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, struct vm_area_struct *vma)
 694 {
 695         pmd_t entry;
 696         entry = mk_pmd(page, vma->vm_page_prot);
 697         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 698         entry = pmd_mkhuge(entry);
 699         return entry;
 700 }
 701
 702 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 703                                         struct vm_area_struct *vma,
 704                                         unsigned long haddr, pmd_t *pmd,
 705                                         struct page *page)
 706 {
 707         pgtable_t pgtable;
 708
 709         VM_BUG_ON(!PageCompound(page));
 710         pgtable = pte_alloc_one(mm, haddr);
 711         if (unlikely(!pgtable))
 712                 return VM_FAULT_OOM;
 713
 714         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 715         /*
 716          * The memory barrier inside __SetPageUptodate makes sure that
 717          * clear_huge_page writes become visible before the set_pmd_at()
 718          * write.
 719          */
 720         __SetPageUptodate(page);
 721
 722         spin_lock(&mm->page_table_lock);
 723         if (unlikely(!pmd_none(*pmd))) {
 724                 spin_unlock(&mm->page_table_lock);
 725                 mem_cgroup_uncharge_page(page);
 726                 put_page(page);
 727                 pte_free(mm, pgtable);
 728         } else {
 729                 pmd_t entry;
 730                 entry = mk_huge_pmd(page, vma);
 731                 page_add_new_anon_rmap(page, vma, haddr);
 732                 set_pmd_at(mm, haddr, pmd, entry);
 733                 pgtable_trans_huge_deposit(mm, pgtable);
 734                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 735                 mm->nr_ptes++;
 736                 spin_unlock(&mm->page_table_lock);
 737         }
 738
 739         return 0;
 740 }
 741
 742 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 743 {
 744         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 745 }
 746
 747 static inline struct page *alloc_hugepage_vma(int defrag,
 748                                               struct vm_area_struct *vma,
 749                                               unsigned long haddr, int nd,
 750                                               gfp_t extra_gfp)
 751 {
 752         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 753                                HPAGE_PMD_ORDER, vma, haddr, nd);
 754 }
 755
 756 #ifndef CONFIG_NUMA
 757 static inline struct page *alloc_hugepage(int defrag)
 758 {
 759         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 760                            HPAGE_PMD_ORDER);
 761 }
 762 #endif
 763
 764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 765                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 766                 struct page *zero_page)
 767 {
 768         pmd_t entry;
 769         if (!pmd_none(*pmd))
 770                 return false;
 771         entry = mk_pmd(zero_page, vma->vm_page_prot);
 772         entry = pmd_wrprotect(entry);
 773         entry = pmd_mkhuge(entry);
 774         set_pmd_at(mm, haddr, pmd, entry);
 775         pgtable_trans_huge_deposit(mm, pgtable);
 776         mm->nr_ptes++;
 777         return true;
 778 }
 779
 780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 781                                unsigned long address, pmd_t *pmd,
 782                                unsigned int flags)
 783 {
 784         struct page *page;
 785         unsigned long haddr = address & HPAGE_PMD_MASK;
 786         pte_t *pte;
 787
 788         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 789                 if (unlikely(anon_vma_prepare(vma)))
 790                         return VM_FAULT_OOM;
 791                 if (unlikely(khugepaged_enter(vma)))
 792                         return VM_FAULT_OOM;
 793                 if (!(flags & FAULT_FLAG_WRITE) &&
 794                                 transparent_hugepage_use_zero_page()) {
 795                         pgtable_t pgtable;
 796                         struct page *zero_page;
 797                         bool set;
 798                         pgtable = pte_alloc_one(mm, haddr);
 799                         if (unlikely(!pgtable))
 800                                 return VM_FAULT_OOM;
 801                         zero_page = get_huge_zero_page();
 802                         if (unlikely(!zero_page)) {
 803                                 pte_free(mm, pgtable);
 804                                 count_vm_event(THP_FAULT_FALLBACK);
 805                                 goto out;
 806                         }
 807                         spin_lock(&mm->page_table_lock);
 808                         set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 809                                         zero_page);
 810                         spin_unlock(&mm->page_table_lock);
 811                         if (!set) {
 812                                 pte_free(mm, pgtable);
 813                                 put_huge_zero_page();
 814                         }
 815                         return 0;
 816                 }
 817                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 818                                           vma, haddr, numa_node_id(), 0);
 819                 if (unlikely(!page)) {
 820                         count_vm_event(THP_FAULT_FALLBACK);
 821                         goto out;
 822                 }
 823                 count_vm_event(THP_FAULT_ALLOC);
 824                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 825                         put_page(page);
 826                         goto out;
 827                 }
 828                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 829                                                           page))) {
 830                         mem_cgroup_uncharge_page(page);
 831                         put_page(page);
 832                         goto out;
 833                 }
 834
 835                 return 0;
 836         }
 837 out:
 838         /*
 839          * Use __pte_alloc instead of pte_alloc_map, because we can't
 840          * run pte_offset_map on the pmd, if an huge pmd could
 841          * materialize from under us from a different thread.
 842          */
 843         if (unlikely(pmd_none(*pmd)) &&
 844             unlikely(__pte_alloc(mm, vma, pmd, address)))
 845                 return VM_FAULT_OOM;
 846         /* if an huge pmd materialized from under us just retry later */
 847         if (unlikely(pmd_trans_huge(*pmd)))
 848                 return 0;
 849         /*
 850          * A regular pmd is established and it can't morph into a huge pmd
 851          * from under us anymore at this point because we hold the mmap_sem
 852          * read mode and khugepaged takes it in write mode. So now it's
 853          * safe to run pte_offset_map().
 854          */
 855         pte = pte_offset_map(pmd, address);
 856         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 857 }
 858
 859 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 860                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 861                   struct vm_area_struct *vma)
 862 {
 863         struct page *src_page;
 864         pmd_t pmd;
 865         pgtable_t pgtable;
 866         int ret;
 867
 868         ret = -ENOMEM;
 869         pgtable = pte_alloc_one(dst_mm, addr);
 870         if (unlikely(!pgtable))
 871                 goto out;
 872
 873         spin_lock(&dst_mm->page_table_lock);
 874         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 875
 876         ret = -EAGAIN;
 877         pmd = *src_pmd;
 878         if (unlikely(!pmd_trans_huge(pmd))) {
 879                 pte_free(dst_mm, pgtable);
 880                 goto out_unlock;
 881         }
 882         /*
 883          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 884          * under splitting since we don't split the page itself, only pmd to
 885          * a page table.
 886          */
 887         if (is_huge_zero_pmd(pmd)) {
 888                 struct page *zero_page;
 889                 bool set;
 890                 /*
 891                  * get_huge_zero_page() will never allocate a new page here,
 892                  * since we already have a zero page to copy. It just takes a
 893                  * reference.
