err;
 312         }
 313
 314         return ret;
 315 }
 316 static struct kobj_attribute enabled_attr =
 317         __ATTR(enabled, 0644, enabled_show, enabled_store);
 318
 319 static ssize_t single_flag_show(struct kobject *kobj,
 320                                 struct kobj_attribute *attr, char *buf,
 321                                 enum transparent_hugepage_flag flag)
 322 {
 323         return sprintf(buf, "%d\n",
 324                        !!test_bit(flag, &transparent_hugepage_flags));
 325 }
 326
 327 static ssize_t single_flag_store(struct kobject *kobj,
 328                                  struct kobj_attribute *attr,
 329                                  const char *buf, size_t count,
 330                                  enum transparent_hugepage_flag flag)
 331 {
 332         unsigned long value;
 333         int ret;
 334
 335         ret = kstrtoul(buf, 10, &value);
 336         if (ret < 0)
 337                 return ret;
 338         if (value > 1)
 339                 return -EINVAL;
 340
 341         if (value)
 342                 set_bit(flag, &transparent_hugepage_flags);
 343         else
 344                 clear_bit(flag, &transparent_hugepage_flags);
 345
 346         return count;
 347 }
 348
 349 /*
 350  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 351  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 352  * memory just to allocate one more hugepage.
 353  */
 354 static ssize_t defrag_show(struct kobject *kobj,
 355                            struct kobj_attribute *attr, char *buf)
 356 {
 357         return double_flag_show(kobj, attr, buf,
 358                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 359                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 360 }
 361 static ssize_t defrag_store(struct kobject *kobj,
 362                             struct kobj_attribute *attr,
 363                             const char *buf, size_t count)
 364 {
 365         return double_flag_store(kobj, attr, buf, count,
 366                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 367                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 368 }
 369 static struct kobj_attribute defrag_attr =
 370         __ATTR(defrag, 0644, defrag_show, defrag_store);
 371
 372 static ssize_t use_zero_page_show(struct kobject *kobj,
 373                 struct kobj_attribute *attr, char *buf)
 374 {
 375         return single_flag_show(kobj, attr, buf,
 376                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static ssize_t use_zero_page_store(struct kobject *kobj,
 379                 struct kobj_attribute *attr, const char *buf, size_t count)
 380 {
 381         return single_flag_store(kobj, attr, buf, count,
 382                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 383 }
 384 static struct kobj_attribute use_zero_page_attr =
 385         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 386 #ifdef CONFIG_DEBUG_VM
 387 static ssize_t debug_cow_show(struct kobject *kobj,
 388                                 struct kobj_attribute *attr, char *buf)
 389 {
 390         return single_flag_show(kobj, attr, buf,
 391                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 392 }
 393 static ssize_t debug_cow_store(struct kobject *kobj,
 394                                struct kobj_attribute *attr,
 395                                const char *buf, size_t count)
 396 {
 397         return single_flag_store(kobj, attr, buf, count,
 398                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 399 }
 400 static struct kobj_attribute debug_cow_attr =
 401         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 402 #endif /* CONFIG_DEBUG_VM */
 403
 404 static struct attribute *hugepage_attr[] = {
 405         &enabled_attr.attr,
 406         &defrag_attr.attr,
 407         &use_zero_page_attr.attr,
 408 #ifdef CONFIG_DEBUG_VM
 409         &debug_cow_attr.attr,
 410 #endif
 411         NULL,
 412 };
 413
 414 static struct attribute_group hugepage_attr_group = {
 415         .attrs = hugepage_attr,
 416 };
 417
 418 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 419                                          struct kobj_attribute *attr,
 420                                          char *buf)
 421 {
 422         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 423 }
 424
 425 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 426                                           struct kobj_attribute *attr,
 427                                           const char *buf, size_t count)
 428 {
 429         unsigned long msecs;
 430         int err;
 431
 432         err = kstrtoul(buf, 10, &msecs);
 433         if (err || msecs > UINT_MAX)
 434                 return -EINVAL;
 435
 436         khugepaged_scan_sleep_millisecs = msecs;
 437         wake_up_interruptible(&khugepaged_wait);
 438
 439         return count;
 440 }
 441 static struct kobj_attribute scan_sleep_millisecs_attr =
 442         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 443                scan_sleep_millisecs_store);
 444
 445 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 446                                           struct kobj_attribute *attr,
 447                                           char *buf)
 448 {
 449         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 450 }
 451
 452 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 453                                            struct kobj_attribute *attr,
 454                                            const char *buf, size_t count)
 455 {
 456         unsigned long msecs;
 457         int err;
 458
 459         err = kstrtoul(buf, 10, &msecs);
 460         if (err || msecs > UINT_MAX)
 461                 return -EINVAL;
 462
 463         khugepaged_alloc_sleep_millisecs = msecs;
 464         wake_up_interruptible(&khugepaged_wait);
 465
 466         return count;
 467 }
 468 static struct kobj_attribute alloc_sleep_millisecs_attr =
 469         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 470                alloc_sleep_millisecs_store);
 471
 472 static ssize_t pages_to_scan_show(struct kobject *kobj,
 473                                   struct kobj_attribute *attr,
 474                                   char *buf)
 475 {
 476         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 477 }
 478 static ssize_t pages_to_scan_store(struct kobject *kobj,
 479                                    struct kobj_attribute *attr,
 480                                    const char *buf, size_t count)
 481 {
 482         int err;
 483         unsigned long pages;
 484
 485         err = kstrtoul(buf, 10, &pages);
 486         if (err || !pages || pages > UINT_MAX)
 487                 return -EINVAL;
 488
 489         khugepaged_pages_to_scan = pages;
 490
 491         return count;
 492 }
 493 static struct kobj_attribute pages_to_scan_attr =
 494         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 495                pages_to_scan_store);
 496
 497 static ssize_t pages_collapsed_show(struct kobject *kobj,
 498                                     struct kobj_attribute *attr,
 499                                     char *buf)
 500 {
 501         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 502 }
 503 static struct kobj_attribute pages_collapsed_attr =
 504         __ATTR_RO(pages_collapsed);
 505
 506 static ssize_t full_scans_show(struct kobject *kobj,
 507                                struct kobj_attribute *attr,
 508                                char *buf)
 509 {
 510         return sprintf(buf, "%u\n", khugepaged_full_scans);
 511 }
 512 static struct kobj_attribute full_scans_attr =
 513         __ATTR_RO(full_scans);
 514
 515 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 516                                       struct kobj_attribute *attr, char *buf)
 517 {
 518         return single_flag_show(kobj, attr, buf,
 519                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 520 }
 521 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 522                                        struct kobj_attribute *attr,
 523                                        const char *buf, size_t count)
 524 {
 525         return single_flag_store(kobj, attr, buf, count,
 526                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 527 }
 528 static struct kobj_attribute khugepaged_defrag_attr =
 529         __ATTR(defrag, 0644, khugepaged_defrag_show,
 530                khugepaged_defrag_store);
 531
 532 /*
 533  * max_ptes_none controls if khugepaged should collapse hugepages over
 534  * any unmapped ptes in turn potentially increasing the memory
 535  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 536  * reduce the available free memory in the system as it
 537  * runs. Increasing max_ptes_none will instead potentially reduce the
 538  * free memory in the system during the khugepaged scan.
 539  */
 540 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 541                                              struct kobj_attribute *attr,
 542                                              char *buf)
 543 {
 544         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 545 }
 546 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 547                                               struct kobj_attribute *attr,
 548                                               const char *buf, size_t count)
 549 {
 550         int err;
 551         unsigned long max_ptes_none;
 552
 553         err = kstrtoul(buf, 10, &max_ptes_none);
 554         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 555                 return -EINVAL;
 556
 557         khugepaged_max_ptes_none = max_ptes_none;
 558
 559         return count;
 560 }
 561 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 562         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 563                khugepaged_max_ptes_none_store);
 564
 565 static struct attribute *khugepaged_attr[] = {
 566         &khugepaged_defrag_attr.attr,
 567         &khugepaged_max_ptes_none_attr.attr,
 568         &pages_to_scan_attr.attr,
 569         &pages_collapsed_attr.attr,
 570         &full_scans_attr.attr,
 571         &scan_sleep_millisecs_attr.attr,
 572         &alloc_sleep_millisecs_attr.attr,
 573         NULL,
 574 };
 575
 576 static struct attribute_group khugepaged_attr_group = {
 577         .attrs = khugepaged_attr,
 578         .name = "khugepaged",
 579 };
 580
 581 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 582 {
 583         int err;
 584
 585         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 586         if (unlikely(!*hugepage_kobj)) {
 587                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 588                 return -ENOMEM;
 589         }
 590
 591         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 592         if (err) {
 593                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 594                 goto delete_obj;
 595         }
 596
 597         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 598         if (err) {
 599                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 600                 goto remove_hp_group;
 601         }
 602
 603         return 0;
 604
 605 remove_hp_group:
 606         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 607 delete_obj:
 608         kobject_put(*hugepage_kobj);
 609         return err;
 610 }
 611
 612 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 613 {
 614         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 615         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 616         kobject_put(hugepage_kobj);
 617 }
 618 #else
 619 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 620 {
 621         return 0;
 622 }
 623
 624 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 625 {
 626 }
 627 #endif /* CONFIG_SYSFS */
 628
 629 static int __init hugepage_init(void)
 630 {
 631         int err;
 632         struct kobject *hugepage_kobj;
 633
 634         if (!has_transparent_hugepage()) {
 635                 transparent_hugepage_flags = 0;
 636                 return -EINVAL;
 637         }
 638
 639         err = hugepage_init_sysfs(&hugepage_kobj);
 640         if (err)
 641                 return err;
 642
 643         err = khugepaged_slab_init();
 644         if (err)
 645                 goto out;
 646
 647         register_shrinker(&huge_zero_page_shrinker);
 648
 649         /*
 650          * By default disable transparent hugepages on smaller systems,
 651          * where the extra memory used could hurt more than TLB overhead
 652          * is likely to save.  The admin can still enable it through /sys.
