mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is enabled for all mappings
  31  * and khugepaged scans all mappings. Defrag is only invoked by
  32  * khugepaged hugepage allocations and by page faults inside
  33  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  34  * allocations.
  35  */
  36 unsigned long transparent_hugepage_flags __read_mostly =
  37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  38         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  39 #endif
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  41         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  42 #endif
  43         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  46
  47 /* default scan 8*512 pte (or vmas) every 30 second */
  48 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  49 static unsigned int khugepaged_pages_collapsed;
  50 static unsigned int khugepaged_full_scans;
  51 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  52 /* during fragmentation poll the hugepage allocator once every minute */
  53 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  54 static struct task_struct *khugepaged_thread __read_mostly;
  55 static DEFINE_MUTEX(khugepaged_mutex);
  56 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  57 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  58 /*
  59  * default collapse hugepages if there is at least one pte mapped like
  60  * it would have happened if the vma was large enough during page
  61  * fault.
  62  */
  63 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  64
  65 static int khugepaged(void *none);
  66 static int khugepaged_slab_init(void);
  67
  68 #define MM_SLOTS_HASH_BITS 10
  69 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  70
  71 static struct kmem_cache *mm_slot_cache __read_mostly;
  72
  73 /**
  74  * struct mm_slot - hash lookup from mm to mm_slot
  75  * @hash: hash collision list
  76  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  77  * @mm: the mm that this information is valid for
  78  */
  79 struct mm_slot {
  80         struct hlist_node hash;
  81         struct list_head mm_node;
  82         struct mm_struct *mm;
  83 };
  84
  85 /**
  86  * struct khugepaged_scan - cursor for scanning
  87  * @mm_head: the head of the mm list to scan
  88  * @mm_slot: the current mm_slot we are scanning
  89  * @address: the next address inside that to be scanned
  90  *
  91  * There is only the one khugepaged_scan instance of this cursor structure.
  92  */
  93 struct khugepaged_scan {
  94         struct list_head mm_head;
  95         struct mm_slot *mm_slot;
  96         unsigned long address;
  97 };
  98 static struct khugepaged_scan khugepaged_scan = {
  99         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 100 };
 101
 102
 103 static int set_recommended_min_free_kbytes(void)
 104 {
 105         struct zone *zone;
 106         int nr_zones = 0;
 107         unsigned long recommended_min;
 108         extern int min_free_kbytes;
 109
 110         if (!khugepaged_enabled())
 111                 return 0;
 112
 113         for_each_populated_zone(zone)
 114                 nr_zones++;
 115
 116         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 117         recommended_min = pageblock_nr_pages * nr_zones * 2;
 118
 119         /*
 120          * Make sure that on average at least two pageblocks are almost free
 121          * of another type, one for a migratetype to fall back to and a
 122          * second to avoid subsequent fallbacks of other types There are 3
 123          * MIGRATE_TYPES we care about.
 124          */
 125         recommended_min += pageblock_nr_pages * nr_zones *
 126                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 127
 128         /* don't ever allow to reserve more than 5% of the lowmem */
 129         recommended_min = min(recommended_min,
 130                               (unsigned long) nr_free_buffer_pages() / 20);
 131         recommended_min <<= (PAGE_SHIFT-10);
 132
 133         if (recommended_min > min_free_kbytes)
 134                 min_free_kbytes = recommended_min;
 135         setup_per_zone_wmarks();
 136         return 0;
 137 }
 138 late_initcall(set_recommended_min_free_kbytes);
 139
 140 static int start_khugepaged(void)
 141 {
 142         int err = 0;
 143         if (khugepaged_enabled()) {
 144                 if (!khugepaged_thread)
 145                         khugepaged_thread = kthread_run(khugepaged, NULL,
 146                                                         "khugepaged");
 147                 if (unlikely(IS_ERR(khugepaged_thread))) {
 148                         printk(KERN_ERR
 149                                "khugepaged: kthread_run(khugepaged) failed\n");
 150                         err = PTR_ERR(khugepaged_thread);
 151                         khugepaged_thread = NULL;
 152                 }
 153
 154                 if (!list_empty(&khugepaged_scan.mm_head))
 155                         wake_up_interruptible(&khugepaged_wait);
 156
 157                 set_recommended_min_free_kbytes();
 158         } else if (khugepaged_thread) {
 159                 kthread_stop(khugepaged_thread);
 160                 khugepaged_thread = NULL;
 161         }
 162
 163         return err;
 164 }
 165
 166 static atomic_t huge_zero_refcount;
 167 static unsigned long huge_zero_pfn __read_mostly;
 168
 169 static inline bool is_huge_zero_pfn(unsigned long pfn)
 170 {
 171         unsigned long zero_pfn = ACCESS_ONCE(huge_zero_pfn);
 172         return zero_pfn && pfn == zero_pfn;
 173 }
 174
 175 static inline bool is_huge_zero_pmd(pmd_t pmd)
 176 {
 177         return is_huge_zero_pfn(pmd_pfn(pmd));
 178 }
 179
 180 static unsigned long get_huge_zero_page(void)
 181 {
 182         struct page *zero_page;
 183 retry:
 184         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 185                 return ACCESS_ONCE(huge_zero_pfn);
 186
 187         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 188                         HPAGE_PMD_ORDER);
 189         if (!zero_page) {
 190                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 191                 return 0;
 192         }
 193         count_vm_event(THP_ZERO_PAGE_ALLOC);
 194         preempt_disable();
 195         if (cmpxchg(&huge_zero_pfn, 0, page_to_pfn(zero_page))) {
 196                 preempt_enable();
 197                 __free_page(zero_page);
 198                 goto retry;
 199         }
 200
 201         /* We take additional reference here. It will be put back by shrinker */
 202         atomic_set(&huge_zero_refcount, 2);
 203         preempt_enable();
 204         return ACCESS_ONCE(huge_zero_pfn);
 205 }
 206
 207 static void put_huge_zero_page(void)
 208 {
 209         /*
 210          * Counter should never go to zero here. Only shrinker can put
 211          * last reference.
 212          */
 213         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 214 }
 215
 216 static int shrink_huge_zero_page(struct shrinker *shrink,
 217                 struct shrink_control *sc)
 218 {
 219         if (!sc->nr_to_scan)
 220                 /* we can free zero page only if last reference remains */
 221                 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 222
 223         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 224                 unsigned long zero_pfn = xchg(&huge_zero_pfn, 0);
 225                 BUG_ON(zero_pfn == 0);
 226                 __free_page(__pfn_to_page(zero_pfn));
 227         }
 228
 229         return 0;
 230 }
 231
 232 static struct shrinker huge_zero_page_shrinker = {
 233         .shrink = shrink_huge_zero_page,
 234         .seeks = DEFAULT_SEEKS,
 235 };
 236
 237 #ifdef CONFIG_SYSFS
 238
 239 static ssize_t double_flag_show(struct kobject *kobj,
 240                                 struct kobj_attribute *attr, char *buf,
 241                                 enum transparent_hugepage_flag enabled,
 242                                 enum transparent_hugepage_flag req_madv)
 243 {
 244         if (test_bit(enabled, &transparent_hugepage_flags)) {
 245                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 246                 return sprintf(buf, "[always] madvise never\n");
 247         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 248                 return sprintf(buf, "always [madvise] never\n");
 249         else
 250                 return sprintf(buf, "always madvise [never]\n");
 251 }
 252 static ssize_t double_flag_store(struct kobject *kobj,
 253                                  struct kobj_attribute *attr,
 254                                  const char *buf, size_t count,
 255                                  enum transparent_hugepage_flag enabled,
 256                                  enum transparent_hugepage_flag req_madv)
 257 {
 258         if (!memcmp("always", buf,
 259                     min(sizeof("always")-1, count))) {
 260                 set_bit(enabled, &transparent_hugepage_flags);
 261                 clear_bit(req_madv, &transparent_hugepage_flags);
 262         } else if (!memcmp("madvise", buf,
 263                            min(sizeof("madvise")-1, count))) {
 264                 clear_bit(enabled, &transparent_hugepage_flags);
 265                 set_bit(req_madv, &transparent_hugepage_flags);
 266         } else if (!memcmp("never", buf,
 267                            min(sizeof("never")-1, count))) {
 268                 clear_bit(enabled, &transparent_hugepage_flags);
 269                 clear_bit(req_madv, &transparent_hugepage_flags);
 270         } else
 271                 return -EINVAL;
 272
 273         return count;
 274 }
 275
 276 static ssize_t enabled_show(struct kobject *kobj,
 277                             struct kobj_attribute *attr, char *buf)
 278 {
 279         return double_flag_show(kobj, attr, buf,
 280                                 TRANSPARENT_HUGEPAGE_FLAG,
 281                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 282 }
 283 static ssize_t enabled_store(struct kobject *kobj,
 284                              struct kobj_attribute *attr,
 285                              const char *buf, size_t count)
 286 {
 287         ssize_t ret;
 288
 289         ret = double_flag_store(kobj, attr, buf, count,
 290                                 TRANSPARENT_HUGEPAGE_FLAG,
 291                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 292
 293         if (ret > 0) {
 294                 int err;
 295
 296                 mutex_lock(&khugepaged_mutex);
 297                 err = start_khugepaged();
 298                 mutex_unlock(&khugepaged_mutex);
 299
 300                 if (err)
 301                         ret = err;
 302         }
 303
 304         return ret;
 305 }
 306 static struct kobj_attribute enabled_attr =
 307         __ATTR(enabled, 0644, enabled_show, enabled_store);
 308
 309 static ssize_t single_flag_show(struct kobject *kobj,
 310                                 struct kobj_attribute *attr, char *buf,
 311                                 enum transparent_hugepage_flag flag)
 312 {
 313         return sprintf(buf, "%d\n",
 314                        !!test_bit(flag, &transparent_hugepage_flags));
 315 }
 316
 317 static ssize_t single_flag_store(struct kobject *kobj,
 318                                  struct kobj_attribute *attr,
 319                                  const char *buf, size_t count,
 320                                  enum transparent_hugepage_flag flag)
 321 {
 322         unsigned long value;
 323         int ret;
 324
 325         ret = kstrtoul(buf, 10, &value);
 326         if (ret < 0)
 327                 return ret;
 328         if (value > 1)
 329                 return -EINVAL;
 330
 331         if (value)
 332                 set_bit(flag, &transparent_hugepage_flags);
 333         else
 334                 clear_bit(flag, &transparent_hugepage_flags);
 335
 336         return count;
 337 }
 338
 339 /*
 340  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 341  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 342  * memory just to allocate one more hugepage.
