mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is enabled for all mappings
  31  * and khugepaged scans all mappings. Defrag is only invoked by
  32  * khugepaged hugepage allocations and by page faults inside
  33  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  34  * allocations.
  35  */
  36 unsigned long transparent_hugepage_flags __read_mostly =
  37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  38         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  39 #endif
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  41         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  42 #endif
  43         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  46
  47 /* default scan 8*512 pte (or vmas) every 30 second */
  48 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  49 static unsigned int khugepaged_pages_collapsed;
  50 static unsigned int khugepaged_full_scans;
  51 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  52 /* during fragmentation poll the hugepage allocator once every minute */
  53 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  54 static struct task_struct *khugepaged_thread __read_mostly;
  55 static DEFINE_MUTEX(khugepaged_mutex);
  56 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  57 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  58 /*
  59  * default collapse hugepages if there is at least one pte mapped like
  60  * it would have happened if the vma was large enough during page
  61  * fault.
  62  */
  63 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  64
  65 static int khugepaged(void *none);
  66 static int khugepaged_slab_init(void);
  67
  68 #define MM_SLOTS_HASH_BITS 10
  69 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  70
  71 static struct kmem_cache *mm_slot_cache __read_mostly;
  72
  73 /**
  74  * struct mm_slot - hash lookup from mm to mm_slot
  75  * @hash: hash collision list
  76  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  77  * @mm: the mm that this information is valid for
  78  */
  79 struct mm_slot {
  80         struct hlist_node hash;
  81         struct list_head mm_node;
  82         struct mm_struct *mm;
  83 };
  84
  85 /**
  86  * struct khugepaged_scan - cursor for scanning
  87  * @mm_head: the head of the mm list to scan
  88  * @mm_slot: the current mm_slot we are scanning
  89  * @address: the next address inside that to be scanned
  90  *
  91  * There is only the one khugepaged_scan instance of this cursor structure.
  92  */
  93 struct khugepaged_scan {
  94         struct list_head mm_head;
  95         struct mm_slot *mm_slot;
  96         unsigned long address;
  97 };
  98 static struct khugepaged_scan khugepaged_scan = {
  99         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 100 };
 101
 102
 103 static int set_recommended_min_free_kbytes(void)
 104 {
 105         struct zone *zone;
 106         int nr_zones = 0;
 107         unsigned long recommended_min;
 108
 109         if (!khugepaged_enabled())
 110                 return 0;
 111
 112         for_each_populated_zone(zone)
 113                 nr_zones++;
 114
 115         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 116         recommended_min = pageblock_nr_pages * nr_zones * 2;
 117
 118         /*
 119          * Make sure that on average at least two pageblocks are almost free
 120          * of another type, one for a migratetype to fall back to and a
 121          * second to avoid subsequent fallbacks of other types There are 3
 122          * MIGRATE_TYPES we care about.
 123          */
 124         recommended_min += pageblock_nr_pages * nr_zones *
 125                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 126
 127         /* don't ever allow to reserve more than 5% of the lowmem */
 128         recommended_min = min(recommended_min,
 129                               (unsigned long) nr_free_buffer_pages() / 20);
 130         recommended_min <<= (PAGE_SHIFT-10);
 131
 132         if (recommended_min > min_free_kbytes)
 133                 min_free_kbytes = recommended_min;
 134         setup_per_zone_wmarks();
 135         return 0;
 136 }
 137 late_initcall(set_recommended_min_free_kbytes);
 138
 139 static int start_khugepaged(void)
 140 {
 141         int err = 0;
 142         if (khugepaged_enabled()) {
 143                 if (!khugepaged_thread)
 144                         khugepaged_thread = kthread_run(khugepaged, NULL,
 145                                                         "khugepaged");
 146                 if (unlikely(IS_ERR(khugepaged_thread))) {
 147                         printk(KERN_ERR
 148                                "khugepaged: kthread_run(khugepaged) failed\n");
 149                         err = PTR_ERR(khugepaged_thread);
 150                         khugepaged_thread = NULL;
 151                 }
 152
 153                 if (!list_empty(&khugepaged_scan.mm_head))
 154                         wake_up_interruptible(&khugepaged_wait);
 155
 156                 set_recommended_min_free_kbytes();
 157         } else if (khugepaged_thread) {
 158                 kthread_stop(khugepaged_thread);
 159                 khugepaged_thread = NULL;
 160         }
 161
 162         return err;
 163 }
 164
 165 static atomic_t huge_zero_refcount;
 166 static struct page *huge_zero_page __read_mostly;
 167
 168 static inline bool is_huge_zero_page(struct page *page)
 169 {
 170         return ACCESS_ONCE(huge_zero_page) == page;
 171 }
 172
 173 static inline bool is_huge_zero_pmd(pmd_t pmd)
 174 {
 175         return is_huge_zero_page(pmd_page(pmd));
 176 }
 177
 178 static struct page *get_huge_zero_page(void)
 179 {
 180         struct page *zero_page;
 181 retry:
 182         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 183                 return ACCESS_ONCE(huge_zero_page);
 184
 185         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 186                         HPAGE_PMD_ORDER);
 187         if (!zero_page) {
 188                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 189                 return NULL;
 190         }
 191         count_vm_event(THP_ZERO_PAGE_ALLOC);
 192         preempt_disable();
 193         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 194                 preempt_enable();
 195                 __free_page(zero_page);
 196                 goto retry;
 197         }
 198
 199         /* We take additional reference here. It will be put back by shrinker */
 200         atomic_set(&huge_zero_refcount, 2);
 201         preempt_enable();
 202         return ACCESS_ONCE(huge_zero_page);
 203 }
 204
 205 static void put_huge_zero_page(void)
 206 {
 207         /*
 208          * Counter should never go to zero here. Only shrinker can put
 209          * last reference.
 210          */
 211         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 212 }
 213
 214 static int shrink_huge_zero_page(struct shrinker *shrink,
 215                 struct shrink_control *sc)
 216 {
 217         if (!sc->nr_to_scan)
 218                 /* we can free zero page only if last reference remains */
 219                 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 220
 221         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 222                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 223                 BUG_ON(zero_page == NULL);
 224                 __free_page(zero_page);
 225         }
 226
 227         return 0;
 228 }
 229
 230 static struct shrinker huge_zero_page_shrinker = {
 231         .shrink = shrink_huge_zero_page,
 232         .seeks = DEFAULT_SEEKS,
 233 };
 234
 235 #ifdef CONFIG_SYSFS
 236
 237 static ssize_t double_flag_show(struct kobject *kobj,
 238                                 struct kobj_attribute *attr, char *buf,
 239                                 enum transparent_hugepage_flag enabled,
 240                                 enum transparent_hugepage_flag req_madv)
 241 {
 242         if (test_bit(enabled, &transparent_hugepage_flags)) {
 243                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 244                 return sprintf(buf, "[always] madvise never\n");
 245         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 246                 return sprintf(buf, "always [madvise] never\n");
 247         else
 248                 return sprintf(buf, "always madvise [never]\n");
 249 }
 250 static ssize_t double_flag_store(struct kobject *kobj,
 251                                  struct kobj_attribute *attr,
 252                                  const char *buf, size_t count,
 253                                  enum transparent_hugepage_flag enabled,
 254                                  enum transparent_hugepage_flag req_madv)
 255 {
 256         if (!memcmp("always", buf,
 257                     min(sizeof("always")-1, count))) {
 258                 set_bit(enabled, &transparent_hugepage_flags);
 259                 clear_bit(req_madv, &transparent_hugepage_flags);
 260         } else if (!memcmp("madvise", buf,
 261                            min(sizeof("madvise")-1, count))) {
 262                 clear_bit(enabled, &transparent_hugepage_flags);
 263                 set_bit(req_madv, &transparent_hugepage_flags);
 264         } else if (!memcmp("never", buf,
 265                            min(sizeof("never")-1, count))) {
 266                 clear_bit(enabled, &transparent_hugepage_flags);
 267                 clear_bit(req_madv, &transparent_hugepage_flags);
 268         } else
 269                 return -EINVAL;
 270
 271         return count;
 272 }
 273
 274 static ssize_t enabled_show(struct kobject *kobj,
 275                             struct kobj_attribute *attr, char *buf)
 276 {
 277         return double_flag_show(kobj, attr, buf,
 278                                 TRANSPARENT_HUGEPAGE_FLAG,
 279                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 280 }
 281 static ssize_t enabled_store(struct kobject *kobj,
 282                              struct kobj_attribute *attr,
 283                              const char *buf, size_t count)
 284 {
 285         ssize_t ret;
 286
 287         ret = double_flag_store(kobj, attr, buf, count,
 288                                 TRANSPARENT_HUGEPAGE_FLAG,
 289                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 290
 291         if (ret > 0) {
 292                 int err;
 293
 294                 mutex_lock(&khugepaged_mutex);
 295                 err = start_khugepaged();
 296                 mutex_unlock(&khugepaged_mutex);
 297
 298                 if (err)
 299                         ret = err;
 300         }
 301
 302         return ret;
 303 }
 304 static struct kobj_attribute enabled_attr =
 305         __ATTR(enabled, 0644, enabled_show, enabled_store);
 306
 307 static ssize_t single_flag_show(struct kobject *kobj,
 308                                 struct kobj_attribute *attr, char *buf,
 309                                 enum transparent_hugepage_flag flag)
 310 {
 311         return sprintf(buf, "%d\n",
 312                        !!test_bit(flag, &transparent_hugepage_flags));
 313 }
 314
 315 static ssize_t single_flag_store(struct kobject *kobj,
 316                                  struct kobj_attribute *attr,
 317                                  const char *buf, size_t count,
 318                                  enum transparent_hugepage_flag flag)
 319 {
 320         unsigned long value;
 321         int ret;
 322
 323         ret = kstrtoul(buf, 10, &value);
 324         if (ret < 0)
 325                 return ret;
 326         if (value > 1)
 327                 return -EINVAL;
 328
 329         if (value)
 330                 set_bit(flag, &transparent_hugepage_flags);
 331         else
 332                 clear_bit(flag, &transparent_hugepage_flags);
 333
 334         return count;
 335 }
 336
 337 /*
 338  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 339  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 340  * memory just to allocate one more hugepage.
