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
  24 #include <asm/tlb.h>
  25 #include <asm/pgalloc.h>
  26 #include "internal.h"
  27
  28 /*
  29  * By default transparent hugepage support is enabled for all mappings
  30  * and khugepaged scans all mappings. Defrag is only invoked by
  31  * khugepaged hugepage allocations and by page faults inside
  32  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  33  * allocations.
  34  */
  35 unsigned long transparent_hugepage_flags __read_mostly =
  36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  37         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  38 #endif
  39 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  40         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  41 #endif
  42         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  43         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  44         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  45
  46 /* default scan 8*512 pte (or vmas) every 30 second */
  47 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  48 static unsigned int khugepaged_pages_collapsed;
  49 static unsigned int khugepaged_full_scans;
  50 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  51 /* during fragmentation poll the hugepage allocator once every minute */
  52 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  53 static struct task_struct *khugepaged_thread __read_mostly;
  54 static DEFINE_MUTEX(khugepaged_mutex);
  55 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  56 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  57 /*
  58  * default collapse hugepages if there is at least one pte mapped like
  59  * it would have happened if the vma was large enough during page
  60  * fault.
  61  */
  62 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  63
  64 static int khugepaged(void *none);
  65 static int mm_slots_hash_init(void);
  66 static int khugepaged_slab_init(void);
  67 static void khugepaged_slab_free(void);
  68
  69 #define MM_SLOTS_HASH_HEADS 1024
  70 static struct hlist_head *mm_slots_hash __read_mostly;
  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         err = mm_slots_hash_init();
 638         if (err) {
 639                 khugepaged_slab_free();
 640                 goto out;
 641         }
 642
 643         register_shrinker(&huge_zero_page_shrinker);
 644
 645         /*
 646          * By default disable transparent hugepages on smaller systems,
 647          * where the extra memory used could hurt more than TLB overhead
 648          * is likely to save.  The admin can still enable it through /sys.
 649          */
 650         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 651                 transparent_hugepage_flags = 0;
 652
 653         start_khugepaged();
 654
 655         return 0;
 656 out:
 657         hugepage_exit_sysfs(hugepage_kobj);
 658         return err;
 659 }
 660 module_init(hugepage_init)
 661
 662 static int __init setup_transparent_hugepage(char *str)
 663 {
 664         int ret = 0;
 665         if (!str)
 666                 goto out;
 667         if (!strcmp(str, "always")) {
 668                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 669                         &transparent_hugepage_flags);
 670                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 671                           &transparent_hugepage_flags);
 672                 ret = 1;
 673         } else if (!strcmp(str, "madvise")) {
 674                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 675                           &transparent_hugepage_flags);
 676                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 677                         &transparent_hugepage_flags);
 678                 ret = 1;
 679         } else if (!strcmp(str, "never")) {
 680                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 681                           &transparent_hugepage_flags);
 682                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 683                           &transparent_hugepage_flags);
 684                 ret = 1;
 685         }
 686 out:
 687         if (!ret)
 688                 printk(KERN_WARNING
 689                        "transparent_hugepage= cannot parse, ignored\n");
 690         return ret;
 691 }
 692 __setup("transparent_hugepage=", setup_transparent_hugepage);
 693
 694 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 695 {
 696         if (likely(vma->vm_flags & VM_WRITE))
 697                 pmd = pmd_mkwrite(pmd);
 698         return pmd;
 699 }
 700
 701 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
 702 {
 703         pmd_t entry;
 704         entry = mk_pmd(page, vma->vm_page_prot);
 705         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 706         entry = pmd_mkhuge(entry);
 707         return entry;
 708 }
 709
 710 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 711                                         struct vm_area_struct *vma,
 712                                         unsigned long haddr, pmd_t *pmd,
 713                                         struct page *page)
 714 {
 715         pgtable_t pgtable;
 716
 717         VM_BUG_ON(!PageCompound(page));
 718         pgtable = pte_alloc_one(mm, haddr);
 719         if (unlikely(!pgtable))
 720                 return VM_FAULT_OOM;
 721
 722         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 723         __SetPageUptodate(page);
 724
 725         spin_lock(&mm->page_table_lock);
 726         if (unlikely(!pmd_none(*pmd))) {
 727                 spin_unlock(&mm->page_table_lock);
 728                 mem_cgroup_uncharge_page(page);
 729                 put_page(page);
 730                 pte_free(mm, pgtable);
 731         } else {
 732                 pmd_t entry;
 733                 entry = mk_huge_pmd(page, vma);
 734                 /*
 735                  * The spinlocking to take the lru_lock inside
 736                  * page_add_new_anon_rmap() acts as a full memory
 737                  * barrier to be sure clear_huge_page writes become
 738                  * visible after the set_pmd_at() write.
 739                  */
 740                 page_add_new_anon_rmap(page, vma, haddr);
 741                 set_pmd_at(mm, haddr, pmd, entry);
 742                 pgtable_trans_huge_deposit(mm, pgtable);
 743                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 744                 mm->nr_ptes++;
 745                 spin_unlock(&mm->page_table_lock);
 746         }
 747
 748         return 0;
 749 }
 750
 751 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 752 {
 753         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 754 }
 755
 756 static inline struct page *alloc_hugepage_vma(int defrag,
 757                                               struct vm_area_struct *vma,
 758                                               unsigned long haddr, int nd,
 759                                               gfp_t extra_gfp)
 760 {
 761         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 762                                HPAGE_PMD_ORDER, vma, haddr, nd);
 763 }
 764
 765 #ifndef CONFIG_NUMA
 766 static inline struct page *alloc_hugepage(int defrag)
 767 {
 768         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 769                            HPAGE_PMD_ORDER);
 770 }
 771 #endif
 772
 773 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 774                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 775                 unsigned long zero_pfn)
 776 {
 777         pmd_t entry;
 778         if (!pmd_none(*pmd))
 779                 return false;
 780         entry = pfn_pmd(zero_pfn, vma->vm_page_prot);
 781         entry = pmd_wrprotect(entry);
 782         entry = pmd_mkhuge(entry);
 783         set_pmd_at(mm, haddr, pmd, entry);
 784         pgtable_trans_huge_deposit(mm, pgtable);
 785         mm->nr_ptes++;
 786         return true;
 787 }
 788
 789 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 790                                unsigned long address, pmd_t *pmd,
 791                                unsigned int flags)
 792 {
 793         struct page *page;
 794         unsigned long haddr = address & HPAGE_PMD_MASK;
 795         pte_t *pte;
 796
 797         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 798                 if (unlikely(anon_vma_prepare(vma)))
 799                         return VM_FAULT_OOM;
 800                 if (unlikely(khugepaged_enter(vma)))
 801                         return VM_FAULT_OOM;
 802                 if (!(flags & FAULT_FLAG_WRITE) &&
 803                                 transparent_hugepage_use_zero_page()) {
 804                         pgtable_t pgtable;
 805                         unsigned long zero_pfn;
 806                         bool set;
 807                         pgtable = pte_alloc_one(mm, haddr);
 808                         if (unlikely(!pgtable))
 809                                 return VM_FAULT_OOM;
 810                         zero_pfn = get_huge_zero_page();
 811                         if (unlikely(!zero_pfn)) {
 812                                 pte_free(mm, pgtable);
 813                                 count_vm_event(THP_FAULT_FALLBACK);
 814                                 goto out;
 815                         }
 816                         spin_lock(&mm->page_table_lock);
 817                         set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 818                                         zero_pfn);
 819                         spin_unlock(&mm->page_table_lock);
 820                         if (!set) {
 821                                 pte_free(mm, pgtable);
 822                                 put_huge_zero_page();
 823                         }
 824                         return 0;
 825                 }
 826                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 827                                           vma, haddr, numa_node_id(), 0);
 828                 if (unlikely(!page)) {
 829                         count_vm_event(THP_FAULT_FALLBACK);
 830                         goto out;
 831                 }
 832                 count_vm_event(THP_FAULT_ALLOC);
 833                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 834                         put_page(page);
 835                         goto out;
 836                 }
 837                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 838                                                           page))) {
 839                         mem_cgroup_uncharge_page(page);
 840                         put_page(page);
 841                         goto out;
 842                 }
 843
 844                 return 0;
 845         }
 846 out:
 847         /*
 848          * Use __pte_alloc instead of pte_alloc_map, because we can't
 849          * run pte_offset_map on the pmd, if an huge pmd could
 850          * materialize from under us from a different thread.
 851          */
 852         if (unlikely(pmd_none(*pmd)) &&
 853             unlikely(__pte_alloc(mm, vma, pmd, address)))
 854                 return VM_FAULT_OOM;
 855         /* if an huge pmd materialized from under us just retry later */
 856         if (unlikely(pmd_trans_huge(*pmd)))
 857                 return 0;
 858         /*
 859          * A regular pmd is established and it can't morph into a huge pmd
 860          * from under us anymore at this point because we hold the mmap_sem
 861          * read mode and khugepaged takes it in write mode. So now it's
 862          * safe to run pte_offset_map().