 343  */
 344 static ssize_t defrag_show(struct kobject *kobj,
 345                            struct kobj_attribute *attr, char *buf)
 346 {
 347         return double_flag_show(kobj, attr, buf,
 348                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 349                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 350 }
 351 static ssize_t defrag_store(struct kobject *kobj,
 352                             struct kobj_attribute *attr,
 353                             const char *buf, size_t count)
 354 {
 355         return double_flag_store(kobj, attr, buf, count,
 356                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 357                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 358 }
 359 static struct kobj_attribute defrag_attr =
 360         __ATTR(defrag, 0644, defrag_show, defrag_store);
 361
 362 static ssize_t use_zero_page_show(struct kobject *kobj,
 363                 struct kobj_attribute *attr, char *buf)
 364 {
 365         return single_flag_show(kobj, attr, buf,
 366                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 367 }
 368 static ssize_t use_zero_page_store(struct kobject *kobj,
 369                 struct kobj_attribute *attr, const char *buf, size_t count)
 370 {
 371         return single_flag_store(kobj, attr, buf, count,
 372                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 373 }
 374 static struct kobj_attribute use_zero_page_attr =
 375         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 376 #ifdef CONFIG_DEBUG_VM
 377 static ssize_t debug_cow_show(struct kobject *kobj,
 378                                 struct kobj_attribute *attr, char *buf)
 379 {
 380         return single_flag_show(kobj, attr, buf,
 381                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 382 }
 383 static ssize_t debug_cow_store(struct kobject *kobj,
 384                                struct kobj_attribute *attr,
 385                                const char *buf, size_t count)
 386 {
 387         return single_flag_store(kobj, attr, buf, count,
 388                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 389 }
 390 static struct kobj_attribute debug_cow_attr =
 391         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 392 #endif /* CONFIG_DEBUG_VM */
 393
 394 static struct attribute *hugepage_attr[] = {
 395         &enabled_attr.attr,
 396         &defrag_attr.attr,
 397         &use_zero_page_attr.attr,
 398 #ifdef CONFIG_DEBUG_VM
 399         &debug_cow_attr.attr,
 400 #endif
 401         NULL,
 402 };
 403
 404 static struct attribute_group hugepage_attr_group = {
 405         .attrs = hugepage_attr,
 406 };
 407
 408 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 409                                          struct kobj_attribute *attr,
 410                                          char *buf)
 411 {
 412         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 413 }
 414
 415 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 416                                           struct kobj_attribute *attr,
 417                                           const char *buf, size_t count)
 418 {
 419         unsigned long msecs;
 420         int err;
 421
 422         err = strict_strtoul(buf, 10, &msecs);
 423         if (err || msecs > UINT_MAX)
 424                 return -EINVAL;
 425
 426         khugepaged_scan_sleep_millisecs = msecs;
 427         wake_up_interruptible(&khugepaged_wait);
 428
 429         return count;
 430 }
 431 static struct kobj_attribute scan_sleep_millisecs_attr =
 432         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 433                scan_sleep_millisecs_store);
 434
 435 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 436                                           struct kobj_attribute *attr,
 437                                           char *buf)
 438 {
 439         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 440 }
 441
 442 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 443                                            struct kobj_attribute *attr,
 444                                            const char *buf, size_t count)
 445 {
 446         unsigned long msecs;
 447         int err;
 448
 449         err = strict_strtoul(buf, 10, &msecs);
 450         if (err || msecs > UINT_MAX)
 451                 return -EINVAL;
 452
 453         khugepaged_alloc_sleep_millisecs = msecs;
 454         wake_up_interruptible(&khugepaged_wait);
 455
 456         return count;
 457 }
 458 static struct kobj_attribute alloc_sleep_millisecs_attr =
 459         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 460                alloc_sleep_millisecs_store);
 461
 462 static ssize_t pages_to_scan_show(struct kobject *kobj,
 463                                   struct kobj_attribute *attr,
 464                                   char *buf)
 465 {
 466         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 467 }
 468 static ssize_t pages_to_scan_store(struct kobject *kobj,
 469                                    struct kobj_attribute *attr,
 470                                    const char *buf, size_t count)
 471 {
 472         int err;
 473         unsigned long pages;
 474
 475         err = strict_strtoul(buf, 10, &pages);
 476         if (err || !pages || pages > UINT_MAX)
 477                 return -EINVAL;
 478
 479         khugepaged_pages_to_scan = pages;
 480
 481         return count;
 482 }
 483 static struct kobj_attribute pages_to_scan_attr =
 484         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 485                pages_to_scan_store);
 486
 487 static ssize_t pages_collapsed_show(struct kobject *kobj,
 488                                     struct kobj_attribute *attr,
 489                                     char *buf)
 490 {
 491         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 492 }
 493 static struct kobj_attribute pages_collapsed_attr =
 494         __ATTR_RO(pages_collapsed);
 495
 496 static ssize_t full_scans_show(struct kobject *kobj,
 497                                struct kobj_attribute *attr,
 498                                char *buf)
 499 {
 500         return sprintf(buf, "%u\n", khugepaged_full_scans);
 501 }
 502 static struct kobj_attribute full_scans_attr =
 503         __ATTR_RO(full_scans);
 504
 505 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 506                                       struct kobj_attribute *attr, char *buf)
 507 {
 508         return single_flag_show(kobj, attr, buf,
 509                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 510 }
 511 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 512                                        struct kobj_attribute *attr,
 513                                        const char *buf, size_t count)
 514 {
 515         return single_flag_store(kobj, attr, buf, count,
 516                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 517 }
 518 static struct kobj_attribute khugepaged_defrag_attr =
 519         __ATTR(defrag, 0644, khugepaged_defrag_show,
 520                khugepaged_defrag_store);
 521
 522 /*
 523  * max_ptes_none controls if khugepaged should collapse hugepages over
 524  * any unmapped ptes in turn potentially increasing the memory
 525  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 526  * reduce the available free memory in the system as it
 527  * runs. Increasing max_ptes_none will instead potentially reduce the
 528  * free memory in the system during the khugepaged scan.
 529  */
 530 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 531                                              struct kobj_attribute *attr,
 532                                              char *buf)
 533 {
 534         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 535 }
 536 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 537                                               struct kobj_attribute *attr,
 538                                               const char *buf, size_t count)
 539 {
 540         int err;
 541         unsigned long max_ptes_none;
 542
 543         err = strict_strtoul(buf, 10, &max_ptes_none);
 544         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 545                 return -EINVAL;
 546
 547         khugepaged_max_ptes_none = max_ptes_none;
 548
 549         return count;
 550 }
 551 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 552         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 553                khugepaged_max_ptes_none_store);
 554
 555 static struct attribute *khugepaged_attr[] = {
 556         &khugepaged_defrag_attr.attr,
 557         &khugepaged_max_ptes_none_attr.attr,
 558         &pages_to_scan_attr.attr,
 559         &pages_collapsed_attr.attr,
 560         &full_scans_attr.attr,
 561         &scan_sleep_millisecs_attr.attr,
 562         &alloc_sleep_millisecs_attr.attr,
 563         NULL,
 564 };
 565
 566 static struct attribute_group khugepaged_attr_group = {
 567         .attrs = khugepaged_attr,
 568         .name = "khugepaged",
 569 };
 570
 571 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 572 {
 573         int err;
 574
 575         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 576         if (unlikely(!*hugepage_kobj)) {
 577                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 578                 return -ENOMEM;
 579         }
 580
 581         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 582         if (err) {
 583                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 584                 goto delete_obj;
 585         }
 586
 587         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 588         if (err) {
 589                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 590                 goto remove_hp_group;
 591         }
 592
 593         return 0;
 594
 595 remove_hp_group:
 596         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 597 delete_obj:
 598         kobject_put(*hugepage_kobj);
 599         return err;
 600 }
 601
 602 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 603 {
 604         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 605         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 606         kobject_put(hugepage_kobj);
 607 }
 608 #else
 609 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 610 {
 611         return 0;
 612 }
 613
 614 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 615 {
 616 }
 617 #endif /* CONFIG_SYSFS */
 618
 619 static int __init hugepage_init(void)
 620 {
 621         int err;
 622         struct kobject *hugepage_kobj;
 623
 624         if (!has_transparent_hugepage()) {
 625                 transparent_hugepage_flags = 0;
 626                 return -EINVAL;
 627         }
 628
 629         err = hugepage_init_sysfs(&hugepage_kobj);
 630         if (err)
 631                 return err;
 632
 633         err = khugepaged_slab_init();
 634         if (err)
 635                 goto out;
 636
 637         register_shrinker(&huge_zero_page_shrinker);
 638
 639         /*
 640          * By default disable transparent hugepages on smaller systems,
 641          * where the extra memory used could hurt more than TLB overhead
 642          * is likely to save.  The admin can still enable it through /sys.
 643          */
 644         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 645                 transparent_hugepage_flags = 0;
 646
 647         start_khugepaged();
 648
 649         return 0;
 650 out:
 651         hugepage_exit_sysfs(hugepage_kobj);
 652         return err;
 653 }
 654 module_init(hugepage_init)
 655
 656 static int __init setup_transparent_hugepage(char *str)
 657 {
 658         int ret = 0;
 659         if (!str)
 660                 goto out;
 661         if (!strcmp(str, "always")) {
 662                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 663                         &transparent_hugepage_flags);
 664                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 665                           &transparent_hugepage_flags);
 666                 ret = 1;
 667         } else if (!strcmp(str, "madvise")) {
 668                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 669                           &transparent_hugepage_flags);
 670                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 671                         &transparent_hugepage_flags);
 672                 ret = 1;
 673         } else if (!strcmp(str, "never")) {
 674                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 675                           &transparent_hugepage_flags);
 676                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 677                           &transparent_hugepage_flags);
 678                 ret = 1;
 679         }
 680 out:
 681         if (!ret)
 682                 printk(KERN_WARNING
 683                        "transparent_hugepage= cannot parse, ignored\n");
 684         return ret;
 685 }
 686 __setup("transparent_hugepage=", setup_transparent_hugepage);
 687
 688 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 689 {
 690         if (likely(vma->vm_flags & VM_WRITE))
 691                 pmd = pmd_mkwrite(pmd);
 692         return pmd;
 693 }
 694
 695 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
 696 {
 697         pmd_t entry;
 698         entry = mk_pmd(page, vma->vm_page_prot);
 699         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 700         entry = pmd_mkhuge(entry);
 701         return entry;
 702 }
 703
 704 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 705                                         struct vm_area_struct *vma,
 706                                         unsigned long haddr, pmd_t *pmd,
 707                                         struct page *page)
 708 {
 709         pgtable_t pgtable;
 710
 711         VM_BUG_ON(!PageCompound(page));
 712         pgtable = pte_alloc_one(mm, haddr);
 713         if (unlikely(!pgtable))
 714                 return VM_FAULT_OOM;
 715
 716         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 717         __SetPageUptodate(page);
 718
 719         spin_lock(&mm->page_table_lock);
 720         if (unlikely(!pmd_none(*pmd))) {
 721                 spin_unlock(&mm->page_table_lock);
 722                 mem_cgroup_uncharge_page(page);
 723                 put_page(page);
 724                 pte_free(mm, pgtable);
 725         } else {
 726                 pmd_t entry;
 727                 entry = mk_huge_pmd(page, vma);
 728                 /*
 729                  * The spinlocking to take the lru_lock inside
 730                  * page_add_new_anon_rmap() acts as a full memory
 731                  * barrier to be sure clear_huge_page writes become
 732                  * visible after the set_pmd_at() write.
