mm/huge_memory.c

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