 341  */
 342 static ssize_t defrag_show(struct kobject *kobj,
 343                            struct kobj_attribute *attr, char *buf)
 344 {
 345         return double_flag_show(kobj, attr, buf,
 346                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 347                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 348 }
 349 static ssize_t defrag_store(struct kobject *kobj,
 350                             struct kobj_attribute *attr,
 351                             const char *buf, size_t count)
 352 {
 353         return double_flag_store(kobj, attr, buf, count,
 354                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 355                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 356 }
 357 static struct kobj_attribute defrag_attr =
 358         __ATTR(defrag, 0644, defrag_show, defrag_store);
 359
 360 static ssize_t use_zero_page_show(struct kobject *kobj,
 361                 struct kobj_attribute *attr, char *buf)
 362 {
 363         return single_flag_show(kobj, attr, buf,
 364                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 365 }
 366 static ssize_t use_zero_page_store(struct kobject *kobj,
 367                 struct kobj_attribute *attr, const char *buf, size_t count)
 368 {
 369         return single_flag_store(kobj, attr, buf, count,
 370                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 371 }
 372 static struct kobj_attribute use_zero_page_attr =
 373         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 374 #ifdef CONFIG_DEBUG_VM
 375 static ssize_t debug_cow_show(struct kobject *kobj,
 376                                 struct kobj_attribute *attr, char *buf)
 377 {
 378         return single_flag_show(kobj, attr, buf,
 379                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 380 }
 381 static ssize_t debug_cow_store(struct kobject *kobj,
 382                                struct kobj_attribute *attr,
 383                                const char *buf, size_t count)
 384 {
 385         return single_flag_store(kobj, attr, buf, count,
 386                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 387 }
 388 static struct kobj_attribute debug_cow_attr =
 389         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 390 #endif /* CONFIG_DEBUG_VM */
 391
 392 static struct attribute *hugepage_attr[] = {
 393         &enabled_attr.attr,
 394         &defrag_attr.attr,
 395         &use_zero_page_attr.attr,
 396 #ifdef CONFIG_DEBUG_VM
 397         &debug_cow_attr.attr,
 398 #endif
 399         NULL,
 400 };
 401
 402 static struct attribute_group hugepage_attr_group = {
 403         .attrs = hugepage_attr,
 404 };
 405
 406 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 407                                          struct kobj_attribute *attr,
 408                                          char *buf)
 409 {
 410         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 411 }
 412
 413 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 414                                           struct kobj_attribute *attr,
 415                                           const char *buf, size_t count)
 416 {
 417         unsigned long msecs;
 418         int err;
 419
 420         err = kstrtoul(buf, 10, &msecs);
 421         if (err || msecs > UINT_MAX)
 422                 return -EINVAL;
 423
 424         khugepaged_scan_sleep_millisecs = msecs;
 425         wake_up_interruptible(&khugepaged_wait);
 426
 427         return count;
 428 }
 429 static struct kobj_attribute scan_sleep_millisecs_attr =
 430         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 431                scan_sleep_millisecs_store);
 432
 433 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 434                                           struct kobj_attribute *attr,
 435                                           char *buf)
 436 {
 437         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 438 }
 439
 440 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 441                                            struct kobj_attribute *attr,
 442                                            const char *buf, size_t count)
 443 {
 444         unsigned long msecs;
 445         int err;
 446
 447         err = kstrtoul(buf, 10, &msecs);
 448         if (err || msecs > UINT_MAX)
 449                 return -EINVAL;
 450
 451         khugepaged_alloc_sleep_millisecs = msecs;
 452         wake_up_interruptible(&khugepaged_wait);
 453
 454         return count;
 455 }
 456 static struct kobj_attribute alloc_sleep_millisecs_attr =
 457         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 458                alloc_sleep_millisecs_store);
 459
 460 static ssize_t pages_to_scan_show(struct kobject *kobj,
 461                                   struct kobj_attribute *attr,
 462                                   char *buf)
 463 {
 464         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 465 }
 466 static ssize_t pages_to_scan_store(struct kobject *kobj,
 467                                    struct kobj_attribute *attr,
 468                                    const char *buf, size_t count)
 469 {
 470         int err;
 471         unsigned long pages;
 472
 473         err = kstrtoul(buf, 10, &pages);
 474         if (err || !pages || pages > UINT_MAX)
 475                 return -EINVAL;
 476
 477         khugepaged_pages_to_scan = pages;
 478
 479         return count;
 480 }
 481 static struct kobj_attribute pages_to_scan_attr =
 482         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 483                pages_to_scan_store);
 484
 485 static ssize_t pages_collapsed_show(struct kobject *kobj,
 486                                     struct kobj_attribute *attr,
 487                                     char *buf)
 488 {
 489         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 490 }
 491 static struct kobj_attribute pages_collapsed_attr =
 492         __ATTR_RO(pages_collapsed);
 493
 494 static ssize_t full_scans_show(struct kobject *kobj,
 495                                struct kobj_attribute *attr,
 496                                char *buf)
 497 {
 498         return sprintf(buf, "%u\n", khugepaged_full_scans);
 499 }
 500 static struct kobj_attribute full_scans_attr =
 501         __ATTR_RO(full_scans);
 502
 503 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 504                                       struct kobj_attribute *attr, char *buf)
 505 {
 506         return single_flag_show(kobj, attr, buf,
 507                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 508 }
 509 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 510                                        struct kobj_attribute *attr,
 511                                        const char *buf, size_t count)
 512 {
 513         return single_flag_store(kobj, attr, buf, count,
 514                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 515 }
 516 static struct kobj_attribute khugepaged_defrag_attr =
 517         __ATTR(defrag, 0644, khugepaged_defrag_show,
 518                khugepaged_defrag_store);
 519
 520 /*
 521  * max_ptes_none controls if khugepaged should collapse hugepages over
 522  * any unmapped ptes in turn potentially increasing the memory
 523  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 524  * reduce the available free memory in the system as it
 525  * runs. Increasing max_ptes_none will instead potentially reduce the
 526  * free memory in the system during the khugepaged scan.
 527  */
 528 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 529                                              struct kobj_attribute *attr,
 530                                              char *buf)
 531 {
 532         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 533 }
 534 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 535                                               struct kobj_attribute *attr,
 536                                               const char *buf, size_t count)
 537 {
 538         int err;
 539         unsigned long max_ptes_none;
 540
 541         err = kstrtoul(buf, 10, &max_ptes_none);
 542         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 543                 return -EINVAL;
 544
 545         khugepaged_max_ptes_none = max_ptes_none;
 546
 547         return count;
 548 }
 549 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 550         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 551                khugepaged_max_ptes_none_store);
 552
 553 static struct attribute *khugepaged_attr[] = {
 554         &khugepaged_defrag_attr.attr,
 555         &khugepaged_max_ptes_none_attr.attr,
 556         &pages_to_scan_attr.attr,
 557         &pages_collapsed_attr.attr,
 558         &full_scans_attr.attr,
 559         &scan_sleep_millisecs_attr.attr,
 560         &alloc_sleep_millisecs_attr.attr,
 561         NULL,
 562 };
 563
 564 static struct attribute_group khugepaged_attr_group = {
 565         .attrs = khugepaged_attr,
 566         .name = "khugepaged",
 567 };
 568
 569 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 570 {
 571         int err;
 572
 573         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 574         if (unlikely(!*hugepage_kobj)) {
 575                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 576                 return -ENOMEM;
 577         }
 578
 579         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 580         if (err) {
 581                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 582                 goto delete_obj;
 583         }
 584
 585         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 586         if (err) {
 587                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 588                 goto remove_hp_group;
 589         }
 590
 591         return 0;
 592
 593 remove_hp_group:
 594         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 595 delete_obj:
 596         kobject_put(*hugepage_kobj);
 597         return err;
 598 }
 599
 600 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 601 {
 602         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 603         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 604         kobject_put(hugepage_kobj);
 605 }
 606 #else
 607 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 608 {
 609         return 0;
 610 }
 611
 612 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 613 {
 614 }
 615 #endif /* CONFIG_SYSFS */
 616
 617 static int __init hugepage_init(void)
 618 {
 619         int err;
 620         struct kobject *hugepage_kobj;
 621
 622         if (!has_transparent_hugepage()) {
 623                 transparent_hugepage_flags = 0;
 624                 return -EINVAL;
 625         }
 626
 627         err = hugepage_init_sysfs(&hugepage_kobj);
 628         if (err)
 629                 return err;
 630
 631         err = khugepaged_slab_init();
 632         if (err)
 633                 goto out;
 634
 635         register_shrinker(&huge_zero_page_shrinker);
 636
 637         /*
 638          * By default disable transparent hugepages on smaller systems,
 639          * where the extra memory used could hurt more than TLB overhead
 640          * is likely to save.  The admin can still enable it through /sys.
 641          */
 642         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 643                 transparent_hugepage_flags = 0;
 644
 645         start_khugepaged();
 646
 647         return 0;
 648 out:
 649         hugepage_exit_sysfs(hugepage_kobj);
 650         return err;
 651 }
 652 module_init(hugepage_init)
 653
 654 static int __init setup_transparent_hugepage(char *str)
 655 {
 656         int ret = 0;
 657         if (!str)
 658                 goto out;
 659         if (!strcmp(str, "always")) {
 660                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 661                         &transparent_hugepage_flags);
 662                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 663                           &transparent_hugepage_flags);
 664                 ret = 1;
 665         } else if (!strcmp(str, "madvise")) {
 666                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 667                           &transparent_hugepage_flags);
 668                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 669                         &transparent_hugepage_flags);
 670                 ret = 1;
 671         } else if (!strcmp(str, "never")) {
 672                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 673                           &transparent_hugepage_flags);
 674                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 675                           &transparent_hugepage_flags);
 676                 ret = 1;
 677         }
 678 out:
 679         if (!ret)
 680                 printk(KERN_WARNING
 681                        "transparent_hugepage= cannot parse, ignored\n");
 682         return ret;
 683 }
 684 __setup("transparent_hugepage=", setup_transparent_hugepage);
 685
 686 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 687 {
 688         if (likely(vma->vm_flags & VM_WRITE))
 689                 pmd = pmd_mkwrite(pmd);
 690         return pmd;
 691 }
 692
 693 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 694 {
 695         pmd_t entry;
 696         entry = mk_pmd(page, prot);
 697         entry = pmd_mkhuge(entry);
 698         return entry;
 699 }
 700
 701 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 702                                         struct vm_area_struct *vma,
 703                                         unsigned long haddr, pmd_t *pmd,
 704                                         struct page *page)
 705 {
 706         pgtable_t pgtable;
 707
 708         VM_BUG_ON(!PageCompound(page));
 709         pgtable = pte_alloc_one(mm, haddr);
 710         if (unlikely(!pgtable))
 711                 return VM_FAULT_OOM;
 712
 713         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 714         /*
 715          * The memory barrier inside __SetPageUptodate makes sure that
 716          * clear_huge_page writes become visible before the set_pmd_at()
 717          * write.