 733                  */
 734                 page_add_new_anon_rmap(page, vma, haddr);
 735                 set_pmd_at(mm, haddr, pmd, entry);
 736                 pgtable_trans_huge_deposit(mm, pgtable);
 737                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 738                 mm->nr_ptes++;
 739                 spin_unlock(&mm->page_table_lock);
 740         }
 741
 742         return 0;
 743 }
 744
 745 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 746 {
 747         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 748 }
 749
 750 static inline struct page *alloc_hugepage_vma(int defrag,
 751                                               struct vm_area_struct *vma,
 752                                               unsigned long haddr, int nd,
 753                                               gfp_t extra_gfp)
 754 {
 755         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 756                                HPAGE_PMD_ORDER, vma, haddr, nd);
 757 }
 758
 759 #ifndef CONFIG_NUMA
 760 static inline struct page *alloc_hugepage(int defrag)
 761 {
 762         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 763                            HPAGE_PMD_ORDER);
 764 }
 765 #endif
 766
 767 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 768                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 769                 unsigned long zero_pfn)
 770 {
 771         pmd_t entry;
 772         if (!pmd_none(*pmd))
 773                 return false;
 774         entry = pfn_pmd(zero_pfn, vma->vm_page_prot);
 775         entry = pmd_wrprotect(entry);
 776         entry = pmd_mkhuge(entry);
 777         set_pmd_at(mm, haddr, pmd, entry);
 778         pgtable_trans_huge_deposit(mm, pgtable);
 779         mm->nr_ptes++;
 780         return true;
 781 }
 782
 783 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 784                                unsigned long address, pmd_t *pmd,
 785                                unsigned int flags)
 786 {
 787         struct page *page;
 788         unsigned long haddr = address & HPAGE_PMD_MASK;
 789         pte_t *pte;
 790
 791         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 792                 if (unlikely(anon_vma_prepare(vma)))
 793                         return VM_FAULT_OOM;
 794                 if (unlikely(khugepaged_enter(vma)))
 795                         return VM_FAULT_OOM;
 796                 if (!(flags & FAULT_FLAG_WRITE) &&
 797                                 transparent_hugepage_use_zero_page()) {
 798                         pgtable_t pgtable;
 799                         unsigned long zero_pfn;
 800                         bool set;
 801                         pgtable = pte_alloc_one(mm, haddr);
 802                         if (unlikely(!pgtable))
 803                                 return VM_FAULT_OOM;
 804                         zero_pfn = get_huge_zero_page();
 805                         if (unlikely(!zero_pfn)) {
 806                                 pte_free(mm, pgtable);
 807                                 count_vm_event(THP_FAULT_FALLBACK);
 808                                 goto out;
 809                         }
 810                         spin_lock(&mm->page_table_lock);
 811                         set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 812                                         zero_pfn);
 813                         spin_unlock(&mm->page_table_lock);
 814                         if (!set) {
 815                                 pte_free(mm, pgtable);
 816                                 put_huge_zero_page();
 817                         }
 818                         return 0;
 819                 }
 820                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 821                                           vma, haddr, numa_node_id(), 0);
 822                 if (unlikely(!page)) {
 823                         count_vm_event(THP_FAULT_FALLBACK);
 824                         goto out;
 825                 }
 826                 count_vm_event(THP_FAULT_ALLOC);
 827                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 828                         put_page(page);
 829                         goto out;
 830                 }
 831                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 832                                                           page))) {
 833                         mem_cgroup_uncharge_page(page);
 834                         put_page(page);
 835                         goto out;
 836                 }
 837
 838                 return 0;
 839         }
 840 out:
 841         /*
 842          * Use __pte_alloc instead of pte_alloc_map, because we can't
 843          * run pte_offset_map on the pmd, if an huge pmd could
 844          * materialize from under us from a different thread.
 845          */
 846         if (unlikely(pmd_none(*pmd)) &&
 847             unlikely(__pte_alloc(mm, vma, pmd, address)))
 848                 return VM_FAULT_OOM;
 849         /* if an huge pmd materialized from under us just retry later */
 850         if (unlikely(pmd_trans_huge(*pmd)))
 851                 return 0;
 852         /*
 853          * A regular pmd is established and it can't morph into a huge pmd
 854          * from under us anymore at this point because we hold the mmap_sem
 855          * read mode and khugepaged takes it in write mode. So now it's
 856          * safe to run pte_offset_map().
 857          */
 858         pte = pte_offset_map(pmd, address);
 859         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 860 }
 861
 862 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 863                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 864                   struct vm_area_struct *vma)
 865 {
 866         struct page *src_page;
 867         pmd_t pmd;
 868         pgtable_t pgtable;
 869         int ret;
 870
 871         ret = -ENOMEM;
 872         pgtable = pte_alloc_one(dst_mm, addr);
 873         if (unlikely(!pgtable))
 874                 goto out;
 875
 876         spin_lock(&dst_mm->page_table_lock);
 877         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 878
 879         ret = -EAGAIN;
 880         pmd = *src_pmd;
 881         if (unlikely(!pmd_trans_huge(pmd))) {
 882                 pte_free(dst_mm, pgtable);
 883                 goto out_unlock;
 884         }
 885         /*
 886          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 887          * under splitting since we don't split the page itself, only pmd to
 888          * a page table.
 889          */
 890         if (is_huge_zero_pmd(pmd)) {
 891                 unsigned long zero_pfn;
 892                 bool set;
 893                 /*
 894                  * get_huge_zero_page() will never allocate a new page here,
 895                  * since we already have a zero page to copy. It just takes a
 896                  * reference.
 897                  */
 898                 zero_pfn = get_huge_zero_page();
 899                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 900                                 zero_pfn);
 901                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 902                 ret = 0;
 903                 goto out_unlock;
 904         }
 905         if (unlikely(pmd_trans_splitting(pmd))) {
 906                 /* split huge page running from under us */
 907                 spin_unlock(&src_mm->page_table_lock);
 908                 spin_unlock(&dst_mm->page_table_lock);
 909                 pte_free(dst_mm, pgtable);
 910
 911                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 912                 goto out;
 913         }
 914         src_page = pmd_page(pmd);
 915         VM_BUG_ON(!PageHead(src_page));
 916         get_page(src_page);
 917         page_dup_rmap(src_page);
 918         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 919
 920         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 921         pmd = pmd_mkold(pmd_wrprotect(pmd));
 922         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 923         pgtable_trans_huge_deposit(dst_mm, pgtable);
 924         dst_mm->nr_ptes++;
 925
 926         ret = 0;
 927 out_unlock:
 928         spin_unlock(&src_mm->page_table_lock);
 929         spin_unlock(&dst_mm->page_table_lock);
 930 out:
 931         return ret;
 932 }
 933
 934 void huge_pmd_set_accessed(struct mm_struct *mm,
 935                            struct vm_area_struct *vma,
 936                            unsigned long address,
 937                            pmd_t *pmd, pmd_t orig_pmd,
 938                            int dirty)
 939 {
 940         pmd_t entry;
 941         unsigned long haddr;
 942
 943         spin_lock(&mm->page_table_lock);
 944         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 945                 goto unlock;
 946
 947         entry = pmd_mkyoung(orig_pmd);
 948         haddr = address & HPAGE_PMD_MASK;
 949         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 950                 update_mmu_cache_pmd(vma, address, pmd);
 951
 952 unlock:
 953         spin_unlock(&mm->page_table_lock);
 954 }
 955
 956 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 957                 struct vm_area_struct *vma, unsigned long address,
 958                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 959 {
 960         pgtable_t pgtable;
 961         pmd_t _pmd;
 962         struct page *page;
 963         int i, ret = 0;
 964         unsigned long mmun_start;       /* For mmu_notifiers */
 965         unsigned long mmun_end;         /* For mmu_notifiers */
 966
 967         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 968         if (!page) {
 969                 ret |= VM_FAULT_OOM;
 970                 goto out;
 971         }
 972
 973         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 974                 put_page(page);
 975                 ret |= VM_FAULT_OOM;
 976                 goto out;
 977         }
 978
 979         clear_user_highpage(page, address);
 980         __SetPageUptodate(page);
 981
 982         mmun_start = haddr;
 983         mmun_end   = haddr + HPAGE_PMD_SIZE;
 984         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 985
 986         spin_lock(&mm->page_table_lock);
 987         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 988                 goto out_free_page;
 989
 990         pmdp_clear_flush(vma, haddr, pmd);
 991         /* leave pmd empty until pte is filled */
 992
 993         pgtable = pgtable_trans_huge_withdraw(mm);
 994         pmd_populate(mm, &_pmd, pgtable);
 995
 996         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 997                 pte_t *pte, entry;
 998                 if (haddr == (address & PAGE_MASK)) {
 999                         entry = mk_pte(page, vma->vm_page_prot);
1000                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1001                         page_add_new_anon_rmap(page, vma, haddr);
1002                 } else {
1003                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1004                         entry = pte_mkspecial(entry);
1005                 }
1006                 pte = pte_offset_map(&_pmd, haddr);
1007                 VM_BUG_ON(!pte_none(*pte));
1008                 set_pte_at(mm, haddr, pte, entry);
1009                 pte_unmap(pte);
1010         }
1011         smp_wmb(); /* make pte visible before pmd */
1012         pmd_populate(mm, pmd, pgtable);
1013         spin_unlock(&mm->page_table_lock);
1014         put_huge_zero_page();
1015         inc_mm_counter(mm, MM_ANONPAGES);
1016
1017         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1018
1019         ret |= VM_FAULT_WRITE;
1020 out:
1021         return ret;
1022 out_free_page:
1023         spin_unlock(&mm->page_table_lock);
1024         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1025         mem_cgroup_uncharge_page(page);
1026         put_page(page);
1027         goto out;
1028 }
1029
1030 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1031                                         struct vm_area_struct *vma,
1032                                         unsigned long address,
1033                                         pmd_t *pmd, pmd_t orig_pmd,
1034                                         struct page *page,
1035                                         unsigned long haddr)
1036 {
1037         pgtable_t pgtable;
1038         pmd_t _pmd;
1039         int ret = 0, i;
1040         struct page **pages;
1041         unsigned long mmun_start;       /* For mmu_notifiers */
1042         unsigned long mmun_end;         /* For mmu_notifiers */
1043
1044         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1045                         GFP_KERNEL);
1046         if (unlikely(!pages)) {
1047                 ret |= VM_FAULT_OOM;
1048                 goto out;
1049         }
1050
1051         for (i = 0; i < HPAGE_PMD_NR; i++) {
1052                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1053                                                __GFP_OTHER_NODE,
1054                                                vma, address, page_to_nid(page));
1055                 if (unlikely(!pages[i] ||
1056                              mem_cgroup_newpage_charge(pages[i], mm,
1057                                                        GFP_KERNEL))) {
1058                         if (pages[i])
1059                                 put_page(pages[i]);
1060                         mem_cgroup_uncharge_start();
1061                         while (--i >= 0) {
1062                                 mem_cgroup_uncharge_page(pages[i]);
1063                                 put_page(pages[i]);
1064                         }
1065                         mem_cgroup_uncharge_end();
1066                         kfree(pages);
1067                         ret |= VM_FAULT_OOM;
1068                         goto out;
1069                 }
1070         }
1071
1072         for (i = 0; i < HPAGE_PMD_NR; i++) {
1073                 copy_user_highpage(pages[i], page + i,
1074                                    haddr + PAGE_SIZE * i, vma);
