mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is disabled in order that avoid
  31  * to risk increase the memory footprint of applications without a guaranteed
  32  * benefit. When transparent hugepage support is enabled, is for all mappings,
  33  * and khugepaged scans all mappings.
  34  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  35  * for all hugepage allocations.
  36  */
  37 unsigned long transparent_hugepage_flags __read_mostly =
  38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  39         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  40 #endif
  41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  42         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  43 #endif
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  46         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  47
  48 /* default scan 8*512 pte (or vmas) every 30 second */
  49 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  50 static unsigned int khugepaged_pages_collapsed;
  51 static unsigned int khugepaged_full_scans;
  52 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  53 /* during fragmentation poll the hugepage allocator once every minute */
  54 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  55 static struct task_struct *khugepaged_thread __read_mostly;
  56 static DEFINE_MUTEX(khugepaged_mutex);
  57 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  58 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  59 /*
  60  * default collapse hugepages if there is at least one pte mapped like
  61  * it would have happened if the vma was large enough during page
  62  * fault.
  63  */
  64 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  65
  66 static int khugepaged(void *none);
  67 static int khugepaged_slab_init(void);
  68
  69 #define MM_SLOTS_HASH_BITS 10
  70 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  71
  72 static struct kmem_cache *mm_slot_cache __read_mostly;
  73
  74 /**
  75  * struct mm_slot - hash lookup from mm to mm_slot
  76  * @hash: hash collision list
  77  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  78  * @mm: the mm that this information is valid for
  79  */
  80 struct mm_slot {
  81         struct hlist_node hash;
  82         struct list_head mm_node;
  83         struct mm_struct *mm;
  84 };
  85
  86 /**
  87  * struct khugepaged_scan - cursor for scanning
  88  * @mm_head: the head of the mm list to scan
  89  * @mm_slot: the current mm_slot we are scanning
  90  * @address: the next address inside that to be scanned
  91  *
  92  * There is only the one khugepaged_scan instance of this cursor structure.
  93  */
  94 struct khugepaged_scan {
  95         struct list_head mm_head;
  96         struct mm_slot *mm_slot;
  97         unsigned long address;
  98 };
  99 static struct khugepaged_scan khugepaged_scan = {
 100         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 101 };
 102
 103
 104 static int set_recommended_min_free_kbytes(void)
 105 {
 106         struct zone *zone;
 107         int nr_zones = 0;
 108         unsigned long recommended_min;
 109
 110         if (!khugepaged_enabled())
 111                 return 0;
 112
 113         for_each_populated_zone(zone)
 114                 nr_zones++;
 115
 116         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 117         recommended_min = pageblock_nr_pages * nr_zones * 2;
 118
 119         /*
 120          * Make sure that on average at least two pageblocks are almost free
 121          * of another type, one for a migratetype to fall back to and a
 122          * second to avoid subsequent fallbacks of other types There are 3
 123          * MIGRATE_TYPES we care about.
 124          */
 125         recommended_min += pageblock_nr_pages * nr_zones *
 126                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 127
 128         /* don't ever allow to reserve more than 5% of the lowmem */
 129         recommended_min = min(recommended_min,
 130                               (unsigned long) nr_free_buffer_pages() / 20);
 131         recommended_min <<= (PAGE_SHIFT-10);
 132
 133         if (recommended_min > min_free_kbytes) {
 134                 if (user_min_free_kbytes >= 0)
 135                         pr_info("raising min_free_kbytes from %d to %lu "
 136                                 "to help transparent hugepage allocations\n",
 137                                 min_free_kbytes, recommended_min);
 138
 139                 min_free_kbytes = recommended_min;
 140         }
 141         setup_per_zone_wmarks();
 142         return 0;
 143 }
 144 late_initcall(set_recommended_min_free_kbytes);
 145
 146 static int start_khugepaged(void)
 147 {
 148         int err = 0;
 149         if (khugepaged_enabled()) {
 150                 if (!khugepaged_thread)
 151                         khugepaged_thread = kthread_run(khugepaged, NULL,
 152                                                         "khugepaged");
 153                 if (unlikely(IS_ERR(khugepaged_thread))) {
 154                         printk(KERN_ERR
 155                                "khugepaged: kthread_run(khugepaged) failed\n");
 156                         err = PTR_ERR(khugepaged_thread);
 157                         khugepaged_thread = NULL;
 158                 }
 159
 160                 if (!list_empty(&khugepaged_scan.mm_head))
 161                         wake_up_interruptible(&khugepaged_wait);
 162
 163                 set_recommended_min_free_kbytes();
 164         } else if (khugepaged_thread) {
 165                 kthread_stop(khugepaged_thread);
 166                 khugepaged_thread = NULL;
 167         }
 168
 169         return err;
 170 }
 171
 172 static atomic_t huge_zero_refcount;
 173 static struct page *huge_zero_page __read_mostly;
 174
 175 static inline bool is_huge_zero_page(struct page *page)
 176 {
 177         return ACCESS_ONCE(huge_zero_page) == page;
 178 }
 179
 180 static inline bool is_huge_zero_pmd(pmd_t pmd)
 181 {
 182         return is_huge_zero_page(pmd_page(pmd));
 183 }
 184
 185 static struct page *get_huge_zero_page(void)
 186 {
 187         struct page *zero_page;
 188 retry:
 189         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 190                 return ACCESS_ONCE(huge_zero_page);
 191
 192         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 193                         HPAGE_PMD_ORDER);
 194         if (!zero_page) {
 195                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 196                 return NULL;
 197         }
 198         count_vm_event(THP_ZERO_PAGE_ALLOC);
 199         preempt_disable();
 200         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 201                 preempt_enable();
 202                 __free_pages(zero_page, compound_order(zero_page));
 203                 goto retry;
 204         }
 205
 206         /* We take additional reference here. It will be put back by shrinker */
 207         atomic_set(&huge_zero_refcount, 2);
 208         preempt_enable();
 209         return ACCESS_ONCE(huge_zero_page);
 210 }
 211
 212 static void put_huge_zero_page(void)
 213 {
 214         /*
 215          * Counter should never go to zero here. Only shrinker can put
 216          * last reference.
 217          */
 218         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 219 }
 220
 221 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 222                                         struct shrink_control *sc)
 223 {
 224         /* we can free zero page only if last reference remains */
 225         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 226 }
 227
 228 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 229                                        struct shrink_control *sc)
 230 {
 231         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 232                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 233                 BUG_ON(zero_page == NULL);
 234                 __free_pages(zero_page, compound_order(zero_page));
 235                 return HPAGE_PMD_NR;
 236         }
 237
 238         return 0;
 239 }
 240
 241 static struct shrinker huge_zero_page_shrinker = {
 242         .count_objects = shrink_huge_zero_page_count,
 243         .scan_objects = shrink_huge_zero_page_scan,
 244         .seeks = DEFAULT_SEEKS,
 245 };
 246
 247 #ifdef CONFIG_SYSFS
 248
 249 static ssize_t double_flag_show(struct kobject *kobj,
 250                                 struct kobj_attribute *attr, char *buf,
 251                                 enum transparent_hugepage_flag enabled,
 252                                 enum transparent_hugepage_flag req_madv)
 253 {
 254         if (test_bit(enabled, &transparent_hugepage_flags)) {
 255                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 256                 return sprintf(buf, "[always] madvise never\n");
 257         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 258                 return sprintf(buf, "always [madvise] never\n");
 259         else
 260                 return sprintf(buf, "always madvise [never]\n");
 261 }
 262 static ssize_t double_flag_store(struct kobject *kobj,
 263                                  struct kobj_attribute *attr,
 264                                  const char *buf, size_t count,
 265                                  enum transparent_hugepage_flag enabled,
 266                                  enum transparent_hugepage_flag req_madv)
 267 {
 268         if (!memcmp("always", buf,
 269                     min(sizeof("always")-1, count))) {
 270                 set_bit(enabled, &transparent_hugepage_flags);
 271                 clear_bit(req_madv, &transparent_hugepage_flags);
 272         } else if (!memcmp("madvise", buf,
 273                            min(sizeof("madvise")-1, count))) {
 274                 clear_bit(enabled, &transparent_hugepage_flags);
 275                 set_bit(req_madv, &transparent_hugepage_flags);
 276         } else if (!memcmp("never", buf,
 277                            min(sizeof("never")-1, count))) {
 278                 clear_bit(enabled, &transparent_hugepage_flags);
 279                 clear_bit(req_madv, &transparent_hugepage_flags);
 280         } else
 281                 return -EINVAL;
 282
 283         return count;
 284 }
 285
 286 static ssize_t enabled_show(struct kobject *kobj,
 287                             struct kobj_attribute *attr, char *buf)
 288 {
 289         return double_flag_show(kobj, attr, buf,
 290                                 TRANSPARENT_HUGEPAGE_FLAG,
 291                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 292 }
 293 static ssize_t enabled_store(struct kobject *kobj,
 294                              struct kobj_attribute *attr,
 295                              const char *buf, size_t count)
 296 {
 297         ssize_t ret;
 298
 299         ret = double_flag_store(kobj, attr, buf, count,
 300                                 TRANSPARENT_HUGEPAGE_FLAG,
 301                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 302
 303         if (ret > 0) {
 304                 int err;
 305
 306                 mutex_lock(&khugepaged_mutex);
 307                 err = start_khugepaged();
 308                 mutex_unlock(&khugepaged_mutex);
 309
 310                 if (err)
 311                         ret = err;
 312         }
 313
 314         return ret;
 315 }
 316 static struct kobj_attribute enabled_attr =
 317         __ATTR(enabled, 0644, enabled_show, enabled_store);
 318
 319 static ssize_t single_flag_show(struct kobject *kobj,
 320                                 struct kobj_attribute *attr, char *buf,
 321                                 enum transparent_hugepage_flag flag)
 322 {
 323         return sprintf(buf, "%d\n",
 324                        !!test_bit(flag, &transparent_hugepage_flags));
 325 }
 326
 327 static ssize_t single_flag_store(struct kobject *kobj,
 328                                  struct kobj_attribute *attr,
 329                                  const char *buf, size_t count,
 330                                  enum transparent_hugepage_flag flag)
 331 {
 332         unsigned long value;
 333         int ret;
 334
 335         ret = kstrtoul(buf, 10, &value);
 336         if (ret < 0)
 337                 return ret;
 338         if (value > 1)
 339                 return -EINVAL;
 340
 341         if (value)
 342                 set_bit(flag, &transparent_hugepage_flags);
 343         else
 344                 clear_bit(flag, &transparent_hugepage_flags);
 345
 346         return count;
 347 }
 348
 349 /*
 350  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 351  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 352  * memory just to allocate one more hugepage.