 863          */
 864         pte = pte_offset_map(pmd, address);
 865         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 866 }
 867
 868 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 869                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 870                   struct vm_area_struct *vma)
 871 {
 872         struct page *src_page;
 873         pmd_t pmd;
 874         pgtable_t pgtable;
 875         int ret;
 876
 877         ret = -ENOMEM;
 878         pgtable = pte_alloc_one(dst_mm, addr);
 879         if (unlikely(!pgtable))
 880                 goto out;
 881
 882         spin_lock(&dst_mm->page_table_lock);
 883         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 884
 885         ret = -EAGAIN;
 886         pmd = *src_pmd;
 887         if (unlikely(!pmd_trans_huge(pmd))) {
 888                 pte_free(dst_mm, pgtable);
 889                 goto out_unlock;
 890         }
 891         /*
 892          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 893          * under splitting since we don't split the page itself, only pmd to
 894          * a page table.
 895          */
 896         if (is_huge_zero_pmd(pmd)) {
 897                 unsigned long zero_pfn;
 898                 bool set;
 899                 /*
 900                  * get_huge_zero_page() will never allocate a new page here,
 901                  * since we already have a zero page to copy. It just takes a
 902                  * reference.
 903                  */
 904                 zero_pfn = get_huge_zero_page();
 905                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 906                                 zero_pfn);
 907                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 908                 ret = 0;
 909                 goto out_unlock;
 910         }
 911         if (unlikely(pmd_trans_splitting(pmd))) {
 912                 /* split huge page running from under us */
 913                 spin_unlock(&src_mm->page_table_lock);
 914                 spin_unlock(&dst_mm->page_table_lock);
 915                 pte_free(dst_mm, pgtable);
 916
 917                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 918                 goto out;
 919         }
 920         src_page = pmd_page(pmd);
 921         VM_BUG_ON(!PageHead(src_page));
 922         get_page(src_page);
 923         page_dup_rmap(src_page);
 924         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 925
 926         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 927         pmd = pmd_mkold(pmd_wrprotect(pmd));
 928         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 929         pgtable_trans_huge_deposit(dst_mm, pgtable);
 930         dst_mm->nr_ptes++;
 931
 932         ret = 0;
 933 out_unlock:
 934         spin_unlock(&src_mm->page_table_lock);
 935         spin_unlock(&dst_mm->page_table_lock);
 936 out:
 937         return ret;
 938 }
 939
 940 void huge_pmd_set_accessed(struct mm_struct *mm,
 941                            struct vm_area_struct *vma,
 942                            unsigned long address,
 943                            pmd_t *pmd, pmd_t orig_pmd,
 944                            int dirty)
 945 {
 946         pmd_t entry;
 947         unsigned long haddr;
 948
 949         spin_lock(&mm->page_table_lock);
 950         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 951                 goto unlock;
 952
 953         entry = pmd_mkyoung(orig_pmd);
 954         haddr = address & HPAGE_PMD_MASK;
 955         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 956                 update_mmu_cache_pmd(vma, address, pmd);
 957
 958 unlock:
 959         spin_unlock(&mm->page_table_lock);
 960 }
 961
 962 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 963                 struct vm_area_struct *vma, unsigned long address,
 964                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 965 {
 966         pgtable_t pgtable;
 967         pmd_t _pmd;
 968         struct page *page;
 969         int i, ret = 0;
 970         unsigned long mmun_start;       /* For mmu_notifiers */
 971         unsigned long mmun_end;         /* For mmu_notifiers */
 972
 973         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 974         if (!page) {
 975                 ret |= VM_FAULT_OOM;
 976                 goto out;
 977         }
 978
 979         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 980                 put_page(page);
 981                 ret |= VM_FAULT_OOM;
 982                 goto out;
 983         }
 984
 985         clear_user_highpage(page, address);
 986         __SetPageUptodate(page);
 987
 988         mmun_start = haddr;
 989         mmun_end   = haddr + HPAGE_PMD_SIZE;
 990         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 991
 992         spin_lock(&mm->page_table_lock);
 993         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 994                 goto out_free_page;
 995
 996         pmdp_clear_flush(vma, haddr, pmd);
 997         /* leave pmd empty until pte is filled */
 998
 999         pgtable = pgtable_trans_huge_withdraw(mm);
1000         pmd_populate(mm, &_pmd, pgtable);
1001
1002         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1003                 pte_t *pte, entry;
1004                 if (haddr == (address & PAGE_MASK)) {
1005                         entry = mk_pte(page, vma->vm_page_prot);
1006                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1007                         page_add_new_anon_rmap(page, vma, haddr);
1008                 } else {
1009                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1010                         entry = pte_mkspecial(entry);
1011                 }
1012                 pte = pte_offset_map(&_pmd, haddr);
1013                 VM_BUG_ON(!pte_none(*pte));
1014                 set_pte_at(mm, haddr, pte, entry);
1015                 pte_unmap(pte);
1016         }
1017         smp_wmb(); /* make pte visible before pmd */
1018         pmd_populate(mm, pmd, pgtable);
1019         spin_unlock(&mm->page_table_lock);
1020         put_huge_zero_page();
1021         inc_mm_counter(mm, MM_ANONPAGES);
1022
1023         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1024
1025         ret |= VM_FAULT_WRITE;
1026 out:
1027         return ret;
1028 out_free_page:
1029         spin_unlock(&mm->page_table_lock);
1030         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1031         mem_cgroup_uncharge_page(page);
1032         put_page(page);
1033         goto out;
1034 }
1035
1036 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1037                                         struct vm_area_struct *vma,
1038                                         unsigned long address,
1039                                         pmd_t *pmd, pmd_t orig_pmd,
1040                                         struct page *page,
1041                                         unsigned long haddr)
1042 {
1043         pgtable_t pgtable;
1044         pmd_t _pmd;
1045         int ret = 0, i;
1046         struct page **pages;
1047         unsigned long mmun_start;       /* For mmu_notifiers */
1048         unsigned long mmun_end;         /* For mmu_notifiers */
1049
1050         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1051                         GFP_KERNEL);
1052         if (unlikely(!pages)) {
1053                 ret |= VM_FAULT_OOM;
1054                 goto out;
1055         }
1056
1057         for (i = 0; i < HPAGE_PMD_NR; i++) {
1058                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1059                                                __GFP_OTHER_NODE,
1060                                                vma, address, page_to_nid(page));
1061                 if (unlikely(!pages[i] ||
1062                              mem_cgroup_newpage_charge(pages[i], mm,
1063                                                        GFP_KERNEL))) {
1064                         if (pages[i])
1065                                 put_page(pages[i]);
1066                         mem_cgroup_uncharge_start();
1067                         while (--i >= 0) {
1068                                 mem_cgroup_uncharge_page(pages[i]);
1069                                 put_page(pages[i]);
1070                         }
1071                         mem_cgroup_uncharge_end();
1072                         kfree(pages);
1073                         ret |= VM_FAULT_OOM;
1074                         goto out;
1075                 }
1076         }
1077
1078         for (i = 0; i < HPAGE_PMD_NR; i++) {
1079                 copy_user_highpage(pages[i], page + i,
1080                                    haddr + PAGE_SIZE * i, vma);
1081                 __SetPageUptodate(pages[i]);
1082                 cond_resched();
1083         }
1084
1085         mmun_start = haddr;
1086         mmun_end   = haddr + HPAGE_PMD_SIZE;
1087         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1088
1089         spin_lock(&mm->page_table_lock);
1090         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1091                 goto out_free_pages;
1092         VM_BUG_ON(!PageHead(page));
1093
1094         pmdp_clear_flush(vma, haddr, pmd);
1095         /* leave pmd empty until pte is filled */
1096
1097         pgtable = pgtable_trans_huge_withdraw(mm);
1098         pmd_populate(mm, &_pmd, pgtable);
1099
1100         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1101                 pte_t *pte, entry;
1102                 entry = mk_pte(pages[i], vma->vm_page_prot);
1103                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1104                 page_add_new_anon_rmap(pages[i], vma, haddr);
1105                 pte = pte_offset_map(&_pmd, haddr);
1106                 VM_BUG_ON(!pte_none(*pte));
1107                 set_pte_at(mm, haddr, pte, entry);
1108                 pte_unmap(pte);
1109         }
1110         kfree(pages);
1111
1112         smp_wmb(); /* make pte visible before pmd */
1113         pmd_populate(mm, pmd, pgtable);
1114         page_remove_rmap(page);
1115         spin_unlock(&mm->page_table_lock);
1116
1117         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1118
1119         ret |= VM_FAULT_WRITE;
1120         put_page(page);
1121
1122 out:
1123         return ret;
1124
1125 out_free_pages:
1126         spin_unlock(&mm->page_table_lock);
1127         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1128         mem_cgroup_uncharge_start();
1129         for (i = 0; i < HPAGE_PMD_NR; i++) {
1130                 mem_cgroup_uncharge_page(pages[i]);
1131                 put_page(pages[i]);
1132         }
1133         mem_cgroup_uncharge_end();
1134         kfree(pages);
1135         goto out;
1136 }
1137
1138 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1139                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1140 {
1141         int ret = 0;
1142         struct page *page = NULL, *new_page;
1143         unsigned long haddr;
1144         unsigned long mmun_start;       /* For mmu_notifiers */
1145         unsigned long mmun_end;         /* For mmu_notifiers */
1146
1147         VM_BUG_ON(!vma->anon_vma);
1148         haddr = address & HPAGE_PMD_MASK;
1149         if (is_huge_zero_pmd(orig_pmd))
1150                 goto alloc;
1151         spin_lock(&mm->page_table_lock);