 653          */
 654         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 655                 transparent_hugepage_flags = 0;
 656
 657         start_khugepaged();
 658
 659         return 0;
 660 out:
 661         hugepage_exit_sysfs(hugepage_kobj);
 662         return err;
 663 }
 664 subsys_initcall(hugepage_init);
 665
 666 static int __init setup_transparent_hugepage(char *str)
 667 {
 668         int ret = 0;
 669         if (!str)
 670                 goto out;
 671         if (!strcmp(str, "always")) {
 672                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 673                         &transparent_hugepage_flags);
 674                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 675                           &transparent_hugepage_flags);
 676                 ret = 1;
 677         } else if (!strcmp(str, "madvise")) {
 678                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 679                           &transparent_hugepage_flags);
 680                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 681                         &transparent_hugepage_flags);
 682                 ret = 1;
 683         } else if (!strcmp(str, "never")) {
 684                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 685                           &transparent_hugepage_flags);
 686                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 687                           &transparent_hugepage_flags);
 688                 ret = 1;
 689         }
 690 out:
 691         if (!ret)
 692                 printk(KERN_WARNING
 693                        "transparent_hugepage= cannot parse, ignored\n");
 694         return ret;
 695 }
 696 __setup("transparent_hugepage=", setup_transparent_hugepage);
 697
 698 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 699 {
 700         if (likely(vma->vm_flags & VM_WRITE))
 701                 pmd = pmd_mkwrite(pmd);
 702         return pmd;
 703 }
 704
 705 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 706 {
 707         pmd_t entry;
 708         entry = mk_pmd(page, prot);
 709         entry = pmd_mkhuge(entry);
 710         return entry;
 711 }
 712
 713 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 714                                         struct vm_area_struct *vma,
 715                                         unsigned long haddr, pmd_t *pmd,
 716                                         struct page *page)
 717 {
 718         pgtable_t pgtable;
 719         spinlock_t *ptl;
 720
 721         VM_BUG_ON_PAGE(!PageCompound(page), page);
 722         pgtable = pte_alloc_one(mm, haddr);
 723         if (unlikely(!pgtable))
 724                 return VM_FAULT_OOM;
 725
 726         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 727         /*
 728          * The memory barrier inside __SetPageUptodate makes sure that
 729          * clear_huge_page writes become visible before the set_pmd_at()
 730          * write.
 731          */
 732         __SetPageUptodate(page);
 733
 734         ptl = pmd_lock(mm, pmd);
 735         if (unlikely(!pmd_none(*pmd))) {
 736                 spin_unlock(ptl);
 737                 mem_cgroup_uncharge_page(page);
 738                 put_page(page);
 739                 pte_free(mm, pgtable);
 740         } else {
 741                 pmd_t entry;
 742                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 743                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 744                 page_add_new_anon_rmap(page, vma, haddr);
 745                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 746                 set_pmd_at(mm, haddr, pmd, entry);
 747                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 748                 atomic_long_inc(&mm->nr_ptes);
 749                 spin_unlock(ptl);
 750         }
 751
 752         return 0;
 753 }
 754
 755 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 756 {
 757         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 758 }
 759
 760 static inline struct page *alloc_hugepage_vma(int defrag,
 761                                               struct vm_area_struct *vma,
 762                                               unsigned long haddr, int nd,
 763                                               gfp_t extra_gfp)
 764 {
 765         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 766                                HPAGE_PMD_ORDER, vma, haddr, nd);
 767 }
 768
 769 /* Caller must hold page table lock. */
 770 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 771                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 772                 struct page *zero_page)
 773 {
 774         pmd_t entry;
 775         if (!pmd_none(*pmd))
 776                 return false;
 777         entry = mk_pmd(zero_page, vma->vm_page_prot);
 778         entry = pmd_wrprotect(entry);
 779         entry = pmd_mkhuge(entry);
 780         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 781         set_pmd_at(mm, haddr, pmd, entry);
 782         atomic_long_inc(&mm->nr_ptes);
 783         return true;
 784 }
 785
 786 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 787                                unsigned long address, pmd_t *pmd,
 788                                unsigned int flags)
 789 {
 790         struct page *page;
 791         unsigned long haddr = address & HPAGE_PMD_MASK;
 792
 793         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 794                 return VM_FAULT_FALLBACK;
 795         if (unlikely(anon_vma_prepare(vma)))
 796                 return VM_FAULT_OOM;
 797         if (unlikely(khugepaged_enter(vma)))
 798                 return VM_FAULT_OOM;
 799         if (!(flags & FAULT_FLAG_WRITE) &&
 800                         transparent_hugepage_use_zero_page()) {
 801                 spinlock_t *ptl;
 802                 pgtable_t pgtable;
 803                 struct page *zero_page;
 804                 bool set;
 805                 pgtable = pte_alloc_one(mm, haddr);
 806                 if (unlikely(!pgtable))
 807                         return VM_FAULT_OOM;
 808                 zero_page = get_huge_zero_page();
 809                 if (unlikely(!zero_page)) {
 810                         pte_free(mm, pgtable);
 811                         count_vm_event(THP_FAULT_FALLBACK);
 812                         return VM_FAULT_FALLBACK;
 813                 }
 814                 ptl = pmd_lock(mm, pmd);
 815                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 816                                 zero_page);
 817                 spin_unlock(ptl);
 818                 if (!set) {
 819                         pte_free(mm, pgtable);
 820                         put_huge_zero_page();
 821                 }
 822                 return 0;
 823         }
 824         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 825                         vma, haddr, numa_node_id(), 0);
 826         if (unlikely(!page)) {
 827                 count_vm_event(THP_FAULT_FALLBACK);
 828                 return VM_FAULT_FALLBACK;
 829         }
 830         if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 831                 put_page(page);
 832                 count_vm_event(THP_FAULT_FALLBACK);
 833                 return VM_FAULT_FALLBACK;
 834         }
 835         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 836                 mem_cgroup_uncharge_page(page);
 837                 put_page(page);
 838                 count_vm_event(THP_FAULT_FALLBACK);
 839                 return VM_FAULT_FALLBACK;
 840         }
 841
 842         count_vm_event(THP_FAULT_ALLOC);
 843         return 0;
 844 }
 845
 846 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 847                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 848                   struct vm_area_struct *vma)
 849 {
 850         spinlock_t *dst_ptl, *src_ptl;
 851         struct page *src_page;
 852         pmd_t pmd;
 853         pgtable_t pgtable;
 854         int ret;
 855
 856         ret = -ENOMEM;
 857         pgtable = pte_alloc_one(dst_mm, addr);
 858         if (unlikely(!pgtable))
 859                 goto out;
 860
 861         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 862         src_ptl = pmd_lockptr(src_mm, src_pmd);
 863         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 864
 865         ret = -EAGAIN;
 866         pmd = *src_pmd;
 867         if (unlikely(!pmd_trans_huge(pmd))) {
 868                 pte_free(dst_mm, pgtable);
 869                 goto out_unlock;
 870         }
 871         /*
 872          * When page table lock is held, the huge zero pmd should not be
 873          * under splitting since we don't split the page itself, only pmd to
 874          * a page table.
 875          */
 876         if (is_huge_zero_pmd(pmd)) {
 877                 struct page *zero_page;
 878                 bool set;
 879                 /*
 880                  * get_huge_zero_page() will never allocate a new page here,
 881                  * since we already have a zero page to copy. It just takes a
 882                  * reference.
 883                  */
 884                 zero_page = get_huge_zero_page();
 885                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 886                                 zero_page);
 887                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 888                 ret = 0;
 889                 goto out_unlock;
 890         }
 891
 892         if (unlikely(pmd_trans_splitting(pmd))) {
 893                 /* split huge page running from under us */
 894                 spin_unlock(src_ptl);
 895                 spin_unlock(dst_ptl);
 896                 pte_free(dst_mm, pgtable);
 897
 898                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 899                 goto out;
 900         }
 901         src_page = pmd_page(pmd);
 902         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 903         get_page(src_page);
 904         page_dup_rmap(src_page);
 905         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 906
 907         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 908         pmd = pmd_mkold(pmd_wrprotect(pmd));
 909         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 910         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 911         atomic_long_inc(&dst_mm->nr_ptes);
 912
 913         ret = 0;
 914 out_unlock:
 915         spin_unlock(src_ptl);
 916         spin_unlock(dst_ptl);
 917 out:
 918         return ret;
 919 }
 920
 921 void huge_pmd_set_accessed(struct mm_struct *mm,
 922                            struct vm_area_struct *vma,
 923                            unsigned long address,
 924                            pmd_t *pmd, pmd_t orig_pmd,
 925                            int dirty)
 926 {
 927         spinlock_t *ptl;
 928         pmd_t entry;
 929         unsigned long haddr;
 930
 931         ptl = pmd_lock(mm, pmd);
 932         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 933                 goto unlock;
 934
 935         entry = pmd_mkyoung(orig_pmd);
 936         haddr = address & HPAGE_PMD_MASK;
 937         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 938                 update_mmu_cache_pmd(vma, address, pmd);
 939
 940 unlock:
 941         spin_unlock(ptl);
 942 }
 943
 944 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 945                 struct vm_area_struct *vma, unsigned long address,
 946                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 947 {
 948         spinlock_t *ptl;
 949         pgtable_t pgtable;
 950         pmd_t _pmd;
 951         struct page *page;
 952         int i, ret = 0;
 953         unsigned long mmun_start;       /* For mmu_notifiers */
 954         unsigned long mmun_end;         /* For mmu_notifiers */
 955
 956         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 957         if (!page) {
 958                 ret |= VM_FAULT_OOM;
 959                 goto out;
 960         }
 961
 962         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 963                 put_page(page);
 964                 ret |= VM_FAULT_OOM;
 965                 goto out;
 966         }
 967
 968         clear_user_highpage(page, address);
 969         __SetPageUptodate(page);
 970
 971         mmun_start = haddr;
 972         mmun_end   = haddr + HPAGE_PMD_SIZE;
 973         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 974
 975         ptl = pmd_lock(mm, pmd);
 976         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 977                 goto out_free_page;
 978
 979         pmdp_clear_flush(vma, haddr, pmd);
 980         /* leave pmd empty until pte is filled */
 981
 982         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
 983         pmd_populate(mm, &_pmd, pgtable);
 984
 985         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 986                 pte_t *pte, entry;
 987                 if (haddr == (address & PAGE_MASK)) {
 988                         entry = mk_pte(page, vma->vm_page_prot);
 989                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 990                         page_add_new_anon_rmap(page, vma, haddr);
 991                 } else {
 992                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
 993                         entry = pte_mkspecial(entry);
 994                 }
 995                 pte = pte_offset_map(&_pmd, haddr);
 996                 VM_BUG_ON(!pte_none(*pte));
 997                 set_pte_at(mm, haddr, pte, entry);
 998                 pte_unmap(pte);