1075                 __SetPageUptodate(pages[i]);
1076                 cond_resched();
1077         }
1078
1079         mmun_start = haddr;
1080         mmun_end   = haddr + HPAGE_PMD_SIZE;
1081         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1082
1083         spin_lock(&mm->page_table_lock);
1084         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1085                 goto out_free_pages;
1086         VM_BUG_ON(!PageHead(page));
1087
1088         pmdp_clear_flush(vma, haddr, pmd);
1089         /* leave pmd empty until pte is filled */
1090
1091         pgtable = pgtable_trans_huge_withdraw(mm);
1092         pmd_populate(mm, &_pmd, pgtable);
1093
1094         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1095                 pte_t *pte, entry;
1096                 entry = mk_pte(pages[i], vma->vm_page_prot);
1097                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1098                 page_add_new_anon_rmap(pages[i], vma, haddr);
1099                 pte = pte_offset_map(&_pmd, haddr);
1100                 VM_BUG_ON(!pte_none(*pte));
1101                 set_pte_at(mm, haddr, pte, entry);
1102                 pte_unmap(pte);
1103         }
1104         kfree(pages);
1105
1106         smp_wmb(); /* make pte visible before pmd */
1107         pmd_populate(mm, pmd, pgtable);
1108         page_remove_rmap(page);
1109         spin_unlock(&mm->page_table_lock);
1110
1111         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1112
1113         ret |= VM_FAULT_WRITE;
1114         put_page(page);
1115
1116 out:
1117         return ret;
1118
1119 out_free_pages:
1120         spin_unlock(&mm->page_table_lock);
1121         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1122         mem_cgroup_uncharge_start();
1123         for (i = 0; i < HPAGE_PMD_NR; i++) {
1124                 mem_cgroup_uncharge_page(pages[i]);
1125                 put_page(pages[i]);
1126         }
1127         mem_cgroup_uncharge_end();
1128         kfree(pages);
1129         goto out;
1130 }
1131
1132 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1133                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1134 {
1135         int ret = 0;
1136         struct page *page = NULL, *new_page;
1137         unsigned long haddr;
1138         unsigned long mmun_start;       /* For mmu_notifiers */
1139         unsigned long mmun_end;         /* For mmu_notifiers */
1140
1141         VM_BUG_ON(!vma->anon_vma);
1142         haddr = address & HPAGE_PMD_MASK;
1143         if (is_huge_zero_pmd(orig_pmd))
1144                 goto alloc;
1145         spin_lock(&mm->page_table_lock);
1146         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1147                 goto out_unlock;
1148
1149         page = pmd_page(orig_pmd);
1150         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1151         if (page_mapcount(page) == 1) {
1152                 pmd_t entry;
1153                 entry = pmd_mkyoung(orig_pmd);
1154                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1155                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1156                         update_mmu_cache_pmd(vma, address, pmd);
1157                 ret |= VM_FAULT_WRITE;
1158                 goto out_unlock;
1159         }
1160         get_page(page);
1161         spin_unlock(&mm->page_table_lock);
1162 alloc:
1163         if (transparent_hugepage_enabled(vma) &&
1164             !transparent_hugepage_debug_cow())
1165                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1166                                               vma, haddr, numa_node_id(), 0);
1167         else
1168                 new_page = NULL;
1169
1170         if (unlikely(!new_page)) {
1171                 count_vm_event(THP_FAULT_FALLBACK);
1172                 if (is_huge_zero_pmd(orig_pmd)) {
1173                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1174                                         address, pmd, orig_pmd, haddr);
1175                 } else {
1176                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1177                                         pmd, orig_pmd, page, haddr);
1178                         if (ret & VM_FAULT_OOM)
1179                                 split_huge_page(page);
1180                         put_page(page);
1181                 }
1182                 goto out;
1183         }
1184         count_vm_event(THP_FAULT_ALLOC);
1185
1186         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1187                 put_page(new_page);
1188                 if (page) {
1189                         split_huge_page(page);
1190                         put_page(page);
1191                 }
1192                 ret |= VM_FAULT_OOM;
1193                 goto out;
1194         }
1195
1196         if (is_huge_zero_pmd(orig_pmd))
1197                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1198         else
1199                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1200         __SetPageUptodate(new_page);
1201
1202         mmun_start = haddr;
1203         mmun_end   = haddr + HPAGE_PMD_SIZE;
1204         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1205
1206         spin_lock(&mm->page_table_lock);
1207         if (page)
1208                 put_page(page);
1209         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1210                 spin_unlock(&mm->page_table_lock);
1211                 mem_cgroup_uncharge_page(new_page);
1212                 put_page(new_page);
1213                 goto out_mn;
1214         } else {
1215                 pmd_t entry;
1216                 entry = mk_huge_pmd(new_page, vma);
1217                 pmdp_clear_flush(vma, haddr, pmd);
1218                 page_add_new_anon_rmap(new_page, vma, haddr);
1219                 set_pmd_at(mm, haddr, pmd, entry);
1220                 update_mmu_cache_pmd(vma, address, pmd);
1221                 if (is_huge_zero_pmd(orig_pmd)) {
1222                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1223                         put_huge_zero_page();
1224                 } else {
1225                         VM_BUG_ON(!PageHead(page));
1226                         page_remove_rmap(page);
1227                         put_page(page);
1228                 }
1229                 ret |= VM_FAULT_WRITE;
1230         }
1231         spin_unlock(&mm->page_table_lock);
1232 out_mn:
1233         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1234 out:
1235         return ret;
1236 out_unlock:
1237         spin_unlock(&mm->page_table_lock);
1238         return ret;
1239 }
1240
1241 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1242                                    unsigned long addr,
1243                                    pmd_t *pmd,
1244                                    unsigned int flags)
1245 {
1246         struct mm_struct *mm = vma->vm_mm;
1247         struct page *page = NULL;
1248
1249         assert_spin_locked(&mm->page_table_lock);
1250
1251         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1252                 goto out;
1253
1254         /* Avoid dumping huge zero page */
1255         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1256                 return ERR_PTR(-EFAULT);
1257
1258         page = pmd_page(*pmd);
1259         VM_BUG_ON(!PageHead(page));
1260         if (flags & FOLL_TOUCH) {
1261                 pmd_t _pmd;
1262                 /*
1263                  * We should set the dirty bit only for FOLL_WRITE but
1264                  * for now the dirty bit in the pmd is meaningless.
1265                  * And if the dirty bit will become meaningful and
1266                  * we'll only set it with FOLL_WRITE, an atomic
1267                  * set_bit will be required on the pmd to set the
1268                  * young bit, instead of the current set_pmd_at.
1269                  */
1270                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1271                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1272         }
1273         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1274                 if (page->mapping && trylock_page(page)) {
1275                         lru_add_drain();
1276                         if (page->mapping)
1277                                 mlock_vma_page(page);
1278                         unlock_page(page);
1279                 }
1280         }
1281         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1282         VM_BUG_ON(!PageCompound(page));
1283         if (flags & FOLL_GET)
1284                 get_page_foll(page);
1285
1286 out:
1287         return page;
1288 }
1289
1290 /* NUMA hinting page fault entry point for trans huge pmds */
1291 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1292                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1293 {
1294         struct page *page;
1295         unsigned long haddr = addr & HPAGE_PMD_MASK;
1296         int target_nid;
1297         int current_nid = -1;
1298         bool migrated;
1299         bool page_locked = false;
1300
1301         spin_lock(&mm->page_table_lock);
1302         if (unlikely(!pmd_same(pmd, *pmdp)))
1303                 goto out_unlock;
1304
1305         page = pmd_page(pmd);
1306         get_page(page);
1307         current_nid = page_to_nid(page);
1308         count_vm_numa_event(NUMA_HINT_FAULTS);
1309         if (current_nid == numa_node_id())
1310                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1311
1312         target_nid = mpol_misplaced(page, vma, haddr);
1313         if (target_nid == -1) {
1314                 put_page(page);
1315                 goto clear_pmdnuma;
1316         }
1317
1318         /* Acquire the page lock to serialise THP migrations */
1319         spin_unlock(&mm->page_table_lock);
1320         lock_page(page);
1321         page_locked = true;
1322
1323         /* Confirm the PTE did not while locked */
1324         spin_lock(&mm->page_table_lock);
1325         if (unlikely(!pmd_same(pmd, *pmdp))) {
1326                 unlock_page(page);
1327                 put_page(page);
1328                 goto out_unlock;
1329         }
1330         spin_unlock(&mm->page_table_lock);
1331
1332         /* Migrate the THP to the requested node */
1333         migrated = migrate_misplaced_transhuge_page(mm, vma,
1334                                 pmdp, pmd, addr,
1335                                 page, target_nid);
1336         if (migrated)
1337                 current_nid = target_nid;
1338         else {
1339                 spin_lock(&mm->page_table_lock);
1340                 if (unlikely(!pmd_same(pmd, *pmdp))) {
1341                         unlock_page(page);
1342                         goto out_unlock;
1343                 }
1344                 goto clear_pmdnuma;
1345         }
1346
1347         task_numa_fault(current_nid, HPAGE_PMD_NR, migrated);
1348         return 0;
1349
1350 clear_pmdnuma:
1351         pmd = pmd_mknonnuma(pmd);
1352         set_pmd_at(mm, haddr, pmdp, pmd);
1353         VM_BUG_ON(pmd_numa(*pmdp));
1354         update_mmu_cache_pmd(vma, addr, pmdp);
1355         if (page_locked)
1356                 unlock_page(page);
1357
1358 out_unlock:
1359         spin_unlock(&mm->page_table_lock);
1360         if (current_nid != -1)
1361                 task_numa_fault(current_nid, HPAGE_PMD_NR, migrated);
1362         return 0;
1363 }
1364
1365 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1366                  pmd_t *pmd, unsigned long addr)
1367 {
1368         int ret = 0;
1369
1370         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1371                 struct page *page;
1372                 pgtable_t pgtable;
1373                 pmd_t orig_pmd;
1374                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1375                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1376                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1377                 if (is_huge_zero_pmd(orig_pmd)) {
1378                         tlb->mm->nr_ptes--;
1379                         spin_unlock(&tlb->mm->page_table_lock);
1380                         put_huge_zero_page();
1381                 } else {
1382                         page = pmd_page(orig_pmd);
1383                         page_remove_rmap(page);
1384                         VM_BUG_ON(page_mapcount(page) < 0);
1385                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1386                         VM_BUG_ON(!PageHead(page));
1387                         tlb->mm->nr_ptes--;
1388                         spin_unlock(&tlb->mm->page_table_lock);
1389                         tlb_remove_page(tlb, page);
1390                 }
1391                 pte_free(tlb->mm, pgtable);
1392                 ret = 1;
1393         }
1394         return ret;
1395 }
1396
1397 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1398                 unsigned long addr, unsigned long end,
1399                 unsigned char *vec)
1400 {
1401         int ret = 0;
1402
1403         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1404                 /*
1405                  * All logical pages in the range are present
1406                  * if backed by a huge page.
1407                  */
1408                 spin_unlock(&vma->vm_mm->page_table_lock);
1409                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1410                 ret = 1;
1411         }
1412
1413         return ret;
1414 }
1415
1416 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1417                   unsigned long old_addr,
1418                   unsigned long new_addr, unsigned long old_end,
1419                   pmd_t *old_pmd, pmd_t *new_pmd)
1420 {
1421         int ret = 0;
1422         pmd_t pmd;
1423
1424         struct mm_struct *mm = vma->vm_mm;
1425
1426         if ((old_addr & ~HPAGE_PMD_MASK) ||
1427             (new_addr & ~HPAGE_PMD_MASK) ||
1428             old_end - old_addr < HPAGE_PMD_SIZE ||
1429             (new_vma->vm_flags & VM_NOHUGEPAGE))
1430                 goto out;
1431
1432         /*
1433          * The destination pmd shouldn't be established, free_pgtables()
1434          * should have release it.