1152         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1153                 goto out_unlock;
1154
1155         page = pmd_page(orig_pmd);
1156         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1157         if (page_mapcount(page) == 1) {
1158                 pmd_t entry;
1159                 entry = pmd_mkyoung(orig_pmd);
1160                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1161                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1162                         update_mmu_cache_pmd(vma, address, pmd);
1163                 ret |= VM_FAULT_WRITE;
1164                 goto out_unlock;
1165         }
1166         get_page(page);
1167         spin_unlock(&mm->page_table_lock);
1168 alloc:
1169         if (transparent_hugepage_enabled(vma) &&
1170             !transparent_hugepage_debug_cow())
1171                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1172                                               vma, haddr, numa_node_id(), 0);
1173         else
1174                 new_page = NULL;
1175
1176         if (unlikely(!new_page)) {
1177                 count_vm_event(THP_FAULT_FALLBACK);
1178                 if (is_huge_zero_pmd(orig_pmd)) {
1179                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1180                                         address, pmd, orig_pmd, haddr);
1181                 } else {
1182                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1183                                         pmd, orig_pmd, page, haddr);
1184                         if (ret & VM_FAULT_OOM)
1185                                 split_huge_page(page);
1186                         put_page(page);
1187                 }
1188                 goto out;
1189         }
1190         count_vm_event(THP_FAULT_ALLOC);
1191
1192         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1193                 put_page(new_page);
1194                 if (page) {
1195                         split_huge_page(page);
1196                         put_page(page);
1197                 }
1198                 ret |= VM_FAULT_OOM;
1199                 goto out;
1200         }
1201
1202         if (is_huge_zero_pmd(orig_pmd))
1203                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1204         else
1205                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1206         __SetPageUptodate(new_page);
1207
1208         mmun_start = haddr;
1209         mmun_end   = haddr + HPAGE_PMD_SIZE;
1210         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1211
1212         spin_lock(&mm->page_table_lock);
1213         if (page)
1214                 put_page(page);
1215         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1216                 spin_unlock(&mm->page_table_lock);
1217                 mem_cgroup_uncharge_page(new_page);
1218                 put_page(new_page);
1219                 goto out_mn;
1220         } else {
1221                 pmd_t entry;
1222                 entry = mk_huge_pmd(new_page, vma);
1223                 pmdp_clear_flush(vma, haddr, pmd);
1224                 page_add_new_anon_rmap(new_page, vma, haddr);
1225                 set_pmd_at(mm, haddr, pmd, entry);
1226                 update_mmu_cache_pmd(vma, address, pmd);
1227                 if (is_huge_zero_pmd(orig_pmd)) {
1228                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1229                         put_huge_zero_page();
1230                 } else {
1231                         VM_BUG_ON(!PageHead(page));
1232                         page_remove_rmap(page);
1233                         put_page(page);
1234                 }
1235                 ret |= VM_FAULT_WRITE;
1236         }
1237         spin_unlock(&mm->page_table_lock);
1238 out_mn:
1239         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1240 out:
1241         return ret;
1242 out_unlock:
1243         spin_unlock(&mm->page_table_lock);
1244         return ret;
1245 }
1246
1247 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1248                                    unsigned long addr,
1249                                    pmd_t *pmd,
1250                                    unsigned int flags)
1251 {
1252         struct mm_struct *mm = vma->vm_mm;
1253         struct page *page = NULL;
1254
1255         assert_spin_locked(&mm->page_table_lock);
1256
1257         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1258                 goto out;
1259
1260         page = pmd_page(*pmd);
1261         VM_BUG_ON(!PageHead(page));
1262         if (flags & FOLL_TOUCH) {
1263                 pmd_t _pmd;
1264                 /*
1265                  * We should set the dirty bit only for FOLL_WRITE but
1266                  * for now the dirty bit in the pmd is meaningless.
1267                  * And if the dirty bit will become meaningful and
1268                  * we'll only set it with FOLL_WRITE, an atomic
1269                  * set_bit will be required on the pmd to set the
1270                  * young bit, instead of the current set_pmd_at.
1271                  */
1272                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1273                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1274         }
1275         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1276                 if (page->mapping && trylock_page(page)) {
1277                         lru_add_drain();
1278                         if (page->mapping)
1279                                 mlock_vma_page(page);
1280                         unlock_page(page);
1281                 }
1282         }
1283         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1284         VM_BUG_ON(!PageCompound(page));
1285         if (flags & FOLL_GET)
1286                 get_page_foll(page);
1287
1288 out:
1289         return page;
1290 }
1291
1292 /* NUMA hinting page fault entry point for trans huge pmds */
1293 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1294                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1295 {
1296         struct page *page;
1297         unsigned long haddr = addr & HPAGE_PMD_MASK;
1298         int target_nid;
1299         int current_nid = -1;
1300         bool migrated;
1301         bool page_locked = false;
1302
1303         spin_lock(&mm->page_table_lock);
1304         if (unlikely(!pmd_same(pmd, *pmdp)))
1305                 goto out_unlock;
1306
1307         page = pmd_page(pmd);
1308         get_page(page);
1309         current_nid = page_to_nid(page);
1310         count_vm_numa_event(NUMA_HINT_FAULTS);
1311         if (current_nid == numa_node_id())
1312                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1313
1314         target_nid = mpol_misplaced(page, vma, haddr);
1315         if (target_nid == -1) {
1316                 put_page(page);
1317                 goto clear_pmdnuma;
1318         }
1319
1320         /* Acquire the page lock to serialise THP migrations */
1321         spin_unlock(&mm->page_table_lock);
1322         lock_page(page);
1323         page_locked = true;
1324
1325         /* Confirm the PTE did not while locked */
1326         spin_lock(&mm->page_table_lock);
1327         if (unlikely(!pmd_same(pmd, *pmdp))) {
1328                 unlock_page(page);
1329                 put_page(page);
1330                 goto out_unlock;
1331         }
1332         spin_unlock(&mm->page_table_lock);
1333
1334         /* Migrate the THP to the requested node */
1335         migrated = migrate_misplaced_transhuge_page(mm, vma,
1336                                 pmdp, pmd, addr,
1337                                 page, target_nid);
1338         if (migrated)
1339                 current_nid = target_nid;
1340         else {
1341                 spin_lock(&mm->page_table_lock);
1342                 if (unlikely(!pmd_same(pmd, *pmdp))) {
1343                         unlock_page(page);
1344                         goto out_unlock;
1345                 }
1346                 goto clear_pmdnuma;
1347         }
1348
1349         task_numa_fault(current_nid, HPAGE_PMD_NR, migrated);
1350         return 0;
1351
1352 clear_pmdnuma:
1353         pmd = pmd_mknonnuma(pmd);
1354         set_pmd_at(mm, haddr, pmdp, pmd);
1355         VM_BUG_ON(pmd_numa(*pmdp));
1356         update_mmu_cache_pmd(vma, addr, pmdp);
1357         if (page_locked)
1358                 unlock_page(page);
1359
1360 out_unlock:
1361         spin_unlock(&mm->page_table_lock);
1362         if (current_nid != -1)
1363                 task_numa_fault(current_nid, HPAGE_PMD_NR, migrated);
1364         return 0;
1365 }
1366
1367 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1368                  pmd_t *pmd, unsigned long addr)
1369 {
1370         int ret = 0;
1371
1372         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1373                 struct page *page;
1374                 pgtable_t pgtable;
1375                 pmd_t orig_pmd;
1376                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1377                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1378                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1379                 if (is_huge_zero_pmd(orig_pmd)) {
1380                         tlb->mm->nr_ptes--;
1381                         spin_unlock(&tlb->mm->page_table_lock);
1382                         put_huge_zero_page();
1383                 } else {
1384                         page = pmd_page(orig_pmd);
1385                         page_remove_rmap(page);
1386                         VM_BUG_ON(page_mapcount(page) < 0);
1387                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1388                         VM_BUG_ON(!PageHead(page));
1389                         tlb->mm->nr_ptes--;
1390                         spin_unlock(&tlb->mm->page_table_lock);
1391                         tlb_remove_page(tlb, page);
1392                 }
1393                 pte_free(tlb->mm, pgtable);
1394                 ret = 1;
1395         }
1396         return ret;
1397 }
1398
1399 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1400                 unsigned long addr, unsigned long end,
1401                 unsigned char *vec)
1402 {
1403         int ret = 0;
1404
1405         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1406                 /*
1407                  * All logical pages in the range are present
1408                  * if backed by a huge page.
1409                  */
1410                 spin_unlock(&vma->vm_mm->page_table_lock);
1411                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1412                 ret = 1;
1413         }
1414
1415         return ret;
1416 }
1417
1418 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1419                   unsigned long old_addr,
1420                   unsigned long new_addr, unsigned long old_end,
1421                   pmd_t *old_pmd, pmd_t *new_pmd)
1422 {
1423         int ret = 0;
1424         pmd_t pmd;
1425
1426         struct mm_struct *mm = vma->vm_mm;
1427
1428         if ((old_addr & ~HPAGE_PMD_MASK) ||
1429             (new_addr & ~HPAGE_PMD_MASK) ||
1430             old_end - old_addr < HPAGE_PMD_SIZE ||
1431             (new_vma->vm_flags & VM_NOHUGEPAGE))
1432                 goto out;
1433
1434         /*
1435          * The destination pmd shouldn't be established, free_pgtables()
1436          * should have release it.