 353  */
 354 static ssize_t defrag_show(struct kobject *kobj,
 355                            struct kobj_attribute *attr, char *buf)
 356 {
 357         return double_flag_show(kobj, attr, buf,
 358                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 359                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 360 }
 361 static ssize_t defrag_store(struct kobject *kobj,
 362                             struct kobj_attribute *attr,
 363                             const char *buf, size_t count)
 364 {
 365         return double_flag_store(kobj, attr, buf, count,
 366                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 367                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 368 }
 369 static struct kobj_attribute defrag_attr =
 370         __ATTR(defrag, 0644, defrag_show, defrag_store);
 371
 372 static ssize_t use_zero_page_show(struct kobject *kobj,
 373                 struct kobj_attribute *attr, char *buf)
 374 {
 375         return single_flag_show(kobj, attr, buf,
 376                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static ssize_t use_zero_page_store(struct kobject *kobj,
 379                 struct kobj_attribute *attr, const char *buf, size_t count)
 380 {
 381         return single_flag_store(kobj, attr, buf, count,
 382                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 383 }
 384 static struct kobj_attribute use_zero_page_attr =
 385         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 386 #ifdef CONFIG_DEBUG_VM
 387 static ssize_t debug_cow_show(struct kobject *kobj,
 388                                 struct kobj_attribute *attr, char *buf)
 389 {
 390         return single_flag_show(kobj, attr, buf,
 391                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 392 }
 393 static ssize_t debug_cow_store(struct kobject *kobj,
 394                                struct kobj_attribute *attr,
 395                                const char *buf, size_t count)
 396 {
 397         return single_flag_store(kobj, attr, buf, count,
 398                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 399 }
 400 static struct kobj_attribute debug_cow_attr =
 401         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 402 #endif /* CONFIG_DEBUG_VM */
 403
 404 static struct attribute *hugepage_attr[] = {
 405         &enabled_attr.attr,
 406         &defrag_attr.attr,
 407         &use_zero_page_attr.attr,
 408 #ifdef CONFIG_DEBUG_VM
 409         &debug_cow_attr.attr,
 410 #endif
 411         NULL,
 412 };
 413
 414 static struct attribute_group hugepage_attr_group = {
 415         .attrs = hugepage_attr,
 416 };
 417
 418 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 419                                          struct kobj_attribute *attr,
 420                                          char *buf)
 421 {
 422         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 423 }
 424
 425 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 426                                           struct kobj_attribute *attr,
 427                                           const char *buf, size_t count)
 428 {
 429         unsigned long msecs;
 430         int err;
 431
 432         err = kstrtoul(buf, 10, &msecs);
 433         if (err || msecs > UINT_MAX)
 434                 return -EINVAL;
 435
 436         khugepaged_scan_sleep_millisecs = msecs;
 437         wake_up_interruptible(&khugepaged_wait);
 438
 439         return count;
 440 }
 441 static struct kobj_attribute scan_sleep_millisecs_attr =
 442         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 443                scan_sleep_millisecs_store);
 444
 445 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 446                                           struct kobj_attribute *attr,
 447                                           char *buf)
 448 {
 449         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 450 }
 451
 452 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 453                                            struct kobj_attribute *attr,
 454                                            const char *buf, size_t count)
 455 {
 456         unsigned long msecs;
 457         int err;
 458
 459         err = kstrtoul(buf, 10, &msecs);
 460         if (err || msecs > UINT_MAX)
 461                 return -EINVAL;
 462
 463         khugepaged_alloc_sleep_millisecs = msecs;
 464         wake_up_interruptible(&khugepaged_wait);
 465
 466         return count;
 467 }
 468 static struct kobj_attribute alloc_sleep_millisecs_attr =
 469         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 470                alloc_sleep_millisecs_store);
 471
 472 static ssize_t pages_to_scan_show(struct kobject *kobj,
 473                                   struct kobj_attribute *attr,
 474                                   char *buf)
 475 {
 476         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 477 }
 478 static ssize_t pages_to_scan_store(struct kobject *kobj,
 479                                    struct kobj_attribute *attr,
 480                                    const char *buf, size_t count)
 481 {
 482         int err;
 483         unsigned long pages;
 484
 485         err = kstrtoul(buf, 10, &pages);
 486         if (err || !pages || pages > UINT_MAX)
 487                 return -EINVAL;
 488
 489         khugepaged_pages_to_scan = pages;
 490
 491         return count;
 492 }
 493 static struct kobj_attribute pages_to_scan_attr =
 494         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 495                pages_to_scan_store);
 496
 497 static ssize_t pages_collapsed_show(struct kobject *kobj,
 498                                     struct kobj_attribute *attr,
 499                                     char *buf)
 500 {
 501         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 502 }
 503 static struct kobj_attribute pages_collapsed_attr =
 504         __ATTR_RO(pages_collapsed);
 505
 506 static ssize_t full_scans_show(struct kobject *kobj,
 507                                struct kobj_attribute *attr,
 508                                char *buf)
 509 {
 510         return sprintf(buf, "%u\n", khugepaged_full_scans);
 511 }
 512 static struct kobj_attribute full_scans_attr =
 513         __ATTR_RO(full_scans);
 514
 515 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 516                                       struct kobj_attribute *attr, char *buf)
 517 {
 518         return single_flag_show(kobj, attr, buf,
 519                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 520 }
 521 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 522                                        struct kobj_attribute *attr,
 523                                        const char *buf, size_t count)
 524 {
 525         return single_flag_store(kobj, attr, buf, count,
 526                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 527 }
 528 static struct kobj_attribute khugepaged_defrag_attr =
 529         __ATTR(defrag, 0644, khugepaged_defrag_show,
 530                khugepaged_defrag_store);
 531
 532 /*
 533  * max_ptes_none controls if khugepaged should collapse hugepages over
 534  * any unmapped ptes in turn potentially increasing the memory
 535  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 536  * reduce the available free memory in the system as it
 537  * runs. Increasing max_ptes_none will instead potentially reduce the
 538  * free memory in the system during the khugepaged scan.
 539  */
 540 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 541                                              struct kobj_attribute *attr,
 542                                              char *buf)
 543 {
 544         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 545 }
 546 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 547                                               struct kobj_attribute *attr,
 548                                               const char *buf, size_t count)
 549 {
 550         int err;
 551         unsigned long max_ptes_none;
 552
 553         err = kstrtoul(buf, 10, &max_ptes_none);
 554         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 555                 return -EINVAL;
 556
 557         khugepaged_max_ptes_none = max_ptes_none;
 558
 559         return count;
 560 }
 561 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 562         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 563                khugepaged_max_ptes_none_store);
 564
 565 static struct attribute *khugepaged_attr[] = {
 566         &khugepaged_defrag_attr.attr,
 567         &khugepaged_max_ptes_none_attr.attr,
 568         &pages_to_scan_attr.attr,
 569         &pages_collapsed_attr.attr,
 570         &full_scans_attr.attr,
 571         &scan_sleep_millisecs_attr.attr,
 572         &alloc_sleep_millisecs_attr.attr,
 573         NULL,
 574 };
 575
 576 static struct attribute_group khugepaged_attr_group = {
 577         .attrs = khugepaged_attr,
 578         .name = "khugepaged",
 579 };
 580
 581 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 582 {
 583         int err;
 584
 585         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 586         if (unlikely(!*hugepage_kobj)) {
 587                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 588                 return -ENOMEM;
 589         }
 590
 591         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 592         if (err) {
 593                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 594                 goto delete_obj;
 595         }
 596
 597         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 598         if (err) {
 599                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 600                 goto remove_hp_group;
 601         }
 602
 603         return 0;
 604
 605 remove_hp_group:
 606         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 607 delete_obj:
 608         kobject_put(*hugepage_kobj);
 609         return err;
 610 }
 611
 612 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 613 {
 614         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 615         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 616         kobject_put(hugepage_kobj);
 617 }
 618 #else
 619 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 620 {
 621         return 0;
 622 }
 623
 624 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 625 {
 626 }
 627 #endif /* CONFIG_SYSFS */
 628
 629 static int __init hugepage_init(void)
 630 {
 631         int err;
 632         struct kobject *hugepage_kobj;
 633
 634         if (!has_transparent_hugepage()) {
 635                 transparent_hugepage_flags = 0;
 636                 return -EINVAL;
 637         }
 638
 639         err = hugepage_init_sysfs(&hugepage_kobj);
 640         if (err)
 641                 return err;
 642
 643         err = khugepaged_slab_init();
 644         if (err)
 645                 goto out;
 646
 647         register_shrinker(&huge_zero_page_shrinker);
 648
 649         /*
 650          * By default disable transparent hugepages on smaller systems,
 651          * where the extra memory used could hurt more than TLB overhead
 652          * is likely to save.  The admin can still enable it through /sys.
