mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is enabled for all mappings
  31  * and khugepaged scans all mappings. Defrag is only invoked by
  32  * khugepaged hugepage allocations and by page faults inside
  33  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  34  * allocations.
  35  */
  36 unsigned long transparent_hugepage_flags __read_mostly =
  37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  38         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  39 #endif
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  41         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  42 #endif
  43         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  46
  47 /* default scan 8*512 pte (or vmas) every 30 second */
  48 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  49 static unsigned int khugepaged_pages_collapsed;
  50 static unsigned int khugepaged_full_scans;
  51 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  52 /* during fragmentation poll the hugepage allocator once every minute */
  53 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  54 static struct task_struct *khugepaged_thread __read_mostly;
  55 static DEFINE_MUTEX(khugepaged_mutex);
  56 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  57 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  58 /*
  59  * default collapse hugepages if there is at least one pte mapped like
  60  * it would have happened if the vma was large enough during page
  61  * fault.
  62  */
  63 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  64
  65 static int khugepaged(void *none);
  66 static int khugepaged_slab_init(void);
  67
  68 #define MM_SLOTS_HASH_BITS 10
  69 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  70
  71 static struct kmem_cache *mm_slot_cache __read_mostly;
  72
  73 /**
  74  * struct mm_slot - hash lookup from mm to mm_slot
  75  * @hash: hash collision list
  76  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  77  * @mm: the mm that this information is valid for
  78  */
  79 struct mm_slot {
  80         struct hlist_node hash;
  81         struct list_head mm_node;
  82         struct mm_struct *mm;
  83 };
  84
  85 /**
  86  * struct khugepaged_scan - cursor for scanning
  87  * @mm_head: the head of the mm list to scan
  88  * @mm_slot: the current mm_slot we are scanning
  89  * @address: the next address inside that to be scanned
  90  *
  91  * There is only the one khugepaged_scan instance of this cursor structure.
  92  */
  93 struct khugepaged_scan {
  94         struct list_head mm_head;
  95         struct mm_slot *mm_slot;
  96         unsigned long address;
  97 };
  98 static struct khugepaged_scan khugepaged_scan = {
  99         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 100 };
 101
 102
 103 static int set_recommended_min_free_kbytes(void)
 104 {
 105         struct zone *zone;
 106         int nr_zones = 0;
 107         unsigned long recommended_min;
 108
 109         if (!khugepaged_enabled())
 110                 return 0;
 111
 112         for_each_populated_zone(zone)
 113                 nr_zones++;
 114
 115         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 116         recommended_min = pageblock_nr_pages * nr_zones * 2;
 117
 118         /*
 119          * Make sure that on average at least two pageblocks are almost free
 120          * of another type, one for a migratetype to fall back to and a
 121          * second to avoid subsequent fallbacks of other types There are 3
 122          * MIGRATE_TYPES we care about.
 123          */
 124         recommended_min += pageblock_nr_pages * nr_zones *
 125                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 126
 127         /* don't ever allow to reserve more than 5% of the lowmem */
 128         recommended_min = min(recommended_min,
 129                               (unsigned long) nr_free_buffer_pages() / 20);
 130         recommended_min <<= (PAGE_SHIFT-10);
 131
 132         if (recommended_min > min_free_kbytes)
 133                 min_free_kbytes = recommended_min;
 134         setup_per_zone_wmarks();
 135         return 0;
 136 }
 137 late_initcall(set_recommended_min_free_kbytes);
 138
 139 static int start_khugepaged(void)
 140 {
 141         int err = 0;
 142         if (khugepaged_enabled()) {
 143                 if (!khugepaged_thread)
 144                         khugepaged_thread = kthread_run(khugepaged, NULL,
 145                                                         "khugepaged");
 146                 if (unlikely(IS_ERR(khugepaged_thread))) {
 147                         printk(KERN_ERR
 148                                "khugepaged: kthread_run(khugepaged) failed\n");
 149                         err = PTR_ERR(khugepaged_thread);
 150                         khugepaged_thread = NULL;
 151                 }
 152
 153                 if (!list_empty(&khugepaged_scan.mm_head))
 154                         wake_up_interruptible(&khugepaged_wait);
 155
 156                 set_recommended_min_free_kbytes();
 157         } else if (khugepaged_thread) {
 158                 kthread_stop(khugepaged_thread);
 159                 khugepaged_thread = NULL;
 160         }
 161
 162         return err;
 163 }
 164
 165 static atomic_t huge_zero_refcount;
 166 static struct page *huge_zero_page __read_mostly;
 167
 168 static inline bool is_huge_zero_page(struct page *page)
 169 {
 170         return ACCESS_ONCE(huge_zero_page) == page;
 171 }
 172
 173 static inline bool is_huge_zero_pmd(pmd_t pmd)
 174 {
 175         return is_huge_zero_page(pmd_page(pmd));
 176 }
 177
 178 static struct page *get_huge_zero_page(void)
 179 {
 180         struct page *zero_page;
 181 retry:
 182         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 183                 return ACCESS_ONCE(huge_zero_page);
 184
 185         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 186                         HPAGE_PMD_ORDER);
 187         if (!zero_page) {
 188                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 189                 return NULL;
 190         }
 191         count_vm_event(THP_ZERO_PAGE_ALLOC);
 192         preempt_disable();
 193         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 194                 preempt_enable();
 195                 __free_page(zero_page);
 196                 goto retry;
 197         }
 198
 199         /* We take additional reference here. It will be put back by shrinker */
 200         atomic_set(&huge_zero_refcount, 2);
 201         preempt_enable();
 202         return ACCESS_ONCE(huge_zero_page);
 203 }
 204
 205 static void put_huge_zero_page(void)
 206 {
 207         /*
 208          * Counter should never go to zero here. Only shrinker can put
 209          * last reference.
 210          */
 211         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 212 }
 213
 214 static int shrink_huge_zero_page(struct shrinker *shrink,
 215                 struct shrink_control *sc)
 216 {
 217         if (!sc->nr_to_scan)
 218                 /* we can free zero page only if last reference remains */
 219                 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 220
 221         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 222                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 223                 BUG_ON(zero_page == NULL);
 224                 __free_page(zero_page);
 225         }
 226
 227         return 0;
 228 }
 229
 230 static struct shrinker huge_zero_page_shrinker = {
 231         .shrink = shrink_huge_zero_page,
 232         .seeks = DEFAULT_SEEKS,
 233 };
 234
 235 #ifdef CONFIG_SYSFS
 236
 237 static ssize_t double_flag_show(struct kobject *kobj,
 238                                 struct kobj_attribute *attr, char *buf,
 239                                 enum transparent_hugepage_flag enabled,
 240                                 enum transparent_hugepage_flag req_madv)
 241 {
 242         if (test_bit(enabled, &transparent_hugepage_flags)) {
 243                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 244                 return sprintf(buf, "[always] madvise never\n");
 245         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 246                 return sprintf(buf, "always [madvise] never\n");
 247         else
 248                 return sprintf(buf, "always madvise [never]\n");
 249 }
 250 static ssize_t double_flag_store(struct kobject *kobj,
 251                                  struct kobj_attribute *attr,
 252                                  const char *buf, size_t count,
 253                                  enum transparent_hugepage_flag enabled,
 254                                  enum transparent_hugepage_flag req_madv)
 255 {
 256         if (!memcmp("always", buf,
 257                     min(sizeof("always")-1, count))) {
 258                 set_bit(enabled, &transparent_hugepage_flags);
 259                 clear_bit(req_madv, &transparent_hugepage_flags);
 260         } else if (!memcmp("madvise", buf,
 261                            min(sizeof("madvise")-1, count))) {
 262                 clear_bit(enabled, &transparent_hugepage_flags);
 263                 set_bit(req_madv, &transparent_hugepage_flags);
 264         } else if (!memcmp("never", buf,
 265                            min(sizeof("never")-1, count))) {
 266                 clear_bit(enabled, &transparent_hugepage_flags);
 267                 clear_bit(req_madv, &transparent_hugepage_flags);
 268         } else
 269                 return -EINVAL;
 270
 271         return count;
 272 }
 273
 274 static ssize_t enabled_show(struct kobject *kobj,
 275                             struct kobj_attribute *attr, char *buf)
 276 {
 277         return double_flag_show(kobj, attr, buf,
 278                                 TRANSPARENT_HUGEPAGE_FLAG,
 279                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 280 }
 281 static ssize_t enabled_store(struct kobject *kobj,
 282                              struct kobj_attribute *attr,
 283                              const char *buf, size_t count)
 284 {
 285         ssize_t ret;
 286
 287         ret = double_flag_store(kobj, attr, buf, count,
 288                                 TRANSPARENT_HUGEPAGE_FLAG,
 289                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 290
 291         if (ret > 0) {
 292                 int err;
 293
 294                 mutex_lock(&khugepaged_mutex);
 295                 err = start_khugepaged();
 296                 mutex_unlock(&khugepaged_mutex);
 297
 298                 if (err)
 299                         ret = err;
 300         }
 301
 302         return ret;
 303 }
 304 static struct kobj_attribute enabled_attr =
 305         __ATTR(enabled, 0644, enabled_show, enabled_store);
 306
 307 static ssize_t single_flag_show(struct kobject *kobj,
 308                                 struct kobj_attribute *attr, char *buf,
 309                                 enum transparent_hugepage_flag flag)
 310 {
 311         return sprintf(buf, "%d\n",
 312                        !!test_bit(flag, &transparent_hugepage_flags));
 313 }
 314
 315 static ssize_t single_flag_store(struct kobject *kobj,
 316                                  struct kobj_attribute *attr,
 317                                  const char *buf, size_t count,
 318                                  enum transparent_hugepage_flag flag)
 319 {
 320         unsigned long value;
 321         int ret;
 322
 323         ret = kstrtoul(buf, 10, &value);
 324         if (ret < 0)
 325                 return ret;
 326         if (value > 1)
 327                 return -EINVAL;
 328
 329         if (value)
 330                 set_bit(flag, &transparent_hugepage_flags);
 331         else
 332                 clear_bit(flag, &transparent_hugepage_flags);
 333
 334         return count;
 335 }
 336
 337 /*
 338  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 339  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 340  * memory just to allocate one more hugepage.