1437          */
1438         if (WARN_ON(!pmd_none(*new_pmd))) {
1439                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1440                 goto out;
1441         }
1442
1443         ret = __pmd_trans_huge_lock(old_pmd, vma);
1444         if (ret == 1) {
1445                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1446                 VM_BUG_ON(!pmd_none(*new_pmd));
1447                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1448                 spin_unlock(&mm->page_table_lock);
1449         }
1450 out:
1451         return ret;
1452 }
1453
1454 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1455                 unsigned long addr, pgprot_t newprot, int prot_numa)
1456 {
1457         struct mm_struct *mm = vma->vm_mm;
1458         int ret = 0;
1459
1460         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1461                 pmd_t entry;
1462                 entry = pmdp_get_and_clear(mm, addr, pmd);
1463                 if (!prot_numa) {
1464                         entry = pmd_modify(entry, newprot);
1465                         BUG_ON(pmd_write(entry));
1466                 } else {
1467                         struct page *page = pmd_page(*pmd);
1468
1469                         /* only check non-shared pages */
1470                         if (page_mapcount(page) == 1 &&
1471                             !pmd_numa(*pmd)) {
1472                                 entry = pmd_mknuma(entry);
1473                         }
1474                 }
1475                 set_pmd_at(mm, addr, pmd, entry);
1476                 spin_unlock(&vma->vm_mm->page_table_lock);
1477                 ret = 1;
1478         }
1479
1480         return ret;
1481 }
1482
1483 /*
1484  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1485  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1486  *
1487  * Note that if it returns 1, this routine returns without unlocking page
1488  * table locks. So callers must unlock them.
1489  */
1490 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1491 {
1492         spin_lock(&vma->vm_mm->page_table_lock);
1493         if (likely(pmd_trans_huge(*pmd))) {
1494                 if (unlikely(pmd_trans_splitting(*pmd))) {
1495                         spin_unlock(&vma->vm_mm->page_table_lock);
1496                         wait_split_huge_page(vma->anon_vma, pmd);
1497                         return -1;
1498                 } else {
1499                         /* Thp mapped by 'pmd' is stable, so we can
1500                          * handle it as it is. */
1501                         return 1;
1502                 }
1503         }
1504         spin_unlock(&vma->vm_mm->page_table_lock);
1505         return 0;
1506 }
1507
1508 pmd_t *page_check_address_pmd(struct page *page,
1509                               struct mm_struct *mm,
1510                               unsigned long address,
1511                               enum page_check_address_pmd_flag flag)
1512 {
1513         pmd_t *pmd, *ret = NULL;
1514
1515         if (address & ~HPAGE_PMD_MASK)
1516                 goto out;
1517
1518         pmd = mm_find_pmd(mm, address);
1519         if (!pmd)
1520                 goto out;
1521         if (pmd_none(*pmd))
1522                 goto out;
1523         if (pmd_page(*pmd) != page)
1524                 goto out;
1525         /*
1526          * split_vma() may create temporary aliased mappings. There is
1527          * no risk as long as all huge pmd are found and have their
1528          * splitting bit set before __split_huge_page_refcount
1529          * runs. Finding the same huge pmd more than once during the
1530          * same rmap walk is not a problem.
1531          */
1532         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1533             pmd_trans_splitting(*pmd))
1534                 goto out;
1535         if (pmd_trans_huge(*pmd)) {
1536                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1537                           !pmd_trans_splitting(*pmd));
1538                 ret = pmd;
1539         }
1540 out:
1541         return ret;
1542 }
1543
1544 static int __split_huge_page_splitting(struct page *page,
1545                                        struct vm_area_struct *vma,
1546                                        unsigned long address)
1547 {
1548         struct mm_struct *mm = vma->vm_mm;
1549         pmd_t *pmd;
1550         int ret = 0;
1551         /* For mmu_notifiers */
1552         const unsigned long mmun_start = address;
1553         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1554
1555         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1556         spin_lock(&mm->page_table_lock);
1557         pmd = page_check_address_pmd(page, mm, address,
1558                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1559         if (pmd) {
1560                 /*
1561                  * We can't temporarily set the pmd to null in order
1562                  * to split it, the pmd must remain marked huge at all
1563                  * times or the VM won't take the pmd_trans_huge paths
1564                  * and it won't wait on the anon_vma->root->rwsem to
1565                  * serialize against split_huge_page*.
1566                  */
1567                 pmdp_splitting_flush(vma, address, pmd);
1568                 ret = 1;
1569         }
1570         spin_unlock(&mm->page_table_lock);
1571         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1572
1573         return ret;
1574 }
1575
1576 static void __split_huge_page_refcount(struct page *page)
1577 {
1578         int i;
1579         struct zone *zone = page_zone(page);
1580         struct lruvec *lruvec;
1581         int tail_count = 0;
1582
1583         /* prevent PageLRU to go away from under us, and freeze lru stats */
1584         spin_lock_irq(&zone->lru_lock);
1585         lruvec = mem_cgroup_page_lruvec(page, zone);
1586
1587         compound_lock(page);
1588         /* complete memcg works before add pages to LRU */
1589         mem_cgroup_split_huge_fixup(page);
1590
1591         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1592                 struct page *page_tail = page + i;
1593
1594                 /* tail_page->_mapcount cannot change */
1595                 BUG_ON(page_mapcount(page_tail) < 0);
1596                 tail_count += page_mapcount(page_tail);
1597                 /* check for overflow */
1598                 BUG_ON(tail_count < 0);
1599                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1600                 /*
1601                  * tail_page->_count is zero and not changing from
1602                  * under us. But get_page_unless_zero() may be running
1603                  * from under us on the tail_page. If we used
1604                  * atomic_set() below instead of atomic_add(), we
1605                  * would then run atomic_set() concurrently with
1606                  * get_page_unless_zero(), and atomic_set() is
1607                  * implemented in C not using locked ops. spin_unlock
1608                  * on x86 sometime uses locked ops because of PPro
1609                  * errata 66, 92, so unless somebody can guarantee
1610                  * atomic_set() here would be safe on all archs (and
1611                  * not only on x86), it's safer to use atomic_add().
1612                  */
1613                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1614                            &page_tail->_count);
1615
1616                 /* after clearing PageTail the gup refcount can be released */
1617                 smp_mb();
1618
1619                 /*
1620                  * retain hwpoison flag of the poisoned tail page:
1621                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1622                  *   by the memory-failure.
1623                  */
1624                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1625                 page_tail->flags |= (page->flags &
1626                                      ((1L << PG_referenced) |
1627                                       (1L << PG_swapbacked) |
1628                                       (1L << PG_mlocked) |
1629                                       (1L << PG_uptodate)));
1630                 page_tail->flags |= (1L << PG_dirty);
1631
1632                 /* clear PageTail before overwriting first_page */
1633                 smp_wmb();
1634
1635                 /*
1636                  * __split_huge_page_splitting() already set the
1637                  * splitting bit in all pmd that could map this
1638                  * hugepage, that will ensure no CPU can alter the
1639                  * mapcount on the head page. The mapcount is only
1640                  * accounted in the head page and it has to be
1641                  * transferred to all tail pages in the below code. So
1642                  * for this code to be safe, the split the mapcount
1643                  * can't change. But that doesn't mean userland can't
1644                  * keep changing and reading the page contents while
1645                  * we transfer the mapcount, so the pmd splitting
1646                  * status is achieved setting a reserved bit in the
1647                  * pmd, not by clearing the present bit.
1648                 */
1649                 page_tail->_mapcount = page->_mapcount;
1650
1651                 BUG_ON(page_tail->mapping);
1652                 page_tail->mapping = page->mapping;
1653
1654                 page_tail->index = page->index + i;
1655                 page_xchg_last_nid(page_tail, page_last_nid(page));
1656
1657                 BUG_ON(!PageAnon(page_tail));
1658                 BUG_ON(!PageUptodate(page_tail));
1659                 BUG_ON(!PageDirty(page_tail));
1660                 BUG_ON(!PageSwapBacked(page_tail));
1661
1662                 lru_add_page_tail(page, page_tail, lruvec);
1663         }
1664         atomic_sub(tail_count, &page->_count);
1665         BUG_ON(atomic_read(&page->_count) <= 0);
1666
1667         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1668         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1669
1670         ClearPageCompound(page);
1671         compound_unlock(page);
1672         spin_unlock_irq(&zone->lru_lock);
1673
1674         for (i = 1; i < HPAGE_PMD_NR; i++) {
1675                 struct page *page_tail = page + i;
1676                 BUG_ON(page_count(page_tail) <= 0);
1677                 /*
1678                  * Tail pages may be freed if there wasn't any mapping
1679                  * like if add_to_swap() is running on a lru page that
1680                  * had its mapping zapped. And freeing these pages
1681                  * requires taking the lru_lock so we do the put_page
1682                  * of the tail pages after the split is complete.