1435          */
1436         if (WARN_ON(!pmd_none(*new_pmd))) {
1437                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1438                 goto out;
1439         }
1440
1441         ret = __pmd_trans_huge_lock(old_pmd, vma);
1442         if (ret == 1) {
1443                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1444                 VM_BUG_ON(!pmd_none(*new_pmd));
1445                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1446                 spin_unlock(&mm->page_table_lock);
1447         }
1448 out:
1449         return ret;
1450 }
1451
1452 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1453                 unsigned long addr, pgprot_t newprot, int prot_numa)
1454 {
1455         struct mm_struct *mm = vma->vm_mm;
1456         int ret = 0;
1457
1458         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1459                 pmd_t entry;
1460                 entry = pmdp_get_and_clear(mm, addr, pmd);
1461                 if (!prot_numa) {
1462                         entry = pmd_modify(entry, newprot);
1463                         BUG_ON(pmd_write(entry));
1464                 } else {
1465                         struct page *page = pmd_page(*pmd);
1466
1467                         /* only check non-shared pages */
1468                         if (page_mapcount(page) == 1 &&
1469                             !pmd_numa(*pmd)) {
1470                                 entry = pmd_mknuma(entry);
1471                         }
1472                 }
1473                 set_pmd_at(mm, addr, pmd, entry);
1474                 spin_unlock(&vma->vm_mm->page_table_lock);
1475                 ret = 1;
1476         }
1477
1478         return ret;
1479 }
1480
1481 /*
1482  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1483  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1484  *
1485  * Note that if it returns 1, this routine returns without unlocking page
1486  * table locks. So callers must unlock them.
1487  */
1488 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1489 {
1490         spin_lock(&vma->vm_mm->page_table_lock);
1491         if (likely(pmd_trans_huge(*pmd))) {
1492                 if (unlikely(pmd_trans_splitting(*pmd))) {
1493                         spin_unlock(&vma->vm_mm->page_table_lock);
1494                         wait_split_huge_page(vma->anon_vma, pmd);
1495                         return -1;
1496                 } else {
1497                         /* Thp mapped by 'pmd' is stable, so we can
1498                          * handle it as it is. */
1499                         return 1;
1500                 }
1501         }
1502         spin_unlock(&vma->vm_mm->page_table_lock);
1503         return 0;
1504 }
1505
1506 pmd_t *page_check_address_pmd(struct page *page,
1507                               struct mm_struct *mm,
1508                               unsigned long address,
1509                               enum page_check_address_pmd_flag flag)
1510 {
1511         pmd_t *pmd, *ret = NULL;
1512
1513         if (address & ~HPAGE_PMD_MASK)
1514                 goto out;
1515
1516         pmd = mm_find_pmd(mm, address);
1517         if (!pmd)
1518                 goto out;
1519         if (pmd_none(*pmd))
1520                 goto out;
1521         if (pmd_page(*pmd) != page)
1522                 goto out;
1523         /*
1524          * split_vma() may create temporary aliased mappings. There is
1525          * no risk as long as all huge pmd are found and have their
1526          * splitting bit set before __split_huge_page_refcount
1527          * runs. Finding the same huge pmd more than once during the
1528          * same rmap walk is not a problem.
1529          */
1530         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1531             pmd_trans_splitting(*pmd))
1532                 goto out;
1533         if (pmd_trans_huge(*pmd)) {
1534                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1535                           !pmd_trans_splitting(*pmd));
1536                 ret = pmd;
1537         }
1538 out:
1539         return ret;
1540 }
1541
1542 static int __split_huge_page_splitting(struct page *page,
1543                                        struct vm_area_struct *vma,
1544                                        unsigned long address)
1545 {
1546         struct mm_struct *mm = vma->vm_mm;
1547         pmd_t *pmd;
1548         int ret = 0;
1549         /* For mmu_notifiers */
1550         const unsigned long mmun_start = address;
1551         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1552
1553         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1554         spin_lock(&mm->page_table_lock);
1555         pmd = page_check_address_pmd(page, mm, address,
1556                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1557         if (pmd) {
1558                 /*
1559                  * We can't temporarily set the pmd to null in order
1560                  * to split it, the pmd must remain marked huge at all
1561                  * times or the VM won't take the pmd_trans_huge paths
1562                  * and it won't wait on the anon_vma->root->rwsem to
1563                  * serialize against split_huge_page*.
1564                  */
1565                 pmdp_splitting_flush(vma, address, pmd);
1566                 ret = 1;
1567         }
1568         spin_unlock(&mm->page_table_lock);
1569         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1570
1571         return ret;
1572 }
1573
1574 static void __split_huge_page_refcount(struct page *page)
1575 {
1576         int i;
1577         struct zone *zone = page_zone(page);
1578         struct lruvec *lruvec;
1579         int tail_count = 0;
1580
1581         /* prevent PageLRU to go away from under us, and freeze lru stats */
1582         spin_lock_irq(&zone->lru_lock);
1583         lruvec = mem_cgroup_page_lruvec(page, zone);
1584
1585         compound_lock(page);
1586         /* complete memcg works before add pages to LRU */
1587         mem_cgroup_split_huge_fixup(page);
1588
1589         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1590                 struct page *page_tail = page + i;
1591
1592                 /* tail_page->_mapcount cannot change */
1593                 BUG_ON(page_mapcount(page_tail) < 0);
1594                 tail_count += page_mapcount(page_tail);
1595                 /* check for overflow */
1596                 BUG_ON(tail_count < 0);
1597                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1598                 /*
1599                  * tail_page->_count is zero and not changing from
1600                  * under us. But get_page_unless_zero() may be running
1601                  * from under us on the tail_page. If we used
1602                  * atomic_set() below instead of atomic_add(), we
1603                  * would then run atomic_set() concurrently with
1604                  * get_page_unless_zero(), and atomic_set() is
1605                  * implemented in C not using locked ops. spin_unlock
1606                  * on x86 sometime uses locked ops because of PPro
1607                  * errata 66, 92, so unless somebody can guarantee
1608                  * atomic_set() here would be safe on all archs (and
1609                  * not only on x86), it's safer to use atomic_add().
1610                  */
1611                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1612                            &page_tail->_count);
1613
1614                 /* after clearing PageTail the gup refcount can be released */
1615                 smp_mb();
1616
1617                 /*
1618                  * retain hwpoison flag of the poisoned tail page:
1619                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1620                  *   by the memory-failure.
1621                  */
1622                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1623                 page_tail->flags |= (page->flags &
1624                                      ((1L << PG_referenced) |
1625                                       (1L << PG_swapbacked) |
1626                                       (1L << PG_mlocked) |
1627                                       (1L << PG_uptodate)));
1628                 page_tail->flags |= (1L << PG_dirty);
1629
1630                 /* clear PageTail before overwriting first_page */
1631                 smp_wmb();
1632
1633                 /*
1634                  * __split_huge_page_splitting() already set the
1635                  * splitting bit in all pmd that could map this
1636                  * hugepage, that will ensure no CPU can alter the
1637                  * mapcount on the head page. The mapcount is only
1638                  * accounted in the head page and it has to be
1639                  * transferred to all tail pages in the below code. So
1640                  * for this code to be safe, the split the mapcount
1641                  * can't change. But that doesn't mean userland can't
1642                  * keep changing and reading the page contents while
1643                  * we transfer the mapcount, so the pmd splitting
1644                  * status is achieved setting a reserved bit in the
1645                  * pmd, not by clearing the present bit.
1646                 */
1647                 page_tail->_mapcount = page->_mapcount;
1648
1649                 BUG_ON(page_tail->mapping);
1650                 page_tail->mapping = page->mapping;
1651
1652                 page_tail->index = page->index + i;
1653                 page_xchg_last_nid(page_tail, page_last_nid(page));
1654
1655                 BUG_ON(!PageAnon(page_tail));
1656                 BUG_ON(!PageUptodate(page_tail));
1657                 BUG_ON(!PageDirty(page_tail));
1658                 BUG_ON(!PageSwapBacked(page_tail));
1659
1660                 lru_add_page_tail(page, page_tail, lruvec);
1661         }
1662         atomic_sub(tail_count, &page->_count);
1663         BUG_ON(atomic_read(&page->_count) <= 0);
1664
1665         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1666         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1667
1668         ClearPageCompound(page);
1669         compound_unlock(page);
1670         spin_unlock_irq(&zone->lru_lock);
1671
1672         for (i = 1; i < HPAGE_PMD_NR; i++) {
1673                 struct page *page_tail = page + i;
1674                 BUG_ON(page_count(page_tail) <= 0);
1675                 /*
1676                  * Tail pages may be freed if there wasn't any mapping
1677                  * like if add_to_swap() is running on a lru page that
1678                  * had its mapping zapped. And freeing these pages
1679                  * requires taking the lru_lock so we do the put_page
1680                  * of the tail pages after the split is complete.
1681                  */
1682                 put_page(page_tail);
1683         }
1684
1685         /*
1686          * Only the head page (now become a regular page) is required
1687          * to be pinned by the caller.
1688          */
1689         BUG_ON(page_count(page) <= 0);
1690 }
1691
1692 static int __split_huge_page_map(struct page *page,
1693                                  struct vm_area_struct *vma,
1694                                  unsigned long address)
1695 {
1696         struct mm_struct *mm = vma->vm_mm;
1697         pmd_t *pmd, _pmd;
1698         int ret = 0, i;
1699         pgtable_t pgtable;
1700         unsigned long haddr;
1701
1702         spin_lock(&mm->page_table_lock);
1703         pmd = page_check_address_pmd(page, mm, address,
1704                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1705         if (pmd) {
1706                 pgtable = pgtable_trans_huge_withdraw(mm);
1707                 pmd_populate(mm, &_pmd, pgtable);
1708
1709                 haddr = address;
1710                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1711                         pte_t *pte, entry;
1712                         BUG_ON(PageCompound(page+i));
1713                         entry = mk_pte(page + i, vma->vm_page_prot);
1714                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1715                         if (!pmd_write(*pmd))
1716                                 entry = pte_wrprotect(entry);
1717                         else
1718                                 BUG_ON(page_mapcount(page) != 1);
1719                         if (!pmd_young(*pmd))
1720                                 entry = pte_mkold(entry);
1721                         if (pmd_numa(*pmd))
1722                                 entry = pte_mknuma(entry);
1723                         pte = pte_offset_map(&_pmd, haddr);
1724                         BUG_ON(!pte_none(*pte));
1725                         set_pte_at(mm, haddr, pte, entry);
1726                         pte_unmap(pte);
1727                 }
1728
1729                 smp_wmb(); /* make pte visible before pmd */
1730                 /*
1731                  * Up to this point the pmd is present and huge and
1732                  * userland has the whole access to the hugepage
1733                  * during the split (which happens in place). If we
1734                  * overwrite the pmd with the not-huge version
1735                  * pointing to the pte here (which of course we could
1736                  * if all CPUs were bug free), userland could trigger
1737                  * a small page size TLB miss on the small sized TLB
1738                  * while the hugepage TLB entry is still established
1739                  * in the huge TLB. Some CPU doesn't like that. See
1740                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1741                  * Erratum 383 on page 93. Intel should be safe but is
1742                  * also warns that it's only safe if the permission
1743                  * and cache attributes of the two entries loaded in
1744                  * the two TLB is identical (which should be the case
1745                  * here). But it is generally safer to never allow
1746                  * small and huge TLB entries for the same virtual
1747                  * address to be loaded simultaneously. So instead of
1748                  * doing "pmd_populate(); flush_tlb_range();" we first
1749                  * mark the current pmd notpresent (atomically because
1750                  * here the pmd_trans_huge and pmd_trans_splitting
1751                  * must remain set at all times on the pmd until the
1752                  * split is complete for this pmd), then we flush the
1753                  * SMP TLB and finally we write the non-huge version
1754                  * of the pmd entry with pmd_populate.