 718          */
 719         __SetPageUptodate(page);
 720
 721         spin_lock(&mm->page_table_lock);
 722         if (unlikely(!pmd_none(*pmd))) {
 723                 spin_unlock(&mm->page_table_lock);
 724                 mem_cgroup_uncharge_page(page);
 725                 put_page(page);
 726                 pte_free(mm, pgtable);
 727         } else {
 728                 pmd_t entry;
 729                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 730                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 731                 page_add_new_anon_rmap(page, vma, haddr);
 732                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 733                 set_pmd_at(mm, haddr, pmd, entry);
 734                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 735                 mm->nr_ptes++;
 736                 spin_unlock(&mm->page_table_lock);
 737         }
 738
 739         return 0;
 740 }
 741
 742 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 743 {
 744         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 745 }
 746
 747 static inline struct page *alloc_hugepage_vma(int defrag,
 748                                               struct vm_area_struct *vma,
 749                                               unsigned long haddr, int nd,
 750                                               gfp_t extra_gfp)
 751 {
 752         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 753                                HPAGE_PMD_ORDER, vma, haddr, nd);
 754 }
 755
 756 #ifndef CONFIG_NUMA
 757 static inline struct page *alloc_hugepage(int defrag)
 758 {
 759         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 760                            HPAGE_PMD_ORDER);
 761 }
 762 #endif
 763
 764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 765                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 766                 struct page *zero_page)
 767 {
 768         pmd_t entry;
 769         if (!pmd_none(*pmd))
 770                 return false;
 771         entry = mk_pmd(zero_page, vma->vm_page_prot);
 772         entry = pmd_wrprotect(entry);
 773         entry = pmd_mkhuge(entry);
 774         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 775         set_pmd_at(mm, haddr, pmd, entry);
 776         mm->nr_ptes++;
 777         return true;
 778 }
 779
 780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 781                                unsigned long address, pmd_t *pmd,
 782                                unsigned int flags)
 783 {
 784         struct page *page;
 785         unsigned long haddr = address & HPAGE_PMD_MASK;
 786         pte_t *pte;
 787
 788         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 789                 goto out;
 790         if (unlikely(anon_vma_prepare(vma)))
 791                 return VM_FAULT_OOM;
 792         if (unlikely(khugepaged_enter(vma)))
 793                 return VM_FAULT_OOM;
 794         if (!(flags & FAULT_FLAG_WRITE) &&
 795                         transparent_hugepage_use_zero_page()) {
 796                 pgtable_t pgtable;
 797                 struct page *zero_page;
 798                 bool set;
 799                 pgtable = pte_alloc_one(mm, haddr);
 800                 if (unlikely(!pgtable))
 801                         return VM_FAULT_OOM;
 802                 zero_page = get_huge_zero_page();
 803                 if (unlikely(!zero_page)) {
 804                         pte_free(mm, pgtable);
 805                         count_vm_event(THP_FAULT_FALLBACK);
 806                         goto out;
 807                 }
 808                 spin_lock(&mm->page_table_lock);
 809                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 810                                 zero_page);
 811                 spin_unlock(&mm->page_table_lock);
 812                 if (!set) {
 813                         pte_free(mm, pgtable);
 814                         put_huge_zero_page();
 815                 }
 816                 return 0;
 817         }
 818         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 819                         vma, haddr, numa_node_id(), 0);
 820         if (unlikely(!page)) {
 821                 count_vm_event(THP_FAULT_FALLBACK);
 822                 goto out;
 823         }
 824         count_vm_event(THP_FAULT_ALLOC);
 825         if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 826                 put_page(page);
 827                 goto out;
 828         }
 829         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 830                 mem_cgroup_uncharge_page(page);
 831                 put_page(page);
 832                 goto out;
 833         }
 834
 835         return 0;
 836 out:
 837         /*
 838          * Use __pte_alloc instead of pte_alloc_map, because we can't
 839          * run pte_offset_map on the pmd, if an huge pmd could
 840          * materialize from under us from a different thread.
 841          */
 842         if (unlikely(pmd_none(*pmd)) &&
 843             unlikely(__pte_alloc(mm, vma, pmd, address)))
 844                 return VM_FAULT_OOM;
 845         /* if an huge pmd materialized from under us just retry later */
 846         if (unlikely(pmd_trans_huge(*pmd)))
 847                 return 0;
 848         /*
 849          * A regular pmd is established and it can't morph into a huge pmd
 850          * from under us anymore at this point because we hold the mmap_sem
 851          * read mode and khugepaged takes it in write mode. So now it's
 852          * safe to run pte_offset_map().
 853          */
 854         pte = pte_offset_map(pmd, address);
 855         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 856 }
 857
 858 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 859                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 860                   struct vm_area_struct *vma)
 861 {
 862         struct page *src_page;
 863         pmd_t pmd;
 864         pgtable_t pgtable;
 865         int ret;
 866
 867         ret = -ENOMEM;
 868         pgtable = pte_alloc_one(dst_mm, addr);
 869         if (unlikely(!pgtable))
 870                 goto out;
 871
 872         spin_lock(&dst_mm->page_table_lock);
 873         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 874
 875         ret = -EAGAIN;
 876         pmd = *src_pmd;
 877         if (unlikely(!pmd_trans_huge(pmd))) {
 878                 pte_free(dst_mm, pgtable);
 879                 goto out_unlock;
 880         }
 881         /*
 882          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 883          * under splitting since we don't split the page itself, only pmd to
 884          * a page table.
 885          */
 886         if (is_huge_zero_pmd(pmd)) {
 887                 struct page *zero_page;
 888                 bool set;
 889                 /*
 890                  * get_huge_zero_page() will never allocate a new page here,
 891                  * since we already have a zero page to copy. It just takes a
 892                  * reference.
 893                  */
 894                 zero_page = get_huge_zero_page();
 895                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 896                                 zero_page);
 897                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 898                 ret = 0;
 899                 goto out_unlock;
 900         }
 901         if (unlikely(pmd_trans_splitting(pmd))) {
 902                 /* split huge page running from under us */
 903                 spin_unlock(&src_mm->page_table_lock);
 904                 spin_unlock(&dst_mm->page_table_lock);
 905                 pte_free(dst_mm, pgtable);
 906
 907                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 908                 goto out;
 909         }
 910         src_page = pmd_page(pmd);
 911         VM_BUG_ON(!PageHead(src_page));
 912         get_page(src_page);
 913         page_dup_rmap(src_page);
 914         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 915
 916         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 917         pmd = pmd_mkold(pmd_wrprotect(pmd));
 918         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 919         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 920         dst_mm->nr_ptes++;
 921
 922         ret = 0;
 923 out_unlock:
 924         spin_unlock(&src_mm->page_table_lock);
 925         spin_unlock(&dst_mm->page_table_lock);
 926 out:
 927         return ret;
 928 }
 929
 930 void huge_pmd_set_accessed(struct mm_struct *mm,
 931                            struct vm_area_struct *vma,
 932                            unsigned long address,
 933                            pmd_t *pmd, pmd_t orig_pmd,
 934                            int dirty)
 935 {
 936         pmd_t entry;
 937         unsigned long haddr;
 938
 939         spin_lock(&mm->page_table_lock);
 940         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 941                 goto unlock;
 942
 943         entry = pmd_mkyoung(orig_pmd);
 944         haddr = address & HPAGE_PMD_MASK;
 945         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 946                 update_mmu_cache_pmd(vma, address, pmd);
 947
 948 unlock:
 949         spin_unlock(&mm->page_table_lock);
 950 }
 951
 952 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 953                 struct vm_area_struct *vma, unsigned long address,
 954                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 955 {
 956         pgtable_t pgtable;
 957         pmd_t _pmd;
 958         struct page *page;
 959         int i, ret = 0;
 960         unsigned long mmun_start;       /* For mmu_notifiers */
 961         unsigned long mmun_end;         /* For mmu_notifiers */
 962
 963         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 964         if (!page) {
 965                 ret |= VM_FAULT_OOM;
 966                 goto out;
 967         }
 968
 969         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 970                 put_page(page);
 971                 ret |= VM_FAULT_OOM;
 972                 goto out;
 973         }
 974
 975         clear_user_highpage(page, address);
 976         __SetPageUptodate(page);
 977
 978         mmun_start = haddr;
 979         mmun_end   = haddr + HPAGE_PMD_SIZE;
 980         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 981
 982         spin_lock(&mm->page_table_lock);
 983         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 984                 goto out_free_page;
 985
 986         pmdp_clear_flush(vma, haddr, pmd);
 987         /* leave pmd empty until pte is filled */
 988
 989         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
 990         pmd_populate(mm, &_pmd, pgtable);
 991
 992         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 993                 pte_t *pte, entry;
 994                 if (haddr == (address & PAGE_MASK)) {
 995                         entry = mk_pte(page, vma->vm_page_prot);
 996                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 997                         page_add_new_anon_rmap(page, vma, haddr);
 998                 } else {
 999                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1000                         entry = pte_mkspecial(entry);
1001                 }
1002                 pte = pte_offset_map(&_pmd, haddr);
1003                 VM_BUG_ON(!pte_none(*pte));
1004                 set_pte_at(mm, haddr, pte, entry);
1005                 pte_unmap(pte);
1006         }
1007         smp_wmb(); /* make pte visible before pmd */
1008         pmd_populate(mm, pmd, pgtable);
1009         spin_unlock(&mm->page_table_lock);
1010         put_huge_zero_page();
1011         inc_mm_counter(mm, MM_ANONPAGES);
1012
1013         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1014
1015         ret |= VM_FAULT_WRITE;
1016 out:
1017         return ret;
1018 out_free_page:
1019         spin_unlock(&mm->page_table_lock);
1020         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1021         mem_cgroup_uncharge_page(page);
1022         put_page(page);
1023         goto out;
1024 }
1025
1026 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1027                                         struct vm_area_struct *vma,
1028                                         unsigned long address,
1029                                         pmd_t *pmd, pmd_t orig_pmd,
1030                                         struct page *page,
1031                                         unsigned long haddr)
1032 {
1033         pgtable_t pgtable;
1034         pmd_t _pmd;
1035         int ret = 0, i;
1036         struct page **pages;
1037         unsigned long mmun_start;       /* For mmu_notifiers */
1038         unsigned long mmun_end;         /* For mmu_notifiers */
1039
1040         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1041                         GFP_KERNEL);
1042         if (unlikely(!pages)) {
1043                 ret |= VM_FAULT_OOM;
1044                 goto out;
1045         }
1046
1047         for (i = 0; i < HPAGE_PMD_NR; i++) {
1048                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1049                                                __GFP_OTHER_NODE,
1050                                                vma, address, page_to_nid(page));
1051                 if (unlikely(!pages[i] ||
1052                              mem_cgroup_newpage_charge(pages[i], mm,
1053                                                        GFP_KERNEL))) {
1054                         if (pages[i])
1055                                 put_page(pages[i]);
1056                         mem_cgroup_uncharge_start();
1057                         while (--i >= 0) {
1058                                 mem_cgroup_uncharge_page(pages[i]);
1059                                 put_page(pages[i]);
1060                         }
1061                         mem_cgroup_uncharge_end();
1062                         kfree(pages);
1063                         ret |= VM_FAULT_OOM;
1064                         goto out;
1065                 }
1066         }
1067
1068         for (i = 0; i < HPAGE_PMD_NR; i++) {
1069                 copy_user_highpage(pages[i], page + i,