 894                  */
 895                 zero_page = get_huge_zero_page();
 896                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 897                                 zero_page);
 898                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 899                 ret = 0;
 900                 goto out_unlock;
 901         }
 902         if (unlikely(pmd_trans_splitting(pmd))) {
 903                 /* split huge page running from under us */
 904                 spin_unlock(&src_mm->page_table_lock);
 905                 spin_unlock(&dst_mm->page_table_lock);
 906                 pte_free(dst_mm, pgtable);
 907
 908                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 909                 goto out;
 910         }
 911         src_page = pmd_page(pmd);
 912         VM_BUG_ON(!PageHead(src_page));
 913         get_page(src_page);
 914         page_dup_rmap(src_page);
 915         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 916
 917         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 918         pmd = pmd_mkold(pmd_wrprotect(pmd));
 919         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 920         pgtable_trans_huge_deposit(dst_mm, pgtable);
 921         dst_mm->nr_ptes++;
 922
 923         ret = 0;
 924 out_unlock:
 925         spin_unlock(&src_mm->page_table_lock);
 926         spin_unlock(&dst_mm->page_table_lock);
 927 out:
 928         return ret;
 929 }
 930
 931 void huge_pmd_set_accessed(struct mm_struct *mm,
 932                            struct vm_area_struct *vma,
 933                            unsigned long address,
 934                            pmd_t *pmd, pmd_t orig_pmd,
 935                            int dirty)
 936 {
 937         pmd_t entry;
 938         unsigned long haddr;
 939
 940         spin_lock(&mm->page_table_lock);
 941         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 942                 goto unlock;
 943
 944         entry = pmd_mkyoung(orig_pmd);
 945         haddr = address & HPAGE_PMD_MASK;
 946         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 947                 update_mmu_cache_pmd(vma, address, pmd);
 948
 949 unlock:
 950         spin_unlock(&mm->page_table_lock);
 951 }
 952
 953 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 954                 struct vm_area_struct *vma, unsigned long address,
 955                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 956 {
 957         pgtable_t pgtable;
 958         pmd_t _pmd;
 959         struct page *page;
 960         int i, ret = 0;
 961         unsigned long mmun_start;       /* For mmu_notifiers */
 962         unsigned long mmun_end;         /* For mmu_notifiers */
 963
 964         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 965         if (!page) {
 966                 ret |= VM_FAULT_OOM;
 967                 goto out;
 968         }
 969
 970         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 971                 put_page(page);
 972                 ret |= VM_FAULT_OOM;
 973                 goto out;
 974         }
 975
 976         clear_user_highpage(page, address);
 977         __SetPageUptodate(page);
 978
 979         mmun_start = haddr;
 980         mmun_end   = haddr + HPAGE_PMD_SIZE;
 981         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 982
 983         spin_lock(&mm->page_table_lock);
 984         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 985                 goto out_free_page;
 986
 987         pmdp_clear_flush(vma, haddr, pmd);
 988         /* leave pmd empty until pte is filled */
 989
 990         pgtable = pgtable_trans_huge_withdraw(mm);
 991         pmd_populate(mm, &_pmd, pgtable);
 992
 993         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 994                 pte_t *pte, entry;
 995                 if (haddr == (address & PAGE_MASK)) {
 996                         entry = mk_pte(page, vma->vm_page_prot);
 997                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 998                         page_add_new_anon_rmap(page, vma, haddr);
 999                 } else {
1000                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1001                         entry = pte_mkspecial(entry);
1002                 }
1003                 pte = pte_offset_map(&_pmd, haddr);
1004                 VM_BUG_ON(!pte_none(*pte));
1005                 set_pte_at(mm, haddr, pte, entry);
1006                 pte_unmap(pte);
1007         }
1008         smp_wmb(); /* make pte visible before pmd */
1009         pmd_populate(mm, pmd, pgtable);
1010         spin_unlock(&mm->page_table_lock);
1011         put_huge_zero_page();
1012         inc_mm_counter(mm, MM_ANONPAGES);
1013
1014         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1015
1016         ret |= VM_FAULT_WRITE;
1017 out:
1018         return ret;
1019 out_free_page:
1020         spin_unlock(&mm->page_table_lock);
1021         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1022         mem_cgroup_uncharge_page(page);
1023         put_page(page);
1024         goto out;
1025 }
1026
1027 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1028                                         struct vm_area_struct *vma,
1029                                         unsigned long address,
1030                                         pmd_t *pmd, pmd_t orig_pmd,
1031                                         struct page *page,
1032                                         unsigned long haddr)
1033 {
1034         pgtable_t pgtable;
1035         pmd_t _pmd;
1036         int ret = 0, i;
1037         struct page **pages;
1038         unsigned long mmun_start;       /* For mmu_notifiers */
1039         unsigned long mmun_end;         /* For mmu_notifiers */
1040
1041         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1042                         GFP_KERNEL);
1043         if (unlikely(!pages)) {
1044                 ret |= VM_FAULT_OOM;
1045                 goto out;
1046         }
1047
1048         for (i = 0; i < HPAGE_PMD_NR; i++) {
1049                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1050                                                __GFP_OTHER_NODE,
1051                                                vma, address, page_to_nid(page));
1052                 if (unlikely(!pages[i] ||
1053                              mem_cgroup_newpage_charge(pages[i], mm,
1054                                                        GFP_KERNEL))) {
1055                         if (pages[i])
1056                                 put_page(pages[i]);
1057                         mem_cgroup_uncharge_start();
1058                         while (--i >= 0) {
1059                                 mem_cgroup_uncharge_page(pages[i]);
1060                                 put_page(pages[i]);
1061                         }
1062                         mem_cgroup_uncharge_end();
1063                         kfree(pages);
1064                         ret |= VM_FAULT_OOM;
1065                         goto out;
1066                 }
1067         }
1068
1069         for (i = 0; i < HPAGE_PMD_NR; i++) {
1070                 copy_user_highpage(pages[i], page + i,
1071                                    haddr + PAGE_SIZE * i, vma);
1072                 __SetPageUptodate(pages[i]);
1073                 cond_resched();
1074         }
1075
1076         mmun_start = haddr;
1077         mmun_end   = haddr + HPAGE_PMD_SIZE;
1078         mmu_notifier_invalidate_range_start(vma, mmun_start, mmun_end,
1079                                             MMU_MIGRATE);
1080
1081         spin_lock(&mm->page_table_lock);
1082         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1083                 goto out_free_pages;
1084         VM_BUG_ON(!PageHead(page));
1085
1086         pmdp_clear_flush_notify(vma, haddr, pmd);
1087         /* leave pmd empty until pte is filled */
1088         mmu_notifier_invalidate_range_free_pages(vma, mmun_start, mmun_end);
1089
1090         pgtable = pgtable_trans_huge_withdraw(mm);
1091         pmd_populate(mm, &_pmd, pgtable);
1092
1093         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1094                 pte_t *pte, entry;
1095                 entry = mk_pte(pages[i], vma->vm_page_prot);
1096                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1097                 page_add_new_anon_rmap(pages[i], vma, haddr);
1098                 pte = pte_offset_map(&_pmd, haddr);
1099                 VM_BUG_ON(!pte_none(*pte));
1100                 set_pte_at(mm, haddr, pte, entry);
1101                 pte_unmap(pte);
1102         }
1103         kfree(pages);
1104
1105         smp_wmb(); /* make pte visible before pmd */
1106         pmd_populate(mm, pmd, pgtable);
1107         page_remove_rmap(page);
1108         spin_unlock(&mm->page_table_lock);
1109
1110         mmu_notifier_invalidate_range_end(vma, mmun_start,
1111                                           mmun_end, MMU_MIGRATE);
1112
1113         ret |= VM_FAULT_WRITE;
1114         put_page(page);
1115
1116 out:
1117         return ret;
1118
1119 out_free_pages:
1120         spin_unlock(&mm->page_table_lock);
1121         mmu_notifier_invalidate_range_end(vma, mmun_start,
1122                                           mmun_end, MMU_MIGRATE);
1123         mem_cgroup_uncharge_start();
1124         for (i = 0; i < HPAGE_PMD_NR; i++) {
1125                 mem_cgroup_uncharge_page(pages[i]);
1126                 put_page(pages[i]);
1127         }
1128         mem_cgroup_uncharge_end();
1129         kfree(pages);
1130         goto out;
1131 }
1132
1133 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1134                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1135 {
1136         int ret = 0;
1137         struct page *page = NULL, *new_page;
1138         unsigned long haddr;
1139         unsigned long mmun_start;       /* For mmu_notifiers */
1140         unsigned long mmun_end;         /* For mmu_notifiers */
1141
1142         VM_BUG_ON(!vma->anon_vma);
1143         haddr = address & HPAGE_PMD_MASK;
1144         if (is_huge_zero_pmd(orig_pmd))
1145                 goto alloc;
1146         spin_lock(&mm->page_table_lock);
1147         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1148                 goto out_unlock;
1149
1150         page = pmd_page(orig_pmd);
1151         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1152         if (page_mapcount(page) == 1) {
1153                 pmd_t entry;
1154                 entry = pmd_mkyoung(orig_pmd);
1155                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1156                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1157                         update_mmu_cache_pmd(vma, address, pmd);
1158                 ret |= VM_FAULT_WRITE;
1159                 goto out_unlock;
1160         }
1161         get_page(page);
1162         spin_unlock(&mm->page_table_lock);
1163 alloc:
1164         if (transparent_hugepage_enabled(vma) &&
1165             !transparent_hugepage_debug_cow())
1166                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1167                                               vma, haddr, numa_node_id(), 0);
1168         else
1169                 new_page = NULL;
1170
1171         if (unlikely(!new_page)) {
1172                 count_vm_event(THP_FAULT_FALLBACK);
1173                 if (!page) {
1174                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1175                                         address, pmd, orig_pmd, haddr);
1176                 } else {
1177                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1178                                         pmd, orig_pmd, page, haddr);
1179                         if (ret & VM_FAULT_OOM)
1180                                 split_huge_page(page);
1181                         put_page(page);
1182                 }
1183                 goto out;
1184         }
1185         count_vm_event(THP_FAULT_ALLOC);
1186
1187         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1188                 put_page(new_page);
1189                 if (page) {
1190                         split_huge_page(page);
1191                         put_page(page);
1192                 }
1193                 ret |= VM_FAULT_OOM;
1194                 goto out;
1195         }
1196
1197         if (!page)
1198                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1199         else
1200                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1201         __SetPageUptodate(new_page);
1202
1203         mmun_start = haddr;
1204         mmun_end   = haddr + HPAGE_PMD_SIZE;
1205         mmu_notifier_invalidate_range_start(vma, mmun_start, mmun_end,
1206                                             MMU_MIGRATE);
1207
1208         spin_lock(&mm->page_table_lock);
1209         if (page)
1210                 put_page(page);
1211         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1212                 spin_unlock(&mm->page_table_lock);
1213                 mem_cgroup_uncharge_page(new_page);
1214                 put_page(new_page);
1215                 goto out_mn;
1216         } else {
1217                 pmd_t entry;
1218                 entry = mk_huge_pmd(new_page, vma);
1219                 pmdp_clear_flush_notify(vma, haddr, pmd);
1220                 mmu_notifier_invalidate_range_free_pages(vma, mmun_start,
1221                                                          mmun_end);
1222                 page_add_new_anon_rmap(new_page, vma, haddr);
1223                 set_pmd_at(mm, haddr, pmd, entry);
1224                 update_mmu_cache_pmd(vma, address, pmd);
1225                 if (!page) {
1226                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1227                         put_huge_zero_page();
1228                 } else {
1229                         VM_BUG_ON(!PageHead(page));
1230                         page_remove_rmap(page);
1231                         put_page(page);
1232                 }
1233                 ret |= VM_FAULT_WRITE;
1234         }
1235         spin_unlock(&mm->page_table_lock);
1236 out_mn:
1237         mmu_notifier_invalidate_range_end(vma, mmun_start,
1238                                           mmun_end, MMU_MIGRATE);
1239 out:
1240         return ret;
1241 out_unlock:
1242         spin_unlock(&mm->page_table_lock);
1243         return ret;
1244 }
1245
1246 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1247                                    unsigned long addr,
1248                                    pmd_t *pmd,
1249                                    unsigned int flags)
1250 {
1251         struct mm_struct *mm = vma->vm_mm;
1252         struct page *page = NULL;
1253
1254         assert_spin_locked(&mm->page_table_lock);
1255
1256         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1257                 goto out;
1258
1259         /* Avoid dumping huge zero page */
1260         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1261                 return ERR_PTR(-EFAULT);
1262
1263         page = pmd_page(*pmd);
1264         VM_BUG_ON(!PageHead(page));
1265         if (flags & FOLL_TOUCH) {
1266                 pmd_t _pmd;
1267                 /*
1268                  * We should set the dirty bit only for FOLL_WRITE but
1269                  * for now the dirty bit in the pmd is meaningless.