 653          */
 654         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 655                 transparent_hugepage_flags = 0;
 656
 657         start_khugepaged();
 658
 659         return 0;
 660 out:
 661         hugepage_exit_sysfs(hugepage_kobj);
 662         return err;
 663 }
 664 subsys_initcall(hugepage_init);
 665
 666 static int __init setup_transparent_hugepage(char *str)
 667 {
 668         int ret = 0;
 669         if (!str)
 670                 goto out;
 671         if (!strcmp(str, "always")) {
 672                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 673                         &transparent_hugepage_flags);
 674                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 675                           &transparent_hugepage_flags);
 676                 ret = 1;
 677         } else if (!strcmp(str, "madvise")) {
 678                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 679                           &transparent_hugepage_flags);
 680                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 681                         &transparent_hugepage_flags);
 682                 ret = 1;
 683         } else if (!strcmp(str, "never")) {
 684                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 685                           &transparent_hugepage_flags);
 686                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 687                           &transparent_hugepage_flags);
 688                 ret = 1;
 689         }
 690 out:
 691         if (!ret)
 692                 printk(KERN_WARNING
 693                        "transparent_hugepage= cannot parse, ignored\n");
 694         return ret;
 695 }
 696 __setup("transparent_hugepage=", setup_transparent_hugepage);
 697
 698 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 699 {
 700         if (likely(vma->vm_flags & VM_WRITE))
 701                 pmd = pmd_mkwrite(pmd);
 702         return pmd;
 703 }
 704
 705 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 706 {
 707         pmd_t entry;
 708         entry = mk_pmd(page, prot);
 709         entry = pmd_mkhuge(entry);
 710         return entry;
 711 }
 712
 713 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 714                                         struct vm_area_struct *vma,
 715                                         unsigned long haddr, pmd_t *pmd,
 716                                         struct page *page)
 717 {
 718         pgtable_t pgtable;
 719         spinlock_t *ptl;
 720
 721         VM_BUG_ON_PAGE(!PageCompound(page), page);
 722         pgtable = pte_alloc_one(mm, haddr);
 723         if (unlikely(!pgtable))
 724                 return VM_FAULT_OOM;
 725
 726         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 727         /*
 728          * The memory barrier inside __SetPageUptodate makes sure that
 729          * clear_huge_page writes become visible before the set_pmd_at()
 730          * write.
 731          */
 732         __SetPageUptodate(page);
 733
 734         ptl = pmd_lock(mm, pmd);
 735         if (unlikely(!pmd_none(*pmd))) {
 736                 spin_unlock(ptl);
 737                 mem_cgroup_uncharge_page(page);
 738                 put_page(page);
 739                 pte_free(mm, pgtable);
 740         } else {
 741                 pmd_t entry;
 742                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 743                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 744                 page_add_new_anon_rmap(page, vma, haddr);
 745                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 746                 set_pmd_at(mm, haddr, pmd, entry);
 747                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 748                 atomic_long_inc(&mm->nr_ptes);
 749                 spin_unlock(ptl);
 750         }
 751
 752         return 0;
 753 }
 754
 755 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 756 {
 757         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 758 }
 759
 760 static inline struct page *alloc_hugepage_vma(int defrag,
 761                                               struct vm_area_struct *vma,
 762                                               unsigned long haddr, int nd,
 763                                               gfp_t extra_gfp)
 764 {
 765         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 766                                HPAGE_PMD_ORDER, vma, haddr, nd);
 767 }
 768
 769 /* Caller must hold page table lock. */
 770 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 771                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 772                 struct page *zero_page)
 773 {
 774         pmd_t entry;
 775         if (!pmd_none(*pmd))
 776                 return false;
 777         entry = mk_pmd(zero_page, vma->vm_page_prot);
 778         entry = pmd_wrprotect(entry);
 779         entry = pmd_mkhuge(entry);
 780         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 781         set_pmd_at(mm, haddr, pmd, entry);
 782         atomic_long_inc(&mm->nr_ptes);
 783         return true;
 784 }
 785
 786 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 787                                unsigned long address, pmd_t *pmd,
 788                                unsigned int flags)
 789 {
 790         struct page *page;
 791         unsigned long haddr = address & HPAGE_PMD_MASK;
 792
 793         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 794                 return VM_FAULT_FALLBACK;
 795         if (unlikely(anon_vma_prepare(vma)))
 796                 return VM_FAULT_OOM;
 797         if (unlikely(khugepaged_enter(vma)))
 798                 return VM_FAULT_OOM;
 799         if (!(flags & FAULT_FLAG_WRITE) &&
 800                         transparent_hugepage_use_zero_page()) {
 801                 spinlock_t *ptl;
 802                 pgtable_t pgtable;
 803                 struct page *zero_page;
 804                 bool set;
 805                 pgtable = pte_alloc_one(mm, haddr);
 806                 if (unlikely(!pgtable))
 807                         return VM_FAULT_OOM;
 808                 zero_page = get_huge_zero_page();
 809                 if (unlikely(!zero_page)) {
 810                         pte_free(mm, pgtable);
 811                         count_vm_event(THP_FAULT_FALLBACK);
 812                         return VM_FAULT_FALLBACK;
 813                 }
 814                 ptl = pmd_lock(mm, pmd);
 815                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 816                                 zero_page);
 817                 spin_unlock(ptl);
 818                 if (!set) {
 819                         pte_free(mm, pgtable);
 820                         put_huge_zero_page();
 821                 }
 822                 return 0;
 823         }
 824         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 825                         vma, haddr, numa_node_id(), 0);
 826         if (unlikely(!page)) {
 827                 count_vm_event(THP_FAULT_FALLBACK);
 828                 return VM_FAULT_FALLBACK;
 829         }
 830         if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 831                 put_page(page);
 832                 count_vm_event(THP_FAULT_FALLBACK);
 833                 return VM_FAULT_FALLBACK;
 834         }
 835         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 836                 mem_cgroup_uncharge_page(page);
 837                 put_page(page);
 838                 count_vm_event(THP_FAULT_FALLBACK);
 839                 return VM_FAULT_FALLBACK;
 840         }
 841
 842         count_vm_event(THP_FAULT_ALLOC);
 843         return 0;
 844 }
 845
 846 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 847                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 848                   struct vm_area_struct *vma)
 849 {
 850         spinlock_t *dst_ptl, *src_ptl;
 851         struct page *src_page;
 852         pmd_t pmd;
 853         pgtable_t pgtable;
 854         int ret;
 855
 856         ret = -ENOMEM;
 857         pgtable = pte_alloc_one(dst_mm, addr);
 858         if (unlikely(!pgtable))
 859                 goto out;
 860
 861         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 862         src_ptl = pmd_lockptr(src_mm, src_pmd);
 863         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 864
 865         ret = -EAGAIN;
 866         pmd = *src_pmd;
 867         if (unlikely(!pmd_trans_huge(pmd))) {
 868                 pte_free(dst_mm, pgtable);
 869                 goto out_unlock;
 870         }
 871         /*
 872          * When page table lock is held, the huge zero pmd should not be
 873          * under splitting since we don't split the page itself, only pmd to
 874          * a page table.
 875          */
 876         if (is_huge_zero_pmd(pmd)) {
 877                 struct page *zero_page;
 878                 bool set;
 879                 /*
 880                  * get_huge_zero_page() will never allocate a new page here,
 881                  * since we already have a zero page to copy. It just takes a
 882                  * reference.