 341  */
 342 static ssize_t defrag_show(struct kobject *kobj,
 343                            struct kobj_attribute *attr, char *buf)
 344 {
 345         return double_flag_show(kobj, attr, buf,
 346                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 347                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 348 }
 349 static ssize_t defrag_store(struct kobject *kobj,
 350                             struct kobj_attribute *attr,
 351                             const char *buf, size_t count)
 352 {
 353         return double_flag_store(kobj, attr, buf, count,
 354                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 355                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 356 }
 357 static struct kobj_attribute defrag_attr =
 358         __ATTR(defrag, 0644, defrag_show, defrag_store);
 359
 360 static ssize_t use_zero_page_show(struct kobject *kobj,
 361                 struct kobj_attribute *attr, char *buf)
 362 {
 363         return single_flag_show(kobj, attr, buf,
 364                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 365 }
 366 static ssize_t use_zero_page_store(struct kobject *kobj,
 367                 struct kobj_attribute *attr, const char *buf, size_t count)
 368 {
 369         return single_flag_store(kobj, attr, buf, count,
 370                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 371 }
 372 static struct kobj_attribute use_zero_page_attr =
 373         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 374 #ifdef CONFIG_DEBUG_VM
 375 static ssize_t debug_cow_show(struct kobject *kobj,
 376                                 struct kobj_attribute *attr, char *buf)
 377 {
 378         return single_flag_show(kobj, attr, buf,
 379                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 380 }
 381 static ssize_t debug_cow_store(struct kobject *kobj,
 382                                struct kobj_attribute *attr,
 383                                const char *buf, size_t count)
 384 {
 385         return single_flag_store(kobj, attr, buf, count,
 386                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 387 }
 388 static struct kobj_attribute debug_cow_attr =
 389         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 390 #endif /* CONFIG_DEBUG_VM */
 391
 392 static struct attribute *hugepage_attr[] = {
 393         &enabled_attr.attr,
 394         &defrag_attr.attr,
 395         &use_zero_page_attr.attr,
 396 #ifdef CONFIG_DEBUG_VM
 397         &debug_cow_attr.attr,
 398 #endif
 399         NULL,
 400 };
 401
 402 static struct attribute_group hugepage_attr_group = {
 403         .attrs = hugepage_attr,
 404 };
 405
 406 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 407                                          struct kobj_attribute *attr,
 408                                          char *buf)
 409 {
 410         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 411 }
 412
 413 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 414                                           struct kobj_attribute *attr,
 415                                           const char *buf, size_t count)
 416 {
 417         unsigned long msecs;
 418         int err;
 419
 420         err = kstrtoul(buf, 10, &msecs);
 421         if (err || msecs > UINT_MAX)
 422                 return -EINVAL;
 423
 424         khugepaged_scan_sleep_millisecs = msecs;
 425         wake_up_interruptible(&khugepaged_wait);
 426
 427         return count;
 428 }
 429 static struct kobj_attribute scan_sleep_millisecs_attr =
 430         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 431                scan_sleep_millisecs_store);
 432
 433 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 434                                           struct kobj_attribute *attr,
 435                                           char *buf)
 436 {
 437         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 438 }
 439
 440 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 441                                            struct kobj_attribute *attr,
 442                                            const char *buf, size_t count)
 443 {
 444         unsigned long msecs;
 445         int err;
 446
 447         err = kstrtoul(buf, 10, &msecs);
 448         if (err || msecs > UINT_MAX)
 449                 return -EINVAL;
 450
 451         khugepaged_alloc_sleep_millisecs = msecs;
 452         wake_up_interruptible(&khugepaged_wait);
 453
 454         return count;
 455 }
 456 static struct kobj_attribute alloc_sleep_millisecs_attr =
 457         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 458                alloc_sleep_millisecs_store);
 459
 460 static ssize_t pages_to_scan_show(struct kobject *kobj,
 461                                   struct kobj_attribute *attr,
 462                                   char *buf)
 463 {
 464         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 465 }
 466 static ssize_t pages_to_scan_store(struct kobject *kobj,
 467                                    struct kobj_attribute *attr,
 468                                    const char *buf, size_t count)
 469 {
 470         int err;
 471         unsigned long pages;
 472
 473         err = kstrtoul(buf, 10, &pages);
 474         if (err || !pages || pages > UINT_MAX)
 475                 return -EINVAL;
 476
 477         khugepaged_pages_to_scan = pages;
 478
 479         return count;
 480 }
 481 static struct kobj_attribute pages_to_scan_attr =
 482         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 483                pages_to_scan_store);
 484
 485 static ssize_t pages_collapsed_show(struct kobject *kobj,
 486                                     struct kobj_attribute *attr,
 487                                     char *buf)
 488 {
 489         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 490 }
 491 static struct kobj_attribute pages_collapsed_attr =
 492         __ATTR_RO(pages_collapsed);
 493
 494 static ssize_t full_scans_show(struct kobject *kobj,
 495                                struct kobj_attribute *attr,
 496                                char *buf)
 497 {
 498         return sprintf(buf, "%u\n", khugepaged_full_scans);
 499 }
 500 static struct kobj_attribute full_scans_attr =
 501         __ATTR_RO(full_scans);
 502
 503 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 504                                       struct kobj_attribute *attr, char *buf)
 505 {
 506         return single_flag_show(kobj, attr, buf,
 507                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 508 }
 509 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 510                                        struct kobj_attribute *attr,
 511                                        const char *buf, size_t count)
 512 {
 513         return single_flag_store(kobj, attr, buf, count,
 514                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 515 }
 516 static struct kobj_attribute khugepaged_defrag_attr =
 517         __ATTR(defrag, 0644, khugepaged_defrag_show,
 518                khugepaged_defrag_store);
 519
 520 /*
 521  * max_ptes_none controls if khugepaged should collapse hugepages over
 522  * any unmapped ptes in turn potentially increasing the memory
 523  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 524  * reduce the available free memory in the system as it
 525  * runs. Increasing max_ptes_none will instead potentially reduce the
 526  * free memory in the system during the khugepaged scan.
 527  */
 528 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 529                                              struct kobj_attribute *attr,
 530                                              char *buf)
 531 {
 532         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 533 }
 534 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 535                                               struct kobj_attribute *attr,
 536                                               const char *buf, size_t count)
 537 {
 538         int err;
 539         unsigned long max_ptes_none;
 540
 541         err = kstrtoul(buf, 10, &max_ptes_none);
 542         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 543                 return -EINVAL;
 544
 545         khugepaged_max_ptes_none = max_ptes_none;
 546
 547         return count;
 548 }
 549 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 550         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 551                khugepaged_max_ptes_none_store);
 552
 553 static struct attribute *khugepaged_attr[] = {
 554         &khugepaged_defrag_attr.attr,
 555         &khugepaged_max_ptes_none_attr.attr,
 556         &pages_to_scan_attr.attr,
 557         &pages_collapsed_attr.attr,
 558         &full_scans_attr.attr,
 559         &scan_sleep_millisecs_attr.attr,
 560         &alloc_sleep_millisecs_attr.attr,
 561         NULL,
 562 };
 563
 564 static struct attribute_group khugepaged_attr_group = {
 565         .attrs = khugepaged_attr,
 566         .name = "khugepaged",
 567 };
 568
 569 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 570 {
 571         int err;
 572
 573         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 574         if (unlikely(!*hugepage_kobj)) {
 575                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 576                 return -ENOMEM;
 577         }
 578
 579         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 580         if (err) {
 581                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 582                 goto delete_obj;
 583         }
 584
 585         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 586         if (err) {
 587                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 588                 goto remove_hp_group;
 589         }
 590
 591         return 0;
 592
 593 remove_hp_group:
 594         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 595 delete_obj:
 596         kobject_put(*hugepage_kobj);
 597         return err;
 598 }
 599
 600 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 601 {
 602         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 603         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 604         kobject_put(hugepage_kobj);
 605 }
 606 #else
 607 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 608 {
 609         return 0;
 610 }
 611
 612 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 613 {
 614 }
 615 #endif /* CONFIG_SYSFS */
 616
 617 static int __init hugepage_init(void)
 618 {
 619         int err;
 620         struct kobject *hugepage_kobj;
 621
 622         if (!has_transparent_hugepage()) {
 623                 transparent_hugepage_flags = 0;
 624                 return -EINVAL;
 625         }
 626
 627         err = hugepage_init_sysfs(&hugepage_kobj);
 628         if (err)
 629                 return err;
 630
 631         err = khugepaged_slab_init();
 632         if (err)
 633                 goto out;
 634
 635         register_shrinker(&huge_zero_page_shrinker);
 636
 637         /*
 638          * By default disable transparent hugepages on smaller systems,
 639          * where the extra memory used could hurt more than TLB overhead
 640          * is likely to save.  The admin can still enable it through /sys.
 641          */
 642         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 643                 transparent_hugepage_flags = 0;
 644
 645         start_khugepaged();
 646
 647         return 0;
 648 out:
 649         hugepage_exit_sysfs(hugepage_kobj);
 650         return err;
 651 }
 652 module_init(hugepage_init)
 653
 654 static int __init setup_transparent_hugepage(char *str)
 655 {
 656         int ret = 0;
 657         if (!str)
 658                 goto out;
 659         if (!strcmp(str, "always")) {
 660                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 661                         &transparent_hugepage_flags);
 662                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 663                           &transparent_hugepage_flags);
 664                 ret = 1;
 665         } else if (!strcmp(str, "madvise")) {
 666                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 667                           &transparent_hugepage_flags);
 668                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 669                         &transparent_hugepage_flags);
 670                 ret = 1;
 671         } else if (!strcmp(str, "never")) {
 672                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 673                           &transparent_hugepage_flags);
 674                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 675                           &transparent_hugepage_flags);
 676                 ret = 1;
 677         }
 678 out:
 679         if (!ret)
 680                 printk(KERN_WARNING
 681                        "transparent_hugepage= cannot parse, ignored\n");
 682         return ret;
 683 }
 684 __setup("transparent_hugepage=", setup_transparent_hugepage);
 685
 686 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 687 {
 688         if (likely(vma->vm_flags & VM_WRITE))
 689                 pmd = pmd_mkwrite(pmd);
 690         return pmd;
 691 }
 692
 693 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 694 {
 695         pmd_t entry;
 696         entry = mk_pmd(page, prot);
 697         entry = pmd_mkhuge(entry);
 698         return entry;
 699 }
 700
 701 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 702                                         struct vm_area_struct *vma,
 703                                         unsigned long haddr, pmd_t *pmd,
 704                                         struct page *page)
 705 {
 706         pgtable_t pgtable;
 707
 708         VM_BUG_ON(!PageCompound(page));
 709         pgtable = pte_alloc_one(mm, haddr);
 710         if (unlikely(!pgtable))
 711                 return VM_FAULT_OOM;
 712
 713         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 714         /*
 715          * The memory barrier inside __SetPageUptodate makes sure that
 716          * clear_huge_page writes become visible before the set_pmd_at()
 717          * write.