1270                  * And if the dirty bit will become meaningful and
1271                  * we'll only set it with FOLL_WRITE, an atomic
1272                  * set_bit will be required on the pmd to set the
1273                  * young bit, instead of the current set_pmd_at.
1274                  */
1275                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1276                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1277         }
1278         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1279                 if (page->mapping && trylock_page(page)) {
1280                         lru_add_drain();
1281                         if (page->mapping)
1282                                 mlock_vma_page(page);
1283                         unlock_page(page);
1284                 }
1285         }
1286         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1287         VM_BUG_ON(!PageCompound(page));
1288         if (flags & FOLL_GET)
1289                 get_page_foll(page);
1290
1291 out:
1292         return page;
1293 }
1294
1295 /* NUMA hinting page fault entry point for trans huge pmds */
1296 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1297                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1298 {
1299         struct anon_vma *anon_vma = NULL;
1300         struct page *page;
1301         unsigned long haddr = addr & HPAGE_PMD_MASK;
1302         int page_nid = -1, this_nid = numa_node_id();
1303         int target_nid;
1304         bool page_locked;
1305         bool migrated = false;
1306
1307         spin_lock(&mm->page_table_lock);
1308         if (unlikely(!pmd_same(pmd, *pmdp)))
1309                 goto out_unlock;
1310
1311         page = pmd_page(pmd);
1312         page_nid = page_to_nid(page);
1313         count_vm_numa_event(NUMA_HINT_FAULTS);
1314         if (page_nid == this_nid)
1315                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1316
1317         /*
1318          * Acquire the page lock to serialise THP migrations but avoid dropping
1319          * page_table_lock if at all possible
1320          */
1321         page_locked = trylock_page(page);
1322         target_nid = mpol_misplaced(page, vma, haddr);
1323         if (target_nid == -1) {
1324                 /* If the page was locked, there are no parallel migrations */
1325                 if (page_locked)
1326                         goto clear_pmdnuma;
1327
1328                 /*
1329                  * Otherwise wait for potential migrations and retry. We do
1330                  * relock and check_same as the page may no longer be mapped.
1331                  * As the fault is being retried, do not account for it.
1332                  */
1333                 spin_unlock(&mm->page_table_lock);
1334                 wait_on_page_locked(page);
1335                 page_nid = -1;
1336                 goto out;
1337         }
1338
1339         /* Page is misplaced, serialise migrations and parallel THP splits */
1340         get_page(page);
1341         spin_unlock(&mm->page_table_lock);
1342         if (!page_locked)
1343                 lock_page(page);
1344         anon_vma = page_lock_anon_vma_read(page);
1345
1346         /* Confirm the PTE did not while locked */
1347         spin_lock(&mm->page_table_lock);
1348         if (unlikely(!pmd_same(pmd, *pmdp))) {
1349                 unlock_page(page);
1350                 put_page(page);
1351                 page_nid = -1;
1352                 goto out_unlock;
1353         }
1354
1355         /* Bail if we fail to protect against THP splits for any reason */
1356         if (unlikely(!anon_vma)) {
1357                 put_page(page);
1358                 page_nid = -1;
1359                 goto clear_pmdnuma;
1360         }
1361
1362         /*
1363          * The page_table_lock above provides a memory barrier
1364          * with change_protection_range.
1365          */
1366         if (mm_tlb_flush_pending(mm))
1367                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1368
1369         /*
1370          * Migrate the THP to the requested node, returns with page unlocked
1371          * and pmd_numa cleared.
1372          */
1373         spin_unlock(&mm->page_table_lock);
1374         migrated = migrate_misplaced_transhuge_page(mm, vma,
1375                                 pmdp, pmd, addr, page, target_nid);
1376         if (migrated)
1377                 page_nid = target_nid;
1378
1379         goto out;
1380 clear_pmdnuma:
1381         BUG_ON(!PageLocked(page));
1382         pmd = pmd_mknonnuma(pmd);
1383         set_pmd_at(mm, haddr, pmdp, pmd);
1384         VM_BUG_ON(pmd_numa(*pmdp));
1385         update_mmu_cache_pmd(vma, addr, pmdp);
1386         unlock_page(page);
1387 out_unlock:
1388         spin_unlock(&mm->page_table_lock);
1389
1390 out:
1391         if (anon_vma)
1392                 page_unlock_anon_vma_read(anon_vma);
1393
1394         if (page_nid != -1)
1395                 task_numa_fault(page_nid, HPAGE_PMD_NR, migrated);
1396
1397         return 0;
1398 }
1399
1400 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1401                  pmd_t *pmd, unsigned long addr)
1402 {
1403         int ret = 0;
1404
1405         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1406                 struct page *page;
1407                 pgtable_t pgtable;
1408                 pmd_t orig_pmd;
1409                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1410                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1411                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1412                 if (is_huge_zero_pmd(orig_pmd)) {
1413                         tlb->mm->nr_ptes--;
1414                         spin_unlock(&tlb->mm->page_table_lock);
1415                         put_huge_zero_page();
1416                 } else {
1417                         page = pmd_page(orig_pmd);
1418                         page_remove_rmap(page);
1419                         VM_BUG_ON(page_mapcount(page) < 0);
1420                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1421                         VM_BUG_ON(!PageHead(page));
1422                         tlb->mm->nr_ptes--;
1423                         spin_unlock(&tlb->mm->page_table_lock);
1424                         tlb_remove_page(tlb, page);
1425                 }
1426                 pte_free(tlb->mm, pgtable);
1427                 ret = 1;
1428         }
1429         return ret;
1430 }
1431
1432 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1433                 unsigned long addr, unsigned long end,
1434                 unsigned char *vec)
1435 {
1436         int ret = 0;
1437
1438         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1439                 /*
1440                  * All logical pages in the range are present
1441                  * if backed by a huge page.
1442                  */
1443                 spin_unlock(&vma->vm_mm->page_table_lock);
1444                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1445                 ret = 1;
1446         }
1447
1448         return ret;
1449 }
1450
1451 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1452                   unsigned long old_addr,
1453                   unsigned long new_addr, unsigned long old_end,
1454                   pmd_t *old_pmd, pmd_t *new_pmd)
1455 {
1456         int ret = 0;
1457         pmd_t pmd;
1458
1459         struct mm_struct *mm = vma->vm_mm;
1460
1461         if ((old_addr & ~HPAGE_PMD_MASK) ||
1462             (new_addr & ~HPAGE_PMD_MASK) ||
1463             old_end - old_addr < HPAGE_PMD_SIZE ||
1464             (new_vma->vm_flags & VM_NOHUGEPAGE))
1465                 goto out;
1466
1467         /*
1468          * The destination pmd shouldn't be established, free_pgtables()
1469          * should have release it.
1470          */
1471         if (WARN_ON(!pmd_none(*new_pmd))) {
1472                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1473                 goto out;
1474         }
1475
1476         ret = __pmd_trans_huge_lock(old_pmd, vma);
1477         if (ret == 1) {
1478                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1479                 VM_BUG_ON(!pmd_none(*new_pmd));
1480                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1481                 spin_unlock(&mm->page_table_lock);
1482         }
1483 out:
1484         return ret;
1485 }
1486
1487 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1488                 unsigned long addr, pgprot_t newprot, int prot_numa)
1489 {
1490         struct mm_struct *mm = vma->vm_mm;
1491         int ret = 0;
1492
1493         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1494                 pmd_t entry;
1495                 entry = pmdp_get_and_clear_notify(mm, addr, pmd);
1496                 if (!prot_numa) {
1497                         entry = pmd_modify(entry, newprot);
1498                         BUG_ON(pmd_write(entry));
1499                 } else {
1500                         struct page *page = pmd_page(*pmd);
1501
1502                         /* only check non-shared pages */
1503                         if (page_mapcount(page) == 1 &&
1504                             !pmd_numa(*pmd)) {
1505                                 entry = pmd_mknuma(entry);
1506                         }
1507                 }
1508                 set_pmd_at(mm, addr, pmd, entry);
1509                 spin_unlock(&vma->vm_mm->page_table_lock);
1510                 ret = 1;
1511         }
1512
1513         return ret;
1514 }
1515
1516 /*
1517  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1518  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1519  *
1520  * Note that if it returns 1, this routine returns without unlocking page
1521  * table locks. So callers must unlock them.