 999         }
1000         smp_wmb(); /* make pte visible before pmd */
1001         pmd_populate(mm, pmd, pgtable);
1002         spin_unlock(ptl);
1003         put_huge_zero_page();
1004         inc_mm_counter(mm, MM_ANONPAGES);
1005
1006         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007
1008         ret |= VM_FAULT_WRITE;
1009 out:
1010         return ret;
1011 out_free_page:
1012         spin_unlock(ptl);
1013         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1014         mem_cgroup_uncharge_page(page);
1015         put_page(page);
1016         goto out;
1017 }
1018
1019 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1020                                         struct vm_area_struct *vma,
1021                                         unsigned long address,
1022                                         pmd_t *pmd, pmd_t orig_pmd,
1023                                         struct page *page,
1024                                         unsigned long haddr)
1025 {
1026         spinlock_t *ptl;
1027         pgtable_t pgtable;
1028         pmd_t _pmd;
1029         int ret = 0, i;
1030         struct page **pages;
1031         unsigned long mmun_start;       /* For mmu_notifiers */
1032         unsigned long mmun_end;         /* For mmu_notifiers */
1033
1034         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1035                         GFP_KERNEL);
1036         if (unlikely(!pages)) {
1037                 ret |= VM_FAULT_OOM;
1038                 goto out;
1039         }
1040
1041         for (i = 0; i < HPAGE_PMD_NR; i++) {
1042                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1043                                                __GFP_OTHER_NODE,
1044                                                vma, address, page_to_nid(page));
1045                 if (unlikely(!pages[i] ||
1046                              mem_cgroup_newpage_charge(pages[i], mm,
1047                                                        GFP_KERNEL))) {
1048                         if (pages[i])
1049                                 put_page(pages[i]);
1050                         mem_cgroup_uncharge_start();
1051                         while (--i >= 0) {
1052                                 mem_cgroup_uncharge_page(pages[i]);
1053                                 put_page(pages[i]);
1054                         }
1055                         mem_cgroup_uncharge_end();
1056                         kfree(pages);
1057                         ret |= VM_FAULT_OOM;
1058                         goto out;
1059                 }
1060         }
1061
1062         for (i = 0; i < HPAGE_PMD_NR; i++) {
1063                 copy_user_highpage(pages[i], page + i,
1064                                    haddr + PAGE_SIZE * i, vma);
1065                 __SetPageUptodate(pages[i]);
1066                 cond_resched();
1067         }
1068
1069         mmun_start = haddr;
1070         mmun_end   = haddr + HPAGE_PMD_SIZE;
1071         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1072
1073         ptl = pmd_lock(mm, pmd);
1074         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1075                 goto out_free_pages;
1076         VM_BUG_ON_PAGE(!PageHead(page), page);
1077
1078         pmdp_clear_flush(vma, haddr, pmd);
1079         /* leave pmd empty until pte is filled */
1080
1081         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1082         pmd_populate(mm, &_pmd, pgtable);
1083
1084         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1085                 pte_t *pte, entry;
1086                 entry = mk_pte(pages[i], vma->vm_page_prot);
1087                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1088                 page_add_new_anon_rmap(pages[i], vma, haddr);
1089                 pte = pte_offset_map(&_pmd, haddr);
1090                 VM_BUG_ON(!pte_none(*pte));
1091                 set_pte_at(mm, haddr, pte, entry);
1092                 pte_unmap(pte);
1093         }
1094         kfree(pages);
1095
1096         smp_wmb(); /* make pte visible before pmd */
1097         pmd_populate(mm, pmd, pgtable);
1098         page_remove_rmap(page);
1099         spin_unlock(ptl);
1100
1101         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1102
1103         ret |= VM_FAULT_WRITE;
1104         put_page(page);
1105
1106 out:
1107         return ret;
1108
1109 out_free_pages:
1110         spin_unlock(ptl);
1111         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1112         mem_cgroup_uncharge_start();
1113         for (i = 0; i < HPAGE_PMD_NR; i++) {
1114                 mem_cgroup_uncharge_page(pages[i]);
1115                 put_page(pages[i]);
1116         }
1117         mem_cgroup_uncharge_end();
1118         kfree(pages);
1119         goto out;
1120 }
1121
1122 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1123                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1124 {
1125         spinlock_t *ptl;
1126         int ret = 0;
1127         struct page *page = NULL, *new_page;
1128         unsigned long haddr;
1129         unsigned long mmun_start;       /* For mmu_notifiers */
1130         unsigned long mmun_end;         /* For mmu_notifiers */
1131
1132         ptl = pmd_lockptr(mm, pmd);
1133         VM_BUG_ON(!vma->anon_vma);
1134         haddr = address & HPAGE_PMD_MASK;
1135         if (is_huge_zero_pmd(orig_pmd))
1136                 goto alloc;
1137         spin_lock(ptl);
1138         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1139                 goto out_unlock;
1140
1141         page = pmd_page(orig_pmd);
1142         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1143         if (page_mapcount(page) == 1) {
1144                 pmd_t entry;
1145                 entry = pmd_mkyoung(orig_pmd);
1146                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1147                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1148                         update_mmu_cache_pmd(vma, address, pmd);
1149                 ret |= VM_FAULT_WRITE;
1150                 goto out_unlock;
1151         }
1152         get_page(page);
1153         spin_unlock(ptl);
1154 alloc:
1155         if (transparent_hugepage_enabled(vma) &&
1156             !transparent_hugepage_debug_cow())
1157                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1158                                               vma, haddr, numa_node_id(), 0);
1159         else
1160                 new_page = NULL;
1161
1162         if (unlikely(!new_page)) {
1163                 if (!page) {
1164                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1165                                         address, pmd, orig_pmd, haddr);
1166                 } else {
1167                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1168                                         pmd, orig_pmd, page, haddr);
1169                         if (ret & VM_FAULT_OOM)
1170                                 split_huge_page(page);
1171                         put_page(page);
1172                 }
1173                 count_vm_event(THP_FAULT_FALLBACK);
1174                 goto out;
1175         }
1176
1177         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1178                 put_page(new_page);
1179                 if (page) {
1180                         split_huge_page(page);
1181                         put_page(page);
1182                 }
1183                 count_vm_event(THP_FAULT_FALLBACK);
1184                 ret |= VM_FAULT_OOM;
1185                 goto out;
1186         }
1187
1188         count_vm_event(THP_FAULT_ALLOC);
1189
1190         if (!page)
1191                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1192         else
1193                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1194         __SetPageUptodate(new_page);
1195
1196         mmun_start = haddr;
1197         mmun_end   = haddr + HPAGE_PMD_SIZE;
1198         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1199
1200         spin_lock(ptl);
1201         if (page)
1202                 put_page(page);
1203         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1204                 spin_unlock(ptl);
1205                 mem_cgroup_uncharge_page(new_page);
1206                 put_page(new_page);
1207                 goto out_mn;
1208         } else {
1209                 pmd_t entry;
1210                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1211                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1212                 pmdp_clear_flush(vma, haddr, pmd);
1213                 page_add_new_anon_rmap(new_page, vma, haddr);
1214                 set_pmd_at(mm, haddr, pmd, entry);
1215                 update_mmu_cache_pmd(vma, address, pmd);
1216                 if (!page) {
1217                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1218                         put_huge_zero_page();
1219                 } else {
1220                         VM_BUG_ON_PAGE(!PageHead(page), page);
1221                         page_remove_rmap(page);
1222                         put_page(page);
1223                 }
1224                 ret |= VM_FAULT_WRITE;
1225         }
1226         spin_unlock(ptl);
1227 out_mn:
1228         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1229 out:
1230         return ret;
1231 out_unlock:
1232         spin_unlock(ptl);
1233         return ret;
1234 }
1235
1236 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1237                                    unsigned long addr,
1238                                    pmd_t *pmd,
1239                                    unsigned int flags)
1240 {
1241         struct mm_struct *mm = vma->vm_mm;
1242         struct page *page = NULL;
1243
1244         assert_spin_locked(pmd_lockptr(mm, pmd));
1245
1246         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1247                 goto out;
1248
1249         /* Avoid dumping huge zero page */
1250         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1251                 return ERR_PTR(-EFAULT);
1252
1253         /* Full NUMA hinting faults to serialise migration in fault paths */
1254         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1255                 goto out;
1256
1257         page = pmd_page(*pmd);
1258         VM_BUG_ON_PAGE(!PageHead(page), page);
1259         if (flags & FOLL_TOUCH) {
1260                 pmd_t _pmd;
1261                 /*
1262                  * We should set the dirty bit only for FOLL_WRITE but
1263                  * for now the dirty bit in the pmd is meaningless.
1264                  * And if the dirty bit will become meaningful and
1265                  * we'll only set it with FOLL_WRITE, an atomic
1266                  * set_bit will be required on the pmd to set the
1267                  * young bit, instead of the current set_pmd_at.
1268                  */
1269                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1270                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1271                                           pmd, _pmd,  1))
1272                         update_mmu_cache_pmd(vma, addr, pmd);
1273         }
1274         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1275                 if (page->mapping && trylock_page(page)) {
1276                         lru_add_drain();
1277                         if (page->mapping)
1278                                 mlock_vma_page(page);
1279                         unlock_page(page);
1280                 }
1281         }
1282         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1283         VM_BUG_ON_PAGE(!PageCompound(page), page);
1284         if (flags & FOLL_GET)
1285                 get_page_foll(page);
1286
1287 out:
1288         return page;
1289 }
1290
1291 /* NUMA hinting page fault entry point for trans huge pmds */
1292 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1293                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1294 {
1295         spinlock_t *ptl;
1296         struct anon_vma *anon_vma = NULL;
1297         struct page *page;
1298         unsigned long haddr = addr & HPAGE_PMD_MASK;
1299         int page_nid = -1, this_nid = numa_node_id();
1300         int target_nid, last_cpupid = -1;
1301         bool page_locked;
1302         bool migrated = false;
1303         int flags = 0;
1304
1305         ptl = pmd_lock(mm, pmdp);
1306         if (unlikely(!pmd_same(pmd, *pmdp)))
1307                 goto out_unlock;
1308
1309         /*
1310          * If there are potential migrations, wait for completion and retry
1311          * without disrupting NUMA hinting information. Do not relock and
1312          * check_same as the page may no longer be mapped.
1313          */
1314         if (unlikely(pmd_trans_migrating(*pmdp))) {
1315                 spin_unlock(ptl);
1316                 wait_migrate_huge_page(vma->anon_vma, pmdp);
1317                 goto out;
1318         }
1319
1320         page = pmd_page(pmd);
1321         BUG_ON(is_huge_zero_page(page));
1322         page_nid = page_to_nid(page);
1323         last_cpupid = page_cpupid_last(page);
1324         count_vm_numa_event(NUMA_HINT_FAULTS);
1325         if (page_nid == this_nid) {
1326                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1327                 flags |= TNF_FAULT_LOCAL;
1328         }
1329
1330         /*
1331          * Avoid grouping on DSO/COW pages in specific and RO pages
1332          * in general, RO pages shouldn't hurt as much anyway since
1333          * they can be in shared cache state.