 883                  */
 884                 zero_page = get_huge_zero_page();
 885                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 886                                 zero_page);
 887                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 888                 ret = 0;
 889                 goto out_unlock;
 890         }
 891
 892         if (unlikely(pmd_trans_splitting(pmd))) {
 893                 /* split huge page running from under us */
 894                 spin_unlock(src_ptl);
 895                 spin_unlock(dst_ptl);
 896                 pte_free(dst_mm, pgtable);
 897
 898                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 899                 goto out;
 900         }
 901         src_page = pmd_page(pmd);
 902         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 903         get_page(src_page);
 904         page_dup_rmap(src_page);
 905         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 906
 907         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 908         pmd = pmd_mkold(pmd_wrprotect(pmd));
 909         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 910         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 911         atomic_long_inc(&dst_mm->nr_ptes);
 912
 913         ret = 0;
 914 out_unlock:
 915         spin_unlock(src_ptl);
 916         spin_unlock(dst_ptl);
 917 out:
 918         return ret;
 919 }
 920
 921 void huge_pmd_set_accessed(struct mm_struct *mm,
 922                            struct vm_area_struct *vma,
 923                            unsigned long address,
 924                            pmd_t *pmd, pmd_t orig_pmd,
 925                            int dirty)
 926 {
 927         spinlock_t *ptl;
 928         pmd_t entry;
 929         unsigned long haddr;
 930
 931         ptl = pmd_lock(mm, pmd);
 932         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 933                 goto unlock;
 934
 935         entry = pmd_mkyoung(orig_pmd);
 936         haddr = address & HPAGE_PMD_MASK;
 937         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 938                 update_mmu_cache_pmd(vma, address, pmd);
 939
 940 unlock:
 941         spin_unlock(ptl);
 942 }
 943
 944 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 945                 struct vm_area_struct *vma, unsigned long address,
 946                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 947 {
 948         spinlock_t *ptl;
 949         pgtable_t pgtable;
 950         pmd_t _pmd;
 951         struct page *page;
 952         int i, ret = 0;
 953         unsigned long mmun_start;       /* For mmu_notifiers */
 954         unsigned long mmun_end;         /* For mmu_notifiers */
 955
 956         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 957         if (!page) {
 958                 ret |= VM_FAULT_OOM;
 959                 goto out;
 960         }
 961
 962         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 963                 put_page(page);
 964                 ret |= VM_FAULT_OOM;
 965                 goto out;
 966         }
 967
 968         clear_user_highpage(page, address);
 969         __SetPageUptodate(page);
 970
 971         mmun_start = haddr;
 972         mmun_end   = haddr + HPAGE_PMD_SIZE;
 973         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 974
 975         ptl = pmd_lock(mm, pmd);
 976         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 977                 goto out_free_page;
 978
 979         pmdp_clear_flush(vma, haddr, pmd);
 980         /* leave pmd empty until pte is filled */
 981
 982         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
 983         pmd_populate(mm, &_pmd, pgtable);
 984
 985         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 986                 pte_t *pte, entry;
 987                 if (haddr == (address & PAGE_MASK)) {
 988                         entry = mk_pte(page, vma->vm_page_prot);
 989                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 990                         page_add_new_anon_rmap(page, vma, haddr);
 991                 } else {
 992                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
 993                         entry = pte_mkspecial(entry);
 994                 }
 995                 pte = pte_offset_map(&_pmd, haddr);
 996                 VM_BUG_ON(!pte_none(*pte));
 997                 set_pte_at(mm, haddr, pte, entry);
 998                 pte_unmap(pte);
 999         }
1000         smp_wmb(); /* make pte visible before pmd */
1001         pmd_populate(mm, pmd, pgtable);
1002         spin_unlock(ptl);
1003         put_huge_zero_page();
1004         inc_mm_counter(mm, MM_ANONPAGES);
1005
1006         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007
1008         ret |= VM_FAULT_WRITE;
1009 out:
1010         return ret;
1011 out_free_page:
1012         spin_unlock(ptl);
1013         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1014         mem_cgroup_uncharge_page(page);
1015         put_page(page);
1016         goto out;
1017 }
1018
1019 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1020                                         struct vm_area_struct *vma,
1021                                         unsigned long address,
1022                                         pmd_t *pmd, pmd_t orig_pmd,
1023                                         struct page *page,
1024                                         unsigned long haddr)
1025 {
1026         spinlock_t *ptl;
1027         pgtable_t pgtable;
1028         pmd_t _pmd;
1029         int ret = 0, i;
1030         struct page **pages;
1031         unsigned long mmun_start;       /* For mmu_notifiers */
1032         unsigned long mmun_end;         /* For mmu_notifiers */
1033
1034         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1035                         GFP_KERNEL);
1036         if (unlikely(!pages)) {
1037                 ret |= VM_FAULT_OOM;
1038                 goto out;
1039         }
1040
1041         for (i = 0; i < HPAGE_PMD_NR; i++) {
1042                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1043                                                __GFP_OTHER_NODE,
1044                                                vma, address, page_to_nid(page));
1045                 if (unlikely(!pages[i] ||
1046                              mem_cgroup_newpage_charge(pages[i], mm,
1047                                                        GFP_KERNEL))) {
1048                         if (pages[i])
1049                                 put_page(pages[i]);
1050                         mem_cgroup_uncharge_start();
1051                         while (--i >= 0) {
1052                                 mem_cgroup_uncharge_page(pages[i]);
1053                                 put_page(pages[i]);
1054                         }
1055                         mem_cgroup_uncharge_end();
1056                         kfree(pages);
1057                         ret |= VM_FAULT_OOM;
1058                         goto out;
1059                 }
1060         }
1061
1062         for (i = 0; i < HPAGE_PMD_NR; i++) {
1063                 copy_user_highpage(pages[i], page + i,
1064                                    haddr + PAGE_SIZE * i, vma);
1065                 __SetPageUptodate(pages[i]);
1066                 cond_resched();
1067         }
1068
1069         mmun_start = haddr;
1070         mmun_end   = haddr + HPAGE_PMD_SIZE;
1071         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1072
1073         ptl = pmd_lock(mm, pmd);
1074         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1075                 goto out_free_pages;
1076         VM_BUG_ON_PAGE(!PageHead(page), page);
1077
1078         pmdp_clear_flush(vma, haddr, pmd);
1079         /* leave pmd empty until pte is filled */
1080
1081         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1082         pmd_populate(mm, &_pmd, pgtable);
1083
1084         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1085                 pte_t *pte, entry;
1086                 entry = mk_pte(pages[i], vma->vm_page_prot);
1087                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1088                 page_add_new_anon_rmap(pages[i], vma, haddr);
1089                 pte = pte_offset_map(&_pmd, haddr);
1090                 VM_BUG_ON(!pte_none(*pte));
1091                 set_pte_at(mm, haddr, pte, entry);
1092                 pte_unmap(pte);
1093         }
1094         kfree(pages);
1095
1096         smp_wmb(); /* make pte visible before pmd */
1097         pmd_populate(mm, pmd, pgtable);
1098         page_remove_rmap(page);
1099         spin_unlock(ptl);
1100
1101         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1102
1103         ret |= VM_FAULT_WRITE;
1104         put_page(page);
1105
1106 out:
1107         return ret;
1108
1109 out_free_pages:
1110         spin_unlock(ptl);
1111         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1112         mem_cgroup_uncharge_start();
1113         for (i = 0; i < HPAGE_PMD_NR; i++) {
1114                 mem_cgroup_uncharge_page(pages[i]);
1115                 put_page(pages[i]);
1116         }
1117         mem_cgroup_uncharge_end();
1118         kfree(pages);
1119         goto out;
1120 }
1121
1122 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1123                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1124 {
1125         spinlock_t *ptl;
1126         int ret = 0;
1127         struct page *page = NULL, *new_page;
1128         unsigned long haddr;
1129         unsigned long mmun_start;       /* For mmu_notifiers */
1130         unsigned long mmun_end;         /* For mmu_notifiers */
1131
1132         ptl = pmd_lockptr(mm, pmd);
1133         VM_BUG_ON(!vma->anon_vma);
1134         haddr = address & HPAGE_PMD_MASK;
1135         if (is_huge_zero_pmd(orig_pmd))
1136                 goto alloc;
1137         spin_lock(ptl);
1138         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1139                 goto out_unlock;
1140
1141         page = pmd_page(orig_pmd);
1142         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1143         if (page_mapcount(page) == 1) {
1144                 pmd_t entry;
1145                 entry = pmd_mkyoung(orig_pmd);
1146                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1147                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1148                         update_mmu_cache_pmd(vma, address, pmd);
1149                 ret |= VM_FAULT_WRITE;
1150                 goto out_unlock;
1151         }
1152         get_page(page);
1153         spin_unlock(ptl);
1154 alloc:
1155         if (transparent_hugepage_enabled(vma) &&
1156             !transparent_hugepage_debug_cow())
1157                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1158                                               vma, haddr, numa_node_id(), 0);
1159         else
1160                 new_page = NULL;
1161
1162         if (unlikely(!new_page)) {
1163                 if (!page) {
1164                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1165                                         address, pmd, orig_pmd, haddr);
1166                 } else {
1167                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1168                                         pmd, orig_pmd, page, haddr);
1169                         if (ret & VM_FAULT_OOM) {
1170                                 split_huge_page(page);
1171                                 ret |= VM_FAULT_FALLBACK;
1172                         }
1173                         put_page(page);
1174                 }
1175                 count_vm_event(THP_FAULT_FALLBACK);
1176                 goto out;
1177         }
1178
1179         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1180                 put_page(new_page);
1181                 if (page) {
1182                         split_huge_page(page);
1183                         put_page(page);
1184                 } else
1185                         split_huge_page_pmd(vma, address, pmd);
1186                 ret |= VM_FAULT_FALLBACK;
1187                 count_vm_event(THP_FAULT_FALLBACK);
1188                 goto out;
1189         }
1190
1191         count_vm_event(THP_FAULT_ALLOC);
1192
1193         if (!page)
1194                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1195         else
1196                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1197         __SetPageUptodate(new_page);
1198
1199         mmun_start = haddr;
1200         mmun_end   = haddr + HPAGE_PMD_SIZE;
1201         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1202
1203         spin_lock(ptl);
1204         if (page)
1205                 put_page(page);
1206         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1207                 spin_unlock(ptl);
1208                 mem_cgroup_uncharge_page(new_page);
1209                 put_page(new_page);
1210                 goto out_mn;
1211         } else {
1212                 pmd_t entry;
1213                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1214                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1215                 pmdp_clear_flush(vma, haddr, pmd);
1216                 page_add_new_anon_rmap(new_page, vma, haddr);
1217                 set_pmd_at(mm, haddr, pmd, entry);
1218                 update_mmu_cache_pmd(vma, address, pmd);
1219                 if (!page) {
1220                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1221                         put_huge_zero_page();
1222                 } else {
1223                         VM_BUG_ON_PAGE(!PageHead(page), page);
1224                         page_remove_rmap(page);
1225                         put_page(page);
1226                 }
1227                 ret |= VM_FAULT_WRITE;
1228         }
1229         spin_unlock(ptl);
1230 out_mn:
1231         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1232 out:
1233         return ret;
1234 out_unlock:
1235         spin_unlock(ptl);
1236         return ret;
1237 }
1238
1239 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1240                                    unsigned long addr,
1241                                    pmd_t *pmd,
1242                                    unsigned int flags)
1243 {
1244         struct mm_struct *mm = vma->vm_mm;
1245         struct page *page = NULL;
1246
1247         assert_spin_locked(pmd_lockptr(mm, pmd));
1248
1249         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1250                 goto out;
1251
1252         /* Avoid dumping huge zero page */
1253         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1254                 return ERR_PTR(-EFAULT);
1255
1256         /* Full NUMA hinting faults to serialise migration in fault paths */
1257         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1258                 goto out;
1259
1260         page = pmd_page(*pmd);
1261         VM_BUG_ON_PAGE(!PageHead(page), page);
1262         if (flags & FOLL_TOUCH) {
1263                 pmd_t _pmd;
1264                 /*
1265                  * We should set the dirty bit only for FOLL_WRITE but
1266                  * for now the dirty bit in the pmd is meaningless.