1683                  */
1684                 put_page(page_tail);
1685         }
1686
1687         /*
1688          * Only the head page (now become a regular page) is required
1689          * to be pinned by the caller.
1690          */
1691         BUG_ON(page_count(page) <= 0);
1692 }
1693
1694 static int __split_huge_page_map(struct page *page,
1695                                  struct vm_area_struct *vma,
1696                                  unsigned long address)
1697 {
1698         struct mm_struct *mm = vma->vm_mm;
1699         pmd_t *pmd, _pmd;
1700         int ret = 0, i;
1701         pgtable_t pgtable;
1702         unsigned long haddr;
1703
1704         spin_lock(&mm->page_table_lock);
1705         pmd = page_check_address_pmd(page, mm, address,
1706                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1707         if (pmd) {
1708                 pgtable = pgtable_trans_huge_withdraw(mm);
1709                 pmd_populate(mm, &_pmd, pgtable);
1710
1711                 haddr = address;
1712                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1713                         pte_t *pte, entry;
1714                         BUG_ON(PageCompound(page+i));
1715                         entry = mk_pte(page + i, vma->vm_page_prot);
1716                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1717                         if (!pmd_write(*pmd))
1718                                 entry = pte_wrprotect(entry);
1719                         else
1720                                 BUG_ON(page_mapcount(page) != 1);
1721                         if (!pmd_young(*pmd))
1722                                 entry = pte_mkold(entry);
1723                         if (pmd_numa(*pmd))
1724                                 entry = pte_mknuma(entry);
1725                         pte = pte_offset_map(&_pmd, haddr);
1726                         BUG_ON(!pte_none(*pte));
1727                         set_pte_at(mm, haddr, pte, entry);
1728                         pte_unmap(pte);
1729                 }
1730
1731                 smp_wmb(); /* make pte visible before pmd */
1732                 /*
1733                  * Up to this point the pmd is present and huge and
1734                  * userland has the whole access to the hugepage
1735                  * during the split (which happens in place). If we
1736                  * overwrite the pmd with the not-huge version
1737                  * pointing to the pte here (which of course we could
1738                  * if all CPUs were bug free), userland could trigger
1739                  * a small page size TLB miss on the small sized TLB
1740                  * while the hugepage TLB entry is still established
1741                  * in the huge TLB. Some CPU doesn't like that. See
1742                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1743                  * Erratum 383 on page 93. Intel should be safe but is
1744                  * also warns that it's only safe if the permission
1745                  * and cache attributes of the two entries loaded in
1746                  * the two TLB is identical (which should be the case
1747                  * here). But it is generally safer to never allow
1748                  * small and huge TLB entries for the same virtual
1749                  * address to be loaded simultaneously. So instead of
1750                  * doing "pmd_populate(); flush_tlb_range();" we first
1751                  * mark the current pmd notpresent (atomically because
1752                  * here the pmd_trans_huge and pmd_trans_splitting
1753                  * must remain set at all times on the pmd until the
1754                  * split is complete for this pmd), then we flush the
1755                  * SMP TLB and finally we write the non-huge version
1756                  * of the pmd entry with pmd_populate.
1757                  */
1758                 pmdp_invalidate(vma, address, pmd);
1759                 pmd_populate(mm, pmd, pgtable);
1760                 ret = 1;
1761         }
1762         spin_unlock(&mm->page_table_lock);
1763
1764         return ret;
1765 }
1766
1767 /* must be called with anon_vma->root->rwsem held */
1768 static void __split_huge_page(struct page *page,
1769                               struct anon_vma *anon_vma)
1770 {
1771         int mapcount, mapcount2;
1772         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1773         struct anon_vma_chain *avc;
1774
1775         BUG_ON(!PageHead(page));
1776         BUG_ON(PageTail(page));
1777
1778         mapcount = 0;
1779         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1780                 struct vm_area_struct *vma = avc->vma;
1781                 unsigned long addr = vma_address(page, vma);
1782                 BUG_ON(is_vma_temporary_stack(vma));
1783                 mapcount += __split_huge_page_splitting(page, vma, addr);
1784         }
1785         /*
1786          * It is critical that new vmas are added to the tail of the
1787          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1788          * and establishes a child pmd before
1789          * __split_huge_page_splitting() freezes the parent pmd (so if
1790          * we fail to prevent copy_huge_pmd() from running until the
1791          * whole __split_huge_page() is complete), we will still see
1792          * the newly established pmd of the child later during the
1793          * walk, to be able to set it as pmd_trans_splitting too.
1794          */
1795         if (mapcount != page_mapcount(page))
1796                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1797                        mapcount, page_mapcount(page));
1798         BUG_ON(mapcount != page_mapcount(page));
1799
1800         __split_huge_page_refcount(page);
1801
1802         mapcount2 = 0;
1803         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1804                 struct vm_area_struct *vma = avc->vma;
1805                 unsigned long addr = vma_address(page, vma);
1806                 BUG_ON(is_vma_temporary_stack(vma));
1807                 mapcount2 += __split_huge_page_map(page, vma, addr);
1808         }
1809         if (mapcount != mapcount2)
1810                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1811                        mapcount, mapcount2, page_mapcount(page));
1812         BUG_ON(mapcount != mapcount2);
1813 }
1814
1815 int split_huge_page(struct page *page)
1816 {
1817         struct anon_vma *anon_vma;
1818         int ret = 1;
1819
1820         BUG_ON(is_huge_zero_pfn(page_to_pfn(page)));
1821         BUG_ON(!PageAnon(page));
1822         anon_vma = page_lock_anon_vma_read(page);
1823         if (!anon_vma)
1824                 goto out;
1825         ret = 0;
1826         if (!PageCompound(page))
1827                 goto out_unlock;
1828
1829         BUG_ON(!PageSwapBacked(page));
1830         __split_huge_page(page, anon_vma);
1831         count_vm_event(THP_SPLIT);
1832
1833         BUG_ON(PageCompound(page));
1834 out_unlock:
1835         page_unlock_anon_vma_read(anon_vma);
1836 out:
1837         return ret;
1838 }
1839
1840 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1841
1842 int hugepage_madvise(struct vm_area_struct *vma,
1843                      unsigned long *vm_flags, int advice)
1844 {
1845         struct mm_struct *mm = vma->vm_mm;
1846
1847         switch (advice) {
1848         case MADV_HUGEPAGE:
1849                 /*
1850                  * Be somewhat over-protective like KSM for now!
1851                  */
1852                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1853                         return -EINVAL;
1854                 if (mm->def_flags & VM_NOHUGEPAGE)
1855                         return -EINVAL;
1856                 *vm_flags &= ~VM_NOHUGEPAGE;
1857                 *vm_flags |= VM_HUGEPAGE;
1858                 /*
1859                  * If the vma become good for khugepaged to scan,
1860                  * register it here without waiting a page fault that
1861                  * may not happen any time soon.
1862                  */
1863                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1864                         return -ENOMEM;
1865                 break;
1866         case MADV_NOHUGEPAGE:
1867                 /*
1868                  * Be somewhat over-protective like KSM for now!
1869                  */
1870                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1871                         return -EINVAL;
1872                 *vm_flags &= ~VM_HUGEPAGE;
1873                 *vm_flags |= VM_NOHUGEPAGE;
1874                 /*
1875                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1876                  * this vma even if we leave the mm registered in khugepaged if
1877                  * it got registered before VM_NOHUGEPAGE was set.
1878                  */
1879                 break;
1880         }
1881
1882         return 0;
1883 }
1884
1885 static int __init khugepaged_slab_init(void)
1886 {
1887         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1888                                           sizeof(struct mm_slot),
1889                                           __alignof__(struct mm_slot), 0, NULL);
1890         if (!mm_slot_cache)
1891                 return -ENOMEM;
1892
1893         return 0;
1894 }
1895
1896 static void __init khugepaged_slab_free(void)
1897 {
1898         kmem_cache_destroy(mm_slot_cache);
1899         mm_slot_cache = NULL;
1900 }
1901
1902 static inline struct mm_slot *alloc_mm_slot(void)
1903 {
1904         if (!mm_slot_cache)     /* initialization failed */
1905                 return NULL;
1906         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1907 }
1908
1909 static inline void free_mm_slot(struct mm_slot *mm_slot)
1910 {
1911         kmem_cache_free(mm_slot_cache, mm_slot);
1912 }
1913
1914 static int __init mm_slots_hash_init(void)
1915 {
1916         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1917                                 GFP_KERNEL);
1918         if (!mm_slots_hash)
1919                 return -ENOMEM;
1920         return 0;
1921 }
1922
1923 #if 0
1924 static void __init mm_slots_hash_free(void)
1925 {
1926         kfree(mm_slots_hash);
1927         mm_slots_hash = NULL;
1928 }
1929 #endif
1930
1931 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1932 {
1933         struct mm_slot *mm_slot;
1934         struct hlist_head *bucket;
1935         struct hlist_node *node;
1936
1937         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1938                                 % MM_SLOTS_HASH_HEADS];
1939         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1940                 if (mm == mm_slot->mm)
1941                         return mm_slot;
1942         }
1943         return NULL;
1944 }
1945
1946 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1947                                     struct mm_slot *mm_slot)
1948 {
1949         struct hlist_head *bucket;
1950
1951         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1952                                 % MM_SLOTS_HASH_HEADS];
1953         mm_slot->mm = mm;
1954         hlist_add_head(&mm_slot->hash, bucket);
1955 }
1956
1957 static inline int khugepaged_test_exit(struct mm_struct *mm)
1958 {
1959         return atomic_read(&mm->mm_users) == 0;
1960 }
1961
1962 int __khugepaged_enter(struct mm_struct *mm)
1963 {
1964         struct mm_slot *mm_slot;
1965         int wakeup;
1966
1967         mm_slot = alloc_mm_slot();
1968         if (!mm_slot)
1969                 return -ENOMEM;
1970
1971         /* __khugepaged_exit() must not run from under us */
1972         VM_BUG_ON(khugepaged_test_exit(mm));
1973         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1974                 free_mm_slot(mm_slot);
1975                 return 0;
1976         }
1977
1978         spin_lock(&khugepaged_mm_lock);
1979         insert_to_mm_slots_hash(mm, mm_slot);
1980         /*
1981          * Insert just behind the scanning cursor, to let the area settle
1982          * down a little.