1334          */
1335         if (!pmd_write(pmd))
1336                 flags |= TNF_NO_GROUP;
1337
1338         /*
1339          * Acquire the page lock to serialise THP migrations but avoid dropping
1340          * page_table_lock if at all possible
1341          */
1342         page_locked = trylock_page(page);
1343         target_nid = mpol_misplaced(page, vma, haddr);
1344         if (target_nid == -1) {
1345                 /* If the page was locked, there are no parallel migrations */
1346                 if (page_locked)
1347                         goto clear_pmdnuma;
1348         }
1349
1350         /* Migration could have started since the pmd_trans_migrating check */
1351         if (!page_locked) {
1352                 spin_unlock(ptl);
1353                 wait_on_page_locked(page);
1354                 page_nid = -1;
1355                 goto out;
1356         }
1357
1358         /*
1359          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1360          * to serialises splits
1361          */
1362         get_page(page);
1363         spin_unlock(ptl);
1364         anon_vma = page_lock_anon_vma_read(page);
1365
1366         /* Confirm the PMD did not change while page_table_lock was released */
1367         spin_lock(ptl);
1368         if (unlikely(!pmd_same(pmd, *pmdp))) {
1369                 unlock_page(page);
1370                 put_page(page);
1371                 page_nid = -1;
1372                 goto out_unlock;
1373         }
1374
1375         /* Bail if we fail to protect against THP splits for any reason */
1376         if (unlikely(!anon_vma)) {
1377                 put_page(page);
1378                 page_nid = -1;
1379                 goto clear_pmdnuma;
1380         }
1381
1382         /*
1383          * Migrate the THP to the requested node, returns with page unlocked
1384          * and pmd_numa cleared.
1385          */
1386         spin_unlock(ptl);
1387         migrated = migrate_misplaced_transhuge_page(mm, vma,
1388                                 pmdp, pmd, addr, page, target_nid);
1389         if (migrated) {
1390                 flags |= TNF_MIGRATED;
1391                 page_nid = target_nid;
1392         }
1393
1394         goto out;
1395 clear_pmdnuma:
1396         BUG_ON(!PageLocked(page));
1397         pmd = pmd_mknonnuma(pmd);
1398         set_pmd_at(mm, haddr, pmdp, pmd);
1399         VM_BUG_ON(pmd_numa(*pmdp));
1400         update_mmu_cache_pmd(vma, addr, pmdp);
1401         unlock_page(page);
1402 out_unlock:
1403         spin_unlock(ptl);
1404
1405 out:
1406         if (anon_vma)
1407                 page_unlock_anon_vma_read(anon_vma);
1408
1409         if (page_nid != -1)
1410                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1411
1412         return 0;
1413 }
1414
1415 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1416                  pmd_t *pmd, unsigned long addr)
1417 {
1418         spinlock_t *ptl;
1419         int ret = 0;
1420
1421         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1422                 struct page *page;
1423                 pgtable_t pgtable;
1424                 pmd_t orig_pmd;
1425                 /*
1426                  * For architectures like ppc64 we look at deposited pgtable
1427                  * when calling pmdp_get_and_clear. So do the
1428                  * pgtable_trans_huge_withdraw after finishing pmdp related
1429                  * operations.
1430                  */
1431                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1432                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1433                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1434                 if (is_huge_zero_pmd(orig_pmd)) {
1435                         atomic_long_dec(&tlb->mm->nr_ptes);
1436                         spin_unlock(ptl);
1437                         put_huge_zero_page();
1438                 } else {
1439                         page = pmd_page(orig_pmd);
1440                         page_remove_rmap(page);
1441                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1442                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1443                         VM_BUG_ON_PAGE(!PageHead(page), page);
1444                         atomic_long_dec(&tlb->mm->nr_ptes);
1445                         spin_unlock(ptl);
1446                         tlb_remove_page(tlb, page);
1447                 }
1448                 pte_free(tlb->mm, pgtable);
1449                 ret = 1;
1450         }
1451         return ret;
1452 }
1453
1454 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1455                 unsigned long addr, unsigned long end,
1456                 unsigned char *vec)
1457 {
1458         spinlock_t *ptl;
1459         int ret = 0;
1460
1461         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1462                 /*
1463                  * All logical pages in the range are present
1464                  * if backed by a huge page.
1465                  */
1466                 spin_unlock(ptl);
1467                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1468                 ret = 1;
1469         }
1470
1471         return ret;
1472 }
1473
1474 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1475                   unsigned long old_addr,
1476                   unsigned long new_addr, unsigned long old_end,
1477                   pmd_t *old_pmd, pmd_t *new_pmd)
1478 {
1479         spinlock_t *old_ptl, *new_ptl;
1480         int ret = 0;
1481         pmd_t pmd;
1482
1483         struct mm_struct *mm = vma->vm_mm;
1484
1485         if ((old_addr & ~HPAGE_PMD_MASK) ||
1486             (new_addr & ~HPAGE_PMD_MASK) ||
1487             old_end - old_addr < HPAGE_PMD_SIZE ||
1488             (new_vma->vm_flags & VM_NOHUGEPAGE))
1489                 goto out;
1490
1491         /*
1492          * The destination pmd shouldn't be established, free_pgtables()
1493          * should have release it.
1494          */
1495         if (WARN_ON(!pmd_none(*new_pmd))) {
1496                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1497                 goto out;
1498         }
1499
1500         /*
1501          * We don't have to worry about the ordering of src and dst
1502          * ptlocks because exclusive mmap_sem prevents deadlock.
1503          */
1504         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1505         if (ret == 1) {
1506                 new_ptl = pmd_lockptr(mm, new_pmd);
1507                 if (new_ptl != old_ptl)
1508                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1509                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1510                 VM_BUG_ON(!pmd_none(*new_pmd));
1511
1512                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1513                         pgtable_t pgtable;
1514                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1515                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1516                 }
1517                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1518                 if (new_ptl != old_ptl)
1519                         spin_unlock(new_ptl);
1520                 spin_unlock(old_ptl);
1521         }
1522 out:
1523         return ret;
1524 }
1525
1526 /*
1527  * Returns
1528  *  - 0 if PMD could not be locked
1529  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1530  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1531  */
1532 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1533                 unsigned long addr, pgprot_t newprot, int prot_numa)
1534 {
1535         struct mm_struct *mm = vma->vm_mm;
1536         spinlock_t *ptl;
1537         int ret = 0;
1538
1539         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1540                 pmd_t entry;
1541                 ret = 1;
1542                 if (!prot_numa) {
1543                         entry = pmdp_get_and_clear(mm, addr, pmd);
1544                         if (pmd_numa(entry))
1545                                 entry = pmd_mknonnuma(entry);
1546                         entry = pmd_modify(entry, newprot);
1547                         ret = HPAGE_PMD_NR;
1548                         BUG_ON(pmd_write(entry));
1549                 } else {
1550                         struct page *page = pmd_page(*pmd);
1551
1552                         /*
1553                          * Do not trap faults against the zero page. The
1554                          * read-only data is likely to be read-cached on the
1555                          * local CPU cache and it is less useful to know about
1556                          * local vs remote hits on the zero page.
1557                          */
1558                         if (!is_huge_zero_page(page) &&
1559                             !pmd_numa(*pmd)) {
1560                                 entry = *pmd;
1561                                 entry = pmd_mknuma(entry);
1562                                 ret = HPAGE_PMD_NR;
1563                         }
1564                 }
1565
1566                 /* Set PMD if cleared earlier */
1567                 if (ret == HPAGE_PMD_NR)
1568                         set_pmd_at(mm, addr, pmd, entry);
1569
1570                 spin_unlock(ptl);
1571         }
1572
1573         return ret;
1574 }
1575
1576 /*
1577  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1578  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1579  *
1580  * Note that if it returns 1, this routine returns without unlocking page
1581  * table locks. So callers must unlock them.
1582  */
1583 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1584                 spinlock_t **ptl)
1585 {
1586         *ptl = pmd_lock(vma->vm_mm, pmd);
1587         if (likely(pmd_trans_huge(*pmd))) {
1588                 if (unlikely(pmd_trans_splitting(*pmd))) {
1589                         spin_unlock(*ptl);
1590                         wait_split_huge_page(vma->anon_vma, pmd);
1591                         return -1;
1592                 } else {
1593                         /* Thp mapped by 'pmd' is stable, so we can
1594                          * handle it as it is. */
1595                         return 1;
1596                 }
1597         }
1598         spin_unlock(*ptl);
1599         return 0;
1600 }
1601
1602 /*
1603  * This function returns whether a given @page is mapped onto the @address
1604  * in the virtual space of @mm.
1605  *
1606  * When it's true, this function returns *pmd with holding the page table lock
1607  * and passing it back to the caller via @ptl.
1608  * If it's false, returns NULL without holding the page table lock.
1609  */
1610 pmd_t *page_check_address_pmd(struct page *page,
1611                               struct mm_struct *mm,
1612                               unsigned long address,
1613                               enum page_check_address_pmd_flag flag,
1614                               spinlock_t **ptl)
1615 {
1616         pmd_t *pmd;
1617
1618         if (address & ~HPAGE_PMD_MASK)
1619                 return NULL;
1620
1621         pmd = mm_find_pmd(mm, address);
1622         if (!pmd)
1623                 return NULL;
1624         *ptl = pmd_lock(mm, pmd);
1625         if (pmd_none(*pmd))
1626                 goto unlock;
1627         if (pmd_page(*pmd) != page)
1628                 goto unlock;
1629         /*
1630          * split_vma() may create temporary aliased mappings. There is
1631          * no risk as long as all huge pmd are found and have their
1632          * splitting bit set before __split_huge_page_refcount
1633          * runs. Finding the same huge pmd more than once during the
1634          * same rmap walk is not a problem.
1635          */
1636         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1637             pmd_trans_splitting(*pmd))
1638                 goto unlock;
1639         if (pmd_trans_huge(*pmd)) {
1640                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1641                           !pmd_trans_splitting(*pmd));
1642                 return pmd;
1643         }
1644 unlock:
1645         spin_unlock(*ptl);
1646         return NULL;
1647 }
1648
1649 static int __split_huge_page_splitting(struct page *page,
1650                                        struct vm_area_struct *vma,
1651                                        unsigned long address)
1652 {
1653         struct mm_struct *mm = vma->vm_mm;
1654         spinlock_t *ptl;
1655         pmd_t *pmd;
1656         int ret = 0;
1657         /* For mmu_notifiers */
1658         const unsigned long mmun_start = address;
1659         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1660
1661         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1662         pmd = page_check_address_pmd(page, mm, address,
1663                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1664         if (pmd) {
1665                 /*
1666                  * We can't temporarily set the pmd to null in order
1667                  * to split it, the pmd must remain marked huge at all
1668                  * times or the VM won't take the pmd_trans_huge paths
1669                  * and it won't wait on the anon_vma->root->rwsem to
1670                  * serialize against split_huge_page*.