1267                  * And if the dirty bit will become meaningful and
1268                  * we'll only set it with FOLL_WRITE, an atomic
1269                  * set_bit will be required on the pmd to set the
1270                  * young bit, instead of the current set_pmd_at.
1271                  */
1272                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1273                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1274                                           pmd, _pmd,  1))
1275                         update_mmu_cache_pmd(vma, addr, pmd);
1276         }
1277         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1278                 if (page->mapping && trylock_page(page)) {
1279                         lru_add_drain();
1280                         if (page->mapping)
1281                                 mlock_vma_page(page);
1282                         unlock_page(page);
1283                 }
1284         }
1285         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1286         VM_BUG_ON_PAGE(!PageCompound(page), page);
1287         if (flags & FOLL_GET)
1288                 get_page_foll(page);
1289
1290 out:
1291         return page;
1292 }
1293
1294 /* NUMA hinting page fault entry point for trans huge pmds */
1295 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1296                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1297 {
1298         spinlock_t *ptl;
1299         struct anon_vma *anon_vma = NULL;
1300         struct page *page;
1301         unsigned long haddr = addr & HPAGE_PMD_MASK;
1302         int page_nid = -1, this_nid = numa_node_id();
1303         int target_nid, last_cpupid = -1;
1304         bool page_locked;
1305         bool migrated = false;
1306         int flags = 0;
1307
1308         ptl = pmd_lock(mm, pmdp);
1309         if (unlikely(!pmd_same(pmd, *pmdp)))
1310                 goto out_unlock;
1311
1312         /*
1313          * If there are potential migrations, wait for completion and retry
1314          * without disrupting NUMA hinting information. Do not relock and
1315          * check_same as the page may no longer be mapped.
1316          */
1317         if (unlikely(pmd_trans_migrating(*pmdp))) {
1318                 spin_unlock(ptl);
1319                 wait_migrate_huge_page(vma->anon_vma, pmdp);
1320                 goto out;
1321         }
1322
1323         page = pmd_page(pmd);
1324         BUG_ON(is_huge_zero_page(page));
1325         page_nid = page_to_nid(page);
1326         last_cpupid = page_cpupid_last(page);
1327         count_vm_numa_event(NUMA_HINT_FAULTS);
1328         if (page_nid == this_nid) {
1329                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1330                 flags |= TNF_FAULT_LOCAL;
1331         }
1332
1333         /*
1334          * Avoid grouping on DSO/COW pages in specific and RO pages
1335          * in general, RO pages shouldn't hurt as much anyway since
1336          * they can be in shared cache state.
1337          */
1338         if (!pmd_write(pmd))
1339                 flags |= TNF_NO_GROUP;
1340
1341         /*
1342          * Acquire the page lock to serialise THP migrations but avoid dropping
1343          * page_table_lock if at all possible
1344          */
1345         page_locked = trylock_page(page);
1346         target_nid = mpol_misplaced(page, vma, haddr);
1347         if (target_nid == -1) {
1348                 /* If the page was locked, there are no parallel migrations */
1349                 if (page_locked)
1350                         goto clear_pmdnuma;
1351         }
1352
1353         /* Migration could have started since the pmd_trans_migrating check */
1354         if (!page_locked) {
1355                 spin_unlock(ptl);
1356                 wait_on_page_locked(page);
1357                 page_nid = -1;
1358                 goto out;
1359         }
1360
1361         /*
1362          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1363          * to serialises splits
1364          */
1365         get_page(page);
1366         spin_unlock(ptl);
1367         anon_vma = page_lock_anon_vma_read(page);
1368
1369         /* Confirm the PMD did not change while page_table_lock was released */
1370         spin_lock(ptl);
1371         if (unlikely(!pmd_same(pmd, *pmdp))) {
1372                 unlock_page(page);
1373                 put_page(page);
1374                 page_nid = -1;
1375                 goto out_unlock;
1376         }
1377
1378         /* Bail if we fail to protect against THP splits for any reason */
1379         if (unlikely(!anon_vma)) {
1380                 put_page(page);
1381                 page_nid = -1;
1382                 goto clear_pmdnuma;
1383         }
1384
1385         /*
1386          * Migrate the THP to the requested node, returns with page unlocked
1387          * and pmd_numa cleared.
1388          */
1389         spin_unlock(ptl);
1390         migrated = migrate_misplaced_transhuge_page(mm, vma,
1391                                 pmdp, pmd, addr, page, target_nid);
1392         if (migrated) {
1393                 flags |= TNF_MIGRATED;
1394                 page_nid = target_nid;
1395         }
1396
1397         goto out;
1398 clear_pmdnuma:
1399         BUG_ON(!PageLocked(page));
1400         pmd = pmd_mknonnuma(pmd);
1401         set_pmd_at(mm, haddr, pmdp, pmd);
1402         VM_BUG_ON(pmd_numa(*pmdp));
1403         update_mmu_cache_pmd(vma, addr, pmdp);
1404         unlock_page(page);
1405 out_unlock:
1406         spin_unlock(ptl);
1407
1408 out:
1409         if (anon_vma)
1410                 page_unlock_anon_vma_read(anon_vma);
1411
1412         if (page_nid != -1)
1413                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1414
1415         return 0;
1416 }
1417
1418 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1419                  pmd_t *pmd, unsigned long addr)
1420 {
1421         spinlock_t *ptl;
1422         int ret = 0;
1423
1424         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1425                 struct page *page;
1426                 pgtable_t pgtable;
1427                 pmd_t orig_pmd;
1428                 /*
1429                  * For architectures like ppc64 we look at deposited pgtable
1430                  * when calling pmdp_get_and_clear. So do the
1431                  * pgtable_trans_huge_withdraw after finishing pmdp related
1432                  * operations.
1433                  */
1434                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1435                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1436                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1437                 if (is_huge_zero_pmd(orig_pmd)) {
1438                         atomic_long_dec(&tlb->mm->nr_ptes);
1439                         spin_unlock(ptl);
1440                         put_huge_zero_page();
1441                 } else {
1442                         page = pmd_page(orig_pmd);
1443                         page_remove_rmap(page);
1444                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1445                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1446                         VM_BUG_ON_PAGE(!PageHead(page), page);
1447                         atomic_long_dec(&tlb->mm->nr_ptes);
1448                         spin_unlock(ptl);
1449                         tlb_remove_page(tlb, page);
1450                 }
1451                 pte_free(tlb->mm, pgtable);
1452                 ret = 1;
1453         }
1454         return ret;
1455 }
1456
1457 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1458                 unsigned long addr, unsigned long end,
1459                 unsigned char *vec)
1460 {
1461         spinlock_t *ptl;
1462         int ret = 0;
1463
1464         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1465                 /*
1466                  * All logical pages in the range are present
1467                  * if backed by a huge page.
1468                  */
1469                 spin_unlock(ptl);
1470                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1471                 ret = 1;
1472         }
1473
1474         return ret;
1475 }
1476
1477 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1478                   unsigned long old_addr,
1479                   unsigned long new_addr, unsigned long old_end,
1480                   pmd_t *old_pmd, pmd_t *new_pmd)
1481 {
1482         spinlock_t *old_ptl, *new_ptl;
1483         int ret = 0;
1484         pmd_t pmd;
1485
1486         struct mm_struct *mm = vma->vm_mm;
1487
1488         if ((old_addr & ~HPAGE_PMD_MASK) ||
1489             (new_addr & ~HPAGE_PMD_MASK) ||
1490             old_end - old_addr < HPAGE_PMD_SIZE ||
1491             (new_vma->vm_flags & VM_NOHUGEPAGE))
1492                 goto out;
1493
1494         /*
1495          * The destination pmd shouldn't be established, free_pgtables()
1496          * should have release it.
1497          */
1498         if (WARN_ON(!pmd_none(*new_pmd))) {
1499                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1500                 goto out;
1501         }
1502
1503         /*
1504          * We don't have to worry about the ordering of src and dst
1505          * ptlocks because exclusive mmap_sem prevents deadlock.