1070                                    haddr + PAGE_SIZE * i, vma);
1071                 __SetPageUptodate(pages[i]);
1072                 cond_resched();
1073         }
1074
1075         mmun_start = haddr;
1076         mmun_end   = haddr + HPAGE_PMD_SIZE;
1077         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1078
1079         spin_lock(&mm->page_table_lock);
1080         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1081                 goto out_free_pages;
1082         VM_BUG_ON(!PageHead(page));
1083
1084         pmdp_clear_flush(vma, haddr, pmd);
1085         /* leave pmd empty until pte is filled */
1086
1087         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1088         pmd_populate(mm, &_pmd, pgtable);
1089
1090         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1091                 pte_t *pte, entry;
1092                 entry = mk_pte(pages[i], vma->vm_page_prot);
1093                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1094                 page_add_new_anon_rmap(pages[i], vma, haddr);
1095                 pte = pte_offset_map(&_pmd, haddr);
1096                 VM_BUG_ON(!pte_none(*pte));
1097                 set_pte_at(mm, haddr, pte, entry);
1098                 pte_unmap(pte);
1099         }
1100         kfree(pages);
1101
1102         smp_wmb(); /* make pte visible before pmd */
1103         pmd_populate(mm, pmd, pgtable);
1104         page_remove_rmap(page);
1105         spin_unlock(&mm->page_table_lock);
1106
1107         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1108
1109         ret |= VM_FAULT_WRITE;
1110         put_page(page);
1111
1112 out:
1113         return ret;
1114
1115 out_free_pages:
1116         spin_unlock(&mm->page_table_lock);
1117         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1118         mem_cgroup_uncharge_start();
1119         for (i = 0; i < HPAGE_PMD_NR; i++) {
1120                 mem_cgroup_uncharge_page(pages[i]);
1121                 put_page(pages[i]);
1122         }
1123         mem_cgroup_uncharge_end();
1124         kfree(pages);
1125         goto out;
1126 }
1127
1128 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1129                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1130 {
1131         int ret = 0;
1132         struct page *page = NULL, *new_page;
1133         unsigned long haddr;
1134         unsigned long mmun_start;       /* For mmu_notifiers */
1135         unsigned long mmun_end;         /* For mmu_notifiers */
1136
1137         VM_BUG_ON(!vma->anon_vma);
1138         haddr = address & HPAGE_PMD_MASK;
1139         if (is_huge_zero_pmd(orig_pmd))
1140                 goto alloc;
1141         spin_lock(&mm->page_table_lock);
1142         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1143                 goto out_unlock;
1144
1145         page = pmd_page(orig_pmd);
1146         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1147         if (page_mapcount(page) == 1) {
1148                 pmd_t entry;
1149                 entry = pmd_mkyoung(orig_pmd);
1150                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1151                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1152                         update_mmu_cache_pmd(vma, address, pmd);
1153                 ret |= VM_FAULT_WRITE;
1154                 goto out_unlock;
1155         }
1156         get_page(page);
1157         spin_unlock(&mm->page_table_lock);
1158 alloc:
1159         if (transparent_hugepage_enabled(vma) &&
1160             !transparent_hugepage_debug_cow())
1161                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1162                                               vma, haddr, numa_node_id(), 0);
1163         else
1164                 new_page = NULL;
1165
1166         if (unlikely(!new_page)) {
1167                 count_vm_event(THP_FAULT_FALLBACK);
1168                 if (is_huge_zero_pmd(orig_pmd)) {
1169                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1170                                         address, pmd, orig_pmd, haddr);
1171                 } else {
1172                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1173                                         pmd, orig_pmd, page, haddr);
1174                         if (ret & VM_FAULT_OOM)
1175                                 split_huge_page(page);
1176                         put_page(page);
1177                 }
1178                 goto out;
1179         }
1180         count_vm_event(THP_FAULT_ALLOC);
1181
1182         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1183                 put_page(new_page);
1184                 if (page) {
1185                         split_huge_page(page);
1186                         put_page(page);
1187                 }
1188                 ret |= VM_FAULT_OOM;
1189                 goto out;
1190         }
1191
1192         if (is_huge_zero_pmd(orig_pmd))
1193                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1194         else
1195                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1196         __SetPageUptodate(new_page);
1197
1198         mmun_start = haddr;
1199         mmun_end   = haddr + HPAGE_PMD_SIZE;
1200         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1201
1202         spin_lock(&mm->page_table_lock);
1203         if (page)
1204                 put_page(page);
1205         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1206                 spin_unlock(&mm->page_table_lock);
1207                 mem_cgroup_uncharge_page(new_page);
1208                 put_page(new_page);
1209                 goto out_mn;
1210         } else {
1211                 pmd_t entry;
1212                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1213                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), 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                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1269                                           pmd, _pmd,  1))
1270                         update_mmu_cache_pmd(vma, addr, pmd);
1271         }
1272         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1273                 if (page->mapping && trylock_page(page)) {
1274                         lru_add_drain();
1275                         if (page->mapping)
1276                                 mlock_vma_page(page);
1277                         unlock_page(page);
1278                 }
1279         }
1280         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1281         VM_BUG_ON(!PageCompound(page));
1282         if (flags & FOLL_GET)
1283                 get_page_foll(page);
1284
1285 out:
1286         return page;
1287 }
1288
1289 /* NUMA hinting page fault entry point for trans huge pmds */
1290 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1291                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1292 {
1293         struct page *page;
1294         unsigned long haddr = addr & HPAGE_PMD_MASK;
1295         int target_nid;
1296         int current_nid = -1;
1297         bool migrated;
1298
1299         spin_lock(&mm->page_table_lock);
1300         if (unlikely(!pmd_same(pmd, *pmdp)))
1301                 goto out_unlock;
1302
1303         page = pmd_page(pmd);
1304         get_page(page);
1305         current_nid = page_to_nid(page);
1306         count_vm_numa_event(NUMA_HINT_FAULTS);
1307         if (current_nid == numa_node_id())
1308                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1309
1310         target_nid = mpol_misplaced(page, vma, haddr);
1311         if (target_nid == -1) {
1312                 put_page(page);
1313                 goto clear_pmdnuma;
1314         }
1315
1316         /* Acquire the page lock to serialise THP migrations */
1317         spin_unlock(&mm->page_table_lock);
1318         lock_page(page);
1319
1320         /* Confirm the PTE did not while locked */
1321         spin_lock(&mm->page_table_lock);
1322         if (unlikely(!pmd_same(pmd, *pmdp))) {
1323                 unlock_page(page);
1324                 put_page(page);
1325                 goto out_unlock;
1326         }
1327         spin_unlock(&mm->page_table_lock);
1328
1329         /* Migrate the THP to the requested node */
1330         migrated = migrate_misplaced_transhuge_page(mm, vma,
1331                                 pmdp, pmd, addr, page, target_nid);
1332         if (!migrated)
1333                 goto check_same;
1334
1335         task_numa_fault(target_nid, HPAGE_PMD_NR, true);
1336         return 0;
1337
1338 check_same:
1339         spin_lock(&mm->page_table_lock);
1340         if (unlikely(!pmd_same(pmd, *pmdp)))
1341                 goto out_unlock;
1342 clear_pmdnuma:
1343         pmd = pmd_mknonnuma(pmd);
1344         set_pmd_at(mm, haddr, pmdp, pmd);
1345         VM_BUG_ON(pmd_numa(*pmdp));
1346         update_mmu_cache_pmd(vma, addr, pmdp);
1347 out_unlock:
1348         spin_unlock(&mm->page_table_lock);
1349         if (current_nid != -1)
1350                 task_numa_fault(current_nid, HPAGE_PMD_NR, false);
1351         return 0;
1352 }
1353
1354 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1355                  pmd_t *pmd, unsigned long addr)
1356 {
1357         int ret = 0;
1358
1359         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1360                 struct page *page;
1361                 pgtable_t pgtable;
1362                 pmd_t orig_pmd;
1363                 /*
1364                  * For architectures like ppc64 we look at deposited pgtable
1365                  * when calling pmdp_get_and_clear. So do the
1366                  * pgtable_trans_huge_withdraw after finishing pmdp related
1367                  * operations.
1368                  */
1369                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1370                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1371                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1372                 if (is_huge_zero_pmd(orig_pmd)) {
1373                         tlb->mm->nr_ptes--;
1374                         spin_unlock(&tlb->mm->page_table_lock);
1375                         put_huge_zero_page();
1376                 } else {
1377                         page = pmd_page(orig_pmd);
1378                         page_remove_rmap(page);
1379                         VM_BUG_ON(page_mapcount(page) < 0);
1380                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1381                         VM_BUG_ON(!PageHead(page));
1382                         tlb->mm->nr_ptes--;
1383                         spin_unlock(&tlb->mm->page_table_lock);
1384                         tlb_remove_page(tlb, page);
1385                 }
1386                 pte_free(tlb->mm, pgtable);
1387                 ret = 1;
1388         }
1389         return ret;
1390 }
1391
1392 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1393                 unsigned long addr, unsigned long end,
1394                 unsigned char *vec)
1395 {
1396         int ret = 0;
1397
1398         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1399                 /*
1400                  * All logical pages in the range are present
1401                  * if backed by a huge page.
1402                  */
1403                 spin_unlock(&vma->vm_mm->page_table_lock);
1404                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1405                 ret = 1;
1406         }
1407
1408         return ret;
1409 }
1410
1411 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1412                   unsigned long old_addr,
1413                   unsigned long new_addr, unsigned long old_end,
1414                   pmd_t *old_pmd, pmd_t *new_pmd)
1415 {
1416         int ret = 0;
1417         pmd_t pmd;
1418
1419         struct mm_struct *mm = vma->vm_mm;
1420
1421         if ((old_addr & ~HPAGE_PMD_MASK) ||
1422             (new_addr & ~HPAGE_PMD_MASK) ||
1423             old_end - old_addr < HPAGE_PMD_SIZE ||
1424             (new_vma->vm_flags & VM_NOHUGEPAGE))
1425                 goto out;
1426
1427         /*
1428          * The destination pmd shouldn't be established, free_pgtables()
1429          * should have release it.
1430          */
1431         if (WARN_ON(!pmd_none(*new_pmd))) {
1432                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1433                 goto out;
1434         }
1435
1436         ret = __pmd_trans_huge_lock(old_pmd, vma);
1437         if (ret == 1) {
1438                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1439                 VM_BUG_ON(!pmd_none(*new_pmd));
1440                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1441                 spin_unlock(&mm->page_table_lock);
1442         }
1443 out:
1444         return ret;
1445 }
1446
1447 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1448                 unsigned long addr, pgprot_t newprot, int prot_numa)
1449 {
1450         struct mm_struct *mm = vma->vm_mm;
1451         int ret = 0;
1452
1453         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1454                 pmd_t entry;
1455                 entry = pmdp_get_and_clear(mm, addr, pmd);
1456                 if (!prot_numa) {
1457                         entry = pmd_modify(entry, newprot);
1458                         BUG_ON(pmd_write(entry));
1459                 } else {
1460                         struct page *page = pmd_page(*pmd);
1461
1462                         /* only check non-shared pages */
1463                         if (page_mapcount(page) == 1 &&
1464                             !pmd_numa(*pmd)) {
1465                                 entry = pmd_mknuma(entry);
1466                         }
1467                 }
1468                 set_pmd_at(mm, addr, pmd, entry);
1469                 spin_unlock(&vma->vm_mm->page_table_lock);
1470                 ret = 1;
1471         }
1472
1473         return ret;
1474 }
1475
1476 /*
1477  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1478  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1479  *
1480  * Note that if it returns 1, this routine returns without unlocking page
1481  * table locks. So callers must unlock them.