1755                  */
1756                 pmdp_invalidate(vma, address, pmd);
1757                 pmd_populate(mm, pmd, pgtable);
1758                 ret = 1;
1759         }
1760         spin_unlock(&mm->page_table_lock);
1761
1762         return ret;
1763 }
1764
1765 /* must be called with anon_vma->root->rwsem held */
1766 static void __split_huge_page(struct page *page,
1767                               struct anon_vma *anon_vma)
1768 {
1769         int mapcount, mapcount2;
1770         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1771         struct anon_vma_chain *avc;
1772
1773         BUG_ON(!PageHead(page));
1774         BUG_ON(PageTail(page));
1775
1776         mapcount = 0;
1777         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1778                 struct vm_area_struct *vma = avc->vma;
1779                 unsigned long addr = vma_address(page, vma);
1780                 BUG_ON(is_vma_temporary_stack(vma));
1781                 mapcount += __split_huge_page_splitting(page, vma, addr);
1782         }
1783         /*
1784          * It is critical that new vmas are added to the tail of the
1785          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1786          * and establishes a child pmd before
1787          * __split_huge_page_splitting() freezes the parent pmd (so if
1788          * we fail to prevent copy_huge_pmd() from running until the
1789          * whole __split_huge_page() is complete), we will still see
1790          * the newly established pmd of the child later during the
1791          * walk, to be able to set it as pmd_trans_splitting too.
1792          */
1793         if (mapcount != page_mapcount(page))
1794                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1795                        mapcount, page_mapcount(page));
1796         BUG_ON(mapcount != page_mapcount(page));
1797
1798         __split_huge_page_refcount(page);
1799
1800         mapcount2 = 0;
1801         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1802                 struct vm_area_struct *vma = avc->vma;
1803                 unsigned long addr = vma_address(page, vma);
1804                 BUG_ON(is_vma_temporary_stack(vma));
1805                 mapcount2 += __split_huge_page_map(page, vma, addr);
1806         }
1807         if (mapcount != mapcount2)
1808                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1809                        mapcount, mapcount2, page_mapcount(page));
1810         BUG_ON(mapcount != mapcount2);
1811 }
1812
1813 int split_huge_page(struct page *page)
1814 {
1815         struct anon_vma *anon_vma;
1816         int ret = 1;
1817
1818         BUG_ON(is_huge_zero_pfn(page_to_pfn(page)));
1819         BUG_ON(!PageAnon(page));
1820
1821         /*
1822          * The caller does not necessarily hold an mmap_sem that would prevent
1823          * the anon_vma disappearing so we first we take a reference to it
1824          * and then lock the anon_vma for write. This is similar to
1825          * page_lock_anon_vma_read except the write lock is taken to serialise
1826          * against parallel split or collapse operations.
1827          */
1828         anon_vma = page_get_anon_vma(page);
1829         if (!anon_vma)
1830                 goto out;
1831         anon_vma_lock_write(anon_vma);
1832
1833         ret = 0;
1834         if (!PageCompound(page))
1835                 goto out_unlock;
1836
1837         BUG_ON(!PageSwapBacked(page));
1838         __split_huge_page(page, anon_vma);
1839         count_vm_event(THP_SPLIT);
1840
1841         BUG_ON(PageCompound(page));
1842 out_unlock:
1843         anon_vma_unlock(anon_vma);
1844         put_anon_vma(anon_vma);
1845 out:
1846         return ret;
1847 }
1848
1849 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1850
1851 int hugepage_madvise(struct vm_area_struct *vma,
1852                      unsigned long *vm_flags, int advice)
1853 {
1854         struct mm_struct *mm = vma->vm_mm;
1855
1856         switch (advice) {
1857         case MADV_HUGEPAGE:
1858                 /*
1859                  * Be somewhat over-protective like KSM for now!
1860                  */
1861                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1862                         return -EINVAL;
1863                 if (mm->def_flags & VM_NOHUGEPAGE)
1864                         return -EINVAL;
1865                 *vm_flags &= ~VM_NOHUGEPAGE;
1866                 *vm_flags |= VM_HUGEPAGE;
1867                 /*
1868                  * If the vma become good for khugepaged to scan,
1869                  * register it here without waiting a page fault that
1870                  * may not happen any time soon.
1871                  */
1872                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1873                         return -ENOMEM;
1874                 break;
1875         case MADV_NOHUGEPAGE:
1876                 /*
1877                  * Be somewhat over-protective like KSM for now!
1878                  */
1879                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1880                         return -EINVAL;
1881                 *vm_flags &= ~VM_HUGEPAGE;
1882                 *vm_flags |= VM_NOHUGEPAGE;
1883                 /*
1884                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1885                  * this vma even if we leave the mm registered in khugepaged if
1886                  * it got registered before VM_NOHUGEPAGE was set.
1887                  */
1888                 break;
1889         }
1890
1891         return 0;
1892 }
1893
1894 static int __init khugepaged_slab_init(void)
1895 {
1896         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1897                                           sizeof(struct mm_slot),
1898                                           __alignof__(struct mm_slot), 0, NULL);
1899         if (!mm_slot_cache)
1900                 return -ENOMEM;
1901
1902         return 0;
1903 }
1904
1905 static inline struct mm_slot *alloc_mm_slot(void)
1906 {
1907         if (!mm_slot_cache)     /* initialization failed */
1908                 return NULL;
1909         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1910 }
1911
1912 static inline void free_mm_slot(struct mm_slot *mm_slot)
1913 {
1914         kmem_cache_free(mm_slot_cache, mm_slot);
1915 }
1916
1917 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1918 {
1919         struct mm_slot *mm_slot;
1920         struct hlist_node *node;
1921
1922         hash_for_each_possible(mm_slots_hash, mm_slot, node, hash, (unsigned long)mm)
1923                 if (mm == mm_slot->mm)
1924                         return mm_slot;
1925
1926         return NULL;
1927 }
1928
1929 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1930                                     struct mm_slot *mm_slot)
1931 {
1932         mm_slot->mm = mm;
1933         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1934 }
1935
1936 static inline int khugepaged_test_exit(struct mm_struct *mm)
1937 {
1938         return atomic_read(&mm->mm_users) == 0;
1939 }
1940
1941 int __khugepaged_enter(struct mm_struct *mm)
1942 {
1943         struct mm_slot *mm_slot;
1944         int wakeup;
1945
1946         mm_slot = alloc_mm_slot();
1947         if (!mm_slot)
1948                 return -ENOMEM;
1949
1950         /* __khugepaged_exit() must not run from under us */
1951         VM_BUG_ON(khugepaged_test_exit(mm));
1952         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1953                 free_mm_slot(mm_slot);
1954                 return 0;
1955         }
1956
1957         spin_lock(&khugepaged_mm_lock);
1958         insert_to_mm_slots_hash(mm, mm_slot);
1959         /*
1960          * Insert just behind the scanning cursor, to let the area settle
1961          * down a little.
1962          */
1963         wakeup = list_empty(&khugepaged_scan.mm_head);
1964         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1965         spin_unlock(&khugepaged_mm_lock);
1966
1967         atomic_inc(&mm->mm_count);
1968         if (wakeup)
1969                 wake_up_interruptible(&khugepaged_wait);
1970
1971         return 0;
1972 }
1973
1974 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1975 {
1976         unsigned long hstart, hend;
1977         if (!vma->anon_vma)
1978                 /*
1979                  * Not yet faulted in so we will register later in the
1980                  * page fault if needed.
1981                  */
1982                 return 0;
1983         if (vma->vm_ops)
1984                 /* khugepaged not yet working on file or special mappings */
1985                 return 0;
1986         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1987         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1988         hend = vma->vm_end & HPAGE_PMD_MASK;
1989         if (hstart < hend)
1990                 return khugepaged_enter(vma);
1991         return 0;
1992 }
1993
1994 void __khugepaged_exit(struct mm_struct *mm)
1995 {
1996         struct mm_slot *mm_slot;
1997         int free = 0;
1998
1999         spin_lock(&khugepaged_mm_lock);
2000         mm_slot = get_mm_slot(mm);
2001         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2002                 hash_del(&mm_slot->hash);
2003                 list_del(&mm_slot->mm_node);
2004                 free = 1;
2005         }
2006         spin_unlock(&khugepaged_mm_lock);
2007
2008         if (free) {
2009                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2010                 free_mm_slot(mm_slot);
2011                 mmdrop(mm);
2012         } else if (mm_slot) {
2013                 /*
2014                  * This is required to serialize against
2015                  * khugepaged_test_exit() (which is guaranteed to run
2016                  * under mmap sem read mode). Stop here (after we
2017                  * return all pagetables will be destroyed) until
2018                  * khugepaged has finished working on the pagetables
2019                  * under the mmap_sem.
2020                  */
2021                 down_write(&mm->mmap_sem);
2022                 up_write(&mm->mmap_sem);
2023         }
2024 }
2025
2026 static void release_pte_page(struct page *page)
2027 {
2028         /* 0 stands for page_is_file_cache(page) == false */
2029         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2030         unlock_page(page);
2031         putback_lru_page(page);
2032 }
2033
2034 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2035 {
2036         while (--_pte >= pte) {
2037                 pte_t pteval = *_pte;
2038                 if (!pte_none(pteval))
2039                         release_pte_page(pte_page(pteval));
2040         }
2041 }
2042
2043 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2044                                         unsigned long address,
2045                                         pte_t *pte)
2046 {
2047         struct page *page;
2048         pte_t *_pte;
2049         int referenced = 0, none = 0;
2050         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2051              _pte++, address += PAGE_SIZE) {
2052                 pte_t pteval = *_pte;
2053                 if (pte_none(pteval)) {
2054                         if (++none <= khugepaged_max_ptes_none)
2055                                 continue;
2056                         else
2057                                 goto out;
2058                 }
2059                 if (!pte_present(pteval) || !pte_write(pteval))
2060                         goto out;
2061                 page = vm_normal_page(vma, address, pteval);
2062                 if (unlikely(!page))
2063                         goto out;
2064
2065                 VM_BUG_ON(PageCompound(page));
2066                 BUG_ON(!PageAnon(page));
2067                 VM_BUG_ON(!PageSwapBacked(page));
2068
2069                 /* cannot use mapcount: can't collapse if there's a gup pin */
2070                 if (page_count(page) != 1)
2071                         goto out;
2072                 /*
2073                  * We can do it before isolate_lru_page because the
2074                  * page can't be freed from under us. NOTE: PG_lock
2075                  * is needed to serialize against split_huge_page
2076                  * when invoked from the VM.
2077                  */
2078                 if (!trylock_page(page))
2079                         goto out;
2080                 /*
2081                  * Isolate the page to avoid collapsing an hugepage
2082                  * currently in use by the VM.