1506          */
1507         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1508         if (ret == 1) {
1509                 new_ptl = pmd_lockptr(mm, new_pmd);
1510                 if (new_ptl != old_ptl)
1511                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1512                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1513                 VM_BUG_ON(!pmd_none(*new_pmd));
1514
1515                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1516                         pgtable_t pgtable;
1517                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1518                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1519                 }
1520                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1521                 if (new_ptl != old_ptl)
1522                         spin_unlock(new_ptl);
1523                 spin_unlock(old_ptl);
1524         }
1525 out:
1526         return ret;
1527 }
1528
1529 /*
1530  * Returns
1531  *  - 0 if PMD could not be locked
1532  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1533  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1534  */
1535 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1536                 unsigned long addr, pgprot_t newprot, int prot_numa)
1537 {
1538         struct mm_struct *mm = vma->vm_mm;
1539         spinlock_t *ptl;
1540         int ret = 0;
1541
1542         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1543                 pmd_t entry;
1544                 ret = 1;
1545                 if (!prot_numa) {
1546                         entry = pmdp_get_and_clear(mm, addr, pmd);
1547                         if (pmd_numa(entry))
1548                                 entry = pmd_mknonnuma(entry);
1549                         entry = pmd_modify(entry, newprot);
1550                         ret = HPAGE_PMD_NR;
1551                         set_pmd_at(mm, addr, pmd, entry);
1552                         BUG_ON(pmd_write(entry));
1553                 } else {
1554                         struct page *page = pmd_page(*pmd);
1555
1556                         /*
1557                          * Do not trap faults against the zero page. The
1558                          * read-only data is likely to be read-cached on the
1559                          * local CPU cache and it is less useful to know about
1560                          * local vs remote hits on the zero page.
1561                          */
1562                         if (!is_huge_zero_page(page) &&
1563                             !pmd_numa(*pmd)) {
1564                                 pmdp_set_numa(mm, addr, pmd);
1565                                 ret = HPAGE_PMD_NR;
1566                         }
1567                 }
1568                 spin_unlock(ptl);
1569         }
1570
1571         return ret;
1572 }
1573
1574 /*
1575  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1576  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1577  *
1578  * Note that if it returns 1, this routine returns without unlocking page
1579  * table locks. So callers must unlock them.
1580  */
1581 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1582                 spinlock_t **ptl)
1583 {
1584         *ptl = pmd_lock(vma->vm_mm, pmd);
1585         if (likely(pmd_trans_huge(*pmd))) {
1586                 if (unlikely(pmd_trans_splitting(*pmd))) {
1587                         spin_unlock(*ptl);
1588                         wait_split_huge_page(vma->anon_vma, pmd);
1589                         return -1;
1590                 } else {
1591                         /* Thp mapped by 'pmd' is stable, so we can
1592                          * handle it as it is. */
1593                         return 1;
1594                 }
1595         }
1596         spin_unlock(*ptl);
1597         return 0;
1598 }
1599
1600 /*
1601  * This function returns whether a given @page is mapped onto the @address
1602  * in the virtual space of @mm.
1603  *
1604  * When it's true, this function returns *pmd with holding the page table lock
1605  * and passing it back to the caller via @ptl.
1606  * If it's false, returns NULL without holding the page table lock.
1607  */
1608 pmd_t *page_check_address_pmd(struct page *page,
1609                               struct mm_struct *mm,
1610                               unsigned long address,
1611                               enum page_check_address_pmd_flag flag,
1612                               spinlock_t **ptl)
1613 {
1614         pgd_t *pgd;
1615         pud_t *pud;
1616         pmd_t *pmd;
1617
1618         if (address & ~HPAGE_PMD_MASK)
1619                 return NULL;
1620
1621         pgd = pgd_offset(mm, address);
1622         if (!pgd_present(*pgd))
1623                 return NULL;
1624         pud = pud_offset(pgd, address);
1625         if (!pud_present(*pud))
1626                 return NULL;
1627         pmd = pmd_offset(pud, address);
1628
1629         *ptl = pmd_lock(mm, pmd);
1630         if (!pmd_present(*pmd))
1631                 goto unlock;
1632         if (pmd_page(*pmd) != page)
1633                 goto unlock;
1634         /*
1635          * split_vma() may create temporary aliased mappings. There is
1636          * no risk as long as all huge pmd are found and have their
1637          * splitting bit set before __split_huge_page_refcount
1638          * runs. Finding the same huge pmd more than once during the
1639          * same rmap walk is not a problem.
1640          */
1641         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1642             pmd_trans_splitting(*pmd))
1643                 goto unlock;
1644         if (pmd_trans_huge(*pmd)) {
1645                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1646                           !pmd_trans_splitting(*pmd));
1647                 return pmd;
1648         }
1649 unlock:
1650         spin_unlock(*ptl);
1651         return NULL;
1652 }
1653
1654 static int __split_huge_page_splitting(struct page *page,
1655                                        struct vm_area_struct *vma,
1656                                        unsigned long address)
1657 {
1658         struct mm_struct *mm = vma->vm_mm;
1659         spinlock_t *ptl;
1660         pmd_t *pmd;
1661         int ret = 0;
1662         /* For mmu_notifiers */
1663         const unsigned long mmun_start = address;
1664         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1665
1666         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1667         pmd = page_check_address_pmd(page, mm, address,
1668                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1669         if (pmd) {
1670                 /*
1671                  * We can't temporarily set the pmd to null in order
1672                  * to split it, the pmd must remain marked huge at all
1673                  * times or the VM won't take the pmd_trans_huge paths
1674                  * and it won't wait on the anon_vma->root->rwsem to
1675                  * serialize against split_huge_page*.
1676                  */
1677                 pmdp_splitting_flush(vma, address, pmd);
1678                 ret = 1;
1679                 spin_unlock(ptl);
1680         }
1681         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1682
1683         return ret;
1684 }
1685
1686 static void __split_huge_page_refcount(struct page *page,
1687                                        struct list_head *list)
1688 {
1689         int i;
1690         struct zone *zone = page_zone(page);
1691         struct lruvec *lruvec;
1692         int tail_count = 0;
1693
1694         /* prevent PageLRU to go away from under us, and freeze lru stats */
1695         spin_lock_irq(&zone->lru_lock);
1696         lruvec = mem_cgroup_page_lruvec(page, zone);
1697
1698         compound_lock(page);
1699         /* complete memcg works before add pages to LRU */
1700         mem_cgroup_split_huge_fixup(page);
1701
1702         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1703                 struct page *page_tail = page + i;
1704
1705                 /* tail_page->_mapcount cannot change */
1706                 BUG_ON(page_mapcount(page_tail) < 0);
1707                 tail_count += page_mapcount(page_tail);
1708                 /* check for overflow */
1709                 BUG_ON(tail_count < 0);
1710                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1711                 /*
1712                  * tail_page->_count is zero and not changing from
1713                  * under us. But get_page_unless_zero() may be running
1714                  * from under us on the tail_page. If we used
1715                  * atomic_set() below instead of atomic_add(), we
1716                  * would then run atomic_set() concurrently with
1717                  * get_page_unless_zero(), and atomic_set() is
1718                  * implemented in C not using locked ops. spin_unlock
1719                  * on x86 sometime uses locked ops because of PPro
1720                  * errata 66, 92, so unless somebody can guarantee
1721                  * atomic_set() here would be safe on all archs (and
1722                  * not only on x86), it's safer to use atomic_add().
1723                  */
1724                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1725                            &page_tail->_count);
1726
1727                 /* after clearing PageTail the gup refcount can be released */
1728                 smp_mb();
1729
1730                 /*
1731                  * retain hwpoison flag of the poisoned tail page:
1732                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1733                  *   by the memory-failure.
1734                  */
1735                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1736                 page_tail->flags |= (page->flags &
1737                                      ((1L << PG_referenced) |
1738                                       (1L << PG_swapbacked) |
1739                                       (1L << PG_mlocked) |
1740                                       (1L << PG_uptodate) |
1741                                       (1L << PG_active) |
1742                                       (1L << PG_unevictable)));
1743                 page_tail->flags |= (1L << PG_dirty);
1744
1745                 /* clear PageTail before overwriting first_page */
1746                 smp_wmb();
1747
1748                 /*
1749                  * __split_huge_page_splitting() already set the
1750                  * splitting bit in all pmd that could map this
1751                  * hugepage, that will ensure no CPU can alter the
1752                  * mapcount on the head page. The mapcount is only
1753                  * accounted in the head page and it has to be
1754                  * transferred to all tail pages in the below code. So
1755                  * for this code to be safe, the split the mapcount
1756                  * can't change. But that doesn't mean userland can't
1757                  * keep changing and reading the page contents while
1758                  * we transfer the mapcount, so the pmd splitting
1759                  * status is achieved setting a reserved bit in the
1760                  * pmd, not by clearing the present bit.
1761                 */
1762                 page_tail->_mapcount = page->_mapcount;
1763
1764                 BUG_ON(page_tail->mapping);
1765                 page_tail->mapping = page->mapping;
1766
1767                 page_tail->index = page->index + i;
1768                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1769
1770                 BUG_ON(!PageAnon(page_tail));
1771                 BUG_ON(!PageUptodate(page_tail));
1772                 BUG_ON(!PageDirty(page_tail));
1773                 BUG_ON(!PageSwapBacked(page_tail));
1774
1775                 lru_add_page_tail(page, page_tail, lruvec, list);
1776         }
1777         atomic_sub(tail_count, &page->_count);
1778         BUG_ON(atomic_read(&page->_count) <= 0);
1779
1780         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1781
1782         ClearPageCompound(page);
1783         compound_unlock(page);
1784         spin_unlock_irq(&zone->lru_lock);
1785
1786         for (i = 1; i < HPAGE_PMD_NR; i++) {
1787                 struct page *page_tail = page + i;
1788                 BUG_ON(page_count(page_tail) <= 0);
1789                 /*
1790                  * Tail pages may be freed if there wasn't any mapping
1791                  * like if add_to_swap() is running on a lru page that
1792                  * had its mapping zapped. And freeing these pages
1793                  * requires taking the lru_lock so we do the put_page
1794                  * of the tail pages after the split is complete.
1795                  */
1796                 put_page(page_tail);
1797         }
1798
1799         /*
1800          * Only the head page (now become a regular page) is required
1801          * to be pinned by the caller.