1671                  */
1672                 pmdp_splitting_flush(vma, address, pmd);
1673                 ret = 1;
1674                 spin_unlock(ptl);
1675         }
1676         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1677
1678         return ret;
1679 }
1680
1681 static void __split_huge_page_refcount(struct page *page,
1682                                        struct list_head *list)
1683 {
1684         int i;
1685         struct zone *zone = page_zone(page);
1686         struct lruvec *lruvec;
1687         int tail_count = 0;
1688
1689         /* prevent PageLRU to go away from under us, and freeze lru stats */
1690         spin_lock_irq(&zone->lru_lock);
1691         lruvec = mem_cgroup_page_lruvec(page, zone);
1692
1693         compound_lock(page);
1694         /* complete memcg works before add pages to LRU */
1695         mem_cgroup_split_huge_fixup(page);
1696
1697         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1698                 struct page *page_tail = page + i;
1699
1700                 /* tail_page->_mapcount cannot change */
1701                 BUG_ON(page_mapcount(page_tail) < 0);
1702                 tail_count += page_mapcount(page_tail);
1703                 /* check for overflow */
1704                 BUG_ON(tail_count < 0);
1705                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1706                 /*
1707                  * tail_page->_count is zero and not changing from
1708                  * under us. But get_page_unless_zero() may be running
1709                  * from under us on the tail_page. If we used
1710                  * atomic_set() below instead of atomic_add(), we
1711                  * would then run atomic_set() concurrently with
1712                  * get_page_unless_zero(), and atomic_set() is
1713                  * implemented in C not using locked ops. spin_unlock
1714                  * on x86 sometime uses locked ops because of PPro
1715                  * errata 66, 92, so unless somebody can guarantee
1716                  * atomic_set() here would be safe on all archs (and
1717                  * not only on x86), it's safer to use atomic_add().
1718                  */
1719                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1720                            &page_tail->_count);
1721
1722                 /* after clearing PageTail the gup refcount can be released */
1723                 smp_mb();
1724
1725                 /*
1726                  * retain hwpoison flag of the poisoned tail page:
1727                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1728                  *   by the memory-failure.
1729                  */
1730                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1731                 page_tail->flags |= (page->flags &
1732                                      ((1L << PG_referenced) |
1733                                       (1L << PG_swapbacked) |
1734                                       (1L << PG_mlocked) |
1735                                       (1L << PG_uptodate) |
1736                                       (1L << PG_active) |
1737                                       (1L << PG_unevictable)));
1738                 page_tail->flags |= (1L << PG_dirty);
1739
1740                 /* clear PageTail before overwriting first_page */
1741                 smp_wmb();
1742
1743                 /*
1744                  * __split_huge_page_splitting() already set the
1745                  * splitting bit in all pmd that could map this
1746                  * hugepage, that will ensure no CPU can alter the
1747                  * mapcount on the head page. The mapcount is only
1748                  * accounted in the head page and it has to be
1749                  * transferred to all tail pages in the below code. So
1750                  * for this code to be safe, the split the mapcount
1751                  * can't change. But that doesn't mean userland can't
1752                  * keep changing and reading the page contents while
1753                  * we transfer the mapcount, so the pmd splitting
1754                  * status is achieved setting a reserved bit in the
1755                  * pmd, not by clearing the present bit.
1756                 */
1757                 page_tail->_mapcount = page->_mapcount;
1758
1759                 BUG_ON(page_tail->mapping);
1760                 page_tail->mapping = page->mapping;
1761
1762                 page_tail->index = page->index + i;
1763                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1764
1765                 BUG_ON(!PageAnon(page_tail));
1766                 BUG_ON(!PageUptodate(page_tail));
1767                 BUG_ON(!PageDirty(page_tail));
1768                 BUG_ON(!PageSwapBacked(page_tail));
1769
1770                 lru_add_page_tail(page, page_tail, lruvec, list);
1771         }
1772         atomic_sub(tail_count, &page->_count);
1773         BUG_ON(atomic_read(&page->_count) <= 0);
1774
1775         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1776
1777         ClearPageCompound(page);
1778         compound_unlock(page);
1779         spin_unlock_irq(&zone->lru_lock);
1780
1781         for (i = 1; i < HPAGE_PMD_NR; i++) {
1782                 struct page *page_tail = page + i;
1783                 BUG_ON(page_count(page_tail) <= 0);
1784                 /*
1785                  * Tail pages may be freed if there wasn't any mapping
1786                  * like if add_to_swap() is running on a lru page that
1787                  * had its mapping zapped. And freeing these pages
1788                  * requires taking the lru_lock so we do the put_page
1789                  * of the tail pages after the split is complete.
1790                  */
1791                 put_page(page_tail);
1792         }
1793
1794         /*
1795          * Only the head page (now become a regular page) is required
1796          * to be pinned by the caller.
1797          */
1798         BUG_ON(page_count(page) <= 0);
1799 }
1800
1801 static int __split_huge_page_map(struct page *page,
1802                                  struct vm_area_struct *vma,
1803                                  unsigned long address)
1804 {
1805         struct mm_struct *mm = vma->vm_mm;
1806         spinlock_t *ptl;
1807         pmd_t *pmd, _pmd;
1808         int ret = 0, i;
1809         pgtable_t pgtable;
1810         unsigned long haddr;
1811
1812         pmd = page_check_address_pmd(page, mm, address,
1813                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1814         if (pmd) {
1815                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1816                 pmd_populate(mm, &_pmd, pgtable);
1817
1818                 haddr = address;
1819                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1820                         pte_t *pte, entry;
1821                         BUG_ON(PageCompound(page+i));
1822                         entry = mk_pte(page + i, vma->vm_page_prot);
1823                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1824                         if (!pmd_write(*pmd))
1825                                 entry = pte_wrprotect(entry);
1826                         else
1827                                 BUG_ON(page_mapcount(page) != 1);
1828                         if (!pmd_young(*pmd))
1829                                 entry = pte_mkold(entry);
1830                         if (pmd_numa(*pmd))
1831                                 entry = pte_mknuma(entry);
1832                         pte = pte_offset_map(&_pmd, haddr);
1833                         BUG_ON(!pte_none(*pte));
1834                         set_pte_at(mm, haddr, pte, entry);
1835                         pte_unmap(pte);
1836                 }
1837
1838                 smp_wmb(); /* make pte visible before pmd */
1839                 /*
1840                  * Up to this point the pmd is present and huge and
1841                  * userland has the whole access to the hugepage
1842                  * during the split (which happens in place). If we
1843                  * overwrite the pmd with the not-huge version
1844                  * pointing to the pte here (which of course we could
1845                  * if all CPUs were bug free), userland could trigger
1846                  * a small page size TLB miss on the small sized TLB
1847                  * while the hugepage TLB entry is still established
1848                  * in the huge TLB. Some CPU doesn't like that. See
1849                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1850                  * Erratum 383 on page 93. Intel should be safe but is
1851                  * also warns that it's only safe if the permission
1852                  * and cache attributes of the two entries loaded in
1853                  * the two TLB is identical (which should be the case
1854                  * here). But it is generally safer to never allow
1855                  * small and huge TLB entries for the same virtual
1856                  * address to be loaded simultaneously. So instead of
1857                  * doing "pmd_populate(); flush_tlb_range();" we first
1858                  * mark the current pmd notpresent (atomically because
1859                  * here the pmd_trans_huge and pmd_trans_splitting
1860                  * must remain set at all times on the pmd until the
1861                  * split is complete for this pmd), then we flush the
1862                  * SMP TLB and finally we write the non-huge version
1863                  * of the pmd entry with pmd_populate.
1864                  */
1865                 pmdp_invalidate(vma, address, pmd);
1866                 pmd_populate(mm, pmd, pgtable);
1867                 ret = 1;
1868                 spin_unlock(ptl);
1869         }
1870
1871         return ret;
1872 }
1873
1874 /* must be called with anon_vma->root->rwsem held */
1875 static void __split_huge_page(struct page *page,
1876                               struct anon_vma *anon_vma,
1877                               struct list_head *list)
1878 {
1879         int mapcount, mapcount2;
1880         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1881         struct anon_vma_chain *avc;
1882
1883         BUG_ON(!PageHead(page));
1884         BUG_ON(PageTail(page));
1885
1886         mapcount = 0;
1887         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1888                 struct vm_area_struct *vma = avc->vma;
1889                 unsigned long addr = vma_address(page, vma);
1890                 BUG_ON(is_vma_temporary_stack(vma));
1891                 mapcount += __split_huge_page_splitting(page, vma, addr);
1892         }
1893         /*
1894          * It is critical that new vmas are added to the tail of the
1895          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1896          * and establishes a child pmd before
1897          * __split_huge_page_splitting() freezes the parent pmd (so if
1898          * we fail to prevent copy_huge_pmd() from running until the
1899          * whole __split_huge_page() is complete), we will still see
1900          * the newly established pmd of the child later during the
1901          * walk, to be able to set it as pmd_trans_splitting too.
1902          */
1903         if (mapcount != page_mapcount(page))
1904                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1905                        mapcount, page_mapcount(page));
1906         BUG_ON(mapcount != page_mapcount(page));
1907
1908         __split_huge_page_refcount(page, list);
1909
1910         mapcount2 = 0;
1911         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1912                 struct vm_area_struct *vma = avc->vma;
1913                 unsigned long addr = vma_address(page, vma);
1914                 BUG_ON(is_vma_temporary_stack(vma));
1915                 mapcount2 += __split_huge_page_map(page, vma, addr);
1916         }
1917         if (mapcount != mapcount2)
1918                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1919                        mapcount, mapcount2, page_mapcount(page));
1920         BUG_ON(mapcount != mapcount2);
1921 }
1922
1923 /*
1924  * Split a hugepage into normal pages. This doesn't change the position of head
1925  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1926  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1927  * from the hugepage.
1928  * Return 0 if the hugepage is split successfully otherwise return 1.
1929  */
1930 int split_huge_page_to_list(struct page *page, struct list_head *list)
1931 {
1932         struct anon_vma *anon_vma;
1933         int ret = 1;
1934
1935         BUG_ON(is_huge_zero_page(page));
1936         BUG_ON(!PageAnon(page));
1937
1938         /*
1939          * The caller does not necessarily hold an mmap_sem that would prevent
1940          * the anon_vma disappearing so we first we take a reference to it
1941          * and then lock the anon_vma for write. This is similar to
1942          * page_lock_anon_vma_read except the write lock is taken to serialise
1943          * against parallel split or collapse operations.