2083                  */
2084                 if (isolate_lru_page(page)) {
2085                         unlock_page(page);
2086                         goto out;
2087                 }
2088                 /* 0 stands for page_is_file_cache(page) == false */
2089                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2090                 VM_BUG_ON(!PageLocked(page));
2091                 VM_BUG_ON(PageLRU(page));
2092
2093                 /* If there is no mapped pte young don't collapse the page */
2094                 if (pte_young(pteval) || PageReferenced(page) ||
2095                     mmu_notifier_test_young(vma->vm_mm, address))
2096                         referenced = 1;
2097         }
2098         if (likely(referenced))
2099                 return 1;
2100 out:
2101         release_pte_pages(pte, _pte);
2102         return 0;
2103 }
2104
2105 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2106                                       struct vm_area_struct *vma,
2107                                       unsigned long address,
2108                                       spinlock_t *ptl)
2109 {
2110         pte_t *_pte;
2111         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2112                 pte_t pteval = *_pte;
2113                 struct page *src_page;
2114
2115                 if (pte_none(pteval)) {
2116                         clear_user_highpage(page, address);
2117                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2118                 } else {
2119                         src_page = pte_page(pteval);
2120                         copy_user_highpage(page, src_page, address, vma);
2121                         VM_BUG_ON(page_mapcount(src_page) != 1);
2122                         release_pte_page(src_page);
2123                         /*
2124                          * ptl mostly unnecessary, but preempt has to
2125                          * be disabled to update the per-cpu stats
2126                          * inside page_remove_rmap().
2127                          */
2128                         spin_lock(ptl);
2129                         /*
2130                          * paravirt calls inside pte_clear here are
2131                          * superfluous.
2132                          */
2133                         pte_clear(vma->vm_mm, address, _pte);
2134                         page_remove_rmap(src_page);
2135                         spin_unlock(ptl);
2136                         free_page_and_swap_cache(src_page);
2137                 }
2138
2139                 address += PAGE_SIZE;
2140                 page++;
2141         }
2142 }
2143
2144 static void khugepaged_alloc_sleep(void)
2145 {
2146         wait_event_freezable_timeout(khugepaged_wait, false,
2147                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2148 }
2149
2150 #ifdef CONFIG_NUMA
2151 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2152 {
2153         if (IS_ERR(*hpage)) {
2154                 if (!*wait)
2155                         return false;
2156
2157                 *wait = false;
2158                 *hpage = NULL;
2159                 khugepaged_alloc_sleep();
2160         } else if (*hpage) {
2161                 put_page(*hpage);
2162                 *hpage = NULL;
2163         }
2164
2165         return true;
2166 }
2167
2168 static struct page
2169 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2170                        struct vm_area_struct *vma, unsigned long address,
2171                        int node)
2172 {
2173         VM_BUG_ON(*hpage);
2174         /*
2175          * Allocate the page while the vma is still valid and under
2176          * the mmap_sem read mode so there is no memory allocation
2177          * later when we take the mmap_sem in write mode. This is more
2178          * friendly behavior (OTOH it may actually hide bugs) to
2179          * filesystems in userland with daemons allocating memory in
2180          * the userland I/O paths.  Allocating memory with the
2181          * mmap_sem in read mode is good idea also to allow greater
2182          * scalability.
2183          */
2184         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2185                                       node, __GFP_OTHER_NODE);
2186
2187         /*
2188          * After allocating the hugepage, release the mmap_sem read lock in
2189          * preparation for taking it in write mode.
2190          */
2191         up_read(&mm->mmap_sem);
2192         if (unlikely(!*hpage)) {
2193                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2194                 *hpage = ERR_PTR(-ENOMEM);
2195                 return NULL;
2196         }
2197
2198         count_vm_event(THP_COLLAPSE_ALLOC);
2199         return *hpage;
2200 }
2201 #else
2202 static struct page *khugepaged_alloc_hugepage(bool *wait)
2203 {
2204         struct page *hpage;
2205
2206         do {
2207                 hpage = alloc_hugepage(khugepaged_defrag());
2208                 if (!hpage) {
2209                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2210                         if (!*wait)
2211                                 return NULL;
2212
2213                         *wait = false;
2214                         khugepaged_alloc_sleep();
2215                 } else
2216                         count_vm_event(THP_COLLAPSE_ALLOC);
2217         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2218
2219         return hpage;
2220 }
2221
2222 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2223 {
2224         if (!*hpage)
2225                 *hpage = khugepaged_alloc_hugepage(wait);
2226
2227         if (unlikely(!*hpage))
2228                 return false;
2229
2230         return true;
2231 }
2232
2233 static struct page
2234 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2235                        struct vm_area_struct *vma, unsigned long address,
2236                        int node)
2237 {
2238         up_read(&mm->mmap_sem);
2239         VM_BUG_ON(!*hpage);
2240         return  *hpage;
2241 }
2242 #endif
2243
2244 static bool hugepage_vma_check(struct vm_area_struct *vma)
2245 {
2246         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2247             (vma->vm_flags & VM_NOHUGEPAGE))
2248                 return false;
2249
2250         if (!vma->anon_vma || vma->vm_ops)
2251                 return false;
2252         if (is_vma_temporary_stack(vma))
2253                 return false;
2254         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2255         return true;
2256 }
2257
2258 static void collapse_huge_page(struct mm_struct *mm,
2259                                    unsigned long address,
2260                                    struct page **hpage,
2261                                    struct vm_area_struct *vma,
2262                                    int node)
2263 {
2264         pmd_t *pmd, _pmd;
2265         pte_t *pte;
2266         pgtable_t pgtable;
2267         struct page *new_page;
2268         spinlock_t *ptl;
2269         int isolated;
2270         unsigned long hstart, hend;
2271         unsigned long mmun_start;       /* For mmu_notifiers */
2272         unsigned long mmun_end;         /* For mmu_notifiers */
2273
2274         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2275
2276         /* release the mmap_sem read lock. */
2277         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2278         if (!new_page)
2279                 return;
2280
2281         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2282                 return;
2283
2284         /*
2285          * Prevent all access to pagetables with the exception of
2286          * gup_fast later hanlded by the ptep_clear_flush and the VM
2287          * handled by the anon_vma lock + PG_lock.
2288          */
2289         down_write(&mm->mmap_sem);
2290         if (unlikely(khugepaged_test_exit(mm)))
2291                 goto out;
2292
2293         vma = find_vma(mm, address);
2294         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2295         hend = vma->vm_end & HPAGE_PMD_MASK;
2296         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2297                 goto out;
2298         if (!hugepage_vma_check(vma))
2299                 goto out;
2300         pmd = mm_find_pmd(mm, address);
2301         if (!pmd)
2302                 goto out;
2303         if (pmd_trans_huge(*pmd))
2304                 goto out;
2305
2306         anon_vma_lock_write(vma->anon_vma);
2307
2308         pte = pte_offset_map(pmd, address);
2309         ptl = pte_lockptr(mm, pmd);
2310
2311         mmun_start = address;
2312         mmun_end   = address + HPAGE_PMD_SIZE;
2313         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2314         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2315         /*
2316          * After this gup_fast can't run anymore. This also removes
2317          * any huge TLB entry from the CPU so we won't allow
2318          * huge and small TLB entries for the same virtual address
2319          * to avoid the risk of CPU bugs in that area.
2320          */
2321         _pmd = pmdp_clear_flush(vma, address, pmd);
2322         spin_unlock(&mm->page_table_lock);
2323         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2324
2325         spin_lock(ptl);
2326         isolated = __collapse_huge_page_isolate(vma, address, pte);
2327         spin_unlock(ptl);
2328
2329         if (unlikely(!isolated)) {
2330                 pte_unmap(pte);
2331                 spin_lock(&mm->page_table_lock);
2332                 BUG_ON(!pmd_none(*pmd));
2333                 set_pmd_at(mm, address, pmd, _pmd);
2334                 spin_unlock(&mm->page_table_lock);
2335                 anon_vma_unlock(vma->anon_vma);
2336                 goto out;
2337         }
2338
2339         /*
2340          * All pages are isolated and locked so anon_vma rmap
2341          * can't run anymore.
2342          */
2343         anon_vma_unlock(vma->anon_vma);
2344
2345         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2346         pte_unmap(pte);
2347         __SetPageUptodate(new_page);
2348         pgtable = pmd_pgtable(_pmd);
2349
2350         _pmd = mk_huge_pmd(new_page, vma);
2351
2352         /*
2353          * spin_lock() below is not the equivalent of smp_wmb(), so
2354          * this is needed to avoid the copy_huge_page writes to become
2355          * visible after the set_pmd_at() write.
2356          */
2357         smp_wmb();
2358
2359         spin_lock(&mm->page_table_lock);
2360         BUG_ON(!pmd_none(*pmd));
2361         page_add_new_anon_rmap(new_page, vma, address);
2362         set_pmd_at(mm, address, pmd, _pmd);
2363         update_mmu_cache_pmd(vma, address, pmd);
2364         pgtable_trans_huge_deposit(mm, pgtable);
2365         spin_unlock(&mm->page_table_lock);
2366
2367         *hpage = NULL;
2368
2369         khugepaged_pages_collapsed++;
2370 out_up_write:
2371         up_write(&mm->mmap_sem);
2372         return;
2373
2374 out:
2375         mem_cgroup_uncharge_page(new_page);
2376         goto out_up_write;
2377 }
2378
2379 static int khugepaged_scan_pmd(struct mm_struct *mm,
2380                                struct vm_area_struct *vma,
2381                                unsigned long address,
2382                                struct page **hpage)
2383 {
2384         pmd_t *pmd;
2385         pte_t *pte, *_pte;
2386         int ret = 0, referenced = 0, none = 0;
2387         struct page *page;
2388         unsigned long _address;
2389         spinlock_t *ptl;
2390         int node = -1;
2391
2392         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2393
2394         pmd = mm_find_pmd(mm, address);
2395         if (!pmd)
2396                 goto out;
2397         if (pmd_trans_huge(*pmd))
2398                 goto out;
2399
2400         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2401         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2402              _pte++, _address += PAGE_SIZE) {
2403                 pte_t pteval = *_pte;
2404                 if (pte_none(pteval)) {
2405                         if (++none <= khugepaged_max_ptes_none)
2406                                 continue;
2407                         else
2408                                 goto out_unmap;
2409                 }
2410                 if (!pte_present(pteval) || !pte_write(pteval))
2411                         goto out_unmap;
2412                 page = vm_normal_page(vma, _address, pteval);
2413                 if (unlikely(!page))
2414                         goto out_unmap;
2415                 /*
2416                  * Chose the node of the first page. This could
2417                  * be more sophisticated and look at more pages,
2418                  * but isn't for now.
