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