 718          */
 719         __SetPageUptodate(page);
 720
 721         spin_lock(&mm->page_table_lock);
 722         if (unlikely(!pmd_none(*pmd))) {
 723                 spin_unlock(&mm->page_table_lock);
 724                 mem_cgroup_uncharge_page(page);
 725                 put_page(page);
 726                 pte_free(mm, pgtable);
 727         } else {
 728                 pmd_t entry;
 729                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 730                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 731                 page_add_new_anon_rmap(page, vma, haddr);
 732                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 733                 set_pmd_at(mm, haddr, pmd, entry);
 734                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 735                 mm->nr_ptes++;
 736                 spin_unlock(&mm->page_table_lock);
 737         }
 738
 739         return 0;
 740 }
 741
 742 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 743 {
 744         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 745 }
 746
 747 static inline struct page *alloc_hugepage_vma(int defrag,
 748                                               struct vm_area_struct *vma,
 749                                               unsigned long haddr, int nd,
 750                                               gfp_t extra_gfp)
 751 {
 752         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 753                                HPAGE_PMD_ORDER, vma, haddr, nd);
 754 }
 755
 756 #ifndef CONFIG_NUMA
 757 static inline struct page *alloc_hugepage(int defrag)
 758 {
 759         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 760                            HPAGE_PMD_ORDER);
 761 }
 762 #endif
 763
 764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 765                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 766                 struct page *zero_page)
 767 {
 768         pmd_t entry;
 769         if (!pmd_none(*pmd))
 770                 return false;
 771         entry = mk_pmd(zero_page, vma->vm_page_prot);
 772         entry = pmd_wrprotect(entry);
 773         entry = pmd_mkhuge(entry);
 774         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 775         set_pmd_at(mm, haddr, pmd, entry);
 776         mm->nr_ptes++;
 777         return true;
 778 }
 779
 780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 781                                unsigned long address, pmd_t *pmd,
 782                                unsigned int flags)
 783 {
 784         struct page *page;
 785         unsigned long haddr = address & HPAGE_PMD_MASK;
 786
 787         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 788                 return VM_FAULT_FALLBACK;
 789         if (unlikely(anon_vma_prepare(vma)))
 790                 return VM_FAULT_OOM;
 791         if (unlikely(khugepaged_enter(vma)))
 792                 return VM_FAULT_OOM;
 793         if (!(flags & FAULT_FLAG_WRITE) &&
 794                         transparent_hugepage_use_zero_page()) {
 795                 pgtable_t pgtable;
 796                 struct page *zero_page;
 797                 bool set;
 798                 pgtable = pte_alloc_one(mm, haddr);
 799                 if (unlikely(!pgtable))
 800                         return VM_FAULT_OOM;
 801                 zero_page = get_huge_zero_page();
 802                 if (unlikely(!zero_page)) {
 803                         pte_free(mm, pgtable);
 804                         count_vm_event(THP_FAULT_FALLBACK);
 805                         return VM_FAULT_FALLBACK;
 806                 }
 807                 spin_lock(&mm->page_table_lock);
 808                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 809                                 zero_page);
 810                 spin_unlock(&mm->page_table_lock);
 811                 if (!set) {
 812                         pte_free(mm, pgtable);
 813                         put_huge_zero_page();
 814                 }
 815                 return 0;
 816         }
 817         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 818                         vma, haddr, numa_node_id(), 0);
 819         if (unlikely(!page)) {
 820                 count_vm_event(THP_FAULT_FALLBACK);
 821                 return VM_FAULT_FALLBACK;
 822         }
 823         count_vm_event(THP_FAULT_ALLOC);
 824         if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 825                 put_page(page);
 826                 return VM_FAULT_FALLBACK;
 827         }
 828         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 829                 mem_cgroup_uncharge_page(page);
 830                 put_page(page);
 831                 return VM_FAULT_FALLBACK;
 832         }
 833
 834         return 0;
 835 }
 836
 837 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 838                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 839                   struct vm_area_struct *vma)
 840 {
 841         struct page *src_page;
 842         pmd_t pmd;
 843         pgtable_t pgtable;
 844         int ret;
 845
 846         ret = -ENOMEM;
 847         pgtable = pte_alloc_one(dst_mm, addr);
 848         if (unlikely(!pgtable))
 849                 goto out;
 850
 851         spin_lock(&dst_mm->page_table_lock);
 852         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 853
 854         ret = -EAGAIN;
 855         pmd = *src_pmd;
 856         if (unlikely(!pmd_trans_huge(pmd))) {
 857                 pte_free(dst_mm, pgtable);
 858                 goto out_unlock;
 859         }
 860         /*
 861          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 862          * under splitting since we don't split the page itself, only pmd to
 863          * a page table.
 864          */
 865         if (is_huge_zero_pmd(pmd)) {
 866                 struct page *zero_page;
 867                 bool set;
 868                 /*
 869                  * get_huge_zero_page() will never allocate a new page here,
 870                  * since we already have a zero page to copy. It just takes a
 871                  * reference.
 872                  */
 873                 zero_page = get_huge_zero_page();
 874                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 875                                 zero_page);
 876                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 877                 ret = 0;
 878                 goto out_unlock;
 879         }
 880         if (unlikely(pmd_trans_splitting(pmd))) {
 881                 /* split huge page running from under us */
 882                 spin_unlock(&src_mm->page_table_lock);
 883                 spin_unlock(&dst_mm->page_table_lock);
 884                 pte_free(dst_mm, pgtable);
 885
 886                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 887                 goto out;
 888         }
 889         src_page = pmd_page(pmd);
 890         VM_BUG_ON(!PageHead(src_page));
 891         get_page(src_page);
 892         page_dup_rmap(src_page);
 893         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 894
 895         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 896         pmd = pmd_mkold(pmd_wrprotect(pmd));
 897         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 898         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 899         dst_mm->nr_ptes++;
 900
 901         ret = 0;
 902 out_unlock:
 903         spin_unlock(&src_mm->page_table_lock);
 904         spin_unlock(&dst_mm->page_table_lock);
 905 out:
 906         return ret;
 907 }
 908
 909 void huge_pmd_set_accessed(struct mm_struct *mm,
 910                            struct vm_area_struct *vma,
 911                            unsigned long address,
 912                            pmd_t *pmd, pmd_t orig_pmd,
 913                            int dirty)
 914 {
 915         pmd_t entry;
 916         unsigned long haddr;
 917
 918         spin_lock(&mm->page_table_lock);
 919         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 920                 goto unlock;
 921
 922         entry = pmd_mkyoung(orig_pmd);
 923         haddr = address & HPAGE_PMD_MASK;
 924         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 925                 update_mmu_cache_pmd(vma, address, pmd);
 926
 927 unlock:
 928         spin_unlock(&mm->page_table_lock);
 929 }
 930
 931 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 932                 struct vm_area_struct *vma, unsigned long address,
 933                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 934 {
 935         pgtable_t pgtable;
 936         pmd_t _pmd;
 937         struct page *page;
 938         int i, ret = 0;
 939         unsigned long mmun_start;       /* For mmu_notifiers */
 940         unsigned long mmun_end;         /* For mmu_notifiers */
 941
 942         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 943         if (!page) {
 944                 ret |= VM_FAULT_OOM;
 945                 goto out;
 946         }
 947
 948         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 949                 put_page(page);
 950                 ret |= VM_FAULT_OOM;
 951                 goto out;
 952         }
 953
 954         clear_user_highpage(page, address);
 955         __SetPageUptodate(page);
 956
 957         mmun_start = haddr;
 958         mmun_end   = haddr + HPAGE_PMD_SIZE;
 959         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 960
 961         spin_lock(&mm->page_table_lock);
 962         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 963                 goto out_free_page;
 964
 965         pmdp_clear_flush(vma, haddr, pmd);
 966         /* leave pmd empty until pte is filled */
 967
 968         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
 969         pmd_populate(mm, &_pmd, pgtable);
 970
 971         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 972                 pte_t *pte, entry;
 973                 if (haddr == (address & PAGE_MASK)) {
 974                         entry = mk_pte(page, vma->vm_page_prot);
 975                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 976                         page_add_new_anon_rmap(page, vma, haddr);
 977                 } else {
 978                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
 979                         entry = pte_mkspecial(entry);
 980                 }
 981                 pte = pte_offset_map(&_pmd, haddr);
 982                 VM_BUG_ON(!pte_none(*pte));
 983                 set_pte_at(mm, haddr, pte, entry);
 984                 pte_unmap(pte);
 985         }
 986         smp_wmb(); /* make pte visible before pmd */
 987         pmd_populate(mm, pmd, pgtable);
 988         spin_unlock(&mm->page_table_lock);
 989         put_huge_zero_page();
 990         inc_mm_counter(mm, MM_ANONPAGES);
 991
 992         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 993
 994         ret |= VM_FAULT_WRITE;
 995 out:
 996         return ret;
 997 out_free_page:
 998         spin_unlock(&mm->page_table_lock);
 999         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1000         mem_cgroup_uncharge_page(page);
1001         put_page(page);
1002         goto out;
1003 }
1004
1005 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1006                                         struct vm_area_struct *vma,
1007                                         unsigned long address,
1008                                         pmd_t *pmd, pmd_t orig_pmd,
1009                                         struct page *page,
1010                                         unsigned long haddr)
1011 {
1012         pgtable_t pgtable;
1013         pmd_t _pmd;
1014         int ret = 0, i;
1015         struct page **pages;
1016         unsigned long mmun_start;       /* For mmu_notifiers */
1017         unsigned long mmun_end;         /* For mmu_notifiers */
1018
1019         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1020                         GFP_KERNEL);
1021         if (unlikely(!pages)) {
1022                 ret |= VM_FAULT_OOM;
1023                 goto out;
1024         }
1025
1026         for (i = 0; i < HPAGE_PMD_NR; i++) {
1027                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1028                                                __GFP_OTHER_NODE,
1029                                                vma, address, page_to_nid(page));
1030                 if (unlikely(!pages[i] ||
1031                              mem_cgroup_newpage_charge(pages[i], mm,
1032                                                        GFP_KERNEL))) {
1033                         if (pages[i])
1034                                 put_page(pages[i]);
1035                         mem_cgroup_uncharge_start();
1036                         while (--i >= 0) {
1037                                 mem_cgroup_uncharge_page(pages[i]);
1038                                 put_page(pages[i]);
1039                         }
1040                         mem_cgroup_uncharge_end();
1041                         kfree(pages);
1042                         ret |= VM_FAULT_OOM;
1043                         goto out;
1044                 }
1045         }
1046
1047         for (i = 0; i < HPAGE_PMD_NR; i++) {
1048                 copy_user_highpage(pages[i], page + i,
1049                                    haddr + PAGE_SIZE * i, vma);
1050                 __SetPageUptodate(pages[i]);
1051                 cond_resched();
1052         }
1053
1054         mmun_start = haddr;
1055         mmun_end   = haddr + HPAGE_PMD_SIZE;
1056         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1057
1058         spin_lock(&mm->page_table_lock);
1059         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1060                 goto out_free_pages;
1061         VM_BUG_ON(!PageHead(page));
1062
1063         pmdp_clear_flush(vma, haddr, pmd);
1064         /* leave pmd empty until pte is filled */
1065
1066         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1067         pmd_populate(mm, &_pmd, pgtable);
1068
1069         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1070                 pte_t *pte, entry;
1071                 entry = mk_pte(pages[i], vma->vm_page_prot);
1072                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1073                 page_add_new_anon_rmap(pages[i], vma, haddr);
1074                 pte = pte_offset_map(&_pmd, haddr);
1075                 VM_BUG_ON(!pte_none(*pte));
1076                 set_pte_at(mm, haddr, pte, entry);
1077                 pte_unmap(pte);
1078         }
1079         kfree(pages);
1080
1081         smp_wmb(); /* make pte visible before pmd */
1082         pmd_populate(mm, pmd, pgtable);
1083         page_remove_rmap(page);
1084         spin_unlock(&mm->page_table_lock);
1085
1086         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1087