1522  */
1523 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1524 {
1525         spin_lock(&vma->vm_mm->page_table_lock);
1526         if (likely(pmd_trans_huge(*pmd))) {
1527                 if (unlikely(pmd_trans_splitting(*pmd))) {
1528                         spin_unlock(&vma->vm_mm->page_table_lock);
1529                         wait_split_huge_page(vma->anon_vma, pmd);
1530                         return -1;
1531                 } else {
1532                         /* Thp mapped by 'pmd' is stable, so we can
1533                          * handle it as it is. */
1534                         return 1;
1535                 }
1536         }
1537         spin_unlock(&vma->vm_mm->page_table_lock);
1538         return 0;
1539 }
1540
1541 pmd_t *page_check_address_pmd(struct page *page,
1542                               struct mm_struct *mm,
1543                               unsigned long address,
1544                               enum page_check_address_pmd_flag flag)
1545 {
1546         pmd_t *pmd, *ret = NULL;
1547
1548         if (address & ~HPAGE_PMD_MASK)
1549                 goto out;
1550
1551         pmd = mm_find_pmd(mm, address);
1552         if (!pmd)
1553                 goto out;
1554         if (pmd_none(*pmd))
1555                 goto out;
1556         if (pmd_page(*pmd) != page)
1557                 goto out;
1558         /*
1559          * split_vma() may create temporary aliased mappings. There is
1560          * no risk as long as all huge pmd are found and have their
1561          * splitting bit set before __split_huge_page_refcount
1562          * runs. Finding the same huge pmd more than once during the
1563          * same rmap walk is not a problem.
1564          */
1565         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1566             pmd_trans_splitting(*pmd))
1567                 goto out;
1568         if (pmd_trans_huge(*pmd)) {
1569                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1570                           !pmd_trans_splitting(*pmd));
1571                 ret = pmd;
1572         }
1573 out:
1574         return ret;
1575 }
1576
1577 static int __split_huge_page_splitting(struct page *page,
1578                                        struct vm_area_struct *vma,
1579                                        unsigned long address)
1580 {
1581         struct mm_struct *mm = vma->vm_mm;
1582         pmd_t *pmd;
1583         int ret = 0;
1584         /* For mmu_notifiers */
1585         const unsigned long mmun_start = address;
1586         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1587
1588         mmu_notifier_invalidate_range_start(vma, mmun_start,
1589                                             mmun_end, MMU_STATUS);
1590         spin_lock(&mm->page_table_lock);
1591         pmd = page_check_address_pmd(page, mm, address,
1592                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1593         if (pmd) {
1594                 /*
1595                  * We can't temporarily set the pmd to null in order
1596                  * to split it, the pmd must remain marked huge at all
1597                  * times or the VM won't take the pmd_trans_huge paths
1598                  * and it won't wait on the anon_vma->root->rwsem to
1599                  * serialize against split_huge_page*.
1600                  */
1601                 pmdp_splitting_flush(vma, address, pmd);
1602
1603                 ret = 1;
1604         }
1605         spin_unlock(&mm->page_table_lock);
1606         mmu_notifier_invalidate_range_end(vma, mmun_start,
1607                                           mmun_end, MMU_STATUS);
1608
1609         return ret;
1610 }
1611
1612 static void __split_huge_page_refcount(struct page *page,
1613                                        struct list_head *list)
1614 {
1615         int i;
1616         struct zone *zone = page_zone(page);
1617         struct lruvec *lruvec;
1618         int tail_count = 0;
1619
1620         /* prevent PageLRU to go away from under us, and freeze lru stats */
1621         spin_lock_irq(&zone->lru_lock);
1622         lruvec = mem_cgroup_page_lruvec(page, zone);
1623
1624         compound_lock(page);
1625         /* complete memcg works before add pages to LRU */
1626         mem_cgroup_split_huge_fixup(page);
1627
1628         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1629                 struct page *page_tail = page + i;
1630
1631                 /* tail_page->_mapcount cannot change */
1632                 BUG_ON(page_mapcount(page_tail) < 0);
1633                 tail_count += page_mapcount(page_tail);
1634                 /* check for overflow */
1635                 BUG_ON(tail_count < 0);
1636                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1637                 /*
1638                  * tail_page->_count is zero and not changing from
1639                  * under us. But get_page_unless_zero() may be running
1640                  * from under us on the tail_page. If we used
1641                  * atomic_set() below instead of atomic_add(), we
1642                  * would then run atomic_set() concurrently with
1643                  * get_page_unless_zero(), and atomic_set() is
1644                  * implemented in C not using locked ops. spin_unlock
1645                  * on x86 sometime uses locked ops because of PPro
1646                  * errata 66, 92, so unless somebody can guarantee
1647                  * atomic_set() here would be safe on all archs (and
1648                  * not only on x86), it's safer to use atomic_add().
1649                  */
1650                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1651                            &page_tail->_count);
1652
1653                 /* after clearing PageTail the gup refcount can be released */
1654                 smp_mb();
1655
1656                 /*
1657                  * retain hwpoison flag of the poisoned tail page:
1658                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1659                  *   by the memory-failure.
1660                  */
1661                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1662                 page_tail->flags |= (page->flags &
1663                                      ((1L << PG_referenced) |
1664                                       (1L << PG_swapbacked) |
1665                                       (1L << PG_mlocked) |
1666                                       (1L << PG_uptodate)));
1667                 page_tail->flags |= (1L << PG_dirty);
1668
1669                 /* clear PageTail before overwriting first_page */
1670                 smp_wmb();
1671
1672                 /*
1673                  * __split_huge_page_splitting() already set the
1674                  * splitting bit in all pmd that could map this
1675                  * hugepage, that will ensure no CPU can alter the
1676                  * mapcount on the head page. The mapcount is only
1677                  * accounted in the head page and it has to be
1678                  * transferred to all tail pages in the below code. So
1679                  * for this code to be safe, the split the mapcount
1680                  * can't change. But that doesn't mean userland can't
1681                  * keep changing and reading the page contents while
1682                  * we transfer the mapcount, so the pmd splitting
1683                  * status is achieved setting a reserved bit in the
1684                  * pmd, not by clearing the present bit.
1685                 */
1686                 page_tail->_mapcount = page->_mapcount;
1687
1688                 BUG_ON(page_tail->mapping);
1689                 page_tail->mapping = page->mapping;
1690
1691                 page_tail->index = page->index + i;
1692                 page_nid_xchg_last(page_tail, page_nid_last(page));
1693
1694                 BUG_ON(!PageAnon(page_tail));
1695                 BUG_ON(!PageUptodate(page_tail));
1696                 BUG_ON(!PageDirty(page_tail));
1697                 BUG_ON(!PageSwapBacked(page_tail));
1698
1699                 lru_add_page_tail(page, page_tail, lruvec, list);
1700         }
1701         atomic_sub(tail_count, &page->_count);
1702         BUG_ON(atomic_read(&page->_count) <= 0);
1703
1704         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1705         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1706
1707         ClearPageCompound(page);
1708         compound_unlock(page);
1709         spin_unlock_irq(&zone->lru_lock);
1710
1711         for (i = 1; i < HPAGE_PMD_NR; i++) {
1712                 struct page *page_tail = page + i;
1713                 BUG_ON(page_count(page_tail) <= 0);
1714                 /*
1715                  * Tail pages may be freed if there wasn't any mapping
1716                  * like if add_to_swap() is running on a lru page that
1717                  * had its mapping zapped. And freeing these pages
1718                  * requires taking the lru_lock so we do the put_page
1719                  * of the tail pages after the split is complete.
1720                  */
1721                 put_page(page_tail);
1722         }
1723
1724         /*
1725          * Only the head page (now become a regular page) is required
1726          * to be pinned by the caller.
1727          */
1728         BUG_ON(page_count(page) <= 0);
1729 }
1730
1731 static int __split_huge_page_map(struct page *page,
1732                                  struct vm_area_struct *vma,
1733                                  unsigned long address)
1734 {
1735         struct mm_struct *mm = vma->vm_mm;
1736         pmd_t *pmd, _pmd;
1737         int ret = 0, i;
1738         pgtable_t pgtable;
1739         unsigned long haddr;
1740
1741         spin_lock(&mm->page_table_lock);
1742         pmd = page_check_address_pmd(page, mm, address,
1743                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1744         if (pmd) {
1745                 pgtable = pgtable_trans_huge_withdraw(mm);
1746                 pmd_populate(mm, &_pmd, pgtable);
1747                 if (pmd_write(*pmd))
1748                         BUG_ON(page_mapcount(page) != 1);
1749
1750                 haddr = address;
1751                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1752                         pte_t *pte, entry;
1753                         BUG_ON(PageCompound(page+i));
1754                         /*
1755                          * Note that pmd_numa is not transferred deliberately
1756                          * to avoid any possibility that pte_numa leaks to
1757                          * a PROT_NONE VMA by accident.