1802          */
1803         BUG_ON(page_count(page) <= 0);
1804 }
1805
1806 static int __split_huge_page_map(struct page *page,
1807                                  struct vm_area_struct *vma,
1808                                  unsigned long address)
1809 {
1810         struct mm_struct *mm = vma->vm_mm;
1811         spinlock_t *ptl;
1812         pmd_t *pmd, _pmd;
1813         int ret = 0, i;
1814         pgtable_t pgtable;
1815         unsigned long haddr;
1816
1817         pmd = page_check_address_pmd(page, mm, address,
1818                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1819         if (pmd) {
1820                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1821                 pmd_populate(mm, &_pmd, pgtable);
1822                 if (pmd_write(*pmd))
1823                         BUG_ON(page_mapcount(page) != 1);
1824
1825                 haddr = address;
1826                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1827                         pte_t *pte, entry;
1828                         BUG_ON(PageCompound(page+i));
1829                         /*
1830                          * Note that pmd_numa is not transferred deliberately
1831                          * to avoid any possibility that pte_numa leaks to
1832                          * a PROT_NONE VMA by accident.
1833                          */
1834                         entry = mk_pte(page + i, vma->vm_page_prot);
1835                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1836                         if (!pmd_write(*pmd))
1837                                 entry = pte_wrprotect(entry);
1838                         if (!pmd_young(*pmd))
1839                                 entry = pte_mkold(entry);
1840                         pte = pte_offset_map(&_pmd, haddr);
1841                         BUG_ON(!pte_none(*pte));
1842                         set_pte_at(mm, haddr, pte, entry);
1843                         pte_unmap(pte);
1844                 }
1845
1846                 smp_wmb(); /* make pte visible before pmd */
1847                 /*
1848                  * Up to this point the pmd is present and huge and
1849                  * userland has the whole access to the hugepage
1850                  * during the split (which happens in place). If we
1851                  * overwrite the pmd with the not-huge version
1852                  * pointing to the pte here (which of course we could
1853                  * if all CPUs were bug free), userland could trigger
1854                  * a small page size TLB miss on the small sized TLB
1855                  * while the hugepage TLB entry is still established
1856                  * in the huge TLB. Some CPU doesn't like that. See
1857                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1858                  * Erratum 383 on page 93. Intel should be safe but is
1859                  * also warns that it's only safe if the permission
1860                  * and cache attributes of the two entries loaded in
1861                  * the two TLB is identical (which should be the case
1862                  * here). But it is generally safer to never allow
1863                  * small and huge TLB entries for the same virtual
1864                  * address to be loaded simultaneously. So instead of
1865                  * doing "pmd_populate(); flush_tlb_range();" we first
1866                  * mark the current pmd notpresent (atomically because
1867                  * here the pmd_trans_huge and pmd_trans_splitting
1868                  * must remain set at all times on the pmd until the
1869                  * split is complete for this pmd), then we flush the
1870                  * SMP TLB and finally we write the non-huge version
1871                  * of the pmd entry with pmd_populate.
1872                  */
1873                 pmdp_invalidate(vma, address, pmd);
1874                 pmd_populate(mm, pmd, pgtable);
1875                 ret = 1;
1876                 spin_unlock(ptl);
1877         }
1878
1879         return ret;
1880 }
1881
1882 /* must be called with anon_vma->root->rwsem held */
1883 static void __split_huge_page(struct page *page,
1884                               struct anon_vma *anon_vma,
1885                               struct list_head *list)
1886 {
1887         int mapcount, mapcount2;
1888         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1889         struct anon_vma_chain *avc;
1890
1891         BUG_ON(!PageHead(page));
1892         BUG_ON(PageTail(page));
1893
1894         mapcount = 0;
1895         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1896                 struct vm_area_struct *vma = avc->vma;
1897                 unsigned long addr = vma_address(page, vma);
1898                 BUG_ON(is_vma_temporary_stack(vma));
1899                 mapcount += __split_huge_page_splitting(page, vma, addr);
1900         }
1901         /*
1902          * It is critical that new vmas are added to the tail of the
1903          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1904          * and establishes a child pmd before
1905          * __split_huge_page_splitting() freezes the parent pmd (so if
1906          * we fail to prevent copy_huge_pmd() from running until the
1907          * whole __split_huge_page() is complete), we will still see
1908          * the newly established pmd of the child later during the
1909          * walk, to be able to set it as pmd_trans_splitting too.
1910          */
1911         if (mapcount != page_mapcount(page))
1912                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1913                        mapcount, page_mapcount(page));
1914         BUG_ON(mapcount != page_mapcount(page));
1915
1916         __split_huge_page_refcount(page, list);
1917
1918         mapcount2 = 0;
1919         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1920                 struct vm_area_struct *vma = avc->vma;
1921                 unsigned long addr = vma_address(page, vma);
1922                 BUG_ON(is_vma_temporary_stack(vma));
1923                 mapcount2 += __split_huge_page_map(page, vma, addr);
1924         }
1925         if (mapcount != mapcount2)
1926                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1927                        mapcount, mapcount2, page_mapcount(page));
1928         BUG_ON(mapcount != mapcount2);
1929 }
1930
1931 /*
1932  * Split a hugepage into normal pages. This doesn't change the position of head
1933  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1934  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1935  * from the hugepage.
1936  * Return 0 if the hugepage is split successfully otherwise return 1.
1937  */
1938 int split_huge_page_to_list(struct page *page, struct list_head *list)
1939 {
1940         struct anon_vma *anon_vma;
1941         int ret = 1;
1942
1943         BUG_ON(is_huge_zero_page(page));
1944         BUG_ON(!PageAnon(page));
1945
1946         /*
1947          * The caller does not necessarily hold an mmap_sem that would prevent
1948          * the anon_vma disappearing so we first we take a reference to it
1949          * and then lock the anon_vma for write. This is similar to
1950          * page_lock_anon_vma_read except the write lock is taken to serialise
1951          * against parallel split or collapse operations.
1952          */
1953         anon_vma = page_get_anon_vma(page);
1954         if (!anon_vma)
1955                 goto out;
1956         anon_vma_lock_write(anon_vma);
1957
1958         ret = 0;
1959         if (!PageCompound(page))
1960                 goto out_unlock;
1961
1962         BUG_ON(!PageSwapBacked(page));
1963         __split_huge_page(page, anon_vma, list);
1964         count_vm_event(THP_SPLIT);
1965
1966         BUG_ON(PageCompound(page));
1967 out_unlock:
1968         anon_vma_unlock_write(anon_vma);
1969         put_anon_vma(anon_vma);
1970 out:
1971         return ret;
1972 }
1973
1974 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1975
1976 int hugepage_madvise(struct vm_area_struct *vma,
1977                      unsigned long *vm_flags, int advice)
1978 {
1979         struct mm_struct *mm = vma->vm_mm;
1980
1981         switch (advice) {
1982         case MADV_HUGEPAGE:
1983                 /*
1984                  * Be somewhat over-protective like KSM for now!
1985                  */
1986                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1987                         return -EINVAL;
1988                 if (mm->def_flags & VM_NOHUGEPAGE)
1989                         return -EINVAL;
1990                 *vm_flags &= ~VM_NOHUGEPAGE;
1991                 *vm_flags |= VM_HUGEPAGE;
1992                 /*
1993                  * If the vma become good for khugepaged to scan,
1994                  * register it here without waiting a page fault that
1995                  * may not happen any time soon.
1996                  */
1997                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1998                         return -ENOMEM;
1999                 break;
2000         case MADV_NOHUGEPAGE:
2001                 /*
2002                  * Be somewhat over-protective like KSM for now!
2003                  */
2004                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
2005                         return -EINVAL;
2006                 *vm_flags &= ~VM_HUGEPAGE;
2007                 *vm_flags |= VM_NOHUGEPAGE;
2008                 /*
2009                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2010                  * this vma even if we leave the mm registered in khugepaged if
2011                  * it got registered before VM_NOHUGEPAGE was set.
2012                  */
2013                 break;
2014         }
2015
2016         return 0;
2017 }
2018
2019 static int __init khugepaged_slab_init(void)
2020 {
2021         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2022                                           sizeof(struct mm_slot),
2023                                           __alignof__(struct mm_slot), 0, NULL);
2024         if (!mm_slot_cache)
2025                 return -ENOMEM;
2026
2027         return 0;
2028 }
2029
2030 static inline struct mm_slot *alloc_mm_slot(void)
2031 {
2032         if (!mm_slot_cache)     /* initialization failed */
2033                 return NULL;
2034         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2035 }
2036
2037 static inline void free_mm_slot(struct mm_slot *mm_slot)
2038 {
2039         kmem_cache_free(mm_slot_cache, mm_slot);
2040 }
2041
2042 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2043 {
2044         struct mm_slot *mm_slot;
2045
2046         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2047                 if (mm == mm_slot->mm)
2048                         return mm_slot;
2049
2050         return NULL;
2051 }
2052
2053 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2054                                     struct mm_slot *mm_slot)
2055 {
2056         mm_slot->mm = mm;
2057         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2058 }
2059
2060 static inline int khugepaged_test_exit(struct mm_struct *mm)
2061 {
2062         return atomic_read(&mm->mm_users) == 0;
2063 }
2064
2065 int __khugepaged_enter(struct mm_struct *mm)
2066 {
2067         struct mm_slot *mm_slot;
2068         int wakeup;
2069
2070         mm_slot = alloc_mm_slot();
2071         if (!mm_slot)
2072                 return -ENOMEM;
2073
2074         /* __khugepaged_exit() must not run from under us */
2075         VM_BUG_ON(khugepaged_test_exit(mm));
2076         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2077                 free_mm_slot(mm_slot);
2078                 return 0;
2079         }
2080
2081         spin_lock(&khugepaged_mm_lock);
2082         insert_to_mm_slots_hash(mm, mm_slot);
2083         /*
2084          * Insert just behind the scanning cursor, to let the area settle
2085          * down a little.
