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