1944          */
1945         anon_vma = page_get_anon_vma(page);
1946         if (!anon_vma)
1947                 goto out;
1948         anon_vma_lock_write(anon_vma);
1949
1950         ret = 0;
1951         if (!PageCompound(page))
1952                 goto out_unlock;
1953
1954         BUG_ON(!PageSwapBacked(page));
1955         __split_huge_page(page, anon_vma, list);
1956         count_vm_event(THP_SPLIT);
1957
1958         BUG_ON(PageCompound(page));
1959 out_unlock:
1960         anon_vma_unlock_write(anon_vma);
1961         put_anon_vma(anon_vma);
1962 out:
1963         return ret;
1964 }
1965
1966 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1967
1968 int hugepage_madvise(struct vm_area_struct *vma,
1969                      unsigned long *vm_flags, int advice)
1970 {
1971         struct mm_struct *mm = vma->vm_mm;
1972
1973         switch (advice) {
1974         case MADV_HUGEPAGE:
1975                 /*
1976                  * Be somewhat over-protective like KSM for now!
1977                  */
1978                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1979                         return -EINVAL;
1980                 if (mm->def_flags & VM_NOHUGEPAGE)
1981                         return -EINVAL;
1982                 *vm_flags &= ~VM_NOHUGEPAGE;
1983                 *vm_flags |= VM_HUGEPAGE;
1984                 /*
1985                  * If the vma become good for khugepaged to scan,
1986                  * register it here without waiting a page fault that
1987                  * may not happen any time soon.
1988                  */
1989                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1990                         return -ENOMEM;
1991                 break;
1992         case MADV_NOHUGEPAGE:
1993                 /*
1994                  * Be somewhat over-protective like KSM for now!
1995                  */
1996                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1997                         return -EINVAL;
1998                 *vm_flags &= ~VM_HUGEPAGE;
1999                 *vm_flags |= VM_NOHUGEPAGE;
2000                 /*
2001                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2002                  * this vma even if we leave the mm registered in khugepaged if
2003                  * it got registered before VM_NOHUGEPAGE was set.
2004                  */
2005                 break;
2006         }
2007
2008         return 0;
2009 }
2010
2011 static int __init khugepaged_slab_init(void)
2012 {
2013         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2014                                           sizeof(struct mm_slot),
2015                                           __alignof__(struct mm_slot), 0, NULL);
2016         if (!mm_slot_cache)
2017                 return -ENOMEM;
2018
2019         return 0;
2020 }
2021
2022 static inline struct mm_slot *alloc_mm_slot(void)
2023 {
2024         if (!mm_slot_cache)     /* initialization failed */
2025                 return NULL;
2026         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2027 }
2028
2029 static inline void free_mm_slot(struct mm_slot *mm_slot)
2030 {
2031         kmem_cache_free(mm_slot_cache, mm_slot);
2032 }
2033
2034 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2035 {
2036         struct mm_slot *mm_slot;
2037
2038         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2039                 if (mm == mm_slot->mm)
2040                         return mm_slot;
2041
2042         return NULL;
2043 }
2044
2045 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2046                                     struct mm_slot *mm_slot)
2047 {
2048         mm_slot->mm = mm;
2049         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2050 }
2051
2052 static inline int khugepaged_test_exit(struct mm_struct *mm)
2053 {
2054         return atomic_read(&mm->mm_users) == 0;
2055 }
2056
2057 int __khugepaged_enter(struct mm_struct *mm)
2058 {
2059         struct mm_slot *mm_slot;
2060         int wakeup;
2061
2062         mm_slot = alloc_mm_slot();
2063         if (!mm_slot)
2064                 return -ENOMEM;
2065
2066         /* __khugepaged_exit() must not run from under us */
2067         VM_BUG_ON(khugepaged_test_exit(mm));
2068         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2069                 free_mm_slot(mm_slot);
2070                 return 0;
2071         }
2072
2073         spin_lock(&khugepaged_mm_lock);
2074         insert_to_mm_slots_hash(mm, mm_slot);
2075         /*
2076          * Insert just behind the scanning cursor, to let the area settle
2077          * down a little.
2078          */
2079         wakeup = list_empty(&khugepaged_scan.mm_head);
2080         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2081         spin_unlock(&khugepaged_mm_lock);
2082
2083         atomic_inc(&mm->mm_count);
2084         if (wakeup)
2085                 wake_up_interruptible(&khugepaged_wait);
2086
2087         return 0;
2088 }
2089
2090 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2091 {
2092         unsigned long hstart, hend;
2093         if (!vma->anon_vma)
2094                 /*
2095                  * Not yet faulted in so we will register later in the
2096                  * page fault if needed.
2097                  */
2098                 return 0;
2099         if (vma->vm_ops)
2100                 /* khugepaged not yet working on file or special mappings */
2101                 return 0;
2102         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2103         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2104         hend = vma->vm_end & HPAGE_PMD_MASK;
2105         if (hstart < hend)
2106                 return khugepaged_enter(vma);
2107         return 0;
2108 }
2109
2110 void __khugepaged_exit(struct mm_struct *mm)
2111 {
2112         struct mm_slot *mm_slot;
2113         int free = 0;
2114
2115         spin_lock(&khugepaged_mm_lock);
2116         mm_slot = get_mm_slot(mm);
2117         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2118                 hash_del(&mm_slot->hash);
2119                 list_del(&mm_slot->mm_node);
2120                 free = 1;
2121         }
2122         spin_unlock(&khugepaged_mm_lock);
2123
2124         if (free) {
2125                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2126                 free_mm_slot(mm_slot);
2127                 mmdrop(mm);
2128         } else if (mm_slot) {
2129                 /*
2130                  * This is required to serialize against
2131                  * khugepaged_test_exit() (which is guaranteed to run
2132                  * under mmap sem read mode). Stop here (after we
2133                  * return all pagetables will be destroyed) until
2134                  * khugepaged has finished working on the pagetables
2135                  * under the mmap_sem.
2136                  */
2137                 down_write(&mm->mmap_sem);
2138                 up_write(&mm->mmap_sem);
2139         }
2140 }
2141
2142 static void release_pte_page(struct page *page)
2143 {
2144         /* 0 stands for page_is_file_cache(page) == false */
2145         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2146         unlock_page(page);
2147         putback_lru_page(page);
2148 }
2149
2150 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2151 {
2152         while (--_pte >= pte) {
2153                 pte_t pteval = *_pte;
2154                 if (!pte_none(pteval))
2155                         release_pte_page(pte_page(pteval));
2156         }
2157 }
2158
2159 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2160                                         unsigned long address,
2161                                         pte_t *pte)
2162 {
2163         struct page *page;
2164         pte_t *_pte;
2165         int referenced = 0, none = 0;
2166         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2167              _pte++, address += PAGE_SIZE) {
2168                 pte_t pteval = *_pte;
2169                 if (pte_none(pteval)) {
2170                         if (++none <= khugepaged_max_ptes_none)
2171                                 continue;
2172                         else
2173                                 goto out;
2174                 }
2175                 if (!pte_present(pteval) || !pte_write(pteval))
2176                         goto out;
2177                 page = vm_normal_page(vma, address, pteval);
2178                 if (unlikely(!page))
2179                         goto out;
2180
2181                 VM_BUG_ON_PAGE(PageCompound(page), page);
2182                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2183                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2184
2185                 /* cannot use mapcount: can't collapse if there's a gup pin */
2186                 if (page_count(page) != 1)
2187                         goto out;
2188                 /*
2189                  * We can do it before isolate_lru_page because the
2190                  * page can't be freed from under us. NOTE: PG_lock
2191                  * is needed to serialize against split_huge_page
2192                  * when invoked from the VM.
2193                  */
2194                 if (!trylock_page(page))
2195                         goto out;
2196                 /*
2197                  * Isolate the page to avoid collapsing an hugepage
2198                  * currently in use by the VM.
2199                  */
2200                 if (isolate_lru_page(page)) {
2201                         unlock_page(page);
2202                         goto out;
2203                 }
2204                 /* 0 stands for page_is_file_cache(page) == false */
2205                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2206                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2207                 VM_BUG_ON_PAGE(PageLRU(page), page);
2208
2209                 /* If there is no mapped pte young don't collapse the page */
2210                 if (pte_young(pteval) || PageReferenced(page) ||
2211                     mmu_notifier_test_young(vma->vm_mm, address))
2212                         referenced = 1;
2213         }
2214         if (likely(referenced))
2215                 return 1;
2216 out:
2217         release_pte_pages(pte, _pte);
2218         return 0;
2219 }
2220
2221 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2222                                       struct vm_area_struct *vma,
2223                                       unsigned long address,
2224                                       spinlock_t *ptl)
2225 {
2226         pte_t *_pte;
2227         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2228                 pte_t pteval = *_pte;
2229                 struct page *src_page;
2230
2231                 if (pte_none(pteval)) {
2232                         clear_user_highpage(page, address);
2233                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2234                 } else {
2235                         src_page = pte_page(pteval);
2236                         copy_user_highpage(page, src_page, address, vma);
2237                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2238                         release_pte_page(src_page);
2239                         /*
2240                          * ptl mostly unnecessary, but preempt has to
2241                          * be disabled to update the per-cpu stats
2242                          * inside page_remove_rmap().
2243                          */
2244                         spin_lock(ptl);
2245                         /*
2246                          * paravirt calls inside pte_clear here are
2247                          * superfluous.
2248                          */
2249                         pte_clear(vma->vm_mm, address, _pte);
2250                         page_remove_rmap(src_page);
2251                         spin_unlock(ptl);
2252                         free_page_and_swap_cache(src_page);
2253                 }
2254
2255                 address += PAGE_SIZE;
2256                 page++;
2257         }
2258 }
2259
2260 static void khugepaged_alloc_sleep(void)
2261 {
2262         wait_event_freezable_timeout(khugepaged_wait, false,
2263                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2264 }
2265
2266 static int khugepaged_node_load[MAX_NUMNODES];
2267
2268 #ifdef CONFIG_NUMA
2269 static int khugepaged_find_target_node(void)
2270 {
2271         static int last_khugepaged_target_node = NUMA_NO_NODE;
2272         int nid, target_node = 0, max_value = 0;
2273
2274         /* find first node with max normal pages hit */
2275         for (nid = 0; nid < MAX_NUMNODES; nid++)
2276                 if (khugepaged_node_load[nid] > max_value) {
2277                         max_value = khugepaged_node_load[nid];
2278                         target_node = nid;
2279                 }
2280
2281         /* do some balance if several nodes have the same hit record */
2282         if (target_node <= last_khugepaged_target_node)
2283                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2284                                 nid++)
2285                         if (max_value == khugepaged_node_load[nid]) {
2286                                 target_node = nid;
2287                                 break;
2288                         }
2289
2290         last_khugepaged_target_node = target_node;
2291         return target_node;
2292 }
2293
2294 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2295 {
2296         if (IS_ERR(*hpage)) {
2297                 if (!*wait)
2298                         return false;
2299
2300                 *wait = false;
2301                 *hpage = NULL;
2302                 khugepaged_alloc_sleep();
2303         } else if (*hpage) {
2304                 put_page(*hpage);
2305                 *hpage = NULL;
2306         }
2307
2308         return true;
2309 }
2310
2311 static struct page
2312 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2313                        struct vm_area_struct *vma, unsigned long address,
2314                        int node)
2315 {
2316         VM_BUG_ON_PAGE(*hpage, *hpage);
2317         /*
2318          * Allocate the page while the vma is still valid and under
2319          * the mmap_sem read mode so there is no memory allocation
2320          * later when we take the mmap_sem in write mode. This is more
2321          * friendly behavior (OTOH it may actually hide bugs) to
2322          * filesystems in userland with daemons allocating memory in
2323          * the userland I/O paths.  Allocating memory with the
2324          * mmap_sem in read mode is good idea also to allow greater
2325          * scalability.
