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