1482  */
1483 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1484 {
1485         spin_lock(&vma->vm_mm->page_table_lock);
1486         if (likely(pmd_trans_huge(*pmd))) {
1487                 if (unlikely(pmd_trans_splitting(*pmd))) {
1488                         spin_unlock(&vma->vm_mm->page_table_lock);
1489                         wait_split_huge_page(vma->anon_vma, pmd);
1490                         return -1;
1491                 } else {
1492                         /* Thp mapped by 'pmd' is stable, so we can
1493                          * handle it as it is. */
1494                         return 1;
1495                 }
1496         }
1497         spin_unlock(&vma->vm_mm->page_table_lock);
1498         return 0;
1499 }
1500
1501 pmd_t *page_check_address_pmd(struct page *page,
1502                               struct mm_struct *mm,
1503                               unsigned long address,
1504                               enum page_check_address_pmd_flag flag)
1505 {
1506         pmd_t *pmd, *ret = NULL;
1507
1508         if (address & ~HPAGE_PMD_MASK)
1509                 goto out;
1510
1511         pmd = mm_find_pmd(mm, address);
1512         if (!pmd)
1513                 goto out;
1514         if (pmd_none(*pmd))
1515                 goto out;
1516         if (pmd_page(*pmd) != page)
1517                 goto out;
1518         /*
1519          * split_vma() may create temporary aliased mappings. There is
1520          * no risk as long as all huge pmd are found and have their
1521          * splitting bit set before __split_huge_page_refcount
1522          * runs. Finding the same huge pmd more than once during the
1523          * same rmap walk is not a problem.
1524          */
1525         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1526             pmd_trans_splitting(*pmd))
1527                 goto out;
1528         if (pmd_trans_huge(*pmd)) {
1529                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1530                           !pmd_trans_splitting(*pmd));
1531                 ret = pmd;
1532         }
1533 out:
1534         return ret;
1535 }
1536
1537 static int __split_huge_page_splitting(struct page *page,
1538                                        struct vm_area_struct *vma,
1539                                        unsigned long address)
1540 {
1541         struct mm_struct *mm = vma->vm_mm;
1542         pmd_t *pmd;
1543         int ret = 0;
1544         /* For mmu_notifiers */
1545         const unsigned long mmun_start = address;
1546         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1547
1548         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1549         spin_lock(&mm->page_table_lock);
1550         pmd = page_check_address_pmd(page, mm, address,
1551                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1552         if (pmd) {
1553                 /*
1554                  * We can't temporarily set the pmd to null in order
1555                  * to split it, the pmd must remain marked huge at all
1556                  * times or the VM won't take the pmd_trans_huge paths
1557                  * and it won't wait on the anon_vma->root->rwsem to
1558                  * serialize against split_huge_page*.
1559                  */
1560                 pmdp_splitting_flush(vma, address, pmd);
1561                 ret = 1;
1562         }
1563         spin_unlock(&mm->page_table_lock);
1564         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1565
1566         return ret;
1567 }
1568
1569 static void __split_huge_page_refcount(struct page *page,
1570                                        struct list_head *list)
1571 {
1572         int i;
1573         struct zone *zone = page_zone(page);
1574         struct lruvec *lruvec;
1575         int tail_count = 0;
1576
1577         /* prevent PageLRU to go away from under us, and freeze lru stats */
1578         spin_lock_irq(&zone->lru_lock);
1579         lruvec = mem_cgroup_page_lruvec(page, zone);
1580
1581         compound_lock(page);
1582         /* complete memcg works before add pages to LRU */
1583         mem_cgroup_split_huge_fixup(page);
1584
1585         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1586                 struct page *page_tail = page + i;
1587
1588                 /* tail_page->_mapcount cannot change */
1589                 BUG_ON(page_mapcount(page_tail) < 0);
1590                 tail_count += page_mapcount(page_tail);
1591                 /* check for overflow */
1592                 BUG_ON(tail_count < 0);
1593                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1594                 /*
1595                  * tail_page->_count is zero and not changing from
1596                  * under us. But get_page_unless_zero() may be running
1597                  * from under us on the tail_page. If we used
1598                  * atomic_set() below instead of atomic_add(), we
1599                  * would then run atomic_set() concurrently with
1600                  * get_page_unless_zero(), and atomic_set() is
1601                  * implemented in C not using locked ops. spin_unlock
1602                  * on x86 sometime uses locked ops because of PPro
1603                  * errata 66, 92, so unless somebody can guarantee
1604                  * atomic_set() here would be safe on all archs (and
1605                  * not only on x86), it's safer to use atomic_add().
1606                  */
1607                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1608                            &page_tail->_count);
1609
1610                 /* after clearing PageTail the gup refcount can be released */
1611                 smp_mb();
1612
1613                 /*
1614                  * retain hwpoison flag of the poisoned tail page:
1615                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1616                  *   by the memory-failure.
1617                  */
1618                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1619                 page_tail->flags |= (page->flags &
1620                                      ((1L << PG_referenced) |
1621                                       (1L << PG_swapbacked) |
1622                                       (1L << PG_mlocked) |
1623                                       (1L << PG_uptodate) |
1624                                       (1L << PG_active) |
1625                                       (1L << PG_unevictable)));
1626                 page_tail->flags |= (1L << PG_dirty);
1627
1628                 /* clear PageTail before overwriting first_page */
1629                 smp_wmb();
1630
1631                 /*
1632                  * __split_huge_page_splitting() already set the
1633                  * splitting bit in all pmd that could map this
1634                  * hugepage, that will ensure no CPU can alter the
1635                  * mapcount on the head page. The mapcount is only
1636                  * accounted in the head page and it has to be
1637                  * transferred to all tail pages in the below code. So
1638                  * for this code to be safe, the split the mapcount
1639                  * can't change. But that doesn't mean userland can't
1640                  * keep changing and reading the page contents while
1641                  * we transfer the mapcount, so the pmd splitting
1642                  * status is achieved setting a reserved bit in the
1643                  * pmd, not by clearing the present bit.
1644                 */
1645                 page_tail->_mapcount = page->_mapcount;
1646
1647                 BUG_ON(page_tail->mapping);
1648                 page_tail->mapping = page->mapping;
1649
1650                 page_tail->index = page->index + i;
1651                 page_nid_xchg_last(page_tail, page_nid_last(page));
1652
1653                 BUG_ON(!PageAnon(page_tail));
1654                 BUG_ON(!PageUptodate(page_tail));
1655                 BUG_ON(!PageDirty(page_tail));
1656                 BUG_ON(!PageSwapBacked(page_tail));
1657
1658                 lru_add_page_tail(page, page_tail, lruvec, list);
1659         }
1660         atomic_sub(tail_count, &page->_count);
1661         BUG_ON(atomic_read(&page->_count) <= 0);
1662
1663         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1664
1665         ClearPageCompound(page);
1666         compound_unlock(page);
1667         spin_unlock_irq(&zone->lru_lock);
1668
1669         for (i = 1; i < HPAGE_PMD_NR; i++) {
1670                 struct page *page_tail = page + i;
1671                 BUG_ON(page_count(page_tail) <= 0);
1672                 /*
1673                  * Tail pages may be freed if there wasn't any mapping
1674                  * like if add_to_swap() is running on a lru page that
1675                  * had its mapping zapped. And freeing these pages
1676                  * requires taking the lru_lock so we do the put_page
1677                  * of the tail pages after the split is complete.
1678                  */
1679                 put_page(page_tail);
1680         }
1681
1682         /*
1683          * Only the head page (now become a regular page) is required
1684          * to be pinned by the caller.
1685          */
1686         BUG_ON(page_count(page) <= 0);
1687 }
1688
1689 static int __split_huge_page_map(struct page *page,
1690                                  struct vm_area_struct *vma,
1691                                  unsigned long address)
1692 {
1693         struct mm_struct *mm = vma->vm_mm;
1694         pmd_t *pmd, _pmd;
1695         int ret = 0, i;
1696         pgtable_t pgtable;
1697         unsigned long haddr;
1698
1699         spin_lock(&mm->page_table_lock);
1700         pmd = page_check_address_pmd(page, mm, address,
1701                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1702         if (pmd) {
1703                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1704                 pmd_populate(mm, &_pmd, pgtable);
1705
1706                 haddr = address;
1707                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1708                         pte_t *pte, entry;
1709                         BUG_ON(PageCompound(page+i));
1710                         entry = mk_pte(page + i, vma->vm_page_prot);
1711                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1712                         if (!pmd_write(*pmd))
1713                                 entry = pte_wrprotect(entry);
1714                         else
1715                                 BUG_ON(page_mapcount(page) != 1);
1716                         if (!pmd_young(*pmd))
1717                                 entry = pte_mkold(entry);
1718                         if (pmd_numa(*pmd))
1719                                 entry = pte_mknuma(entry);
1720                         pte = pte_offset_map(&_pmd, haddr);
1721                         BUG_ON(!pte_none(*pte));
1722                         set_pte_at(mm, haddr, pte, entry);
1723                         pte_unmap(pte);
1724                 }
1725
1726                 smp_wmb(); /* make pte visible before pmd */
1727                 /*
1728                  * Up to this point the pmd is present and huge and
1729                  * userland has the whole access to the hugepage
1730                  * during the split (which happens in place). If we
1731                  * overwrite the pmd with the not-huge version
1732                  * pointing to the pte here (which of course we could
1733                  * if all CPUs were bug free), userland could trigger
1734                  * a small page size TLB miss on the small sized TLB
1735                  * while the hugepage TLB entry is still established
1736                  * in the huge TLB. Some CPU doesn't like that. See
1737                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1738                  * Erratum 383 on page 93. Intel should be safe but is
1739                  * also warns that it's only safe if the permission
1740                  * and cache attributes of the two entries loaded in
1741                  * the two TLB is identical (which should be the case
1742                  * here). But it is generally safer to never allow
1743                  * small and huge TLB entries for the same virtual
1744                  * address to be loaded simultaneously. So instead of
1745                  * doing "pmd_populate(); flush_tlb_range();" we first
1746                  * mark the current pmd notpresent (atomically because
1747                  * here the pmd_trans_huge and pmd_trans_splitting
1748                  * must remain set at all times on the pmd until the
1749                  * split is complete for this pmd), then we flush the
1750                  * SMP TLB and finally we write the non-huge version
1751                  * of the pmd entry with pmd_populate.