1088         ret |= VM_FAULT_WRITE;
1089         put_page(page);
1090
1091 out:
1092         return ret;
1093
1094 out_free_pages:
1095         spin_unlock(&mm->page_table_lock);
1096         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1097         mem_cgroup_uncharge_start();
1098         for (i = 0; i < HPAGE_PMD_NR; i++) {
1099                 mem_cgroup_uncharge_page(pages[i]);
1100                 put_page(pages[i]);
1101         }
1102         mem_cgroup_uncharge_end();
1103         kfree(pages);
1104         goto out;
1105 }
1106
1107 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1108                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1109 {
1110         int ret = 0;
1111         struct page *page = NULL, *new_page;
1112         unsigned long haddr;
1113         unsigned long mmun_start;       /* For mmu_notifiers */
1114         unsigned long mmun_end;         /* For mmu_notifiers */
1115
1116         VM_BUG_ON(!vma->anon_vma);
1117         haddr = address & HPAGE_PMD_MASK;
1118         if (is_huge_zero_pmd(orig_pmd))
1119                 goto alloc;
1120         spin_lock(&mm->page_table_lock);
1121         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1122                 goto out_unlock;
1123
1124         page = pmd_page(orig_pmd);
1125         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1126         if (page_mapcount(page) == 1) {
1127                 pmd_t entry;
1128                 entry = pmd_mkyoung(orig_pmd);
1129                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1130                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1131                         update_mmu_cache_pmd(vma, address, pmd);
1132                 ret |= VM_FAULT_WRITE;
1133                 goto out_unlock;
1134         }
1135         get_page(page);
1136         spin_unlock(&mm->page_table_lock);
1137 alloc:
1138         if (transparent_hugepage_enabled(vma) &&
1139             !transparent_hugepage_debug_cow())
1140                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1141                                               vma, haddr, numa_node_id(), 0);
1142         else
1143                 new_page = NULL;
1144
1145         if (unlikely(!new_page)) {
1146                 count_vm_event(THP_FAULT_FALLBACK);
1147                 if (is_huge_zero_pmd(orig_pmd)) {
1148                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1149                                         address, pmd, orig_pmd, haddr);
1150                 } else {
1151                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1152                                         pmd, orig_pmd, page, haddr);
1153                         if (ret & VM_FAULT_OOM)
1154                                 split_huge_page(page);
1155                         put_page(page);
1156                 }
1157                 goto out;
1158         }
1159         count_vm_event(THP_FAULT_ALLOC);
1160
1161         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1162                 put_page(new_page);
1163                 if (page) {
1164                         split_huge_page(page);
1165                         put_page(page);
1166                 }
1167                 ret |= VM_FAULT_OOM;
1168                 goto out;
1169         }
1170
1171         if (is_huge_zero_pmd(orig_pmd))
1172                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1173         else
1174                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1175         __SetPageUptodate(new_page);
1176
1177         mmun_start = haddr;
1178         mmun_end   = haddr + HPAGE_PMD_SIZE;
1179         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1180
1181         spin_lock(&mm->page_table_lock);
1182         if (page)
1183                 put_page(page);
1184         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1185                 spin_unlock(&mm->page_table_lock);
1186                 mem_cgroup_uncharge_page(new_page);
1187                 put_page(new_page);
1188                 goto out_mn;
1189         } else {
1190                 pmd_t entry;
1191                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1192                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1193                 pmdp_clear_flush(vma, haddr, pmd);
1194                 page_add_new_anon_rmap(new_page, vma, haddr);
1195                 set_pmd_at(mm, haddr, pmd, entry);
1196                 update_mmu_cache_pmd(vma, address, pmd);
1197                 if (is_huge_zero_pmd(orig_pmd)) {
1198                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1199                         put_huge_zero_page();
1200                 } else {
1201                         VM_BUG_ON(!PageHead(page));
1202                         page_remove_rmap(page);
1203                         put_page(page);
1204                 }
1205                 ret |= VM_FAULT_WRITE;
1206         }
1207         spin_unlock(&mm->page_table_lock);
1208 out_mn:
1209         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1210 out:
1211         return ret;
1212 out_unlock:
1213         spin_unlock(&mm->page_table_lock);
1214         return ret;
1215 }
1216
1217 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1218                                    unsigned long addr,
1219                                    pmd_t *pmd,
1220                                    unsigned int flags)
1221 {
1222         struct mm_struct *mm = vma->vm_mm;
1223         struct page *page = NULL;
1224
1225         assert_spin_locked(&mm->page_table_lock);
1226
1227         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1228                 goto out;
1229
1230         /* Avoid dumping huge zero page */
1231         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1232                 return ERR_PTR(-EFAULT);
1233
1234         page = pmd_page(*pmd);
1235         VM_BUG_ON(!PageHead(page));
1236         if (flags & FOLL_TOUCH) {
1237                 pmd_t _pmd;
1238                 /*
1239                  * We should set the dirty bit only for FOLL_WRITE but
1240                  * for now the dirty bit in the pmd is meaningless.
1241                  * And if the dirty bit will become meaningful and
1242                  * we'll only set it with FOLL_WRITE, an atomic
1243                  * set_bit will be required on the pmd to set the
1244                  * young bit, instead of the current set_pmd_at.
1245                  */
1246                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1247                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1248                                           pmd, _pmd,  1))
1249                         update_mmu_cache_pmd(vma, addr, pmd);
1250         }
1251         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1252                 if (page->mapping && trylock_page(page)) {
1253                         lru_add_drain();
1254                         if (page->mapping)
1255                                 mlock_vma_page(page);
1256                         unlock_page(page);
1257                 }
1258         }
1259         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1260         VM_BUG_ON(!PageCompound(page));
1261         if (flags & FOLL_GET)
1262                 get_page_foll(page);
1263
1264 out:
1265         return page;
1266 }
1267
1268 /* NUMA hinting page fault entry point for trans huge pmds */
1269 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1270                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1271 {
1272         struct page *page;
1273         unsigned long haddr = addr & HPAGE_PMD_MASK;
1274         int target_nid;
1275         int current_nid = -1;
1276         bool migrated;
1277
1278         spin_lock(&mm->page_table_lock);
1279         if (unlikely(!pmd_same(pmd, *pmdp)))
1280                 goto out_unlock;
1281
1282         page = pmd_page(pmd);
1283         get_page(page);
1284         current_nid = page_to_nid(page);
1285         count_vm_numa_event(NUMA_HINT_FAULTS);
1286         if (current_nid == numa_node_id())
1287                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1288
1289         target_nid = mpol_misplaced(page, vma, haddr);
1290         if (target_nid == -1) {
1291                 put_page(page);
1292                 goto clear_pmdnuma;
1293         }
1294
1295         /* Acquire the page lock to serialise THP migrations */
1296         spin_unlock(&mm->page_table_lock);
1297         lock_page(page);
1298
1299         /* Confirm the PTE did not while locked */
1300         spin_lock(&mm->page_table_lock);
1301         if (unlikely(!pmd_same(pmd, *pmdp))) {
1302                 unlock_page(page);
1303                 put_page(page);
1304                 goto out_unlock;
1305         }
1306         spin_unlock(&mm->page_table_lock);
1307
1308         /* Migrate the THP to the requested node */
1309         migrated = migrate_misplaced_transhuge_page(mm, vma,
1310                                 pmdp, pmd, addr, page, target_nid);
1311         if (!migrated)
1312                 goto check_same;
1313
1314         task_numa_fault(target_nid, HPAGE_PMD_NR, true);
1315         return 0;
1316
1317 check_same:
1318         spin_lock(&mm->page_table_lock);
1319         if (unlikely(!pmd_same(pmd, *pmdp)))
1320                 goto out_unlock;
1321 clear_pmdnuma:
1322         pmd = pmd_mknonnuma(pmd);
1323         set_pmd_at(mm, haddr, pmdp, pmd);
1324         VM_BUG_ON(pmd_numa(*pmdp));
1325         update_mmu_cache_pmd(vma, addr, pmdp);
1326 out_unlock:
1327         spin_unlock(&mm->page_table_lock);
1328         if (current_nid != -1)
1329                 task_numa_fault(current_nid, HPAGE_PMD_NR, false);
1330         return 0;
1331 }
1332
1333 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1334                  pmd_t *pmd, unsigned long addr)
1335 {
1336         int ret = 0;
1337
1338         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1339                 struct page *page;
1340                 pgtable_t pgtable;
1341                 pmd_t orig_pmd;
1342                 /*
1343                  * For architectures like ppc64 we look at deposited pgtable
1344                  * when calling pmdp_get_and_clear. So do the
1345                  * pgtable_trans_huge_withdraw after finishing pmdp related
1346                  * operations.
1347                  */
1348                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1349                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1350                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1351                 if (is_huge_zero_pmd(orig_pmd)) {
1352                         tlb->mm->nr_ptes--;
1353                         spin_unlock(&tlb->mm->page_table_lock);
1354                         put_huge_zero_page();
1355                 } else {
1356                         page = pmd_page(orig_pmd);
1357                         page_remove_rmap(page);
1358                         VM_BUG_ON(page_mapcount(page) < 0);
1359                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1360                         VM_BUG_ON(!PageHead(page));
1361                         tlb->mm->nr_ptes--;
1362                         spin_unlock(&tlb->mm->page_table_lock);
1363                         tlb_remove_page(tlb, page);
1364                 }
1365                 pte_free(tlb->mm, pgtable);
1366                 ret = 1;
1367         }
1368         return ret;
1369 }
1370
1371 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1372                 unsigned long addr, unsigned long end,
1373                 unsigned char *vec)
1374 {
1375         int ret = 0;
1376
1377         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1378                 /*
1379                  * All logical pages in the range are present
1380                  * if backed by a huge page.
1381                  */
1382                 spin_unlock(&vma->vm_mm->page_table_lock);
1383                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1384                 ret = 1;
1385         }
1386
1387         return ret;
1388 }
1389
1390 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1391                   unsigned long old_addr,
1392                   unsigned long new_addr, unsigned long old_end,
1393                   pmd_t *old_pmd, pmd_t *new_pmd)
1394 {
1395         int ret = 0;
1396         pmd_t pmd;
1397
1398         struct mm_struct *mm = vma->vm_mm;
1399
1400         if ((old_addr & ~HPAGE_PMD_MASK) ||
1401             (new_addr & ~HPAGE_PMD_MASK) ||
1402             old_end - old_addr < HPAGE_PMD_SIZE ||
1403             (new_vma->vm_flags & VM_NOHUGEPAGE))
1404                 goto out;
1405
1406         /*
1407          * The destination pmd shouldn't be established, free_pgtables()
1408          * should have release it.
1409          */
1410         if (WARN_ON(!pmd_none(*new_pmd))) {
1411                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1412                 goto out;
1413         }
1414
1415         ret = __pmd_trans_huge_lock(old_pmd, vma);
1416         if (ret == 1) {
1417                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1418                 VM_BUG_ON(!pmd_none(*new_pmd));
1419                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1420                 spin_unlock(&mm->page_table_lock);
1421         }
1422 out:
1423         return ret;
1424 }
1425
1426 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1427                 unsigned long addr, pgprot_t newprot, int prot_numa)
1428 {
1429         struct mm_struct *mm = vma->vm_mm;
1430         int ret = 0;
1431
1432         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1433                 pmd_t entry;
1434                 entry = pmdp_get_and_clear(mm, addr, pmd);
1435                 if (!prot_numa) {
1436                         entry = pmd_modify(entry, newprot);
1437                         BUG_ON(pmd_write(entry));
1438                 } else {
1439                         struct page *page = pmd_page(*pmd);
1440
1441                         /* only check non-shared pages */
1442                         if (page_mapcount(page) == 1 &&
1443                             !pmd_numa(*pmd)) {
1444                                 entry = pmd_mknuma(entry);
1445                         }
1446                 }
1447                 set_pmd_at(mm, addr, pmd, entry);
1448                 spin_unlock(&vma->vm_mm->page_table_lock);
1449                 ret = 1;
1450         }
1451
1452         return ret;
1453 }
1454
1455 /*
1456  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1457  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1458  *
1459  * Note that if it returns 1, this routine returns without unlocking page
1460  * table locks. So callers must unlock them.