2326          */
2327         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2328                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2329         /*
2330          * After allocating the hugepage, release the mmap_sem read lock in
2331          * preparation for taking it in write mode.
2332          */
2333         up_read(&mm->mmap_sem);
2334         if (unlikely(!*hpage)) {
2335                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2336                 *hpage = ERR_PTR(-ENOMEM);
2337                 return NULL;
2338         }
2339
2340         count_vm_event(THP_COLLAPSE_ALLOC);
2341         return *hpage;
2342 }
2343 #else
2344 static int khugepaged_find_target_node(void)
2345 {
2346         return 0;
2347 }
2348
2349 static inline struct page *alloc_hugepage(int defrag)
2350 {
2351         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2352                            HPAGE_PMD_ORDER);
2353 }
2354
2355 static struct page *khugepaged_alloc_hugepage(bool *wait)
2356 {
2357         struct page *hpage;
2358
2359         do {
2360                 hpage = alloc_hugepage(khugepaged_defrag());
2361                 if (!hpage) {
2362                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2363                         if (!*wait)
2364                                 return NULL;
2365
2366                         *wait = false;
2367                         khugepaged_alloc_sleep();
2368                 } else
2369                         count_vm_event(THP_COLLAPSE_ALLOC);
2370         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2371
2372         return hpage;
2373 }
2374
2375 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2376 {
2377         if (!*hpage)
2378                 *hpage = khugepaged_alloc_hugepage(wait);
2379
2380         if (unlikely(!*hpage))
2381                 return false;
2382
2383         return true;
2384 }
2385
2386 static struct page
2387 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2388                        struct vm_area_struct *vma, unsigned long address,
2389                        int node)
2390 {
2391         up_read(&mm->mmap_sem);
2392         VM_BUG_ON(!*hpage);
2393         return  *hpage;
2394 }
2395 #endif
2396
2397 static bool hugepage_vma_check(struct vm_area_struct *vma)
2398 {
2399         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2400             (vma->vm_flags & VM_NOHUGEPAGE))
2401                 return false;
2402
2403         if (!vma->anon_vma || vma->vm_ops)
2404                 return false;
2405         if (is_vma_temporary_stack(vma))
2406                 return false;
2407         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2408         return true;
2409 }
2410
2411 static void collapse_huge_page(struct mm_struct *mm,
2412                                    unsigned long address,
2413                                    struct page **hpage,
2414                                    struct vm_area_struct *vma,
2415                                    int node)
2416 {
2417         pmd_t *pmd, _pmd;
2418         pte_t *pte;
2419         pgtable_t pgtable;
2420         struct page *new_page;
2421         spinlock_t *pmd_ptl, *pte_ptl;
2422         int isolated;
2423         unsigned long hstart, hend;
2424         unsigned long mmun_start;       /* For mmu_notifiers */
2425         unsigned long mmun_end;         /* For mmu_notifiers */
2426
2427         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2428
2429         /* release the mmap_sem read lock. */
2430         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2431         if (!new_page)
2432                 return;
2433
2434         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2435                 return;
2436
2437         /*
2438          * Prevent all access to pagetables with the exception of
2439          * gup_fast later hanlded by the ptep_clear_flush and the VM
2440          * handled by the anon_vma lock + PG_lock.
2441          */
2442         down_write(&mm->mmap_sem);
2443         if (unlikely(khugepaged_test_exit(mm)))
2444                 goto out;
2445
2446         vma = find_vma(mm, address);
2447         if (!vma)
2448                 goto out;
2449         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2450         hend = vma->vm_end & HPAGE_PMD_MASK;
2451         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2452                 goto out;
2453         if (!hugepage_vma_check(vma))
2454                 goto out;
2455         pmd = mm_find_pmd(mm, address);
2456         if (!pmd)
2457                 goto out;
2458         if (pmd_trans_huge(*pmd))
2459                 goto out;
2460
2461         anon_vma_lock_write(vma->anon_vma);
2462
2463         pte = pte_offset_map(pmd, address);
2464         pte_ptl = pte_lockptr(mm, pmd);
2465
2466         mmun_start = address;
2467         mmun_end   = address + HPAGE_PMD_SIZE;
2468         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2469         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2470         /*
2471          * After this gup_fast can't run anymore. This also removes
2472          * any huge TLB entry from the CPU so we won't allow
2473          * huge and small TLB entries for the same virtual address
2474          * to avoid the risk of CPU bugs in that area.
2475          */
2476         _pmd = pmdp_clear_flush(vma, address, pmd);
2477         spin_unlock(pmd_ptl);
2478         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2479
2480         spin_lock(pte_ptl);
2481         isolated = __collapse_huge_page_isolate(vma, address, pte);
2482         spin_unlock(pte_ptl);
2483
2484         if (unlikely(!isolated)) {
2485                 pte_unmap(pte);
2486                 spin_lock(pmd_ptl);
2487                 BUG_ON(!pmd_none(*pmd));
2488                 /*
2489                  * We can only use set_pmd_at when establishing
2490                  * hugepmds and never for establishing regular pmds that
2491                  * points to regular pagetables. Use pmd_populate for that
2492                  */
2493                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2494                 spin_unlock(pmd_ptl);
2495                 anon_vma_unlock_write(vma->anon_vma);
2496                 goto out;
2497         }
2498
2499         /*
2500          * All pages are isolated and locked so anon_vma rmap
2501          * can't run anymore.
2502          */
2503         anon_vma_unlock_write(vma->anon_vma);
2504
2505         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2506         pte_unmap(pte);
2507         __SetPageUptodate(new_page);
2508         pgtable = pmd_pgtable(_pmd);
2509
2510         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2511         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2512
2513         /*
2514          * spin_lock() below is not the equivalent of smp_wmb(), so
2515          * this is needed to avoid the copy_huge_page writes to become
2516          * visible after the set_pmd_at() write.
2517          */
2518         smp_wmb();
2519
2520         spin_lock(pmd_ptl);
2521         BUG_ON(!pmd_none(*pmd));
2522         page_add_new_anon_rmap(new_page, vma, address);
2523         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2524         set_pmd_at(mm, address, pmd, _pmd);
2525         update_mmu_cache_pmd(vma, address, pmd);
2526         spin_unlock(pmd_ptl);
2527
2528         *hpage = NULL;
2529
2530         khugepaged_pages_collapsed++;
2531 out_up_write:
2532         up_write(&mm->mmap_sem);
2533         return;
2534
2535 out:
2536         mem_cgroup_uncharge_page(new_page);
2537         goto out_up_write;
2538 }
2539
2540 static int khugepaged_scan_pmd(struct mm_struct *mm,
2541                                struct vm_area_struct *vma,
2542                                unsigned long address,
2543                                struct page **hpage)
2544 {
2545         pmd_t *pmd;
2546         pte_t *pte, *_pte;
2547         int ret = 0, referenced = 0, none = 0;
2548         struct page *page;
2549         unsigned long _address;
2550         spinlock_t *ptl;
2551         int node = NUMA_NO_NODE;
2552
2553         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2554
2555         pmd = mm_find_pmd(mm, address);
2556         if (!pmd)
2557                 goto out;
2558         if (pmd_trans_huge(*pmd))
2559                 goto out;
2560
2561         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2562         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2563         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2564              _pte++, _address += PAGE_SIZE) {
2565                 pte_t pteval = *_pte;
2566                 if (pte_none(pteval)) {
2567                         if (++none <= khugepaged_max_ptes_none)
2568                                 continue;
2569                         else
2570                                 goto out_unmap;
2571                 }
2572                 if (!pte_present(pteval) || !pte_write(pteval))
2573                         goto out_unmap;
2574                 page = vm_normal_page(vma, _address, pteval);
2575                 if (unlikely(!page))
2576                         goto out_unmap;
2577                 /*
2578                  * Record which node the original page is from and save this
2579                  * information to khugepaged_node_load[].
2580                  * Khupaged will allocate hugepage from the node has the max
2581                  * hit record.
2582                  */
2583                 node = page_to_nid(page);
2584                 khugepaged_node_load[node]++;
2585                 VM_BUG_ON_PAGE(PageCompound(page), page);
2586                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2587                         goto out_unmap;
2588                 /* cannot use mapcount: can't collapse if there's a gup pin */
2589                 if (page_count(page) != 1)
2590                         goto out_unmap;
2591                 if (pte_young(pteval) || PageReferenced(page) ||
2592                     mmu_notifier_test_young(vma->vm_mm, address))
2593                         referenced = 1;
2594         }
2595         if (referenced)
2596                 ret = 1;
2597 out_unmap:
2598         pte_unmap_unlock(pte, ptl);
2599         if (ret) {
2600                 node = khugepaged_find_target_node();
2601                 /* collapse_huge_page will return with the mmap_sem released */
2602                 collapse_huge_page(mm, address, hpage, vma, node);
2603         }
2604 out:
2605         return ret;
2606 }
2607
2608 static void collect_mm_slot(struct mm_slot *mm_slot)
2609 {
2610         struct mm_struct *mm = mm_slot->mm;
2611
2612         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2613
2614         if (khugepaged_test_exit(mm)) {
2615                 /* free mm_slot */
2616                 hash_del(&mm_slot->hash);
2617                 list_del(&mm_slot->mm_node);
2618
2619                 /*
2620                  * Not strictly needed because the mm exited already.