1752                  */
1753                 pmdp_invalidate(vma, address, pmd);
1754                 pmd_populate(mm, pmd, pgtable);
1755                 ret = 1;
1756         }
1757         spin_unlock(&mm->page_table_lock);
1758
1759         return ret;
1760 }
1761
1762 /* must be called with anon_vma->root->rwsem held */
1763 static void __split_huge_page(struct page *page,
1764                               struct anon_vma *anon_vma,
1765                               struct list_head *list)
1766 {
1767         int mapcount, mapcount2;
1768         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1769         struct anon_vma_chain *avc;
1770
1771         BUG_ON(!PageHead(page));
1772         BUG_ON(PageTail(page));
1773
1774         mapcount = 0;
1775         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1776                 struct vm_area_struct *vma = avc->vma;
1777                 unsigned long addr = vma_address(page, vma);
1778                 BUG_ON(is_vma_temporary_stack(vma));
1779                 mapcount += __split_huge_page_splitting(page, vma, addr);
1780         }
1781         /*
1782          * It is critical that new vmas are added to the tail of the
1783          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1784          * and establishes a child pmd before
1785          * __split_huge_page_splitting() freezes the parent pmd (so if
1786          * we fail to prevent copy_huge_pmd() from running until the
1787          * whole __split_huge_page() is complete), we will still see
1788          * the newly established pmd of the child later during the
1789          * walk, to be able to set it as pmd_trans_splitting too.
1790          */
1791         if (mapcount != page_mapcount(page))
1792                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1793                        mapcount, page_mapcount(page));
1794         BUG_ON(mapcount != page_mapcount(page));
1795
1796         __split_huge_page_refcount(page, list);
1797
1798         mapcount2 = 0;
1799         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1800                 struct vm_area_struct *vma = avc->vma;
1801                 unsigned long addr = vma_address(page, vma);
1802                 BUG_ON(is_vma_temporary_stack(vma));
1803                 mapcount2 += __split_huge_page_map(page, vma, addr);
1804         }
1805         if (mapcount != mapcount2)
1806                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1807                        mapcount, mapcount2, page_mapcount(page));
1808         BUG_ON(mapcount != mapcount2);
1809 }
1810
1811 /*
1812  * Split a hugepage into normal pages. This doesn't change the position of head
1813  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1814  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1815  * from the hugepage.
1816  * Return 0 if the hugepage is split successfully otherwise return 1.
1817  */
1818 int split_huge_page_to_list(struct page *page, struct list_head *list)
1819 {
1820         struct anon_vma *anon_vma;
1821         int ret = 1;
1822
1823         BUG_ON(is_huge_zero_page(page));
1824         BUG_ON(!PageAnon(page));
1825
1826         /*
1827          * The caller does not necessarily hold an mmap_sem that would prevent
1828          * the anon_vma disappearing so we first we take a reference to it
1829          * and then lock the anon_vma for write. This is similar to
1830          * page_lock_anon_vma_read except the write lock is taken to serialise
1831          * against parallel split or collapse operations.
1832          */
1833         anon_vma = page_get_anon_vma(page);
1834         if (!anon_vma)
1835                 goto out;
1836         anon_vma_lock_write(anon_vma);
1837
1838         ret = 0;
1839         if (!PageCompound(page))
1840                 goto out_unlock;
1841
1842         BUG_ON(!PageSwapBacked(page));
1843         __split_huge_page(page, anon_vma, list);
1844         count_vm_event(THP_SPLIT);
1845
1846         BUG_ON(PageCompound(page));
1847 out_unlock:
1848         anon_vma_unlock_write(anon_vma);
1849         put_anon_vma(anon_vma);
1850 out:
1851         return ret;
1852 }
1853
1854 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1855
1856 int hugepage_madvise(struct vm_area_struct *vma,
1857                      unsigned long *vm_flags, int advice)
1858 {
1859         struct mm_struct *mm = vma->vm_mm;
1860
1861         switch (advice) {
1862         case MADV_HUGEPAGE:
1863                 /*
1864                  * Be somewhat over-protective like KSM for now!
1865                  */
1866                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1867                         return -EINVAL;
1868                 if (mm->def_flags & VM_NOHUGEPAGE)
1869                         return -EINVAL;
1870                 *vm_flags &= ~VM_NOHUGEPAGE;
1871                 *vm_flags |= VM_HUGEPAGE;
1872                 /*
1873                  * If the vma become good for khugepaged to scan,
1874                  * register it here without waiting a page fault that
1875                  * may not happen any time soon.
1876                  */
1877                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1878                         return -ENOMEM;
1879                 break;
1880         case MADV_NOHUGEPAGE:
1881                 /*
1882                  * Be somewhat over-protective like KSM for now!
1883                  */
1884                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1885                         return -EINVAL;
1886                 *vm_flags &= ~VM_HUGEPAGE;
1887                 *vm_flags |= VM_NOHUGEPAGE;
1888                 /*
1889                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1890                  * this vma even if we leave the mm registered in khugepaged if
1891                  * it got registered before VM_NOHUGEPAGE was set.
1892                  */
1893                 break;
1894         }
1895
1896         return 0;
1897 }
1898
1899 static int __init khugepaged_slab_init(void)
1900 {
1901         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1902                                           sizeof(struct mm_slot),
1903                                           __alignof__(struct mm_slot), 0, NULL);
1904         if (!mm_slot_cache)
1905                 return -ENOMEM;
1906
1907         return 0;
1908 }
1909
1910 static inline struct mm_slot *alloc_mm_slot(void)
1911 {
1912         if (!mm_slot_cache)     /* initialization failed */
1913                 return NULL;
1914         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1915 }
1916
1917 static inline void free_mm_slot(struct mm_slot *mm_slot)
1918 {
1919         kmem_cache_free(mm_slot_cache, mm_slot);
1920 }
1921
1922 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1923 {
1924         struct mm_slot *mm_slot;
1925
1926         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1927                 if (mm == mm_slot->mm)
1928                         return mm_slot;
1929
1930         return NULL;
1931 }
1932
1933 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1934                                     struct mm_slot *mm_slot)
1935 {
1936         mm_slot->mm = mm;
1937         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1938 }
1939
1940 static inline int khugepaged_test_exit(struct mm_struct *mm)
1941 {
1942         return atomic_read(&mm->mm_users) == 0;
1943 }
1944
1945 int __khugepaged_enter(struct mm_struct *mm)
1946 {
1947         struct mm_slot *mm_slot;
1948         int wakeup;
1949
1950         mm_slot = alloc_mm_slot();
1951         if (!mm_slot)
1952                 return -ENOMEM;
1953
1954         /* __khugepaged_exit() must not run from under us */
1955         VM_BUG_ON(khugepaged_test_exit(mm));
1956         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1957                 free_mm_slot(mm_slot);
1958                 return 0;
1959         }
1960
1961         spin_lock(&khugepaged_mm_lock);
1962         insert_to_mm_slots_hash(mm, mm_slot);
1963         /*
1964          * Insert just behind the scanning cursor, to let the area settle
1965          * down a little.
1966          */
1967         wakeup = list_empty(&khugepaged_scan.mm_head);
1968         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1969         spin_unlock(&khugepaged_mm_lock);
1970
1971         atomic_inc(&mm->mm_count);
1972         if (wakeup)
1973                 wake_up_interruptible(&khugepaged_wait);
1974
1975         return 0;
1976 }
1977
1978 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1979 {
1980         unsigned long hstart, hend;
1981         if (!vma->anon_vma)
1982                 /*
1983                  * Not yet faulted in so we will register later in the
1984                  * page fault if needed.
1985                  */
1986                 return 0;
1987         if (vma->vm_ops)
1988                 /* khugepaged not yet working on file or special mappings */
1989                 return 0;
1990         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1991         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1992         hend = vma->vm_end & HPAGE_PMD_MASK;
1993         if (hstart < hend)
1994                 return khugepaged_enter(vma);
1995         return 0;
1996 }
1997
1998 void __khugepaged_exit(struct mm_struct *mm)
1999 {
2000         struct mm_slot *mm_slot;
2001         int free = 0;
2002
2003         spin_lock(&khugepaged_mm_lock);
2004         mm_slot = get_mm_slot(mm);
2005         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2006                 hash_del(&mm_slot->hash);
2007                 list_del(&mm_slot->mm_node);
2008                 free = 1;
2009         }
2010         spin_unlock(&khugepaged_mm_lock);
2011
2012         if (free) {
2013                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2014                 free_mm_slot(mm_slot);
2015                 mmdrop(mm);
2016         } else if (mm_slot) {
2017                 /*
2018                  * This is required to serialize against
2019                  * khugepaged_test_exit() (which is guaranteed to run
2020                  * under mmap sem read mode). Stop here (after we
2021                  * return all pagetables will be destroyed) until
2022                  * khugepaged has finished working on the pagetables
2023                  * under the mmap_sem.
2024                  */
2025                 down_write(&mm->mmap_sem);
2026                 up_write(&mm->mmap_sem);
2027         }
2028 }
2029
2030 static void release_pte_page(struct page *page)
2031 {
2032         /* 0 stands for page_is_file_cache(page) == false */
2033         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2034         unlock_page(page);
2035         putback_lru_page(page);
2036 }
2037
2038 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2039 {
2040         while (--_pte >= pte) {
2041                 pte_t pteval = *_pte;
2042                 if (!pte_none(pteval))
2043                         release_pte_page(pte_page(pteval));
2044         }
2045 }
2046
2047 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2048                                         unsigned long address,
2049                                         pte_t *pte)
2050 {
2051         struct page *page;
2052         pte_t *_pte;
2053         int referenced = 0, none = 0;
2054         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2055              _pte++, address += PAGE_SIZE) {
2056                 pte_t pteval = *_pte;
2057                 if (pte_none(pteval)) {
2058                         if (++none <= khugepaged_max_ptes_none)
2059                                 continue;
2060                         else
2061                                 goto out;
2062                 }
2063                 if (!pte_present(pteval) || !pte_write(pteval))
2064                         goto out;
2065                 page = vm_normal_page(vma, address, pteval);
2066                 if (unlikely(!page))
2067                         goto out;
2068
2069                 VM_BUG_ON(PageCompound(page));
2070                 BUG_ON(!PageAnon(page));
2071                 VM_BUG_ON(!PageSwapBacked(page));
2072
2073                 /* cannot use mapcount: can't collapse if there's a gup pin */
2074                 if (page_count(page) != 1)
2075                         goto out;
2076                 /*
2077                  * We can do it before isolate_lru_page because the
2078                  * page can't be freed from under us. NOTE: PG_lock
2079                  * is needed to serialize against split_huge_page
2080                  * when invoked from the VM.
2081                  */
2082                 if (!trylock_page(page))
2083                         goto out;
2084                 /*
2085                  * Isolate the page to avoid collapsing an hugepage
2086                  * currently in use by the VM.