1983          */
1984         wakeup = list_empty(&khugepaged_scan.mm_head);
1985         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1986         spin_unlock(&khugepaged_mm_lock);
1987
1988         atomic_inc(&mm->mm_count);
1989         if (wakeup)
1990                 wake_up_interruptible(&khugepaged_wait);
1991
1992         return 0;
1993 }
1994
1995 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1996 {
1997         unsigned long hstart, hend;
1998         if (!vma->anon_vma)
1999                 /*
2000                  * Not yet faulted in so we will register later in the
2001                  * page fault if needed.
2002                  */
2003                 return 0;
2004         if (vma->vm_ops)
2005                 /* khugepaged not yet working on file or special mappings */
2006                 return 0;
2007         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2008         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2009         hend = vma->vm_end & HPAGE_PMD_MASK;
2010         if (hstart < hend)
2011                 return khugepaged_enter(vma);
2012         return 0;
2013 }
2014
2015 void __khugepaged_exit(struct mm_struct *mm)
2016 {
2017         struct mm_slot *mm_slot;
2018         int free = 0;
2019
2020         spin_lock(&khugepaged_mm_lock);
2021         mm_slot = get_mm_slot(mm);
2022         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2023                 hlist_del(&mm_slot->hash);
2024                 list_del(&mm_slot->mm_node);
2025                 free = 1;
2026         }
2027         spin_unlock(&khugepaged_mm_lock);
2028
2029         if (free) {
2030                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2031                 free_mm_slot(mm_slot);
2032                 mmdrop(mm);
2033         } else if (mm_slot) {
2034                 /*
2035                  * This is required to serialize against
2036                  * khugepaged_test_exit() (which is guaranteed to run
2037                  * under mmap sem read mode). Stop here (after we
2038                  * return all pagetables will be destroyed) until
2039                  * khugepaged has finished working on the pagetables
2040                  * under the mmap_sem.
2041                  */
2042                 down_write(&mm->mmap_sem);
2043                 up_write(&mm->mmap_sem);
2044         }
2045 }
2046
2047 static void release_pte_page(struct page *page)
2048 {
2049         /* 0 stands for page_is_file_cache(page) == false */
2050         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2051         unlock_page(page);
2052         putback_lru_page(page);
2053 }
2054
2055 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2056 {
2057         while (--_pte >= pte) {
2058                 pte_t pteval = *_pte;
2059                 if (!pte_none(pteval))
2060                         release_pte_page(pte_page(pteval));
2061         }
2062 }
2063
2064 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2065                                         unsigned long address,
2066                                         pte_t *pte)
2067 {
2068         struct page *page;
2069         pte_t *_pte;
2070         int referenced = 0, none = 0;
2071         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2072              _pte++, address += PAGE_SIZE) {
2073                 pte_t pteval = *_pte;
2074                 if (pte_none(pteval)) {
2075                         if (++none <= khugepaged_max_ptes_none)
2076                                 continue;
2077                         else
2078                                 goto out;
2079                 }
2080                 if (!pte_present(pteval) || !pte_write(pteval))
2081                         goto out;
2082                 page = vm_normal_page(vma, address, pteval);
2083                 if (unlikely(!page))
2084                         goto out;
2085
2086                 VM_BUG_ON(PageCompound(page));
2087                 BUG_ON(!PageAnon(page));
2088                 VM_BUG_ON(!PageSwapBacked(page));
2089
2090                 /* cannot use mapcount: can't collapse if there's a gup pin */
2091                 if (page_count(page) != 1)
2092                         goto out;
2093                 /*
2094                  * We can do it before isolate_lru_page because the
2095                  * page can't be freed from under us. NOTE: PG_lock
2096                  * is needed to serialize against split_huge_page
2097                  * when invoked from the VM.
2098                  */
2099                 if (!trylock_page(page))
2100                         goto out;
2101                 /*
2102                  * Isolate the page to avoid collapsing an hugepage
2103                  * currently in use by the VM.
2104                  */
2105                 if (isolate_lru_page(page)) {
2106                         unlock_page(page);
2107                         goto out;
2108                 }
2109                 /* 0 stands for page_is_file_cache(page) == false */
2110                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2111                 VM_BUG_ON(!PageLocked(page));
2112                 VM_BUG_ON(PageLRU(page));
2113
2114                 /* If there is no mapped pte young don't collapse the page */
2115                 if (pte_young(pteval) || PageReferenced(page) ||
2116                     mmu_notifier_test_young(vma->vm_mm, address))
2117                         referenced = 1;
2118         }
2119         if (likely(referenced))
2120                 return 1;
2121 out:
2122         release_pte_pages(pte, _pte);
2123         return 0;
2124 }
2125
2126 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2127                                       struct vm_area_struct *vma,
2128                                       unsigned long address,
2129                                       spinlock_t *ptl)
2130 {
2131         pte_t *_pte;
2132         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2133                 pte_t pteval = *_pte;
2134                 struct page *src_page;
2135
2136                 if (pte_none(pteval)) {
2137                         clear_user_highpage(page, address);
2138                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2139                 } else {
2140                         src_page = pte_page(pteval);
2141                         copy_user_highpage(page, src_page, address, vma);
2142                         VM_BUG_ON(page_mapcount(src_page) != 1);
2143                         release_pte_page(src_page);
2144                         /*
2145                          * ptl mostly unnecessary, but preempt has to
2146                          * be disabled to update the per-cpu stats
2147                          * inside page_remove_rmap().
2148                          */
2149                         spin_lock(ptl);
2150                         /*
2151                          * paravirt calls inside pte_clear here are
2152                          * superfluous.
2153                          */
2154                         pte_clear(vma->vm_mm, address, _pte);
2155                         page_remove_rmap(src_page);
2156                         spin_unlock(ptl);
2157                         free_page_and_swap_cache(src_page);
2158                 }
2159
2160                 address += PAGE_SIZE;
2161                 page++;
2162         }
2163 }
2164
2165 static void khugepaged_alloc_sleep(void)
2166 {
2167         wait_event_freezable_timeout(khugepaged_wait, false,
2168                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2169 }
2170
2171 #ifdef CONFIG_NUMA
2172 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2173 {
2174         if (IS_ERR(*hpage)) {
2175                 if (!*wait)
2176                         return false;
2177
2178                 *wait = false;
2179                 *hpage = NULL;
2180                 khugepaged_alloc_sleep();
2181         } else if (*hpage) {
2182                 put_page(*hpage);
2183                 *hpage = NULL;
2184         }
2185
2186         return true;
2187 }
2188
2189 static struct page
2190 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2191                        struct vm_area_struct *vma, unsigned long address,
2192                        int node)
2193 {
2194         VM_BUG_ON(*hpage);
2195         /*
2196          * Allocate the page while the vma is still valid and under
2197          * the mmap_sem read mode so there is no memory allocation
2198          * later when we take the mmap_sem in write mode. This is more
2199          * friendly behavior (OTOH it may actually hide bugs) to
2200          * filesystems in userland with daemons allocating memory in
2201          * the userland I/O paths.  Allocating memory with the
2202          * mmap_sem in read mode is good idea also to allow greater
2203          * scalability.
2204          */
2205         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2206                                       node, __GFP_OTHER_NODE);
2207
2208         /*
2209          * After allocating the hugepage, release the mmap_sem read lock in
2210          * preparation for taking it in write mode.