1758                          */
1759                         entry = mk_pte(page + i, vma->vm_page_prot);
1760                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1761                         if (!pmd_write(*pmd))
1762                                 entry = pte_wrprotect(entry);
1763                         if (!pmd_young(*pmd))
1764                                 entry = pte_mkold(entry);
1765                         pte = pte_offset_map(&_pmd, haddr);
1766                         BUG_ON(!pte_none(*pte));
1767                         set_pte_at(mm, haddr, pte, entry);
1768                         pte_unmap(pte);
1769                 }
1770
1771                 smp_wmb(); /* make pte visible before pmd */
1772                 /*
1773                  * Up to this point the pmd is present and huge and
1774                  * userland has the whole access to the hugepage
1775                  * during the split (which happens in place). If we
1776                  * overwrite the pmd with the not-huge version
1777                  * pointing to the pte here (which of course we could
1778                  * if all CPUs were bug free), userland could trigger
1779                  * a small page size TLB miss on the small sized TLB
1780                  * while the hugepage TLB entry is still established
1781                  * in the huge TLB. Some CPU doesn't like that. See
1782                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1783                  * Erratum 383 on page 93. Intel should be safe but is
1784                  * also warns that it's only safe if the permission
1785                  * and cache attributes of the two entries loaded in
1786                  * the two TLB is identical (which should be the case
1787                  * here). But it is generally safer to never allow
1788                  * small and huge TLB entries for the same virtual
1789                  * address to be loaded simultaneously. So instead of
1790                  * doing "pmd_populate(); flush_tlb_range();" we first
1791                  * mark the current pmd notpresent (atomically because
1792                  * here the pmd_trans_huge and pmd_trans_splitting
1793                  * must remain set at all times on the pmd until the
1794                  * split is complete for this pmd), then we flush the
1795                  * SMP TLB and finally we write the non-huge version
1796                  * of the pmd entry with pmd_populate.
1797                  */
1798                 pmdp_invalidate(vma, address, pmd);
1799                 pmd_populate(mm, pmd, pgtable);
1800                 ret = 1;
1801         }
1802         spin_unlock(&mm->page_table_lock);
1803
1804         return ret;
1805 }
1806
1807 /* must be called with anon_vma->root->rwsem held */
1808 static void __split_huge_page(struct page *page,
1809                               struct anon_vma *anon_vma,
1810                               struct list_head *list)
1811 {
1812         int mapcount, mapcount2;
1813         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1814         struct anon_vma_chain *avc;
1815
1816         BUG_ON(!PageHead(page));
1817         BUG_ON(PageTail(page));
1818
1819         mapcount = 0;
1820         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1821                 struct vm_area_struct *vma = avc->vma;
1822                 unsigned long addr = vma_address(page, vma);
1823                 BUG_ON(is_vma_temporary_stack(vma));
1824                 mapcount += __split_huge_page_splitting(page, vma, addr);
1825         }
1826         /*
1827          * It is critical that new vmas are added to the tail of the
1828          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1829          * and establishes a child pmd before
1830          * __split_huge_page_splitting() freezes the parent pmd (so if
1831          * we fail to prevent copy_huge_pmd() from running until the
1832          * whole __split_huge_page() is complete), we will still see
1833          * the newly established pmd of the child later during the
1834          * walk, to be able to set it as pmd_trans_splitting too.
1835          */
1836         if (mapcount != page_mapcount(page))
1837                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1838                        mapcount, page_mapcount(page));
1839         BUG_ON(mapcount != page_mapcount(page));
1840
1841         __split_huge_page_refcount(page, list);
1842
1843         mapcount2 = 0;
1844         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1845                 struct vm_area_struct *vma = avc->vma;
1846                 unsigned long addr = vma_address(page, vma);
1847                 BUG_ON(is_vma_temporary_stack(vma));
1848                 mapcount2 += __split_huge_page_map(page, vma, addr);
1849         }
1850         if (mapcount != mapcount2)
1851                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1852                        mapcount, mapcount2, page_mapcount(page));
1853         BUG_ON(mapcount != mapcount2);
1854 }
1855
1856 /*
1857  * Split a hugepage into normal pages. This doesn't change the position of head
1858  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1859  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1860  * from the hugepage.
1861  * Return 0 if the hugepage is split successfully otherwise return 1.
1862  */
1863 int split_huge_page_to_list(struct page *page, struct list_head *list)
1864 {
1865         struct anon_vma *anon_vma;
1866         int ret = 1;
1867
1868         BUG_ON(is_huge_zero_page(page));
1869         BUG_ON(!PageAnon(page));
1870
1871         /*
1872          * The caller does not necessarily hold an mmap_sem that would prevent
1873          * the anon_vma disappearing so we first we take a reference to it
1874          * and then lock the anon_vma for write. This is similar to
1875          * page_lock_anon_vma_read except the write lock is taken to serialise
1876          * against parallel split or collapse operations.
1877          */
1878         anon_vma = page_get_anon_vma(page);
1879         if (!anon_vma)
1880                 goto out;
1881         anon_vma_lock_write(anon_vma);
1882
1883         ret = 0;
1884         if (!PageCompound(page))
1885                 goto out_unlock;
1886
1887         BUG_ON(!PageSwapBacked(page));
1888         __split_huge_page(page, anon_vma, list);
1889         count_vm_event(THP_SPLIT);
1890
1891         BUG_ON(PageCompound(page));
1892 out_unlock:
1893         anon_vma_unlock_write(anon_vma);
1894         put_anon_vma(anon_vma);
1895 out:
1896         return ret;
1897 }
1898
1899 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1900
1901 int hugepage_madvise(struct vm_area_struct *vma,
1902                      unsigned long *vm_flags, int advice)
1903 {
1904         struct mm_struct *mm = vma->vm_mm;
1905
1906         switch (advice) {
1907         case MADV_HUGEPAGE:
1908                 /*
1909                  * Be somewhat over-protective like KSM for now!
1910                  */
1911                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1912                         return -EINVAL;
1913                 if (mm->def_flags & VM_NOHUGEPAGE)
1914                         return -EINVAL;
1915                 *vm_flags &= ~VM_NOHUGEPAGE;
1916                 *vm_flags |= VM_HUGEPAGE;
1917                 /*
1918                  * If the vma become good for khugepaged to scan,
1919                  * register it here without waiting a page fault that
1920                  * may not happen any time soon.
1921                  */
1922                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1923                         return -ENOMEM;
1924                 break;
1925         case MADV_NOHUGEPAGE:
1926                 /*
1927                  * Be somewhat over-protective like KSM for now!
1928                  */
1929                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1930                         return -EINVAL;
1931                 *vm_flags &= ~VM_HUGEPAGE;
1932                 *vm_flags |= VM_NOHUGEPAGE;
1933                 /*
1934                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1935                  * this vma even if we leave the mm registered in khugepaged if
1936                  * it got registered before VM_NOHUGEPAGE was set.
1937                  */
1938                 break;
1939         }
1940
1941         return 0;
1942 }
1943
1944 static int __init khugepaged_slab_init(void)
1945 {
1946         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1947                                           sizeof(struct mm_slot),
1948                                           __alignof__(struct mm_slot), 0, NULL);
1949         if (!mm_slot_cache)
1950                 return -ENOMEM;
1951
1952         return 0;
1953 }
1954
1955 static inline struct mm_slot *alloc_mm_slot(void)
1956 {
1957         if (!mm_slot_cache)     /* initialization failed */
1958                 return NULL;
1959         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1960 }
1961
1962 static inline void free_mm_slot(struct mm_slot *mm_slot)
1963 {
1964         kmem_cache_free(mm_slot_cache, mm_slot);
1965 }
1966
1967 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1968 {
1969         struct mm_slot *mm_slot;
1970
1971         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1972                 if (mm == mm_slot->mm)
1973                         return mm_slot;
1974
1975         return NULL;
1976 }
1977
1978 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1979                                     struct mm_slot *mm_slot)
1980 {
1981         mm_slot->mm = mm;
1982         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1983 }
1984
1985 static inline int khugepaged_test_exit(struct mm_struct *mm)
1986 {
1987         return atomic_read(&mm->mm_users) == 0;
1988 }
1989
1990 int __khugepaged_enter(struct mm_struct *mm)
1991 {
1992         struct mm_slot *mm_slot;
1993         int wakeup;
1994
1995         mm_slot = alloc_mm_slot();
1996         if (!mm_slot)
1997                 return -ENOMEM;
1998
1999         /* __khugepaged_exit() must not run from under us */
2000         VM_BUG_ON(khugepaged_test_exit(mm));
2001         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2002                 free_mm_slot(mm_slot);
2003                 return 0;
2004         }
2005
2006         spin_lock(&khugepaged_mm_lock);
2007         insert_to_mm_slots_hash(mm, mm_slot);
2008         /*
2009          * Insert just behind the scanning cursor, to let the area settle
2010          * down a little.