1461  */
1462 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1463 {
1464         spin_lock(&vma->vm_mm->page_table_lock);
1465         if (likely(pmd_trans_huge(*pmd))) {
1466                 if (unlikely(pmd_trans_splitting(*pmd))) {
1467                         spin_unlock(&vma->vm_mm->page_table_lock);
1468                         wait_split_huge_page(vma->anon_vma, pmd);
1469                         return -1;
1470                 } else {
1471                         /* Thp mapped by 'pmd' is stable, so we can
1472                          * handle it as it is. */
1473                         return 1;
1474                 }
1475         }
1476         spin_unlock(&vma->vm_mm->page_table_lock);
1477         return 0;
1478 }
1479
1480 pmd_t *page_check_address_pmd(struct page *page,
1481                               struct mm_struct *mm,
1482                               unsigned long address,
1483                               enum page_check_address_pmd_flag flag)
1484 {
1485         pmd_t *pmd, *ret = NULL;
1486
1487         if (address & ~HPAGE_PMD_MASK)
1488                 goto out;
1489
1490         pmd = mm_find_pmd(mm, address);
1491         if (!pmd)
1492                 goto out;
1493         if (pmd_none(*pmd))
1494                 goto out;
1495         if (pmd_page(*pmd) != page)
1496                 goto out;
1497         /*
1498          * split_vma() may create temporary aliased mappings. There is
1499          * no risk as long as all huge pmd are found and have their
1500          * splitting bit set before __split_huge_page_refcount
1501          * runs. Finding the same huge pmd more than once during the
1502          * same rmap walk is not a problem.
1503          */
1504         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1505             pmd_trans_splitting(*pmd))
1506                 goto out;
1507         if (pmd_trans_huge(*pmd)) {
1508                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1509                           !pmd_trans_splitting(*pmd));
1510                 ret = pmd;
1511         }
1512 out:
1513         return ret;
1514 }
1515
1516 static int __split_huge_page_splitting(struct page *page,
1517                                        struct vm_area_struct *vma,
1518                                        unsigned long address)
1519 {
1520         struct mm_struct *mm = vma->vm_mm;
1521         pmd_t *pmd;
1522         int ret = 0;
1523         /* For mmu_notifiers */
1524         const unsigned long mmun_start = address;
1525         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1526
1527         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1528         spin_lock(&mm->page_table_lock);
1529         pmd = page_check_address_pmd(page, mm, address,
1530                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1531         if (pmd) {
1532                 /*
1533                  * We can't temporarily set the pmd to null in order
1534                  * to split it, the pmd must remain marked huge at all
1535                  * times or the VM won't take the pmd_trans_huge paths
1536                  * and it won't wait on the anon_vma->root->rwsem to
1537                  * serialize against split_huge_page*.
1538                  */
1539                 pmdp_splitting_flush(vma, address, pmd);
1540                 ret = 1;
1541         }
1542         spin_unlock(&mm->page_table_lock);
1543         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1544
1545         return ret;
1546 }
1547
1548 static void __split_huge_page_refcount(struct page *page,
1549                                        struct list_head *list)
1550 {
1551         int i;
1552         struct zone *zone = page_zone(page);
1553         struct lruvec *lruvec;
1554         int tail_count = 0;
1555
1556         /* prevent PageLRU to go away from under us, and freeze lru stats */
1557         spin_lock_irq(&zone->lru_lock);
1558         lruvec = mem_cgroup_page_lruvec(page, zone);
1559
1560         compound_lock(page);
1561         /* complete memcg works before add pages to LRU */
1562         mem_cgroup_split_huge_fixup(page);
1563
1564         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1565                 struct page *page_tail = page + i;
1566
1567                 /* tail_page->_mapcount cannot change */
1568                 BUG_ON(page_mapcount(page_tail) < 0);
1569                 tail_count += page_mapcount(page_tail);
1570                 /* check for overflow */
1571                 BUG_ON(tail_count < 0);
1572                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1573                 /*
1574                  * tail_page->_count is zero and not changing from
1575                  * under us. But get_page_unless_zero() may be running
1576                  * from under us on the tail_page. If we used
1577                  * atomic_set() below instead of atomic_add(), we
1578                  * would then run atomic_set() concurrently with
1579                  * get_page_unless_zero(), and atomic_set() is
1580                  * implemented in C not using locked ops. spin_unlock
1581                  * on x86 sometime uses locked ops because of PPro
1582                  * errata 66, 92, so unless somebody can guarantee
1583                  * atomic_set() here would be safe on all archs (and
1584                  * not only on x86), it's safer to use atomic_add().
1585                  */
1586                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1587                            &page_tail->_count);
1588
1589                 /* after clearing PageTail the gup refcount can be released */
1590                 smp_mb();
1591
1592                 /*
1593                  * retain hwpoison flag of the poisoned tail page:
1594                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1595                  *   by the memory-failure.
1596                  */
1597                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1598                 page_tail->flags |= (page->flags &
1599                                      ((1L << PG_referenced) |
1600                                       (1L << PG_swapbacked) |
1601                                       (1L << PG_mlocked) |
1602                                       (1L << PG_uptodate) |
1603                                       (1L << PG_active) |
1604                                       (1L << PG_unevictable)));
1605                 page_tail->flags |= (1L << PG_dirty);
1606
1607                 /* clear PageTail before overwriting first_page */
1608                 smp_wmb();
1609
1610                 /*
1611                  * __split_huge_page_splitting() already set the
1612                  * splitting bit in all pmd that could map this
1613                  * hugepage, that will ensure no CPU can alter the
1614                  * mapcount on the head page. The mapcount is only
1615                  * accounted in the head page and it has to be
1616                  * transferred to all tail pages in the below code. So
1617                  * for this code to be safe, the split the mapcount
1618                  * can't change. But that doesn't mean userland can't
1619                  * keep changing and reading the page contents while
1620                  * we transfer the mapcount, so the pmd splitting
1621                  * status is achieved setting a reserved bit in the
1622                  * pmd, not by clearing the present bit.
1623                 */
1624                 page_tail->_mapcount = page->_mapcount;
1625
1626                 BUG_ON(page_tail->mapping);
1627                 page_tail->mapping = page->mapping;
1628
1629                 page_tail->index = page->index + i;
1630                 page_nid_xchg_last(page_tail, page_nid_last(page));
1631
1632                 BUG_ON(!PageAnon(page_tail));
1633                 BUG_ON(!PageUptodate(page_tail));
1634                 BUG_ON(!PageDirty(page_tail));
1635                 BUG_ON(!PageSwapBacked(page_tail));
1636
1637                 lru_add_page_tail(page, page_tail, lruvec, list);
1638         }
1639         atomic_sub(tail_count, &page->_count);
1640         BUG_ON(atomic_read(&page->_count) <= 0);
1641
1642         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1643
1644         ClearPageCompound(page);
1645         compound_unlock(page);
1646         spin_unlock_irq(&zone->lru_lock);
1647
1648         for (i = 1; i < HPAGE_PMD_NR; i++) {
1649                 struct page *page_tail = page + i;
1650                 BUG_ON(page_count(page_tail) <= 0);
1651                 /*
1652                  * Tail pages may be freed if there wasn't any mapping
1653                  * like if add_to_swap() is running on a lru page that
1654                  * had its mapping zapped. And freeing these pages
1655                  * requires taking the lru_lock so we do the put_page
1656                  * of the tail pages after the split is complete.
1657                  */
1658                 put_page(page_tail);
1659         }
1660
1661         /*
1662          * Only the head page (now become a regular page) is required
1663          * to be pinned by the caller.
1664          */
1665         BUG_ON(page_count(page) <= 0);
1666 }
1667
1668 static int __split_huge_page_map(struct page *page,
1669                                  struct vm_area_struct *vma,
1670                                  unsigned long address)
1671 {
1672         struct mm_struct *mm = vma->vm_mm;
1673         pmd_t *pmd, _pmd;
1674         int ret = 0, i;
1675         pgtable_t pgtable;
1676         unsigned long haddr;
1677
1678         spin_lock(&mm->page_table_lock);
1679         pmd = page_check_address_pmd(page, mm, address,
1680                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1681         if (pmd) {
1682                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1683                 pmd_populate(mm, &_pmd, pgtable);
1684
1685                 haddr = address;
1686                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1687                         pte_t *pte, entry;
1688                         BUG_ON(PageCompound(page+i));
1689                         entry = mk_pte(page + i, vma->vm_page_prot);
1690                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1691                         if (!pmd_write(*pmd))
1692                                 entry = pte_wrprotect(entry);
1693                         else
1694                                 BUG_ON(page_mapcount(page) != 1);
1695                         if (!pmd_young(*pmd))
1696                                 entry = pte_mkold(entry);
1697                         if (pmd_numa(*pmd))
1698                                 entry = pte_mknuma(entry);
1699                         pte = pte_offset_map(&_pmd, haddr);
1700                         BUG_ON(!pte_none(*pte));
1701                         set_pte_at(mm, haddr, pte, entry);
1702                         pte_unmap(pte);
1703                 }
1704
1705                 smp_wmb(); /* make pte visible before pmd */
1706                 /*
1707                  * Up to this point the pmd is present and huge and
1708                  * userland has the whole access to the hugepage
1709                  * during the split (which happens in place). If we
1710                  * overwrite the pmd with the not-huge version
1711                  * pointing to the pte here (which of course we could
1712                  * if all CPUs were bug free), userland could trigger
1713                  * a small page size TLB miss on the small sized TLB
1714                  * while the hugepage TLB entry is still established
1715                  * in the huge TLB. Some CPU doesn't like that. See
1716                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1717                  * Erratum 383 on page 93. Intel should be safe but is
1718                  * also warns that it's only safe if the permission
1719                  * and cache attributes of the two entries loaded in
1720                  * the two TLB is identical (which should be the case
1721                  * here). But it is generally safer to never allow
1722                  * small and huge TLB entries for the same virtual
1723                  * address to be loaded simultaneously. So instead of
1724                  * doing "pmd_populate(); flush_tlb_range();" we first
1725                  * mark the current pmd notpresent (atomically because
1726                  * here the pmd_trans_huge and pmd_trans_splitting
1727                  * must remain set at all times on the pmd until the
1728                  * split is complete for this pmd), then we flush the
1729                  * SMP TLB and finally we write the non-huge version
1730                  * of the pmd entry with pmd_populate.