2211          */
2212         up_read(&mm->mmap_sem);
2213         if (unlikely(!*hpage)) {
2214                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2215                 *hpage = ERR_PTR(-ENOMEM);
2216                 return NULL;
2217         }
2218
2219         count_vm_event(THP_COLLAPSE_ALLOC);
2220         return *hpage;
2221 }
2222 #else
2223 static struct page *khugepaged_alloc_hugepage(bool *wait)
2224 {
2225         struct page *hpage;
2226
2227         do {
2228                 hpage = alloc_hugepage(khugepaged_defrag());
2229                 if (!hpage) {
2230                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2231                         if (!*wait)
2232                                 return NULL;
2233
2234                         *wait = false;
2235                         khugepaged_alloc_sleep();
2236                 } else
2237                         count_vm_event(THP_COLLAPSE_ALLOC);
2238         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2239
2240         return hpage;
2241 }
2242
2243 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2244 {
2245         if (!*hpage)
2246                 *hpage = khugepaged_alloc_hugepage(wait);
2247
2248         if (unlikely(!*hpage))
2249                 return false;
2250
2251         return true;
2252 }
2253
2254 static struct page
2255 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2256                        struct vm_area_struct *vma, unsigned long address,
2257                        int node)
2258 {
2259         up_read(&mm->mmap_sem);
2260         VM_BUG_ON(!*hpage);
2261         return  *hpage;
2262 }
2263 #endif
2264
2265 static bool hugepage_vma_check(struct vm_area_struct *vma)
2266 {
2267         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2268             (vma->vm_flags & VM_NOHUGEPAGE))
2269                 return false;
2270
2271         if (!vma->anon_vma || vma->vm_ops)
2272                 return false;
2273         if (is_vma_temporary_stack(vma))
2274                 return false;
2275         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2276         return true;
2277 }
2278
2279 static void collapse_huge_page(struct mm_struct *mm,
2280                                    unsigned long address,
2281                                    struct page **hpage,
2282                                    struct vm_area_struct *vma,
2283                                    int node)
2284 {
2285         pmd_t *pmd, _pmd;
2286         pte_t *pte;
2287         pgtable_t pgtable;
2288         struct page *new_page;
2289         spinlock_t *ptl;
2290         int isolated;
2291         unsigned long hstart, hend;
2292         unsigned long mmun_start;       /* For mmu_notifiers */
2293         unsigned long mmun_end;         /* For mmu_notifiers */
2294
2295         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2296
2297         /* release the mmap_sem read lock. */
2298         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2299         if (!new_page)
2300                 return;
2301
2302         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2303                 return;
2304
2305         /*
2306          * Prevent all access to pagetables with the exception of
2307          * gup_fast later hanlded by the ptep_clear_flush and the VM
2308          * handled by the anon_vma lock + PG_lock.
2309          */
2310         down_write(&mm->mmap_sem);
2311         if (unlikely(khugepaged_test_exit(mm)))
2312                 goto out;
2313
2314         vma = find_vma(mm, address);
2315         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2316         hend = vma->vm_end & HPAGE_PMD_MASK;
2317         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2318                 goto out;
2319         if (!hugepage_vma_check(vma))
2320                 goto out;
2321         pmd = mm_find_pmd(mm, address);
2322         if (!pmd)
2323                 goto out;
2324         if (pmd_trans_huge(*pmd))
2325                 goto out;
2326
2327         anon_vma_lock_write(vma->anon_vma);
2328
2329         pte = pte_offset_map(pmd, address);
2330         ptl = pte_lockptr(mm, pmd);
2331
2332         mmun_start = address;
2333         mmun_end   = address + HPAGE_PMD_SIZE;
2334         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2335         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2336         /*
2337          * After this gup_fast can't run anymore. This also removes
2338          * any huge TLB entry from the CPU so we won't allow
2339          * huge and small TLB entries for the same virtual address
2340          * to avoid the risk of CPU bugs in that area.
2341          */
2342         _pmd = pmdp_clear_flush(vma, address, pmd);
2343         spin_unlock(&mm->page_table_lock);
2344         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2345
2346         spin_lock(ptl);
2347         isolated = __collapse_huge_page_isolate(vma, address, pte);
2348         spin_unlock(ptl);
2349
2350         if (unlikely(!isolated)) {
2351                 pte_unmap(pte);
2352                 spin_lock(&mm->page_table_lock);
2353                 BUG_ON(!pmd_none(*pmd));
2354                 set_pmd_at(mm, address, pmd, _pmd);
2355                 spin_unlock(&mm->page_table_lock);
2356                 anon_vma_unlock(vma->anon_vma);
2357                 goto out;
2358         }
2359
2360         /*
2361          * All pages are isolated and locked so anon_vma rmap
2362          * can't run anymore.
2363          */
2364         anon_vma_unlock(vma->anon_vma);
2365
2366         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2367         pte_unmap(pte);
2368         __SetPageUptodate(new_page);
2369         pgtable = pmd_pgtable(_pmd);
2370
2371         _pmd = mk_huge_pmd(new_page, vma);
2372
2373         /*
2374          * spin_lock() below is not the equivalent of smp_wmb(), so
2375          * this is needed to avoid the copy_huge_page writes to become
2376          * visible after the set_pmd_at() write.
2377          */
2378         smp_wmb();
2379
2380         spin_lock(&mm->page_table_lock);
2381         BUG_ON(!pmd_none(*pmd));
2382         page_add_new_anon_rmap(new_page, vma, address);
2383         set_pmd_at(mm, address, pmd, _pmd);
2384         update_mmu_cache_pmd(vma, address, pmd);
2385         pgtable_trans_huge_deposit(mm, pgtable);
2386         spin_unlock(&mm->page_table_lock);
2387
2388         *hpage = NULL;
2389
2390         khugepaged_pages_collapsed++;
2391 out_up_write:
2392         up_write(&mm->mmap_sem);
2393         return;
2394
2395 out:
2396         mem_cgroup_uncharge_page(new_page);
2397         goto out_up_write;
2398 }
2399
2400 static int khugepaged_scan_pmd(struct mm_struct *mm,
2401                                struct vm_area_struct *vma,
2402                                unsigned long address,
2403                                struct page **hpage)
2404 {
2405         pmd_t *pmd;
2406         pte_t *pte, *_pte;
2407         int ret = 0, referenced = 0, none = 0;
2408         struct page *page;
2409         unsigned long _address;
2410         spinlock_t *ptl;
2411         int node = -1;
2412
2413         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2414
2415         pmd = mm_find_pmd(mm, address);
2416         if (!pmd)
2417                 goto out;
2418         if (pmd_trans_huge(*pmd))
2419                 goto out;
2420
2421         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2422         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2423              _pte++, _address += PAGE_SIZE) {
2424                 pte_t pteval = *_pte;
2425                 if (pte_none(pteval)) {
2426                         if (++none <= khugepaged_max_ptes_none)
2427                                 continue;
2428                         else
2429                                 goto out_unmap;
2430                 }
2431                 if (!pte_present(pteval) || !pte_write(pteval))
2432                         goto out_unmap;
2433                 page = vm_normal_page(vma, _address, pteval);
2434                 if (unlikely(!page))
2435                         goto out_unmap;
2436                 /*
2437                  * Chose the node of the first page. This could
2438                  * be more sophisticated and look at more pages,
2439                  * but isn't for now.
2440                  */
2441                 if (node == -1)
2442                         node = page_to_nid(page);
2443                 VM_BUG_ON(PageCompound(page));
2444                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2445                         goto out_unmap;
2446                 /* cannot use mapcount: can't collapse if there's a gup pin */
2447                 if (page_count(page) != 1)
2448                         goto out_unmap;
2449                 if (pte_young(pteval) || PageReferenced(page) ||
2450                     mmu_notifier_test_young(vma->vm_mm, address))
2451                         referenced = 1;
2452         }
2453         if (referenced)
2454                 ret = 1;
2455 out_unmap:
2456         pte_unmap_unlock(pte, ptl);
2457         if (ret)
2458                 /* collapse_huge_page will return with the mmap_sem released */
2459                 collapse_huge_page(mm, address, hpage, vma, node);
2460 out:
2461         return ret;
2462 }
2463
2464 static void collect_mm_slot(struct mm_slot *mm_slot)
2465 {
2466         struct mm_struct *mm = mm_slot->mm;
2467
2468         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2469
2470         if (khugepaged_test_exit(mm)) {
2471                 /* free mm_slot */
2472                 hlist_del(&mm_slot->hash);
2473                 list_del(&mm_slot->mm_node);
2474
2475                 /*
2476                  * Not strictly needed because the mm exited already.