2087                  */
2088                 if (isolate_lru_page(page)) {
2089                         unlock_page(page);
2090                         goto out;
2091                 }
2092                 /* 0 stands for page_is_file_cache(page) == false */
2093                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2094                 VM_BUG_ON(!PageLocked(page));
2095                 VM_BUG_ON(PageLRU(page));
2096
2097                 /* If there is no mapped pte young don't collapse the page */
2098                 if (pte_young(pteval) || PageReferenced(page) ||
2099                     mmu_notifier_test_young(vma->vm_mm, address))
2100                         referenced = 1;
2101         }
2102         if (likely(referenced))
2103                 return 1;
2104 out:
2105         release_pte_pages(pte, _pte);
2106         return 0;
2107 }
2108
2109 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2110                                       struct vm_area_struct *vma,
2111                                       unsigned long address,
2112                                       spinlock_t *ptl)
2113 {
2114         pte_t *_pte;
2115         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2116                 pte_t pteval = *_pte;
2117                 struct page *src_page;
2118
2119                 if (pte_none(pteval)) {
2120                         clear_user_highpage(page, address);
2121                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2122                 } else {
2123                         src_page = pte_page(pteval);
2124                         copy_user_highpage(page, src_page, address, vma);
2125                         VM_BUG_ON(page_mapcount(src_page) != 1);
2126                         release_pte_page(src_page);
2127                         /*
2128                          * ptl mostly unnecessary, but preempt has to
2129                          * be disabled to update the per-cpu stats
2130                          * inside page_remove_rmap().
2131                          */
2132                         spin_lock(ptl);
2133                         /*
2134                          * paravirt calls inside pte_clear here are
2135                          * superfluous.
2136                          */
2137                         pte_clear(vma->vm_mm, address, _pte);
2138                         page_remove_rmap(src_page);
2139                         spin_unlock(ptl);
2140                         free_page_and_swap_cache(src_page);
2141                 }
2142
2143                 address += PAGE_SIZE;
2144                 page++;
2145         }
2146 }
2147
2148 static void khugepaged_alloc_sleep(void)
2149 {
2150         wait_event_freezable_timeout(khugepaged_wait, false,
2151                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2152 }
2153
2154 #ifdef CONFIG_NUMA
2155 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2156 {
2157         if (IS_ERR(*hpage)) {
2158                 if (!*wait)
2159                         return false;
2160
2161                 *wait = false;
2162                 *hpage = NULL;
2163                 khugepaged_alloc_sleep();
2164         } else if (*hpage) {
2165                 put_page(*hpage);
2166                 *hpage = NULL;
2167         }
2168
2169         return true;
2170 }
2171
2172 static struct page
2173 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2174                        struct vm_area_struct *vma, unsigned long address,
2175                        int node)
2176 {
2177         VM_BUG_ON(*hpage);
2178         /*
2179          * Allocate the page while the vma is still valid and under
2180          * the mmap_sem read mode so there is no memory allocation
2181          * later when we take the mmap_sem in write mode. This is more
2182          * friendly behavior (OTOH it may actually hide bugs) to
2183          * filesystems in userland with daemons allocating memory in
2184          * the userland I/O paths.  Allocating memory with the
2185          * mmap_sem in read mode is good idea also to allow greater
2186          * scalability.
2187          */
2188         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2189                                       node, __GFP_OTHER_NODE);
2190
2191         /*
2192          * After allocating the hugepage, release the mmap_sem read lock in
2193          * preparation for taking it in write mode.
2194          */
2195         up_read(&mm->mmap_sem);
2196         if (unlikely(!*hpage)) {
2197                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2198                 *hpage = ERR_PTR(-ENOMEM);
2199                 return NULL;
2200         }
2201
2202         count_vm_event(THP_COLLAPSE_ALLOC);
2203         return *hpage;
2204 }
2205 #else
2206 static struct page *khugepaged_alloc_hugepage(bool *wait)
2207 {
2208         struct page *hpage;
2209
2210         do {
2211                 hpage = alloc_hugepage(khugepaged_defrag());
2212                 if (!hpage) {
2213                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2214                         if (!*wait)
2215                                 return NULL;
2216
2217                         *wait = false;
2218                         khugepaged_alloc_sleep();
2219                 } else
2220                         count_vm_event(THP_COLLAPSE_ALLOC);
2221         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2222
2223         return hpage;
2224 }
2225
2226 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2227 {
2228         if (!*hpage)
2229                 *hpage = khugepaged_alloc_hugepage(wait);
2230
2231         if (unlikely(!*hpage))
2232                 return false;
2233
2234         return true;
2235 }
2236
2237 static struct page
2238 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2239                        struct vm_area_struct *vma, unsigned long address,
2240                        int node)
2241 {
2242         up_read(&mm->mmap_sem);
2243         VM_BUG_ON(!*hpage);
2244         return  *hpage;
2245 }
2246 #endif
2247
2248 static bool hugepage_vma_check(struct vm_area_struct *vma)
2249 {
2250         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2251             (vma->vm_flags & VM_NOHUGEPAGE))
2252                 return false;
2253
2254         if (!vma->anon_vma || vma->vm_ops)
2255                 return false;
2256         if (is_vma_temporary_stack(vma))
2257                 return false;
2258         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2259         return true;
2260 }
2261
2262 static void collapse_huge_page(struct mm_struct *mm,
2263                                    unsigned long address,
2264                                    struct page **hpage,
2265                                    struct vm_area_struct *vma,
2266                                    int node)
2267 {
2268         pmd_t *pmd, _pmd;
2269         pte_t *pte;
2270         pgtable_t pgtable;
2271         struct page *new_page;
2272         spinlock_t *ptl;
2273         int isolated;
2274         unsigned long hstart, hend;
2275         unsigned long mmun_start;       /* For mmu_notifiers */
2276         unsigned long mmun_end;         /* For mmu_notifiers */
2277
2278         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2279
2280         /* release the mmap_sem read lock. */
2281         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2282         if (!new_page)
2283                 return;
2284
2285         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2286                 return;
2287
2288         /*
2289          * Prevent all access to pagetables with the exception of
2290          * gup_fast later hanlded by the ptep_clear_flush and the VM
2291          * handled by the anon_vma lock + PG_lock.
2292          */
2293         down_write(&mm->mmap_sem);
2294         if (unlikely(khugepaged_test_exit(mm)))
2295                 goto out;
2296
2297         vma = find_vma(mm, address);
2298         if (!vma)
2299                 goto out;
2300         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2301         hend = vma->vm_end & HPAGE_PMD_MASK;
2302         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2303                 goto out;
2304         if (!hugepage_vma_check(vma))
2305                 goto out;
2306         pmd = mm_find_pmd(mm, address);
2307         if (!pmd)
2308                 goto out;
2309         if (pmd_trans_huge(*pmd))
2310                 goto out;
2311
2312         anon_vma_lock_write(vma->anon_vma);
2313
2314         pte = pte_offset_map(pmd, address);
2315         ptl = pte_lockptr(mm, pmd);
2316
2317         mmun_start = address;
2318         mmun_end   = address + HPAGE_PMD_SIZE;
2319         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2320         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2321         /*
2322          * After this gup_fast can't run anymore. This also removes
2323          * any huge TLB entry from the CPU so we won't allow
2324          * huge and small TLB entries for the same virtual address
2325          * to avoid the risk of CPU bugs in that area.
2326          */
2327         _pmd = pmdp_clear_flush(vma, address, pmd);
2328         spin_unlock(&mm->page_table_lock);
2329         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2330
2331         spin_lock(ptl);
2332         isolated = __collapse_huge_page_isolate(vma, address, pte);
2333         spin_unlock(ptl);
2334
2335         if (unlikely(!isolated)) {
2336                 pte_unmap(pte);
2337                 spin_lock(&mm->page_table_lock);
2338                 BUG_ON(!pmd_none(*pmd));
2339                 /*
2340                  * We can only use set_pmd_at when establishing
2341                  * hugepmds and never for establishing regular pmds that
2342                  * points to regular pagetables. Use pmd_populate for that
2343                  */
2344                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2345                 spin_unlock(&mm->page_table_lock);
2346                 anon_vma_unlock_write(vma->anon_vma);
2347                 goto out;
2348         }
2349
2350         /*
2351          * All pages are isolated and locked so anon_vma rmap
2352          * can't run anymore.
2353          */
2354         anon_vma_unlock_write(vma->anon_vma);
2355
2356         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2357         pte_unmap(pte);
2358         __SetPageUptodate(new_page);
2359         pgtable = pmd_pgtable(_pmd);
2360
2361         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2362         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2363
2364         /*
2365          * spin_lock() below is not the equivalent of smp_wmb(), so
2366          * this is needed to avoid the copy_huge_page writes to become
2367          * visible after the set_pmd_at() write.
2368          */
2369         smp_wmb();
2370
2371         spin_lock(&mm->page_table_lock);
2372         BUG_ON(!pmd_none(*pmd));
2373         page_add_new_anon_rmap(new_page, vma, address);
2374         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2375         set_pmd_at(mm, address, pmd, _pmd);
2376         update_mmu_cache_pmd(vma, address, pmd);
2377         spin_unlock(&mm->page_table_lock);
2378
2379         *hpage = NULL;
2380
2381         khugepaged_pages_collapsed++;
2382 out_up_write:
2383         up_write(&mm->mmap_sem);
2384         return;
2385
2386 out:
2387         mem_cgroup_uncharge_page(new_page);
2388         goto out_up_write;
2389 }
2390
2391 static int khugepaged_scan_pmd(struct mm_struct *mm,
2392                                struct vm_area_struct *vma,
2393                                unsigned long address,
2394                                struct page **hpage)
2395 {
2396         pmd_t *pmd;
2397         pte_t *pte, *_pte;
2398         int ret = 0, referenced = 0, none = 0;
2399         struct page *page;
2400         unsigned long _address;
2401         spinlock_t *ptl;
2402         int node = NUMA_NO_NODE;
2403
2404         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2405
2406         pmd = mm_find_pmd(mm, address);
2407         if (!pmd)
2408                 goto out;
2409         if (pmd_trans_huge(*pmd))
2410                 goto out;
2411
2412         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2413         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2414              _pte++, _address += PAGE_SIZE) {
2415                 pte_t pteval = *_pte;
2416                 if (pte_none(pteval)) {
2417                         if (++none <= khugepaged_max_ptes_none)
2418                                 continue;
2419                         else
2420                                 goto out_unmap;
2421                 }
2422                 if (!pte_present(pteval) || !pte_write(pteval))
2423                         goto out_unmap;
2424                 page = vm_normal_page(vma, _address, pteval);
2425                 if (unlikely(!page))
2426                         goto out_unmap;
2427                 /*
2428                  * Chose the node of the first page. This could
2429                  * be more sophisticated and look at more pages,
2430                  * but isn't for now.
2431                  */
2432                 if (node == NUMA_NO_NODE)
2433                         node = page_to_nid(page);
2434                 VM_BUG_ON(PageCompound(page));
2435                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2436                         goto out_unmap;
2437                 /* cannot use mapcount: can't collapse if there's a gup pin */
2438                 if (page_count(page) != 1)
2439                         goto out_unmap;
2440                 if (pte_young(pteval) || PageReferenced(page) ||
2441                     mmu_notifier_test_young(vma->vm_mm, address))
2442                         referenced = 1;
2443         }
2444         if (referenced)
2445                 ret = 1;
2446 out_unmap:
2447         pte_unmap_unlock(pte, ptl);
2448         if (ret)
2449                 /* collapse_huge_page will return with the mmap_sem released */
2450                 collapse_huge_page(mm, address, hpage, vma, node);
2451 out:
2452         return ret;
2453 }
2454
2455 static void collect_mm_slot(struct mm_slot *mm_slot)
2456 {
2457         struct mm_struct *mm = mm_slot->mm;
2458
2459         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2460
2461         if (khugepaged_test_exit(mm)) {
2462                 /* free mm_slot */
2463                 hash_del(&mm_slot->hash);
2464                 list_del(&mm_slot->mm_node);
2465
2466                 /*
2467                  * Not strictly needed because the mm exited already.