1731                  */
1732                 pmdp_invalidate(vma, address, pmd);
1733                 pmd_populate(mm, pmd, pgtable);
1734                 ret = 1;
1735         }
1736         spin_unlock(&mm->page_table_lock);
1737
1738         return ret;
1739 }
1740
1741 /* must be called with anon_vma->root->rwsem held */
1742 static void __split_huge_page(struct page *page,
1743                               struct anon_vma *anon_vma,
1744                               struct list_head *list)
1745 {
1746         int mapcount, mapcount2;
1747         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1748         struct anon_vma_chain *avc;
1749
1750         BUG_ON(!PageHead(page));
1751         BUG_ON(PageTail(page));
1752
1753         mapcount = 0;
1754         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1755                 struct vm_area_struct *vma = avc->vma;
1756                 unsigned long addr = vma_address(page, vma);
1757                 BUG_ON(is_vma_temporary_stack(vma));
1758                 mapcount += __split_huge_page_splitting(page, vma, addr);
1759         }
1760         /*
1761          * It is critical that new vmas are added to the tail of the
1762          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1763          * and establishes a child pmd before
1764          * __split_huge_page_splitting() freezes the parent pmd (so if
1765          * we fail to prevent copy_huge_pmd() from running until the
1766          * whole __split_huge_page() is complete), we will still see
1767          * the newly established pmd of the child later during the
1768          * walk, to be able to set it as pmd_trans_splitting too.
1769          */
1770         if (mapcount != page_mapcount(page))
1771                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1772                        mapcount, page_mapcount(page));
1773         BUG_ON(mapcount != page_mapcount(page));
1774
1775         __split_huge_page_refcount(page, list);
1776
1777         mapcount2 = 0;
1778         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1779                 struct vm_area_struct *vma = avc->vma;
1780                 unsigned long addr = vma_address(page, vma);
1781                 BUG_ON(is_vma_temporary_stack(vma));
1782                 mapcount2 += __split_huge_page_map(page, vma, addr);
1783         }
1784         if (mapcount != mapcount2)
1785                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1786                        mapcount, mapcount2, page_mapcount(page));
1787         BUG_ON(mapcount != mapcount2);
1788 }
1789
1790 /*
1791  * Split a hugepage into normal pages. This doesn't change the position of head
1792  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1793  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1794  * from the hugepage.
1795  * Return 0 if the hugepage is split successfully otherwise return 1.
1796  */
1797 int split_huge_page_to_list(struct page *page, struct list_head *list)
1798 {
1799         struct anon_vma *anon_vma;
1800         int ret = 1;
1801
1802         BUG_ON(is_huge_zero_page(page));
1803         BUG_ON(!PageAnon(page));
1804
1805         /*
1806          * The caller does not necessarily hold an mmap_sem that would prevent
1807          * the anon_vma disappearing so we first we take a reference to it
1808          * and then lock the anon_vma for write. This is similar to
1809          * page_lock_anon_vma_read except the write lock is taken to serialise
1810          * against parallel split or collapse operations.
1811          */
1812         anon_vma = page_get_anon_vma(page);
1813         if (!anon_vma)
1814                 goto out;
1815         anon_vma_lock_write(anon_vma);
1816
1817         ret = 0;
1818         if (!PageCompound(page))
1819                 goto out_unlock;
1820
1821         BUG_ON(!PageSwapBacked(page));
1822         __split_huge_page(page, anon_vma, list);
1823         count_vm_event(THP_SPLIT);
1824
1825         BUG_ON(PageCompound(page));
1826 out_unlock:
1827         anon_vma_unlock_write(anon_vma);
1828         put_anon_vma(anon_vma);
1829 out:
1830         return ret;
1831 }
1832
1833 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1834
1835 int hugepage_madvise(struct vm_area_struct *vma,
1836                      unsigned long *vm_flags, int advice)
1837 {
1838         struct mm_struct *mm = vma->vm_mm;
1839
1840         switch (advice) {
1841         case MADV_HUGEPAGE:
1842                 /*
1843                  * Be somewhat over-protective like KSM for now!
1844                  */
1845                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1846                         return -EINVAL;
1847                 if (mm->def_flags & VM_NOHUGEPAGE)
1848                         return -EINVAL;
1849                 *vm_flags &= ~VM_NOHUGEPAGE;
1850                 *vm_flags |= VM_HUGEPAGE;
1851                 /*
1852                  * If the vma become good for khugepaged to scan,
1853                  * register it here without waiting a page fault that
1854                  * may not happen any time soon.
1855                  */
1856                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1857                         return -ENOMEM;
1858                 break;
1859         case MADV_NOHUGEPAGE:
1860                 /*
1861                  * Be somewhat over-protective like KSM for now!
1862                  */
1863                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1864                         return -EINVAL;
1865                 *vm_flags &= ~VM_HUGEPAGE;
1866                 *vm_flags |= VM_NOHUGEPAGE;
1867                 /*
1868                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1869                  * this vma even if we leave the mm registered in khugepaged if
1870                  * it got registered before VM_NOHUGEPAGE was set.
1871                  */
1872                 break;
1873         }
1874
1875         return 0;
1876 }
1877
1878 static int __init khugepaged_slab_init(void)
1879 {
1880         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1881                                           sizeof(struct mm_slot),
1882                                           __alignof__(struct mm_slot), 0, NULL);
1883         if (!mm_slot_cache)
1884                 return -ENOMEM;
1885
1886         return 0;
1887 }
1888
1889 static inline struct mm_slot *alloc_mm_slot(void)
1890 {
1891         if (!mm_slot_cache)     /* initialization failed */
1892                 return NULL;
1893         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1894 }
1895
1896 static inline void free_mm_slot(struct mm_slot *mm_slot)
1897 {
1898         kmem_cache_free(mm_slot_cache, mm_slot);
1899 }
1900
1901 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1902 {
1903         struct mm_slot *mm_slot;
1904
1905         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1906                 if (mm == mm_slot->mm)
1907                         return mm_slot;
1908
1909         return NULL;
1910 }
1911
1912 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1913                                     struct mm_slot *mm_slot)
1914 {
1915         mm_slot->mm = mm;
1916         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1917 }
1918
1919 static inline int khugepaged_test_exit(struct mm_struct *mm)
1920 {
1921         return atomic_read(&mm->mm_users) == 0;
1922 }
1923
1924 int __khugepaged_enter(struct mm_struct *mm)
1925 {
1926         struct mm_slot *mm_slot;
1927         int wakeup;
1928
1929         mm_slot = alloc_mm_slot();
1930         if (!mm_slot)
1931                 return -ENOMEM;
1932
1933         /* __khugepaged_exit() must not run from under us */
1934         VM_BUG_ON(khugepaged_test_exit(mm));
1935         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1936                 free_mm_slot(mm_slot);
1937                 return 0;
1938         }
1939
1940         spin_lock(&khugepaged_mm_lock);
1941         insert_to_mm_slots_hash(mm, mm_slot);
1942         /*
1943          * Insert just behind the scanning cursor, to let the area settle
1944          * down a little.
1945          */
1946         wakeup = list_empty(&khugepaged_scan.mm_head);
1947         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1948         spin_unlock(&khugepaged_mm_lock);
1949
1950         atomic_inc(&mm->mm_count);
1951         if (wakeup)
1952                 wake_up_interruptible(&khugepaged_wait);
1953
1954         return 0;
1955 }
1956
1957 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1958 {
1959         unsigned long hstart, hend;
1960         if (!vma->anon_vma)
1961                 /*
1962                  * Not yet faulted in so we will register later in the
1963                  * page fault if needed.
1964                  */
1965                 return 0;
1966         if (vma->vm_ops)
1967                 /* khugepaged not yet working on file or special mappings */
1968                 return 0;
1969         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1970         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1971         hend = vma->vm_end & HPAGE_PMD_MASK;
1972         if (hstart < hend)
1973                 return khugepaged_enter(vma);
1974         return 0;
1975 }
1976
1977 void __khugepaged_exit(struct mm_struct *mm)
1978 {
1979         struct mm_slot *mm_slot;
1980         int free = 0;
1981
1982         spin_lock(&khugepaged_mm_lock);
1983         mm_slot = get_mm_slot(mm);
1984         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1985                 hash_del(&mm_slot->hash);
1986                 list_del(&mm_slot->mm_node);
1987                 free = 1;
1988         }
1989         spin_unlock(&khugepaged_mm_lock);
1990
1991         if (free) {
1992                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1993                 free_mm_slot(mm_slot);
1994                 mmdrop(mm);
1995         } else if (mm_slot) {
1996                 /*
1997                  * This is required to serialize against
1998                  * khugepaged_test_exit() (which is guaranteed to run
1999                  * under mmap sem read mode). Stop here (after we
2000                  * return all pagetables will be destroyed) until
2001                  * khugepaged has finished working on the pagetables
2002                  * under the mmap_sem.
2003                  */
2004                 down_write(&mm->mmap_sem);
2005                 up_write(&mm->mmap_sem);
2006         }
2007 }
2008
2009 static void release_pte_page(struct page *page)
2010 {
2011         /* 0 stands for page_is_file_cache(page) == false */
2012         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2013         unlock_page(page);
2014         putback_lru_page(page);
2015 }
2016
2017 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2018 {
2019         while (--_pte >= pte) {
2020                 pte_t pteval = *_pte;
2021                 if (!pte_none(pteval))
2022                         release_pte_page(pte_page(pteval));
2023         }
2024 }
2025
2026 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2027                                         unsigned long address,
2028                                         pte_t *pte)
2029 {
2030         struct page *page;
2031         pte_t *_pte;
2032         int referenced = 0, none = 0;
2033         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2034              _pte++, address += PAGE_SIZE) {
2035                 pte_t pteval = *_pte;
2036                 if (pte_none(pteval)) {
2037                         if (++none <= khugepaged_max_ptes_none)
2038                                 continue;
2039                         else
2040                                 goto out;
2041                 }
2042                 if (!pte_present(pteval) || !pte_write(pteval))
2043                         goto out;
2044                 page = vm_normal_page(vma, address, pteval);
2045                 if (unlikely(!page))
2046                         goto out;
2047
2048                 VM_BUG_ON(PageCompound(page));
2049                 BUG_ON(!PageAnon(page));
2050                 VM_BUG_ON(!PageSwapBacked(page));
2051
2052                 /* cannot use mapcount: can't collapse if there's a gup pin */
2053                 if (page_count(page) != 1)
2054                         goto out;
2055                 /*
2056                  * We can do it before isolate_lru_page because the
2057                  * page can't be freed from under us. NOTE: PG_lock
2058                  * is needed to serialize against split_huge_page
2059                  * when invoked from the VM.
2060                  */
2061                 if (!trylock_page(page))
2062                         goto out;
2063                 /*
2064                  * Isolate the page to avoid collapsing an hugepage
2065                  * currently in use by the VM.