2477                  *
2478                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2479                  */
2480
2481                 /* khugepaged_mm_lock actually not necessary for the below */
2482                 free_mm_slot(mm_slot);
2483                 mmdrop(mm);
2484         }
2485 }
2486
2487 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2488                                             struct page **hpage)
2489         __releases(&khugepaged_mm_lock)
2490         __acquires(&khugepaged_mm_lock)
2491 {
2492         struct mm_slot *mm_slot;
2493         struct mm_struct *mm;
2494         struct vm_area_struct *vma;
2495         int progress = 0;
2496
2497         VM_BUG_ON(!pages);
2498         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2499
2500         if (khugepaged_scan.mm_slot)
2501                 mm_slot = khugepaged_scan.mm_slot;
2502         else {
2503                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2504                                      struct mm_slot, mm_node);
2505                 khugepaged_scan.address = 0;
2506                 khugepaged_scan.mm_slot = mm_slot;
2507         }
2508         spin_unlock(&khugepaged_mm_lock);
2509
2510         mm = mm_slot->mm;
2511         down_read(&mm->mmap_sem);
2512         if (unlikely(khugepaged_test_exit(mm)))
2513                 vma = NULL;
2514         else
2515                 vma = find_vma(mm, khugepaged_scan.address);
2516
2517         progress++;
2518         for (; vma; vma = vma->vm_next) {
2519                 unsigned long hstart, hend;
2520
2521                 cond_resched();
2522                 if (unlikely(khugepaged_test_exit(mm))) {
2523                         progress++;
2524                         break;
2525                 }
2526                 if (!hugepage_vma_check(vma)) {
2527 skip:
2528                         progress++;
2529                         continue;
2530                 }
2531                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2532                 hend = vma->vm_end & HPAGE_PMD_MASK;
2533                 if (hstart >= hend)
2534                         goto skip;
2535                 if (khugepaged_scan.address > hend)
2536                         goto skip;
2537                 if (khugepaged_scan.address < hstart)
2538                         khugepaged_scan.address = hstart;
2539                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2540
2541                 while (khugepaged_scan.address < hend) {
2542                         int ret;
2543                         cond_resched();
2544                         if (unlikely(khugepaged_test_exit(mm)))
2545                                 goto breakouterloop;
2546
2547                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2548                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2549                                   hend);
2550                         ret = khugepaged_scan_pmd(mm, vma,
2551                                                   khugepaged_scan.address,
2552                                                   hpage);
2553                         /* move to next address */
2554                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2555                         progress += HPAGE_PMD_NR;
2556                         if (ret)
2557                                 /* we released mmap_sem so break loop */
2558                                 goto breakouterloop_mmap_sem;
2559                         if (progress >= pages)
2560                                 goto breakouterloop;
2561                 }
2562         }
2563 breakouterloop:
2564         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2565 breakouterloop_mmap_sem:
2566
2567         spin_lock(&khugepaged_mm_lock);
2568         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2569         /*
2570          * Release the current mm_slot if this mm is about to die, or
2571          * if we scanned all vmas of this mm.
2572          */
2573         if (khugepaged_test_exit(mm) || !vma) {
2574                 /*
2575                  * Make sure that if mm_users is reaching zero while
2576                  * khugepaged runs here, khugepaged_exit will find
2577                  * mm_slot not pointing to the exiting mm.
2578                  */
2579                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2580                         khugepaged_scan.mm_slot = list_entry(
2581                                 mm_slot->mm_node.next,
2582                                 struct mm_slot, mm_node);
2583                         khugepaged_scan.address = 0;
2584                 } else {
2585                         khugepaged_scan.mm_slot = NULL;
2586                         khugepaged_full_scans++;
2587                 }
2588
2589                 collect_mm_slot(mm_slot);
2590         }
2591
2592         return progress;
2593 }
2594
2595 static int khugepaged_has_work(void)
2596 {
2597         return !list_empty(&khugepaged_scan.mm_head) &&
2598                 khugepaged_enabled();
2599 }
2600
2601 static int khugepaged_wait_event(void)
2602 {
2603         return !list_empty(&khugepaged_scan.mm_head) ||
2604                 kthread_should_stop();
2605 }
2606
2607 static void khugepaged_do_scan(void)
2608 {
2609         struct page *hpage = NULL;
2610         unsigned int progress = 0, pass_through_head = 0;
2611         unsigned int pages = khugepaged_pages_to_scan;
2612         bool wait = true;
2613
2614         barrier(); /* write khugepaged_pages_to_scan to local stack */
2615
2616         while (progress < pages) {
2617                 if (!khugepaged_prealloc_page(&hpage, &wait))
2618                         break;
2619
2620                 cond_resched();
2621
2622                 if (unlikely(kthread_should_stop() || freezing(current)))
2623                         break;
2624
2625                 spin_lock(&khugepaged_mm_lock);
2626                 if (!khugepaged_scan.mm_slot)
2627                         pass_through_head++;
2628                 if (khugepaged_has_work() &&
2629                     pass_through_head < 2)
2630                         progress += khugepaged_scan_mm_slot(pages - progress,
2631                                                             &hpage);
2632                 else
2633                         progress = pages;
2634                 spin_unlock(&khugepaged_mm_lock);
2635         }
2636
2637         if (!IS_ERR_OR_NULL(hpage))
2638                 put_page(hpage);
2639 }
2640
2641 static void khugepaged_wait_work(void)
2642 {
2643         try_to_freeze();
2644
2645         if (khugepaged_has_work()) {
2646                 if (!khugepaged_scan_sleep_millisecs)
2647                         return;
2648
2649                 wait_event_freezable_timeout(khugepaged_wait,
2650                                              kthread_should_stop(),
2651                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2652                 return;
2653         }
2654
2655         if (khugepaged_enabled())
2656                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2657 }
2658
2659 static int khugepaged(void *none)
2660 {
2661         struct mm_slot *mm_slot;
2662
2663         set_freezable();
2664         set_user_nice(current, 19);
2665
2666         while (!kthread_should_stop()) {
2667                 khugepaged_do_scan();
2668                 khugepaged_wait_work();
2669         }
2670
2671         spin_lock(&khugepaged_mm_lock);
2672         mm_slot = khugepaged_scan.mm_slot;
2673         khugepaged_scan.mm_slot = NULL;
2674         if (mm_slot)
2675                 collect_mm_slot(mm_slot);
2676         spin_unlock(&khugepaged_mm_lock);
2677         return 0;
2678 }
2679
2680 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2681                 unsigned long haddr, pmd_t *pmd)
2682 {
2683         struct mm_struct *mm = vma->vm_mm;
2684         pgtable_t pgtable;
2685         pmd_t _pmd;
2686         int i;
2687
2688         pmdp_clear_flush(vma, haddr, pmd);
2689         /* leave pmd empty until pte is filled */
2690
2691         pgtable = pgtable_trans_huge_withdraw(mm);
2692         pmd_populate(mm, &_pmd, pgtable);
2693
2694         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2695                 pte_t *pte, entry;
2696                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2697                 entry = pte_mkspecial(entry);
2698                 pte = pte_offset_map(&_pmd, haddr);
2699                 VM_BUG_ON(!pte_none(*pte));
2700                 set_pte_at(mm, haddr, pte, entry);
2701                 pte_unmap(pte);
2702         }
2703         smp_wmb(); /* make pte visible before pmd */
2704         pmd_populate(mm, pmd, pgtable);
2705         put_huge_zero_page();
2706 }
2707
2708 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2709                 pmd_t *pmd)
2710 {
2711         struct page *page;
2712         struct mm_struct *mm = vma->vm_mm;
2713         unsigned long haddr = address & HPAGE_PMD_MASK;
2714         unsigned long mmun_start;       /* For mmu_notifiers */
2715         unsigned long mmun_end;         /* For mmu_notifiers */
2716
2717         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2718
2719         mmun_start = haddr;
2720         mmun_end   = haddr + HPAGE_PMD_SIZE;
2721         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2722         spin_lock(&mm->page_table_lock);
2723         if (unlikely(!pmd_trans_huge(*pmd))) {
2724                 spin_unlock(&mm->page_table_lock);
2725                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2726                 return;
2727         }
2728         if (is_huge_zero_pmd(*pmd)) {
2729                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2730                 spin_unlock(&mm->page_table_lock);
2731                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2732                 return;
2733         }
2734         page = pmd_page(*pmd);
2735         VM_BUG_ON(!page_count(page));
2736         get_page(page);
2737         spin_unlock(&mm->page_table_lock);
2738         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2739
2740         split_huge_page(page);
2741
2742         put_page(page);
2743         BUG_ON(pmd_trans_huge(*pmd));
2744 }
2745
2746 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2747                 pmd_t *pmd)
2748 {
2749         struct vm_area_struct *vma;
2750
2751         vma = find_vma(mm, address);
2752         BUG_ON(vma == NULL);
2753         split_huge_page_pmd(vma, address, pmd);
2754 }
2755
2756 static void split_huge_page_address(struct mm_struct *mm,
2757                                     unsigned long address)
2758 {
2759         pmd_t *pmd;
2760
2761         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2762
2763         pmd = mm_find_pmd(mm, address);
2764         if (!pmd)
2765                 return;
2766         /*
2767          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2768          * materialize from under us.
2769          */
2770         split_huge_page_pmd_mm(mm, address, pmd);
2771 }
2772
2773 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2774                              unsigned long start,
2775                              unsigned long end,
2776                              long adjust_next)
2777 {
2778         /*
2779          * If the new start address isn't hpage aligned and it could
2780          * previously contain an hugepage: check if we need to split
2781          * an huge pmd.
2782          */
2783         if (start & ~HPAGE_PMD_MASK &&
2784             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2785             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2786                 split_huge_page_address(vma->vm_mm, start);
2787
2788         /*
2789          * If the new end address isn't hpage aligned and it could
2790          * previously contain an hugepage: check if we need to split
2791          * an huge pmd.
2792          */
2793         if (end & ~HPAGE_PMD_MASK &&
2794             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2795             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2796                 split_huge_page_address(vma->vm_mm, end);
2797
2798         /*
2799          * If we're also updating the vma->vm_next->vm_start, if the new
2800          * vm_next->vm_start isn't page aligned and it could previously
2801          * contain an hugepage: check if we need to split an huge pmd.
2802          */
2803         if (adjust_next > 0) {
2804                 struct vm_area_struct *next = vma->vm_next;
2805                 unsigned long nstart = next->vm_start;
2806                 nstart += adjust_next << PAGE_SHIFT;
2807                 if (nstart & ~HPAGE_PMD_MASK &&
2808                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2809                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2810                         split_huge_page_address(next->vm_mm, nstart);
2811         }
2812 }