2011          */
2012         wakeup = list_empty(&khugepaged_scan.mm_head);
2013         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2014         spin_unlock(&khugepaged_mm_lock);
2015
2016         atomic_inc(&mm->mm_count);
2017         if (wakeup)
2018                 wake_up_interruptible(&khugepaged_wait);
2019
2020         return 0;
2021 }
2022
2023 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2024 {
2025         unsigned long hstart, hend;
2026         if (!vma->anon_vma)
2027                 /*
2028                  * Not yet faulted in so we will register later in the
2029                  * page fault if needed.
2030                  */
2031                 return 0;
2032         if (vma->vm_ops)
2033                 /* khugepaged not yet working on file or special mappings */
2034                 return 0;
2035         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2036         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2037         hend = vma->vm_end & HPAGE_PMD_MASK;
2038         if (hstart < hend)
2039                 return khugepaged_enter(vma);
2040         return 0;
2041 }
2042
2043 void __khugepaged_exit(struct mm_struct *mm)
2044 {
2045         struct mm_slot *mm_slot;
2046         int free = 0;
2047
2048         spin_lock(&khugepaged_mm_lock);
2049         mm_slot = get_mm_slot(mm);
2050         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2051                 hash_del(&mm_slot->hash);
2052                 list_del(&mm_slot->mm_node);
2053                 free = 1;
2054         }
2055         spin_unlock(&khugepaged_mm_lock);
2056
2057         if (free) {
2058                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2059                 free_mm_slot(mm_slot);
2060                 mmdrop(mm);
2061         } else if (mm_slot) {
2062                 /*
2063                  * This is required to serialize against
2064                  * khugepaged_test_exit() (which is guaranteed to run
2065                  * under mmap sem read mode). Stop here (after we
2066                  * return all pagetables will be destroyed) until
2067                  * khugepaged has finished working on the pagetables
2068                  * under the mmap_sem.
2069                  */
2070                 down_write(&mm->mmap_sem);
2071                 up_write(&mm->mmap_sem);
2072         }
2073 }
2074
2075 static void release_pte_page(struct page *page)
2076 {
2077         /* 0 stands for page_is_file_cache(page) == false */
2078         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2079         unlock_page(page);
2080         putback_lru_page(page);
2081 }
2082
2083 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2084 {
2085         while (--_pte >= pte) {
2086                 pte_t pteval = *_pte;
2087                 if (!pte_none(pteval))
2088                         release_pte_page(pte_page(pteval));
2089         }
2090 }
2091
2092 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2093                                         unsigned long address,
2094                                         pte_t *pte)
2095 {
2096         struct page *page;
2097         pte_t *_pte;
2098         int referenced = 0, none = 0;
2099         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2100              _pte++, address += PAGE_SIZE) {
2101                 pte_t pteval = *_pte;
2102                 if (pte_none(pteval)) {
2103                         if (++none <= khugepaged_max_ptes_none)
2104                                 continue;
2105                         else
2106                                 goto out;
2107                 }
2108                 if (!pte_present(pteval) || !pte_write(pteval))
2109                         goto out;
2110                 page = vm_normal_page(vma, address, pteval);
2111                 if (unlikely(!page))
2112                         goto out;
2113
2114                 VM_BUG_ON(PageCompound(page));
2115                 BUG_ON(!PageAnon(page));
2116                 VM_BUG_ON(!PageSwapBacked(page));
2117
2118                 /* cannot use mapcount: can't collapse if there's a gup pin */
2119                 if (page_count(page) != 1)
2120                         goto out;
2121                 /*
2122                  * We can do it before isolate_lru_page because the
2123                  * page can't be freed from under us. NOTE: PG_lock
2124                  * is needed to serialize against split_huge_page
2125                  * when invoked from the VM.
2126                  */
2127                 if (!trylock_page(page))
2128                         goto out;
2129                 /*
2130                  * Isolate the page to avoid collapsing an hugepage
2131                  * currently in use by the VM.
2132                  */
2133                 if (isolate_lru_page(page)) {
2134                         unlock_page(page);
2135                         goto out;
2136                 }
2137                 /* 0 stands for page_is_file_cache(page) == false */
2138                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2139                 VM_BUG_ON(!PageLocked(page));
2140                 VM_BUG_ON(PageLRU(page));
2141
2142                 /* If there is no mapped pte young don't collapse the page */
2143                 if (pte_young(pteval) || PageReferenced(page) ||
2144                     mmu_notifier_test_young(vma->vm_mm, address))
2145                         referenced = 1;
2146         }
2147         if (likely(referenced))
2148                 return 1;
2149 out:
2150         release_pte_pages(pte, _pte);
2151         return 0;
2152 }
2153
2154 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2155                                       struct vm_area_struct *vma,
2156                                       unsigned long address,
2157                                       spinlock_t *ptl)
2158 {
2159         pte_t *_pte;
2160         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2161                 pte_t pteval = *_pte;
2162                 struct page *src_page;
2163
2164                 if (pte_none(pteval)) {
2165                         clear_user_highpage(page, address);
2166                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2167                 } else {
2168                         src_page = pte_page(pteval);
2169                         copy_user_highpage(page, src_page, address, vma);
2170                         VM_BUG_ON(page_mapcount(src_page) != 1);
2171                         release_pte_page(src_page);
2172                         /*
2173                          * ptl mostly unnecessary, but preempt has to
2174                          * be disabled to update the per-cpu stats
2175                          * inside page_remove_rmap().
2176                          */
2177                         spin_lock(ptl);
2178                         /*
2179                          * paravirt calls inside pte_clear here are
2180                          * superfluous.
2181                          */
2182                         pte_clear(vma->vm_mm, address, _pte);
2183                         page_remove_rmap(src_page);
2184                         spin_unlock(ptl);
2185                         free_page_and_swap_cache(src_page);
2186                 }
2187
2188                 address += PAGE_SIZE;
2189                 page++;
2190         }
2191 }
2192
2193 static void khugepaged_alloc_sleep(void)
2194 {
2195         wait_event_freezable_timeout(khugepaged_wait, false,
2196                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2197 }
2198
2199 #ifdef CONFIG_NUMA
2200 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2201 {
2202         if (IS_ERR(*hpage)) {
2203                 if (!*wait)
2204                         return false;
2205
2206                 *wait = false;
2207                 *hpage = NULL;
2208                 khugepaged_alloc_sleep();
2209         } else if (*hpage) {
2210                 put_page(*hpage);
2211                 *hpage = NULL;
2212         }
2213
2214         return true;
2215 }
2216
2217 static struct page
2218 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2219                        struct vm_area_struct *vma, unsigned long address,
2220                        int node)
2221 {
2222         VM_BUG_ON(*hpage);
2223         /*
2224          * Allocate the page while the vma is still valid and under
2225          * the mmap_sem read mode so there is no memory allocation
2226          * later when we take the mmap_sem in write mode. This is more
2227          * friendly behavior (OTOH it may actually hide bugs) to
2228          * filesystems in userland with daemons allocating memory in
2229          * the userland I/O paths.  Allocating memory with the
2230          * mmap_sem in read mode is good idea also to allow greater
2231          * scalability.
2232          */
2233         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2234                                       node, __GFP_OTHER_NODE);
2235
2236         /*
2237          * After allocating the hugepage, release the mmap_sem read lock in
2238          * preparation for taking it in write mode.
2239          */
2240         up_read(&mm->mmap_sem);
2241         if (unlikely(!*hpage)) {
2242                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2243                 *hpage = ERR_PTR(-ENOMEM);
2244                 return NULL;
2245         }
2246
2247         count_vm_event(THP_COLLAPSE_ALLOC);
2248         return *hpage;
2249 }
2250 #else
2251 static struct page *khugepaged_alloc_hugepage(bool *wait)
2252 {
2253         struct page *hpage;
2254
2255         do {
2256                 hpage = alloc_hugepage(khugepaged_defrag());
2257                 if (!hpage) {
2258                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2259                         if (!*wait)
2260                                 return NULL;
2261
2262                         *wait = false;
2263                         khugepaged_alloc_sleep();
2264                 } else
2265                         count_vm_event(THP_COLLAPSE_ALLOC);
2266         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2267
2268         return hpage;
2269 }
2270
2271 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2272 {
2273         if (!*hpage)
2274                 *hpage = khugepaged_alloc_hugepage(wait);
2275
2276         if (unlikely(!*hpage))
2277                 return false;
2278
2279         return true;
2280 }
2281
2282 static struct page
2283 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2284                        struct vm_area_struct *vma, unsigned long address,
2285                        int node)
2286 {
2287         up_read(&mm->mmap_sem);
2288         VM_BUG_ON(!*hpage);
2289         return  *hpage;
2290 }
2291 #endif
2292
2293 static bool hugepage_vma_check(struct vm_area_struct *vma)
2294 {
2295         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2296             (vma->vm_flags & VM_NOHUGEPAGE))
2297                 return false;
2298
2299         if (!vma->anon_vma || vma->vm_ops)
2300                 return false;
2301         if (is_vma_temporary_stack(vma))
2302                 return false;
2303         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2304         return true;
2305 }
2306
2307 static void collapse_huge_page(struct mm_struct *mm,
2308                                    unsigned long address,
2309                                    struct page **hpage,
2310                                    struct vm_area_struct *vma,
2311                                    int node)
2312 {
2313         pmd_t *pmd, _pmd;
2314         pte_t *pte;
2315         pgtable_t pgtable;
2316         struct page *new_page;
2317         spinlock_t *ptl;
2318         int isolated;
2319         unsigned long hstart, hend;
2320         unsigned long mmun_start;       /* For mmu_notifiers */
2321         unsigned long mmun_end;         /* For mmu_notifiers */
2322
2323         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2324
2325         /* release the mmap_sem read lock. */
2326         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2327         if (!new_page)
2328                 return;
2329
2330         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2331                 return;
2332
2333         /*
2334          * Prevent all access to pagetables with the exception of
2335          * gup_fast later hanlded by the ptep_clear_flush and the VM
2336          * handled by the anon_vma lock + PG_lock.