2621                  *
2622                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2623                  */
2624
2625                 /* khugepaged_mm_lock actually not necessary for the below */
2626                 free_mm_slot(mm_slot);
2627                 mmdrop(mm);
2628         }
2629 }
2630
2631 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2632                                             struct page **hpage)
2633         __releases(&khugepaged_mm_lock)
2634         __acquires(&khugepaged_mm_lock)
2635 {
2636         struct mm_slot *mm_slot;
2637         struct mm_struct *mm;
2638         struct vm_area_struct *vma;
2639         int progress = 0;
2640
2641         VM_BUG_ON(!pages);
2642         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2643
2644         if (khugepaged_scan.mm_slot)
2645                 mm_slot = khugepaged_scan.mm_slot;
2646         else {
2647                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2648                                      struct mm_slot, mm_node);
2649                 khugepaged_scan.address = 0;
2650                 khugepaged_scan.mm_slot = mm_slot;
2651         }
2652         spin_unlock(&khugepaged_mm_lock);
2653
2654         mm = mm_slot->mm;
2655         down_read(&mm->mmap_sem);
2656         if (unlikely(khugepaged_test_exit(mm)))
2657                 vma = NULL;
2658         else
2659                 vma = find_vma(mm, khugepaged_scan.address);
2660
2661         progress++;
2662         for (; vma; vma = vma->vm_next) {
2663                 unsigned long hstart, hend;
2664
2665                 cond_resched();
2666                 if (unlikely(khugepaged_test_exit(mm))) {
2667                         progress++;
2668                         break;
2669                 }
2670                 if (!hugepage_vma_check(vma)) {
2671 skip:
2672                         progress++;
2673                         continue;
2674                 }
2675                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2676                 hend = vma->vm_end & HPAGE_PMD_MASK;
2677                 if (hstart >= hend)
2678                         goto skip;
2679                 if (khugepaged_scan.address > hend)
2680                         goto skip;
2681                 if (khugepaged_scan.address < hstart)
2682                         khugepaged_scan.address = hstart;
2683                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2684
2685                 while (khugepaged_scan.address < hend) {
2686                         int ret;
2687                         cond_resched();
2688                         if (unlikely(khugepaged_test_exit(mm)))
2689                                 goto breakouterloop;
2690
2691                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2692                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2693                                   hend);
2694                         ret = khugepaged_scan_pmd(mm, vma,
2695                                                   khugepaged_scan.address,
2696                                                   hpage);
2697                         /* move to next address */
2698                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2699                         progress += HPAGE_PMD_NR;
2700                         if (ret)
2701                                 /* we released mmap_sem so break loop */
2702                                 goto breakouterloop_mmap_sem;
2703                         if (progress >= pages)
2704                                 goto breakouterloop;
2705                 }
2706         }
2707 breakouterloop:
2708         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2709 breakouterloop_mmap_sem:
2710
2711         spin_lock(&khugepaged_mm_lock);
2712         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2713         /*
2714          * Release the current mm_slot if this mm is about to die, or
2715          * if we scanned all vmas of this mm.
2716          */
2717         if (khugepaged_test_exit(mm) || !vma) {
2718                 /*
2719                  * Make sure that if mm_users is reaching zero while
2720                  * khugepaged runs here, khugepaged_exit will find
2721                  * mm_slot not pointing to the exiting mm.
2722                  */
2723                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2724                         khugepaged_scan.mm_slot = list_entry(
2725                                 mm_slot->mm_node.next,
2726                                 struct mm_slot, mm_node);
2727                         khugepaged_scan.address = 0;
2728                 } else {
2729                         khugepaged_scan.mm_slot = NULL;
2730                         khugepaged_full_scans++;
2731                 }
2732
2733                 collect_mm_slot(mm_slot);
2734         }
2735
2736         return progress;
2737 }
2738
2739 static int khugepaged_has_work(void)
2740 {
2741         return !list_empty(&khugepaged_scan.mm_head) &&
2742                 khugepaged_enabled();
2743 }
2744
2745 static int khugepaged_wait_event(void)
2746 {
2747         return !list_empty(&khugepaged_scan.mm_head) ||
2748                 kthread_should_stop();
2749 }
2750
2751 static void khugepaged_do_scan(void)
2752 {
2753         struct page *hpage = NULL;
2754         unsigned int progress = 0, pass_through_head = 0;
2755         unsigned int pages = khugepaged_pages_to_scan;
2756         bool wait = true;
2757
2758         barrier(); /* write khugepaged_pages_to_scan to local stack */
2759
2760         while (progress < pages) {
2761                 if (!khugepaged_prealloc_page(&hpage, &wait))
2762                         break;
2763
2764                 cond_resched();
2765
2766                 if (unlikely(kthread_should_stop() || freezing(current)))
2767                         break;
2768
2769                 spin_lock(&khugepaged_mm_lock);
2770                 if (!khugepaged_scan.mm_slot)
2771                         pass_through_head++;
2772                 if (khugepaged_has_work() &&
2773                     pass_through_head < 2)
2774                         progress += khugepaged_scan_mm_slot(pages - progress,
2775                                                             &hpage);
2776                 else
2777                         progress = pages;
2778                 spin_unlock(&khugepaged_mm_lock);
2779         }
2780
2781         if (!IS_ERR_OR_NULL(hpage))
2782                 put_page(hpage);
2783 }
2784
2785 static void khugepaged_wait_work(void)
2786 {
2787         try_to_freeze();
2788
2789         if (khugepaged_has_work()) {
2790                 if (!khugepaged_scan_sleep_millisecs)
2791                         return;
2792
2793                 wait_event_freezable_timeout(khugepaged_wait,
2794                                              kthread_should_stop(),
2795                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2796                 return;
2797         }
2798
2799         if (khugepaged_enabled())
2800                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2801 }
2802
2803 static int khugepaged(void *none)
2804 {
2805         struct mm_slot *mm_slot;
2806
2807         set_freezable();
2808         set_user_nice(current, 19);
2809
2810         while (!kthread_should_stop()) {
2811                 khugepaged_do_scan();
2812                 khugepaged_wait_work();
2813         }
2814
2815         spin_lock(&khugepaged_mm_lock);
2816         mm_slot = khugepaged_scan.mm_slot;
2817         khugepaged_scan.mm_slot = NULL;
2818         if (mm_slot)
2819                 collect_mm_slot(mm_slot);
2820         spin_unlock(&khugepaged_mm_lock);
2821         return 0;
2822 }
2823
2824 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2825                 unsigned long haddr, pmd_t *pmd)
2826 {
2827         struct mm_struct *mm = vma->vm_mm;
2828         pgtable_t pgtable;
2829         pmd_t _pmd;
2830         int i;
2831
2832         pmdp_clear_flush(vma, haddr, pmd);
2833         /* leave pmd empty until pte is filled */
2834
2835         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2836         pmd_populate(mm, &_pmd, pgtable);
2837
2838         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2839                 pte_t *pte, entry;
2840                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2841                 entry = pte_mkspecial(entry);
2842                 pte = pte_offset_map(&_pmd, haddr);
2843                 VM_BUG_ON(!pte_none(*pte));
2844                 set_pte_at(mm, haddr, pte, entry);
2845                 pte_unmap(pte);
2846         }
2847         smp_wmb(); /* make pte visible before pmd */
2848         pmd_populate(mm, pmd, pgtable);
2849         put_huge_zero_page();
2850 }
2851
2852 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2853                 pmd_t *pmd)
2854 {
2855         spinlock_t *ptl;
2856         struct page *page;
2857         struct mm_struct *mm = vma->vm_mm;
2858         unsigned long haddr = address & HPAGE_PMD_MASK;
2859         unsigned long mmun_start;       /* For mmu_notifiers */
2860         unsigned long mmun_end;         /* For mmu_notifiers */
2861
2862         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2863
2864         mmun_start = haddr;
2865         mmun_end   = haddr + HPAGE_PMD_SIZE;
2866 again:
2867         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2868         ptl = pmd_lock(mm, pmd);
2869         if (unlikely(!pmd_trans_huge(*pmd))) {
2870                 spin_unlock(ptl);
2871                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2872                 return;
2873         }
2874         if (is_huge_zero_pmd(*pmd)) {
2875                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2876                 spin_unlock(ptl);
2877                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2878                 return;
2879         }
2880         page = pmd_page(*pmd);
2881         VM_BUG_ON_PAGE(!page_count(page), page);
2882         get_page(page);
2883         spin_unlock(ptl);
2884         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2885
2886         split_huge_page(page);
2887
2888         put_page(page);
2889
2890         /*
2891          * We don't always have down_write of mmap_sem here: a racing
2892          * do_huge_pmd_wp_page() might have copied-on-write to another
2893          * huge page before our split_huge_page() got the anon_vma lock.
2894          */
2895         if (unlikely(pmd_trans_huge(*pmd)))
2896                 goto again;
2897 }
2898
2899 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2900                 pmd_t *pmd)
2901 {
2902         struct vm_area_struct *vma;
2903
2904         vma = find_vma(mm, address);
2905         BUG_ON(vma == NULL);
2906         split_huge_page_pmd(vma, address, pmd);
2907 }
2908
2909 static void split_huge_page_address(struct mm_struct *mm,
2910                                     unsigned long address)
2911 {
2912         pmd_t *pmd;
2913
2914         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2915
2916         pmd = mm_find_pmd(mm, address);
2917         if (!pmd)
2918                 return;
2919         /*
2920          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2921          * materialize from under us.
2922          */
2923         split_huge_page_pmd_mm(mm, address, pmd);
2924 }
2925
2926 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2927                              unsigned long start,
2928                              unsigned long end,
2929                              long adjust_next)
2930 {
2931         /*
2932          * If the new start address isn't hpage aligned and it could
2933          * previously contain an hugepage: check if we need to split
2934          * an huge pmd.
2935          */
2936         if (start & ~HPAGE_PMD_MASK &&
2937             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2938             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2939                 split_huge_page_address(vma->vm_mm, start);
2940
2941         /*
2942          * If the new end address isn't hpage aligned and it could
2943          * previously contain an hugepage: check if we need to split
2944          * an huge pmd.
2945          */
2946         if (end & ~HPAGE_PMD_MASK &&
2947             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2948             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2949                 split_huge_page_address(vma->vm_mm, end);
2950
2951         /*
2952          * If we're also updating the vma->vm_next->vm_start, if the new
2953          * vm_next->vm_start isn't page aligned and it could previously
2954          * contain an hugepage: check if we need to split an huge pmd.
2955          */
2956         if (adjust_next > 0) {
2957                 struct vm_area_struct *next = vma->vm_next;
2958                 unsigned long nstart = next->vm_start;
2959                 nstart += adjust_next << PAGE_SHIFT;
2960                 if (nstart & ~HPAGE_PMD_MASK &&
2961                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2962                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2963                         split_huge_page_address(next->vm_mm, nstart);
2964         }
2965 }