2468                  *
2469                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2470                  */
2471
2472                 /* khugepaged_mm_lock actually not necessary for the below */
2473                 free_mm_slot(mm_slot);
2474                 mmdrop(mm);
2475         }
2476 }
2477
2478 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2479                                             struct page **hpage)
2480         __releases(&khugepaged_mm_lock)
2481         __acquires(&khugepaged_mm_lock)
2482 {
2483         struct mm_slot *mm_slot;
2484         struct mm_struct *mm;
2485         struct vm_area_struct *vma;
2486         int progress = 0;
2487
2488         VM_BUG_ON(!pages);
2489         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2490
2491         if (khugepaged_scan.mm_slot)
2492                 mm_slot = khugepaged_scan.mm_slot;
2493         else {
2494                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2495                                      struct mm_slot, mm_node);
2496                 khugepaged_scan.address = 0;
2497                 khugepaged_scan.mm_slot = mm_slot;
2498         }
2499         spin_unlock(&khugepaged_mm_lock);
2500
2501         mm = mm_slot->mm;
2502         down_read(&mm->mmap_sem);
2503         if (unlikely(khugepaged_test_exit(mm)))
2504                 vma = NULL;
2505         else
2506                 vma = find_vma(mm, khugepaged_scan.address);
2507
2508         progress++;
2509         for (; vma; vma = vma->vm_next) {
2510                 unsigned long hstart, hend;
2511
2512                 cond_resched();
2513                 if (unlikely(khugepaged_test_exit(mm))) {
2514                         progress++;
2515                         break;
2516                 }
2517                 if (!hugepage_vma_check(vma)) {
2518 skip:
2519                         progress++;
2520                         continue;
2521                 }
2522                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2523                 hend = vma->vm_end & HPAGE_PMD_MASK;
2524                 if (hstart >= hend)
2525                         goto skip;
2526                 if (khugepaged_scan.address > hend)
2527                         goto skip;
2528                 if (khugepaged_scan.address < hstart)
2529                         khugepaged_scan.address = hstart;
2530                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2531
2532                 while (khugepaged_scan.address < hend) {
2533                         int ret;
2534                         cond_resched();
2535                         if (unlikely(khugepaged_test_exit(mm)))
2536                                 goto breakouterloop;
2537
2538                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2539                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2540                                   hend);
2541                         ret = khugepaged_scan_pmd(mm, vma,
2542                                                   khugepaged_scan.address,
2543                                                   hpage);
2544                         /* move to next address */
2545                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2546                         progress += HPAGE_PMD_NR;
2547                         if (ret)
2548                                 /* we released mmap_sem so break loop */
2549                                 goto breakouterloop_mmap_sem;
2550                         if (progress >= pages)
2551                                 goto breakouterloop;
2552                 }
2553         }
2554 breakouterloop:
2555         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2556 breakouterloop_mmap_sem:
2557
2558         spin_lock(&khugepaged_mm_lock);
2559         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2560         /*
2561          * Release the current mm_slot if this mm is about to die, or
2562          * if we scanned all vmas of this mm.
2563          */
2564         if (khugepaged_test_exit(mm) || !vma) {
2565                 /*
2566                  * Make sure that if mm_users is reaching zero while
2567                  * khugepaged runs here, khugepaged_exit will find
2568                  * mm_slot not pointing to the exiting mm.
2569                  */
2570                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2571                         khugepaged_scan.mm_slot = list_entry(
2572                                 mm_slot->mm_node.next,
2573                                 struct mm_slot, mm_node);
2574                         khugepaged_scan.address = 0;
2575                 } else {
2576                         khugepaged_scan.mm_slot = NULL;
2577                         khugepaged_full_scans++;
2578                 }
2579
2580                 collect_mm_slot(mm_slot);
2581         }
2582
2583         return progress;
2584 }
2585
2586 static int khugepaged_has_work(void)
2587 {
2588         return !list_empty(&khugepaged_scan.mm_head) &&
2589                 khugepaged_enabled();
2590 }
2591
2592 static int khugepaged_wait_event(void)
2593 {
2594         return !list_empty(&khugepaged_scan.mm_head) ||
2595                 kthread_should_stop();
2596 }
2597
2598 static void khugepaged_do_scan(void)
2599 {
2600         struct page *hpage = NULL;
2601         unsigned int progress = 0, pass_through_head = 0;
2602         unsigned int pages = khugepaged_pages_to_scan;
2603         bool wait = true;
2604
2605         barrier(); /* write khugepaged_pages_to_scan to local stack */
2606
2607         while (progress < pages) {
2608                 if (!khugepaged_prealloc_page(&hpage, &wait))
2609                         break;
2610
2611                 cond_resched();
2612
2613                 if (unlikely(kthread_should_stop() || freezing(current)))
2614                         break;
2615
2616                 spin_lock(&khugepaged_mm_lock);
2617                 if (!khugepaged_scan.mm_slot)
2618                         pass_through_head++;
2619                 if (khugepaged_has_work() &&
2620                     pass_through_head < 2)
2621                         progress += khugepaged_scan_mm_slot(pages - progress,
2622                                                             &hpage);
2623                 else
2624                         progress = pages;
2625                 spin_unlock(&khugepaged_mm_lock);
2626         }
2627
2628         if (!IS_ERR_OR_NULL(hpage))
2629                 put_page(hpage);
2630 }
2631
2632 static void khugepaged_wait_work(void)
2633 {
2634         try_to_freeze();
2635
2636         if (khugepaged_has_work()) {
2637                 if (!khugepaged_scan_sleep_millisecs)
2638                         return;
2639
2640                 wait_event_freezable_timeout(khugepaged_wait,
2641                                              kthread_should_stop(),
2642                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2643                 return;
2644         }
2645
2646         if (khugepaged_enabled())
2647                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2648 }
2649
2650 static int khugepaged(void *none)
2651 {
2652         struct mm_slot *mm_slot;
2653
2654         set_freezable();
2655         set_user_nice(current, 19);
2656
2657         while (!kthread_should_stop()) {
2658                 khugepaged_do_scan();
2659                 khugepaged_wait_work();
2660         }
2661
2662         spin_lock(&khugepaged_mm_lock);
2663         mm_slot = khugepaged_scan.mm_slot;
2664         khugepaged_scan.mm_slot = NULL;
2665         if (mm_slot)
2666                 collect_mm_slot(mm_slot);
2667         spin_unlock(&khugepaged_mm_lock);
2668         return 0;
2669 }
2670
2671 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2672                 unsigned long haddr, pmd_t *pmd)
2673 {
2674         struct mm_struct *mm = vma->vm_mm;
2675         pgtable_t pgtable;
2676         pmd_t _pmd;
2677         int i;
2678
2679         pmdp_clear_flush(vma, haddr, pmd);
2680         /* leave pmd empty until pte is filled */
2681
2682         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2683         pmd_populate(mm, &_pmd, pgtable);
2684
2685         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2686                 pte_t *pte, entry;
2687                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2688                 entry = pte_mkspecial(entry);
2689                 pte = pte_offset_map(&_pmd, haddr);
2690                 VM_BUG_ON(!pte_none(*pte));
2691                 set_pte_at(mm, haddr, pte, entry);
2692                 pte_unmap(pte);
2693         }
2694         smp_wmb(); /* make pte visible before pmd */
2695         pmd_populate(mm, pmd, pgtable);
2696         put_huge_zero_page();
2697 }
2698
2699 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2700                 pmd_t *pmd)
2701 {
2702         struct page *page;
2703         struct mm_struct *mm = vma->vm_mm;
2704         unsigned long haddr = address & HPAGE_PMD_MASK;
2705         unsigned long mmun_start;       /* For mmu_notifiers */
2706         unsigned long mmun_end;         /* For mmu_notifiers */
2707
2708         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2709
2710         mmun_start = haddr;
2711         mmun_end   = haddr + HPAGE_PMD_SIZE;
2712         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2713         spin_lock(&mm->page_table_lock);
2714         if (unlikely(!pmd_trans_huge(*pmd))) {
2715                 spin_unlock(&mm->page_table_lock);
2716                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2717                 return;
2718         }
2719         if (is_huge_zero_pmd(*pmd)) {
2720                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2721                 spin_unlock(&mm->page_table_lock);
2722                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2723                 return;
2724         }
2725         page = pmd_page(*pmd);
2726         VM_BUG_ON(!page_count(page));
2727         get_page(page);
2728         spin_unlock(&mm->page_table_lock);
2729         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2730
2731         split_huge_page(page);
2732
2733         put_page(page);
2734         BUG_ON(pmd_trans_huge(*pmd));
2735 }
2736
2737 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2738                 pmd_t *pmd)
2739 {
2740         struct vm_area_struct *vma;
2741
2742         vma = find_vma(mm, address);
2743         BUG_ON(vma == NULL);
2744         split_huge_page_pmd(vma, address, pmd);
2745 }
2746
2747 static void split_huge_page_address(struct mm_struct *mm,
2748                                     unsigned long address)
2749 {
2750         pmd_t *pmd;
2751
2752         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2753
2754         pmd = mm_find_pmd(mm, address);
2755         if (!pmd)
2756                 return;
2757         /*
2758          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2759          * materialize from under us.
2760          */
2761         split_huge_page_pmd_mm(mm, address, pmd);
2762 }
2763
2764 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2765                              unsigned long start,
2766                              unsigned long end,
2767                              long adjust_next)
2768 {
2769         /*
2770          * If the new start address isn't hpage aligned and it could
2771          * previously contain an hugepage: check if we need to split
2772          * an huge pmd.
2773          */
2774         if (start & ~HPAGE_PMD_MASK &&
2775             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2776             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2777                 split_huge_page_address(vma->vm_mm, start);
2778
2779         /*
2780          * If the new end address isn't hpage aligned and it could
2781          * previously contain an hugepage: check if we need to split
2782          * an huge pmd.
2783          */
2784         if (end & ~HPAGE_PMD_MASK &&
2785             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2786             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2787                 split_huge_page_address(vma->vm_mm, end);
2788
2789         /*
2790          * If we're also updating the vma->vm_next->vm_start, if the new
2791          * vm_next->vm_start isn't page aligned and it could previously
2792          * contain an hugepage: check if we need to split an huge pmd.
2793          */
2794         if (adjust_next > 0) {
2795                 struct vm_area_struct *next = vma->vm_next;
2796                 unsigned long nstart = next->vm_start;
2797                 nstart += adjust_next << PAGE_SHIFT;
2798                 if (nstart & ~HPAGE_PMD_MASK &&
2799                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2800                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2801                         split_huge_page_address(next->vm_mm, nstart);
2802         }
2803 }