2337          */
2338         down_write(&mm->mmap_sem);
2339         if (unlikely(khugepaged_test_exit(mm)))
2340                 goto out;
2341
2342         vma = find_vma(mm, address);
2343         if (!vma)
2344                 goto out;
2345         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2346         hend = vma->vm_end & HPAGE_PMD_MASK;
2347         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2348                 goto out;
2349         if (!hugepage_vma_check(vma))
2350                 goto out;
2351         pmd = mm_find_pmd(mm, address);
2352         if (!pmd)
2353                 goto out;
2354         if (pmd_trans_huge(*pmd))
2355                 goto out;
2356
2357         anon_vma_lock_write(vma->anon_vma);
2358
2359         pte = pte_offset_map(pmd, address);
2360         ptl = pte_lockptr(mm, pmd);
2361
2362         mmun_start = address;
2363         mmun_end   = address + HPAGE_PMD_SIZE;
2364         mmu_notifier_invalidate_range_start(vma, mmun_start,
2365                                             mmun_end, MMU_MIGRATE);
2366         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2367         /*
2368          * After this gup_fast can't run anymore. This also removes
2369          * any huge TLB entry from the CPU so we won't allow
2370          * huge and small TLB entries for the same virtual address
2371          * to avoid the risk of CPU bugs in that area.
2372          */
2373         _pmd = pmdp_clear_flush(vma, address, pmd);
2374         spin_unlock(&mm->page_table_lock);
2375         mmu_notifier_invalidate_range_end(vma, mmun_start,
2376                                           mmun_end, MMU_MIGRATE);
2377
2378         spin_lock(ptl);
2379         isolated = __collapse_huge_page_isolate(vma, address, pte);
2380         spin_unlock(ptl);
2381
2382         if (unlikely(!isolated)) {
2383                 pte_unmap(pte);
2384                 spin_lock(&mm->page_table_lock);
2385                 BUG_ON(!pmd_none(*pmd));
2386                 /*
2387                  * We can only use set_pmd_at when establishing
2388                  * hugepmds and never for establishing regular pmds that
2389                  * points to regular pagetables. Use pmd_populate for that
2390                  */
2391                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2392                 spin_unlock(&mm->page_table_lock);
2393                 anon_vma_unlock_write(vma->anon_vma);
2394                 goto out;
2395         }
2396
2397         /*
2398          * All pages are isolated and locked so anon_vma rmap
2399          * can't run anymore.
2400          */
2401         anon_vma_unlock_write(vma->anon_vma);
2402
2403         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2404         pte_unmap(pte);
2405         __SetPageUptodate(new_page);
2406         pgtable = pmd_pgtable(_pmd);
2407
2408         _pmd = mk_huge_pmd(new_page, vma);
2409
2410         /*
2411          * spin_lock() below is not the equivalent of smp_wmb(), so
2412          * this is needed to avoid the copy_huge_page writes to become
2413          * visible after the set_pmd_at() write.
2414          */
2415         smp_wmb();
2416
2417         spin_lock(&mm->page_table_lock);
2418         BUG_ON(!pmd_none(*pmd));
2419         page_add_new_anon_rmap(new_page, vma, address);
2420         set_pmd_at(mm, address, pmd, _pmd);
2421         update_mmu_cache_pmd(vma, address, pmd);
2422         pgtable_trans_huge_deposit(mm, pgtable);
2423         spin_unlock(&mm->page_table_lock);
2424
2425         *hpage = NULL;
2426
2427         khugepaged_pages_collapsed++;
2428 out_up_write:
2429         up_write(&mm->mmap_sem);
2430         return;
2431
2432 out:
2433         mem_cgroup_uncharge_page(new_page);
2434         goto out_up_write;
2435 }
2436
2437 static int khugepaged_scan_pmd(struct mm_struct *mm,
2438                                struct vm_area_struct *vma,
2439                                unsigned long address,
2440                                struct page **hpage)
2441 {
2442         pmd_t *pmd;
2443         pte_t *pte, *_pte;
2444         int ret = 0, referenced = 0, none = 0;
2445         struct page *page;
2446         unsigned long _address;
2447         spinlock_t *ptl;
2448         int node = NUMA_NO_NODE;
2449
2450         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2451
2452         pmd = mm_find_pmd(mm, address);
2453         if (!pmd)
2454                 goto out;
2455         if (pmd_trans_huge(*pmd))
2456                 goto out;
2457
2458         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2459         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2460              _pte++, _address += PAGE_SIZE) {
2461                 pte_t pteval = *_pte;
2462                 if (pte_none(pteval)) {
2463                         if (++none <= khugepaged_max_ptes_none)
2464                                 continue;
2465                         else
2466                                 goto out_unmap;
2467                 }
2468                 if (!pte_present(pteval) || !pte_write(pteval))
2469                         goto out_unmap;
2470                 page = vm_normal_page(vma, _address, pteval);
2471                 if (unlikely(!page))
2472                         goto out_unmap;
2473                 /*
2474                  * Chose the node of the first page. This could
2475                  * be more sophisticated and look at more pages,
2476                  * but isn't for now.
2477                  */
2478                 if (node == NUMA_NO_NODE)
2479                         node = page_to_nid(page);
2480                 VM_BUG_ON(PageCompound(page));
2481                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2482                         goto out_unmap;
2483                 /* cannot use mapcount: can't collapse if there's a gup pin */
2484                 if (page_count(page) != 1)
2485                         goto out_unmap;
2486                 if (pte_young(pteval) || PageReferenced(page) ||
2487                     mmu_notifier_test_young(vma->vm_mm, address))
2488                         referenced = 1;
2489         }
2490         if (referenced)
2491                 ret = 1;
2492 out_unmap:
2493         pte_unmap_unlock(pte, ptl);
2494         if (ret)
2495                 /* collapse_huge_page will return with the mmap_sem released */
2496                 collapse_huge_page(mm, address, hpage, vma, node);
2497 out:
2498         return ret;
2499 }
2500
2501 static void collect_mm_slot(struct mm_slot *mm_slot)
2502 {
2503         struct mm_struct *mm = mm_slot->mm;
2504
2505         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2506
2507         if (khugepaged_test_exit(mm)) {
2508                 /* free mm_slot */
2509                 hash_del(&mm_slot->hash);
2510                 list_del(&mm_slot->mm_node);
2511
2512                 /*
2513                  * Not strictly needed because the mm exited already.
2514                  *
2515                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2516                  */
2517
2518                 /* khugepaged_mm_lock actually not necessary for the below */
2519                 free_mm_slot(mm_slot);
2520                 mmdrop(mm);
2521         }
2522 }
2523
2524 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2525                                             struct page **hpage)
2526         __releases(&khugepaged_mm_lock)
2527         __acquires(&khugepaged_mm_lock)
2528 {
2529         struct mm_slot *mm_slot;
2530         struct mm_struct *mm;
2531         struct vm_area_struct *vma;
2532         int progress = 0;
2533
2534         VM_BUG_ON(!pages);
2535         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2536
2537         if (khugepaged_scan.mm_slot)
2538                 mm_slot = khugepaged_scan.mm_slot;
2539         else {
2540                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2541                                      struct mm_slot, mm_node);
2542                 khugepaged_scan.address = 0;
2543                 khugepaged_scan.mm_slot = mm_slot;
2544         }
2545         spin_unlock(&khugepaged_mm_lock);
2546
2547         mm = mm_slot->mm;
2548         down_read(&mm->mmap_sem);
2549         if (unlikely(khugepaged_test_exit(mm)))
2550                 vma = NULL;
2551         else
2552                 vma = find_vma(mm, khugepaged_scan.address);
2553
2554         progress++;
2555         for (; vma; vma = vma->vm_next) {
2556                 unsigned long hstart, hend;
2557
2558                 cond_resched();
2559                 if (unlikely(khugepaged_test_exit(mm))) {
2560                         progress++;
2561                         break;
2562                 }
2563                 if (!hugepage_vma_check(vma)) {
2564 skip:
2565                         progress++;
2566                         continue;
2567                 }
2568                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2569                 hend = vma->vm_end & HPAGE_PMD_MASK;
2570                 if (hstart >= hend)
2571                         goto skip;
2572                 if (khugepaged_scan.address > hend)
2573                         goto skip;
2574                 if (khugepaged_scan.address < hstart)
2575                         khugepaged_scan.address = hstart;
2576                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2577
2578                 while (khugepaged_scan.address < hend) {
2579                         int ret;
2580                         cond_resched();
2581                         if (unlikely(khugepaged_test_exit(mm)))
2582                                 goto breakouterloop;
2583
2584                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2585                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2586                                   hend);
2587                         ret = khugepaged_scan_pmd(mm, vma,
2588                                                   khugepaged_scan.address,
2589                                                   hpage);
2590                         /* move to next address */
2591                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2592                         progress += HPAGE_PMD_NR;
2593                         if (ret)
2594                                 /* we released mmap_sem so break loop */
2595                                 goto breakouterloop_mmap_sem;
2596                         if (progress >= pages)
2597                                 goto breakouterloop;
2598                 }
2599         }
2600 breakouterloop:
2601         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2602 breakouterloop_mmap_sem:
2603
2604         spin_lock(&khugepaged_mm_lock);
2605         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2606         /*
2607          * Release the current mm_slot if this mm is about to die, or
2608          * if we scanned all vmas of this mm.