2419                  */
2420                 if (node == -1)
2421                         node = page_to_nid(page);
2422                 VM_BUG_ON(PageCompound(page));
2423                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2424                         goto out_unmap;
2425                 /* cannot use mapcount: can't collapse if there's a gup pin */
2426                 if (page_count(page) != 1)
2427                         goto out_unmap;
2428                 if (pte_young(pteval) || PageReferenced(page) ||
2429                     mmu_notifier_test_young(vma->vm_mm, address))
2430                         referenced = 1;
2431         }
2432         if (referenced)
2433                 ret = 1;
2434 out_unmap:
2435         pte_unmap_unlock(pte, ptl);
2436         if (ret)
2437                 /* collapse_huge_page will return with the mmap_sem released */
2438                 collapse_huge_page(mm, address, hpage, vma, node);
2439 out:
2440         return ret;
2441 }
2442
2443 static void collect_mm_slot(struct mm_slot *mm_slot)
2444 {
2445         struct mm_struct *mm = mm_slot->mm;
2446
2447         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2448
2449         if (khugepaged_test_exit(mm)) {
2450                 /* free mm_slot */
2451                 hash_del(&mm_slot->hash);
2452                 list_del(&mm_slot->mm_node);
2453
2454                 /*
2455                  * Not strictly needed because the mm exited already.
2456                  *
2457                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2458                  */
2459
2460                 /* khugepaged_mm_lock actually not necessary for the below */
2461                 free_mm_slot(mm_slot);
2462                 mmdrop(mm);
2463         }
2464 }
2465
2466 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2467                                             struct page **hpage)
2468         __releases(&khugepaged_mm_lock)
2469         __acquires(&khugepaged_mm_lock)
2470 {
2471         struct mm_slot *mm_slot;
2472         struct mm_struct *mm;
2473         struct vm_area_struct *vma;
2474         int progress = 0;
2475
2476         VM_BUG_ON(!pages);
2477         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2478
2479         if (khugepaged_scan.mm_slot)
2480                 mm_slot = khugepaged_scan.mm_slot;
2481         else {
2482                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2483                                      struct mm_slot, mm_node);
2484                 khugepaged_scan.address = 0;
2485                 khugepaged_scan.mm_slot = mm_slot;
2486         }
2487         spin_unlock(&khugepaged_mm_lock);
2488
2489         mm = mm_slot->mm;
2490         down_read(&mm->mmap_sem);
2491         if (unlikely(khugepaged_test_exit(mm)))
2492                 vma = NULL;
2493         else
2494                 vma = find_vma(mm, khugepaged_scan.address);
2495
2496         progress++;
2497         for (; vma; vma = vma->vm_next) {
2498                 unsigned long hstart, hend;
2499
2500                 cond_resched();
2501                 if (unlikely(khugepaged_test_exit(mm))) {
2502                         progress++;
2503                         break;
2504                 }
2505                 if (!hugepage_vma_check(vma)) {
2506 skip:
2507                         progress++;
2508                         continue;
2509                 }
2510                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2511                 hend = vma->vm_end & HPAGE_PMD_MASK;
2512                 if (hstart >= hend)
2513                         goto skip;
2514                 if (khugepaged_scan.address > hend)
2515                         goto skip;
2516                 if (khugepaged_scan.address < hstart)
2517                         khugepaged_scan.address = hstart;
2518                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2519
2520                 while (khugepaged_scan.address < hend) {
2521                         int ret;
2522                         cond_resched();
2523                         if (unlikely(khugepaged_test_exit(mm)))
2524                                 goto breakouterloop;
2525
2526                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2527                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2528                                   hend);
2529                         ret = khugepaged_scan_pmd(mm, vma,
2530                                                   khugepaged_scan.address,
2531                                                   hpage);
2532                         /* move to next address */
2533                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2534                         progress += HPAGE_PMD_NR;
2535                         if (ret)
2536                                 /* we released mmap_sem so break loop */
2537                                 goto breakouterloop_mmap_sem;
2538                         if (progress >= pages)
2539                                 goto breakouterloop;
2540                 }
2541         }
2542 breakouterloop:
2543         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2544 breakouterloop_mmap_sem:
2545
2546         spin_lock(&khugepaged_mm_lock);
2547         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2548         /*
2549          * Release the current mm_slot if this mm is about to die, or
2550          * if we scanned all vmas of this mm.
2551          */
2552         if (khugepaged_test_exit(mm) || !vma) {
2553                 /*
2554                  * Make sure that if mm_users is reaching zero while
2555                  * khugepaged runs here, khugepaged_exit will find
2556                  * mm_slot not pointing to the exiting mm.
2557                  */
2558                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2559                         khugepaged_scan.mm_slot = list_entry(
2560                                 mm_slot->mm_node.next,
2561                                 struct mm_slot, mm_node);
2562                         khugepaged_scan.address = 0;
2563                 } else {
2564                         khugepaged_scan.mm_slot = NULL;
2565                         khugepaged_full_scans++;
2566                 }
2567
2568                 collect_mm_slot(mm_slot);
2569         }
2570
2571         return progress;
2572 }
2573
2574 static int khugepaged_has_work(void)
2575 {
2576         return !list_empty(&khugepaged_scan.mm_head) &&
2577                 khugepaged_enabled();
2578 }
2579
2580 static int khugepaged_wait_event(void)
2581 {
2582         return !list_empty(&khugepaged_scan.mm_head) ||
2583                 kthread_should_stop();
2584 }
2585
2586 static void khugepaged_do_scan(void)
2587 {
2588         struct page *hpage = NULL;
2589         unsigned int progress = 0, pass_through_head = 0;
2590         unsigned int pages = khugepaged_pages_to_scan;
2591         bool wait = true;
2592
2593         barrier(); /* write khugepaged_pages_to_scan to local stack */
2594
2595         while (progress < pages) {
2596                 if (!khugepaged_prealloc_page(&hpage, &wait))
2597                         break;
2598
2599                 cond_resched();
2600
2601                 if (unlikely(kthread_should_stop() || freezing(current)))
2602                         break;
2603
2604                 spin_lock(&khugepaged_mm_lock);
2605                 if (!khugepaged_scan.mm_slot)
2606                         pass_through_head++;
2607                 if (khugepaged_has_work() &&
2608                     pass_through_head < 2)
2609                         progress += khugepaged_scan_mm_slot(pages - progress,
2610                                                             &hpage);
2611                 else
2612                         progress = pages;
2613                 spin_unlock(&khugepaged_mm_lock);
2614         }
2615
2616         if (!IS_ERR_OR_NULL(hpage))
2617                 put_page(hpage);
2618 }
2619
2620 static void khugepaged_wait_work(void)
2621 {
2622         try_to_freeze();
2623
2624         if (khugepaged_has_work()) {
2625                 if (!khugepaged_scan_sleep_millisecs)
2626                         return;
2627
2628                 wait_event_freezable_timeout(khugepaged_wait,
2629                                              kthread_should_stop(),
2630                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2631                 return;
2632         }
2633
2634         if (khugepaged_enabled())
2635                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2636 }
2637
2638 static int khugepaged(void *none)
2639 {
2640         struct mm_slot *mm_slot;
2641
2642         set_freezable();
2643         set_user_nice(current, 19);
2644
2645         while (!kthread_should_stop()) {
2646                 khugepaged_do_scan();
2647                 khugepaged_wait_work();
2648         }
2649
2650         spin_lock(&khugepaged_mm_lock);
2651         mm_slot = khugepaged_scan.mm_slot;
2652         khugepaged_scan.mm_slot = NULL;
2653         if (mm_slot)
2654                 collect_mm_slot(mm_slot);
2655         spin_unlock(&khugepaged_mm_lock);
2656         return 0;
2657 }
2658
2659 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2660                 unsigned long haddr, pmd_t *pmd)
2661 {
2662         struct mm_struct *mm = vma->vm_mm;
2663         pgtable_t pgtable;
2664         pmd_t _pmd;
2665         int i;
2666
2667         pmdp_clear_flush(vma, haddr, pmd);
2668         /* leave pmd empty until pte is filled */
2669
2670         pgtable = pgtable_trans_huge_withdraw(mm);
2671         pmd_populate(mm, &_pmd, pgtable);
2672
2673         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2674                 pte_t *pte, entry;
2675                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2676                 entry = pte_mkspecial(entry);
2677                 pte = pte_offset_map(&_pmd, haddr);
2678                 VM_BUG_ON(!pte_none(*pte));
2679                 set_pte_at(mm, haddr, pte, entry);
2680                 pte_unmap(pte);
2681         }
2682         smp_wmb(); /* make pte visible before pmd */
2683         pmd_populate(mm, pmd, pgtable);
2684         put_huge_zero_page();
2685 }
2686
2687 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2688                 pmd_t *pmd)
2689 {
2690         struct page *page;
2691         struct mm_struct *mm = vma->vm_mm;
2692         unsigned long haddr = address & HPAGE_PMD_MASK;
2693         unsigned long mmun_start;       /* For mmu_notifiers */
2694         unsigned long mmun_end;         /* For mmu_notifiers */
2695
2696         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2697
2698         mmun_start = haddr;
2699         mmun_end   = haddr + HPAGE_PMD_SIZE;
2700         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2701         spin_lock(&mm->page_table_lock);
2702         if (unlikely(!pmd_trans_huge(*pmd))) {
2703                 spin_unlock(&mm->page_table_lock);
2704                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2705                 return;
2706         }
2707         if (is_huge_zero_pmd(*pmd)) {
2708                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2709                 spin_unlock(&mm->page_table_lock);
2710                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2711                 return;
2712         }
2713         page = pmd_page(*pmd);
2714         VM_BUG_ON(!page_count(page));
2715         get_page(page);
2716         spin_unlock(&mm->page_table_lock);
2717         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2718
2719         split_huge_page(page);
2720
2721         put_page(page);
2722         BUG_ON(pmd_trans_huge(*pmd));
2723 }
2724
2725 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2726                 pmd_t *pmd)
2727 {
2728         struct vm_area_struct *vma;
2729
2730         vma = find_vma(mm, address);
2731         BUG_ON(vma == NULL);
2732         split_huge_page_pmd(vma, address, pmd);
2733 }
2734
2735 static void split_huge_page_address(struct mm_struct *mm,
2736                                     unsigned long address)
2737 {
2738         pmd_t *pmd;
2739
2740         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2741
2742         pmd = mm_find_pmd(mm, address);
2743         if (!pmd)
2744                 return;
2745         /*
2746          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2747          * materialize from under us.
2748          */
2749         split_huge_page_pmd_mm(mm, address, pmd);
2750 }
2751
2752 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2753                              unsigned long start,
2754                              unsigned long end,
2755                              long adjust_next)
2756 {
2757         /*
2758          * If the new start address isn't hpage aligned and it could
2759          * previously contain an hugepage: check if we need to split
2760          * an huge pmd.
2761          */
2762         if (start & ~HPAGE_PMD_MASK &&
2763             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2764             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2765                 split_huge_page_address(vma->vm_mm, start);
2766
2767         /*
2768          * If the new end address isn't hpage aligned and it could
2769          * previously contain an hugepage: check if we need to split
2770          * an huge pmd.
2771          */
2772         if (end & ~HPAGE_PMD_MASK &&
2773             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2774             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2775                 split_huge_page_address(vma->vm_mm, end);
2776
2777         /*
2778          * If we're also updating the vma->vm_next->vm_start, if the new
2779          * vm_next->vm_start isn't page aligned and it could previously
2780          * contain an hugepage: check if we need to split an huge pmd.
2781          */
2782         if (adjust_next > 0) {
2783                 struct vm_area_struct *next = vma->vm_next;
2784                 unsigned long nstart = next->vm_start;
2785                 nstart += adjust_next << PAGE_SHIFT;
2786                 if (nstart & ~HPAGE_PMD_MASK &&
2787                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2788                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2789                         split_huge_page_address(next->vm_mm, nstart);
2790         }
2791 }