2066                  */
2067                 if (isolate_lru_page(page)) {
2068                         unlock_page(page);
2069                         goto out;
2070                 }
2071                 /* 0 stands for page_is_file_cache(page) == false */
2072                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2073                 VM_BUG_ON(!PageLocked(page));
2074                 VM_BUG_ON(PageLRU(page));
2075
2076                 /* If there is no mapped pte young don't collapse the page */
2077                 if (pte_young(pteval) || PageReferenced(page) ||
2078                     mmu_notifier_test_young(vma->vm_mm, address))
2079                         referenced = 1;
2080         }
2081         if (likely(referenced))
2082                 return 1;
2083 out:
2084         release_pte_pages(pte, _pte);
2085         return 0;
2086 }
2087
2088 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2089                                       struct vm_area_struct *vma,
2090                                       unsigned long address,
2091                                       spinlock_t *ptl)
2092 {
2093         pte_t *_pte;
2094         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2095                 pte_t pteval = *_pte;
2096                 struct page *src_page;
2097
2098                 if (pte_none(pteval)) {
2099                         clear_user_highpage(page, address);
2100                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2101                 } else {
2102                         src_page = pte_page(pteval);
2103                         copy_user_highpage(page, src_page, address, vma);
2104                         VM_BUG_ON(page_mapcount(src_page) != 1);
2105                         release_pte_page(src_page);
2106                         /*
2107                          * ptl mostly unnecessary, but preempt has to
2108                          * be disabled to update the per-cpu stats
2109                          * inside page_remove_rmap().
2110                          */
2111                         spin_lock(ptl);
2112                         /*
2113                          * paravirt calls inside pte_clear here are
2114                          * superfluous.
2115                          */
2116                         pte_clear(vma->vm_mm, address, _pte);
2117                         page_remove_rmap(src_page);
2118                         spin_unlock(ptl);
2119                         free_page_and_swap_cache(src_page);
2120                 }
2121
2122                 address += PAGE_SIZE;
2123                 page++;
2124         }
2125 }
2126
2127 static void khugepaged_alloc_sleep(void)
2128 {
2129         wait_event_freezable_timeout(khugepaged_wait, false,
2130                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2131 }
2132
2133 #ifdef CONFIG_NUMA
2134 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2135 {
2136         if (IS_ERR(*hpage)) {
2137                 if (!*wait)
2138                         return false;
2139
2140                 *wait = false;
2141                 *hpage = NULL;
2142                 khugepaged_alloc_sleep();
2143         } else if (*hpage) {
2144                 put_page(*hpage);
2145                 *hpage = NULL;
2146         }
2147
2148         return true;
2149 }
2150
2151 static struct page
2152 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2153                        struct vm_area_struct *vma, unsigned long address,
2154                        int node)
2155 {
2156         VM_BUG_ON(*hpage);
2157         /*
2158          * Allocate the page while the vma is still valid and under
2159          * the mmap_sem read mode so there is no memory allocation
2160          * later when we take the mmap_sem in write mode. This is more
2161          * friendly behavior (OTOH it may actually hide bugs) to
2162          * filesystems in userland with daemons allocating memory in
2163          * the userland I/O paths.  Allocating memory with the
2164          * mmap_sem in read mode is good idea also to allow greater
2165          * scalability.
2166          */
2167         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2168                                       node, __GFP_OTHER_NODE);
2169
2170         /*
2171          * After allocating the hugepage, release the mmap_sem read lock in
2172          * preparation for taking it in write mode.
2173          */
2174         up_read(&mm->mmap_sem);
2175         if (unlikely(!*hpage)) {
2176                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2177                 *hpage = ERR_PTR(-ENOMEM);
2178                 return NULL;
2179         }
2180
2181         count_vm_event(THP_COLLAPSE_ALLOC);
2182         return *hpage;
2183 }
2184 #else
2185 static struct page *khugepaged_alloc_hugepage(bool *wait)
2186 {
2187         struct page *hpage;
2188
2189         do {
2190                 hpage = alloc_hugepage(khugepaged_defrag());
2191                 if (!hpage) {
2192                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2193                         if (!*wait)
2194                                 return NULL;
2195
2196                         *wait = false;
2197                         khugepaged_alloc_sleep();
2198                 } else
2199                         count_vm_event(THP_COLLAPSE_ALLOC);
2200         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2201
2202         return hpage;
2203 }
2204
2205 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2206 {
2207         if (!*hpage)
2208                 *hpage = khugepaged_alloc_hugepage(wait);
2209
2210         if (unlikely(!*hpage))
2211                 return false;
2212
2213         return true;
2214 }
2215
2216 static struct page
2217 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2218                        struct vm_area_struct *vma, unsigned long address,
2219                        int node)
2220 {
2221         up_read(&mm->mmap_sem);
2222         VM_BUG_ON(!*hpage);
2223         return  *hpage;
2224 }
2225 #endif
2226
2227 static bool hugepage_vma_check(struct vm_area_struct *vma)
2228 {
2229         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2230             (vma->vm_flags & VM_NOHUGEPAGE))
2231                 return false;
2232
2233         if (!vma->anon_vma || vma->vm_ops)
2234                 return false;
2235         if (is_vma_temporary_stack(vma))
2236                 return false;
2237         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2238         return true;
2239 }
2240
2241 static void collapse_huge_page(struct mm_struct *mm,
2242                                    unsigned long address,
2243                                    struct page **hpage,
2244                                    struct vm_area_struct *vma,
2245                                    int node)
2246 {
2247         pmd_t *pmd, _pmd;
2248         pte_t *pte;
2249         pgtable_t pgtable;
2250         struct page *new_page;
2251         spinlock_t *ptl;
2252         int isolated;
2253         unsigned long hstart, hend;
2254         unsigned long mmun_start;       /* For mmu_notifiers */
2255         unsigned long mmun_end;         /* For mmu_notifiers */
2256
2257         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2258
2259         /* release the mmap_sem read lock. */
2260         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2261         if (!new_page)
2262                 return;
2263
2264         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2265                 return;
2266
2267         /*
2268          * Prevent all access to pagetables with the exception of
2269          * gup_fast later hanlded by the ptep_clear_flush and the VM
2270          * handled by the anon_vma lock + PG_lock.
2271          */
2272         down_write(&mm->mmap_sem);
2273         if (unlikely(khugepaged_test_exit(mm)))
2274                 goto out;
2275
2276         vma = find_vma(mm, address);
2277         if (!vma)
2278                 goto out;
2279         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2280         hend = vma->vm_end & HPAGE_PMD_MASK;
2281         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2282                 goto out;
2283         if (!hugepage_vma_check(vma))
2284                 goto out;
2285         pmd = mm_find_pmd(mm, address);
2286         if (!pmd)
2287                 goto out;
2288         if (pmd_trans_huge(*pmd))
2289                 goto out;
2290
2291         anon_vma_lock_write(vma->anon_vma);
2292
2293         pte = pte_offset_map(pmd, address);
2294         ptl = pte_lockptr(mm, pmd);
2295
2296         mmun_start = address;
2297         mmun_end   = address + HPAGE_PMD_SIZE;
2298         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2299         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2300         /*
2301          * After this gup_fast can't run anymore. This also removes
2302          * any huge TLB entry from the CPU so we won't allow
2303          * huge and small TLB entries for the same virtual address
2304          * to avoid the risk of CPU bugs in that area.
2305          */
2306         _pmd = pmdp_clear_flush(vma, address, pmd);
2307         spin_unlock(&mm->page_table_lock);
2308         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2309
2310         spin_lock(ptl);
2311         isolated = __collapse_huge_page_isolate(vma, address, pte);
2312         spin_unlock(ptl);
2313
2314         if (unlikely(!isolated)) {
2315                 pte_unmap(pte);
2316                 spin_lock(&mm->page_table_lock);
2317                 BUG_ON(!pmd_none(*pmd));
2318                 /*
2319                  * We can only use set_pmd_at when establishing
2320                  * hugepmds and never for establishing regular pmds that
2321                  * points to regular pagetables. Use pmd_populate for that
2322                  */
2323                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2324                 spin_unlock(&mm->page_table_lock);
2325                 anon_vma_unlock_write(vma->anon_vma);
2326                 goto out;
2327         }
2328
2329         /*
2330          * All pages are isolated and locked so anon_vma rmap
2331          * can't run anymore.
2332          */
2333         anon_vma_unlock_write(vma->anon_vma);
2334
2335         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2336         pte_unmap(pte);
2337         __SetPageUptodate(new_page);
2338         pgtable = pmd_pgtable(_pmd);
2339
2340         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2341         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2342
2343         /*
2344          * spin_lock() below is not the equivalent of smp_wmb(), so
2345          * this is needed to avoid the copy_huge_page writes to become
2346          * visible after the set_pmd_at() write.
2347          */
2348         smp_wmb();
2349
2350         spin_lock(&mm->page_table_lock);
2351         BUG_ON(!pmd_none(*pmd));
2352         page_add_new_anon_rmap(new_page, vma, address);
2353         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2354         set_pmd_at(mm, address, pmd, _pmd);
2355         update_mmu_cache_pmd(vma, address, pmd);
2356         spin_unlock(&mm->page_table_lock);
2357
2358         *hpage = NULL;
2359
2360         khugepaged_pages_collapsed++;
2361 out_up_write:
2362         up_write(&mm->mmap_sem);
2363         return;
2364
2365 out:
2366         mem_cgroup_uncharge_page(new_page);
2367         goto out_up_write;
2368 }
2369
2370 static int khugepaged_scan_pmd(struct mm_struct *mm,
2371                                struct vm_area_struct *vma,
2372                                unsigned long address,
2373                                struct page **hpage)
2374 {
2375         pmd_t *pmd;
2376         pte_t *pte, *_pte;
2377         int ret = 0, referenced = 0, none = 0;
2378         struct page *page;
2379         unsigned long _address;
2380         spinlock_t *ptl;
2381         int node = NUMA_NO_NODE;
2382
2383         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2384
2385         pmd = mm_find_pmd(mm, address);
2386         if (!pmd)
2387                 goto out;
2388         if (pmd_trans_huge(*pmd))
2389                 goto out;
2390
2391         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2392         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2393              _pte++, _address += PAGE_SIZE) {
2394                 pte_t pteval = *_pte;
2395                 if (pte_none(pteval)) {
2396                         if (++none <= khugepaged_max_ptes_none)
2397                                 continue;
2398                         else
2399                                 goto out_unmap;
2400                 }
2401                 if (!pte_present(pteval) || !pte_write(pteval))
2402                         goto out_unmap;
2403                 page = vm_normal_page(vma, _address, pteval);
2404                 if (unlikely(!page))
2405                         goto out_unmap;
2406                 /*
2407                  * Chose the node of the first page. This could
2408                  * be more sophisticated and look at more pages,
2409                  * but isn't for now.
2410                  */
2411                 if (node == NUMA_NO_NODE)
2412                         node = page_to_nid(page);
2413                 VM_BUG_ON(PageCompound(page));
2414                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2415                         goto out_unmap;
2416                 /* cannot use mapcount: can't collapse if there's a gup pin */
2417                 if (page_count(page) != 1)
2418                         goto out_unmap;
2419                 if (pte_young(pteval) || PageReferenced(page) ||
2420                     mmu_notifier_test_young(vma->vm_mm, address))
2421                         referenced = 1;
2422         }
2423         if (referenced)
2424                 ret = 1;
2425 out_unmap:
2426         pte_unmap_unlock(pte, ptl);
2427         if (ret)
2428                 /* collapse_huge_page will return with the mmap_sem released */
2429                 collapse_huge_page(mm, address, hpage, vma, node);
2430 out:
2431         return ret;
2432 }
2433
2434 static void collect_mm_slot(struct mm_slot *mm_slot)
2435 {
2436         struct mm_struct *mm = mm_slot->mm;
2437
2438         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2439
2440         if (khugepaged_test_exit(mm)) {
2441                 /* free mm_slot */
2442                 hash_del(&mm_slot->hash);
2443                 list_del(&mm_slot->mm_node);
2444
2445                 /*
2446                  * Not strictly needed because the mm exited already.