2609          */
2610         if (khugepaged_test_exit(mm) || !vma) {
2611                 /*
2612                  * Make sure that if mm_users is reaching zero while
2613                  * khugepaged runs here, khugepaged_exit will find
2614                  * mm_slot not pointing to the exiting mm.
2615                  */
2616                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2617                         khugepaged_scan.mm_slot = list_entry(
2618                                 mm_slot->mm_node.next,
2619                                 struct mm_slot, mm_node);
2620                         khugepaged_scan.address = 0;
2621                 } else {
2622                         khugepaged_scan.mm_slot = NULL;
2623                         khugepaged_full_scans++;
2624                 }
2625
2626                 collect_mm_slot(mm_slot);
2627         }
2628
2629         return progress;
2630 }
2631
2632 static int khugepaged_has_work(void)
2633 {
2634         return !list_empty(&khugepaged_scan.mm_head) &&
2635                 khugepaged_enabled();
2636 }
2637
2638 static int khugepaged_wait_event(void)
2639 {
2640         return !list_empty(&khugepaged_scan.mm_head) ||
2641                 kthread_should_stop();
2642 }
2643
2644 static void khugepaged_do_scan(void)
2645 {
2646         struct page *hpage = NULL;
2647         unsigned int progress = 0, pass_through_head = 0;
2648         unsigned int pages = khugepaged_pages_to_scan;
2649         bool wait = true;
2650
2651         barrier(); /* write khugepaged_pages_to_scan to local stack */
2652
2653         while (progress < pages) {
2654                 if (!khugepaged_prealloc_page(&hpage, &wait))
2655                         break;
2656
2657                 cond_resched();
2658
2659                 if (unlikely(kthread_should_stop() || freezing(current)))
2660                         break;
2661
2662                 spin_lock(&khugepaged_mm_lock);
2663                 if (!khugepaged_scan.mm_slot)
2664                         pass_through_head++;
2665                 if (khugepaged_has_work() &&
2666                     pass_through_head < 2)
2667                         progress += khugepaged_scan_mm_slot(pages - progress,
2668                                                             &hpage);
2669                 else
2670                         progress = pages;
2671                 spin_unlock(&khugepaged_mm_lock);
2672         }
2673
2674         if (!IS_ERR_OR_NULL(hpage))
2675                 put_page(hpage);
2676 }
2677
2678 static void khugepaged_wait_work(void)
2679 {
2680         try_to_freeze();
2681
2682         if (khugepaged_has_work()) {
2683                 if (!khugepaged_scan_sleep_millisecs)
2684                         return;
2685
2686                 wait_event_freezable_timeout(khugepaged_wait,
2687                                              kthread_should_stop(),
2688                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2689                 return;
2690         }
2691
2692         if (khugepaged_enabled())
2693                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2694 }
2695
2696 static int khugepaged(void *none)
2697 {
2698         struct mm_slot *mm_slot;
2699
2700         set_freezable();
2701         set_user_nice(current, 19);
2702
2703         while (!kthread_should_stop()) {
2704                 khugepaged_do_scan();
2705                 khugepaged_wait_work();
2706         }
2707
2708         spin_lock(&khugepaged_mm_lock);
2709         mm_slot = khugepaged_scan.mm_slot;
2710         khugepaged_scan.mm_slot = NULL;
2711         if (mm_slot)
2712                 collect_mm_slot(mm_slot);
2713         spin_unlock(&khugepaged_mm_lock);
2714         return 0;
2715 }
2716
2717 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2718                 unsigned long haddr, pmd_t *pmd)
2719 {
2720         struct mm_struct *mm = vma->vm_mm;
2721         pgtable_t pgtable;
2722         pmd_t _pmd;
2723         int i;
2724
2725         pmdp_clear_flush_notify(vma, haddr, pmd);
2726         /* leave pmd empty until pte is filled */
2727
2728         pgtable = pgtable_trans_huge_withdraw(mm);
2729         pmd_populate(mm, &_pmd, pgtable);
2730
2731         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2732                 pte_t *pte, entry;
2733                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2734                 entry = pte_mkspecial(entry);
2735                 pte = pte_offset_map(&_pmd, haddr);
2736                 VM_BUG_ON(!pte_none(*pte));
2737                 set_pte_at(mm, haddr, pte, entry);
2738                 pte_unmap(pte);
2739         }
2740         smp_wmb(); /* make pte visible before pmd */
2741         pmd_populate(mm, pmd, pgtable);
2742         put_huge_zero_page();
2743 }
2744
2745 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2746                 pmd_t *pmd)
2747 {
2748         struct page *page;
2749         struct mm_struct *mm = vma->vm_mm;
2750         unsigned long haddr = address & HPAGE_PMD_MASK;
2751         unsigned long mmun_start;       /* For mmu_notifiers */
2752         unsigned long mmun_end;         /* For mmu_notifiers */
2753
2754         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2755
2756         mmun_start = haddr;
2757         mmun_end   = haddr + HPAGE_PMD_SIZE;
2758 again:
2759         mmu_notifier_invalidate_range_start(vma, mmun_start,
2760                                             mmun_end, MMU_MIGRATE);
2761         spin_lock(&mm->page_table_lock);
2762         if (unlikely(!pmd_trans_huge(*pmd))) {
2763                 spin_unlock(&mm->page_table_lock);
2764                 mmu_notifier_invalidate_range_end(vma, mmun_start,
2765                                                   mmun_end, MMU_MIGRATE);
2766                 return;
2767         }
2768         if (is_huge_zero_pmd(*pmd)) {
2769                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2770                 spin_unlock(&mm->page_table_lock);
2771                 mmu_notifier_invalidate_range_end(vma, mmun_start,
2772                                                   mmun_end, MMU_MIGRATE);
2773                 return;
2774         }
2775         page = pmd_page(*pmd);
2776         VM_BUG_ON(!page_count(page));
2777         get_page(page);
2778         spin_unlock(&mm->page_table_lock);
2779         mmu_notifier_invalidate_range_end(vma, mmun_start,
2780                                           mmun_end, MMU_MIGRATE);
2781
2782         split_huge_page(page);
2783
2784         put_page(page);
2785
2786         /*
2787          * We don't always have down_write of mmap_sem here: a racing
2788          * do_huge_pmd_wp_page() might have copied-on-write to another
2789          * huge page before our split_huge_page() got the anon_vma lock.
2790          */
2791         if (unlikely(pmd_trans_huge(*pmd)))
2792                 goto again;
2793 }
2794
2795 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2796                 pmd_t *pmd)
2797 {
2798         struct vm_area_struct *vma;
2799
2800         vma = find_vma(mm, address);
2801         BUG_ON(vma == NULL);
2802         split_huge_page_pmd(vma, address, pmd);
2803 }
2804
2805 static void split_huge_page_address(struct mm_struct *mm,
2806                                     unsigned long address)
2807 {
2808         pmd_t *pmd;
2809
2810         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2811
2812         pmd = mm_find_pmd(mm, address);
2813         if (!pmd)
2814                 return;
2815         /*
2816          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2817          * materialize from under us.
2818          */
2819         split_huge_page_pmd_mm(mm, address, pmd);
2820 }
2821
2822 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2823                              unsigned long start,
2824                              unsigned long end,
2825                              long adjust_next)
2826 {
2827         /*
2828          * If the new start address isn't hpage aligned and it could
2829          * previously contain an hugepage: check if we need to split
2830          * an huge pmd.
2831          */
2832         if (start & ~HPAGE_PMD_MASK &&
2833             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2834             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2835                 split_huge_page_address(vma->vm_mm, start);
2836
2837         /*
2838          * If the new end address isn't hpage aligned and it could
2839          * previously contain an hugepage: check if we need to split
2840          * an huge pmd.
2841          */
2842         if (end & ~HPAGE_PMD_MASK &&
2843             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2844             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2845                 split_huge_page_address(vma->vm_mm, end);
2846
2847         /*
2848          * If we're also updating the vma->vm_next->vm_start, if the new
2849          * vm_next->vm_start isn't page aligned and it could previously
2850          * contain an hugepage: check if we need to split an huge pmd.
2851          */
2852         if (adjust_next > 0) {
2853                 struct vm_area_struct *next = vma->vm_next;
2854                 unsigned long nstart = next->vm_start;
2855                 nstart += adjust_next << PAGE_SHIFT;
2856                 if (nstart & ~HPAGE_PMD_MASK &&
2857                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2858                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2859                         split_huge_page_address(next->vm_mm, nstart);
2860         }
2861 }