2447                  *
2448                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2449                  */
2450
2451                 /* khugepaged_mm_lock actually not necessary for the below */
2452                 free_mm_slot(mm_slot);
2453                 mmdrop(mm);
2454         }
2455 }
2456
2457 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2458                                             struct page **hpage)
2459         __releases(&khugepaged_mm_lock)
2460         __acquires(&khugepaged_mm_lock)
2461 {
2462         struct mm_slot *mm_slot;
2463         struct mm_struct *mm;
2464         struct vm_area_struct *vma;
2465         int progress = 0;
2466
2467         VM_BUG_ON(!pages);
2468         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2469
2470         if (khugepaged_scan.mm_slot)
2471                 mm_slot = khugepaged_scan.mm_slot;
2472         else {
2473                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2474                                      struct mm_slot, mm_node);
2475                 khugepaged_scan.address = 0;
2476                 khugepaged_scan.mm_slot = mm_slot;
2477         }
2478         spin_unlock(&khugepaged_mm_lock);
2479
2480         mm = mm_slot->mm;
2481         down_read(&mm->mmap_sem);
2482         if (unlikely(khugepaged_test_exit(mm)))
2483                 vma = NULL;
2484         else
2485                 vma = find_vma(mm, khugepaged_scan.address);
2486
2487         progress++;
2488         for (; vma; vma = vma->vm_next) {
2489                 unsigned long hstart, hend;
2490
2491                 cond_resched();
2492                 if (unlikely(khugepaged_test_exit(mm))) {
2493                         progress++;
2494                         break;
2495                 }
2496                 if (!hugepage_vma_check(vma)) {
2497 skip:
2498                         progress++;
2499                         continue;
2500                 }
2501                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2502                 hend = vma->vm_end & HPAGE_PMD_MASK;
2503                 if (hstart >= hend)
2504                         goto skip;
2505                 if (khugepaged_scan.address > hend)
2506                         goto skip;
2507                 if (khugepaged_scan.address < hstart)
2508                         khugepaged_scan.address = hstart;
2509                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2510
2511                 while (khugepaged_scan.address < hend) {
2512                         int ret;
2513                         cond_resched();
2514                         if (unlikely(khugepaged_test_exit(mm)))
2515                                 goto breakouterloop;
2516
2517                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2518                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2519                                   hend);
2520                         ret = khugepaged_scan_pmd(mm, vma,
2521                                                   khugepaged_scan.address,
2522                                                   hpage);
2523                         /* move to next address */
2524                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2525                         progress += HPAGE_PMD_NR;
2526                         if (ret)
2527                                 /* we released mmap_sem so break loop */
2528                                 goto breakouterloop_mmap_sem;
2529                         if (progress >= pages)
2530                                 goto breakouterloop;
2531                 }
2532         }
2533 breakouterloop:
2534         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2535 breakouterloop_mmap_sem:
2536
2537         spin_lock(&khugepaged_mm_lock);
2538         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2539         /*
2540          * Release the current mm_slot if this mm is about to die, or
2541          * if we scanned all vmas of this mm.
2542          */
2543         if (khugepaged_test_exit(mm) || !vma) {
2544                 /*
2545                  * Make sure that if mm_users is reaching zero while
2546                  * khugepaged runs here, khugepaged_exit will find
2547                  * mm_slot not pointing to the exiting mm.
2548                  */
2549                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2550                         khugepaged_scan.mm_slot = list_entry(
2551                                 mm_slot->mm_node.next,
2552                                 struct mm_slot, mm_node);
2553                         khugepaged_scan.address = 0;
2554                 } else {
2555                         khugepaged_scan.mm_slot = NULL;
2556                         khugepaged_full_scans++;
2557                 }
2558
2559                 collect_mm_slot(mm_slot);
2560         }
2561
2562         return progress;
2563 }
2564
2565 static int khugepaged_has_work(void)
2566 {
2567         return !list_empty(&khugepaged_scan.mm_head) &&
2568                 khugepaged_enabled();
2569 }
2570
2571 static int khugepaged_wait_event(void)
2572 {
2573         return !list_empty(&khugepaged_scan.mm_head) ||
2574                 kthread_should_stop();
2575 }
2576
2577 static void khugepaged_do_scan(void)
2578 {
2579         struct page *hpage = NULL;
2580         unsigned int progress = 0, pass_through_head = 0;
2581         unsigned int pages = khugepaged_pages_to_scan;
2582         bool wait = true;
2583
2584         barrier(); /* write khugepaged_pages_to_scan to local stack */
2585
2586         while (progress < pages) {
2587                 if (!khugepaged_prealloc_page(&hpage, &wait))
2588                         break;
2589
2590                 cond_resched();
2591
2592                 if (unlikely(kthread_should_stop() || freezing(current)))
2593                         break;
2594
2595                 spin_lock(&khugepaged_mm_lock);
2596                 if (!khugepaged_scan.mm_slot)
2597                         pass_through_head++;
2598                 if (khugepaged_has_work() &&
2599                     pass_through_head < 2)
2600                         progress += khugepaged_scan_mm_slot(pages - progress,
2601                                                             &hpage);
2602                 else
2603                         progress = pages;
2604                 spin_unlock(&khugepaged_mm_lock);
2605         }
2606
2607         if (!IS_ERR_OR_NULL(hpage))
2608                 put_page(hpage);
2609 }
2610
2611 static void khugepaged_wait_work(void)
2612 {
2613         try_to_freeze();
2614
2615         if (khugepaged_has_work()) {
2616                 if (!khugepaged_scan_sleep_millisecs)
2617                         return;
2618
2619                 wait_event_freezable_timeout(khugepaged_wait,
2620                                              kthread_should_stop(),
2621                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2622                 return;
2623         }
2624
2625         if (khugepaged_enabled())
2626                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2627 }
2628
2629 static int khugepaged(void *none)
2630 {
2631         struct mm_slot *mm_slot;
2632
2633         set_freezable();
2634         set_user_nice(current, 19);
2635
2636         while (!kthread_should_stop()) {
2637                 khugepaged_do_scan();
2638                 khugepaged_wait_work();
2639         }
2640
2641         spin_lock(&khugepaged_mm_lock);
2642         mm_slot = khugepaged_scan.mm_slot;
2643         khugepaged_scan.mm_slot = NULL;
2644         if (mm_slot)
2645                 collect_mm_slot(mm_slot);
2646         spin_unlock(&khugepaged_mm_lock);
2647         return 0;
2648 }
2649
2650 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2651                 unsigned long haddr, pmd_t *pmd)
2652 {
2653         struct mm_struct *mm = vma->vm_mm;
2654         pgtable_t pgtable;
2655         pmd_t _pmd;
2656         int i;
2657
2658         pmdp_clear_flush(vma, haddr, pmd);
2659         /* leave pmd empty until pte is filled */
2660
2661         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2662         pmd_populate(mm, &_pmd, pgtable);
2663
2664         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2665                 pte_t *pte, entry;
2666                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2667                 entry = pte_mkspecial(entry);
2668                 pte = pte_offset_map(&_pmd, haddr);
2669                 VM_BUG_ON(!pte_none(*pte));
2670                 set_pte_at(mm, haddr, pte, entry);
2671                 pte_unmap(pte);
2672         }
2673         smp_wmb(); /* make pte visible before pmd */
2674         pmd_populate(mm, pmd, pgtable);
2675         put_huge_zero_page();
2676 }
2677
2678 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2679                 pmd_t *pmd)
2680 {
2681         struct page *page;
2682         struct mm_struct *mm = vma->vm_mm;
2683         unsigned long haddr = address & HPAGE_PMD_MASK;
2684         unsigned long mmun_start;       /* For mmu_notifiers */
2685         unsigned long mmun_end;         /* For mmu_notifiers */
2686
2687         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2688
2689         mmun_start = haddr;
2690         mmun_end   = haddr + HPAGE_PMD_SIZE;
2691         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2692         spin_lock(&mm->page_table_lock);
2693         if (unlikely(!pmd_trans_huge(*pmd))) {
2694                 spin_unlock(&mm->page_table_lock);
2695                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2696                 return;
2697         }
2698         if (is_huge_zero_pmd(*pmd)) {
2699                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2700                 spin_unlock(&mm->page_table_lock);
2701                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2702                 return;
2703         }
2704         page = pmd_page(*pmd);
2705         VM_BUG_ON(!page_count(page));
2706         get_page(page);
2707         spin_unlock(&mm->page_table_lock);
2708         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2709
2710         split_huge_page(page);
2711
2712         put_page(page);
2713         BUG_ON(pmd_trans_huge(*pmd));
2714 }
2715
2716 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2717                 pmd_t *pmd)
2718 {
2719         struct vm_area_struct *vma;
2720
2721         vma = find_vma(mm, address);
2722         BUG_ON(vma == NULL);
2723         split_huge_page_pmd(vma, address, pmd);
2724 }
2725
2726 static void split_huge_page_address(struct mm_struct *mm,
2727                                     unsigned long address)
2728 {
2729         pmd_t *pmd;
2730
2731         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2732
2733         pmd = mm_find_pmd(mm, address);
2734         if (!pmd)
2735                 return;
2736         /*
2737          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2738          * materialize from under us.
2739          */
2740         split_huge_page_pmd_mm(mm, address, pmd);
2741 }
2742
2743 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2744                              unsigned long start,
2745                              unsigned long end,
2746                              long adjust_next)
2747 {
2748         /*
2749          * If the new start address isn't hpage aligned and it could
2750          * previously contain an hugepage: check if we need to split
2751          * an huge pmd.
2752          */
2753         if (start & ~HPAGE_PMD_MASK &&
2754             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2755             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2756                 split_huge_page_address(vma->vm_mm, start);
2757
2758         /*
2759          * If the new end address isn't hpage aligned and it could
2760          * previously contain an hugepage: check if we need to split
2761          * an huge pmd.
2762          */
2763         if (end & ~HPAGE_PMD_MASK &&
2764             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2765             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2766                 split_huge_page_address(vma->vm_mm, end);
2767
2768         /*
2769          * If we're also updating the vma->vm_next->vm_start, if the new
2770          * vm_next->vm_start isn't page aligned and it could previously
2771          * contain an hugepage: check if we need to split an huge pmd.
2772          */
2773         if (adjust_next > 0) {
2774                 struct vm_area_struct *next = vma->vm_next;
2775                 unsigned long nstart = next->vm_start;
2776                 nstart += adjust_next << PAGE_SHIFT;
2777                 if (nstart & ~HPAGE_PMD_MASK &&
2778                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2779                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2780                         split_huge_page_address(next->vm_mm, nstart);
2781         }
2782 }