2086          */
2087         wakeup = list_empty(&khugepaged_scan.mm_head);
2088         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2089         spin_unlock(&khugepaged_mm_lock);
2090
2091         atomic_inc(&mm->mm_count);
2092         if (wakeup)
2093                 wake_up_interruptible(&khugepaged_wait);
2094
2095         return 0;
2096 }
2097
2098 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2099 {
2100         unsigned long hstart, hend;
2101         if (!vma->anon_vma)
2102                 /*
2103                  * Not yet faulted in so we will register later in the
2104                  * page fault if needed.
2105                  */
2106                 return 0;
2107         if (vma->vm_ops)
2108                 /* khugepaged not yet working on file or special mappings */
2109                 return 0;
2110         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2111         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2112         hend = vma->vm_end & HPAGE_PMD_MASK;
2113         if (hstart < hend)
2114                 return khugepaged_enter(vma);
2115         return 0;
2116 }
2117
2118 void __khugepaged_exit(struct mm_struct *mm)
2119 {
2120         struct mm_slot *mm_slot;
2121         int free = 0;
2122
2123         spin_lock(&khugepaged_mm_lock);
2124         mm_slot = get_mm_slot(mm);
2125         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2126                 hash_del(&mm_slot->hash);
2127                 list_del(&mm_slot->mm_node);
2128                 free = 1;
2129         }
2130         spin_unlock(&khugepaged_mm_lock);
2131
2132         if (free) {
2133                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2134                 free_mm_slot(mm_slot);
2135                 mmdrop(mm);
2136         } else if (mm_slot) {
2137                 /*
2138                  * This is required to serialize against
2139                  * khugepaged_test_exit() (which is guaranteed to run
2140                  * under mmap sem read mode). Stop here (after we
2141                  * return all pagetables will be destroyed) until
2142                  * khugepaged has finished working on the pagetables
2143                  * under the mmap_sem.
2144                  */
2145                 down_write(&mm->mmap_sem);
2146                 up_write(&mm->mmap_sem);
2147         }
2148 }
2149
2150 static void release_pte_page(struct page *page)
2151 {
2152         /* 0 stands for page_is_file_cache(page) == false */
2153         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2154         unlock_page(page);
2155         putback_lru_page(page);
2156 }
2157
2158 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2159 {
2160         while (--_pte >= pte) {
2161                 pte_t pteval = *_pte;
2162                 if (!pte_none(pteval))
2163                         release_pte_page(pte_page(pteval));
2164         }
2165 }
2166
2167 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2168                                         unsigned long address,
2169                                         pte_t *pte)
2170 {
2171         struct page *page;
2172         pte_t *_pte;
2173         int referenced = 0, none = 0;
2174         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2175              _pte++, address += PAGE_SIZE) {
2176                 pte_t pteval = *_pte;
2177                 if (pte_none(pteval)) {
2178                         if (++none <= khugepaged_max_ptes_none)
2179                                 continue;
2180                         else
2181                                 goto out;
2182                 }
2183                 if (!pte_present(pteval) || !pte_write(pteval))
2184                         goto out;
2185                 page = vm_normal_page(vma, address, pteval);
2186                 if (unlikely(!page))
2187                         goto out;
2188
2189                 VM_BUG_ON_PAGE(PageCompound(page), page);
2190                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2191                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2192
2193                 /* cannot use mapcount: can't collapse if there's a gup pin */
2194                 if (page_count(page) != 1)
2195                         goto out;
2196                 /*
2197                  * We can do it before isolate_lru_page because the
2198                  * page can't be freed from under us. NOTE: PG_lock
2199                  * is needed to serialize against split_huge_page
2200                  * when invoked from the VM.
2201                  */
2202                 if (!trylock_page(page))
2203                         goto out;
2204                 /*
2205                  * Isolate the page to avoid collapsing an hugepage
2206                  * currently in use by the VM.
2207                  */
2208                 if (isolate_lru_page(page)) {
2209                         unlock_page(page);
2210                         goto out;
2211                 }
2212                 /* 0 stands for page_is_file_cache(page) == false */
2213                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2214                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2215                 VM_BUG_ON_PAGE(PageLRU(page), page);
2216
2217                 /* If there is no mapped pte young don't collapse the page */
2218                 if (pte_young(pteval) || PageReferenced(page) ||
2219                     mmu_notifier_test_young(vma->vm_mm, address))
2220                         referenced = 1;
2221         }
2222         if (likely(referenced))
2223                 return 1;
2224 out:
2225         release_pte_pages(pte, _pte);
2226         return 0;
2227 }
2228
2229 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2230                                       struct vm_area_struct *vma,
2231                                       unsigned long address,
2232                                       spinlock_t *ptl)
2233 {
2234         pte_t *_pte;
2235         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2236                 pte_t pteval = *_pte;
2237                 struct page *src_page;
2238
2239                 if (pte_none(pteval)) {
2240                         clear_user_highpage(page, address);
2241                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2242                 } else {
2243                         src_page = pte_page(pteval);
2244                         copy_user_highpage(page, src_page, address, vma);
2245                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2246                         release_pte_page(src_page);
2247                         /*
2248                          * ptl mostly unnecessary, but preempt has to
2249                          * be disabled to update the per-cpu stats
2250                          * inside page_remove_rmap().
2251                          */
2252                         spin_lock(ptl);
2253                         /*
2254                          * paravirt calls inside pte_clear here are
2255                          * superfluous.
2256                          */
2257                         pte_clear(vma->vm_mm, address, _pte);
2258                         page_remove_rmap(src_page);
2259                         spin_unlock(ptl);
2260                         free_page_and_swap_cache(src_page);
2261                 }
2262
2263                 address += PAGE_SIZE;
2264                 page++;
2265         }
2266 }
2267
2268 static void khugepaged_alloc_sleep(void)
2269 {
2270         wait_event_freezable_timeout(khugepaged_wait, false,
2271                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2272 }
2273
2274 static int khugepaged_node_load[MAX_NUMNODES];
2275
2276 #ifdef CONFIG_NUMA
2277 static int khugepaged_find_target_node(void)
2278 {
2279         static int last_khugepaged_target_node = NUMA_NO_NODE;
2280         int nid, target_node = 0, max_value = 0;
2281
2282         /* find first node with max normal pages hit */
2283         for (nid = 0; nid < MAX_NUMNODES; nid++)
2284                 if (khugepaged_node_load[nid] > max_value) {
2285                         max_value = khugepaged_node_load[nid];
2286                         target_node = nid;
2287                 }
2288
2289         /* do some balance if several nodes have the same hit record */
2290         if (target_node <= last_khugepaged_target_node)
2291                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2292                                 nid++)
2293                         if (max_value == khugepaged_node_load[nid]) {
2294                                 target_node = nid;
2295                                 break;
2296                         }
2297
2298         last_khugepaged_target_node = target_node;
2299         return target_node;
2300 }
2301
2302 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2303 {
2304         if (IS_ERR(*hpage)) {
2305                 if (!*wait)
2306                         return false;
2307
2308                 *wait = false;
2309                 *hpage = NULL;
2310                 khugepaged_alloc_sleep();
2311         } else if (*hpage) {
2312                 put_page(*hpage);
2313                 *hpage = NULL;
2314         }
2315
2316         return true;
2317 }
2318
2319 static struct page
2320 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2321                        struct vm_area_struct *vma, unsigned long address,
2322                        int node)
2323 {
2324         VM_BUG_ON_PAGE(*hpage, *hpage);
2325         /*
2326          * Allocate the page while the vma is still valid and under
2327          * the mmap_sem read mode so there is no memory allocation
2328          * later when we take the mmap_sem in write mode. This is more
2329          * friendly behavior (OTOH it may actually hide bugs) to
2330          * filesystems in userland with daemons allocating memory in
2331          * the userland I/O paths.  Allocating memory with the
2332          * mmap_sem in read mode is good idea also to allow greater
2333          * scalability.
2334          */
2335         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2336                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2337         /*
2338          * After allocating the hugepage, release the mmap_sem read lock in
2339          * preparation for taking it in write mode.
2340          */
2341         up_read(&mm->mmap_sem);
2342         if (unlikely(!*hpage)) {
2343                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2344                 *hpage = ERR_PTR(-ENOMEM);
2345                 return NULL;
2346         }
2347
2348         count_vm_event(THP_COLLAPSE_ALLOC);
2349         return *hpage;
2350 }
2351 #else
2352 static int khugepaged_find_target_node(void)
2353 {
2354         return 0;
2355 }
2356
2357 static inline struct page *alloc_hugepage(int defrag)
2358 {
2359         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2360                            HPAGE_PMD_ORDER);
2361 }
2362
2363 static struct page *khugepaged_alloc_hugepage(bool *wait)
2364 {
2365         struct page *hpage;
2366
2367         do {
2368                 hpage = alloc_hugepage(khugepaged_defrag());
2369                 if (!hpage) {
2370                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2371                         if (!*wait)
2372                                 return NULL;
2373
2374                         *wait = false;
2375                         khugepaged_alloc_sleep();
2376                 } else
2377                         count_vm_event(THP_COLLAPSE_ALLOC);
2378         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2379
2380         return hpage;
2381 }
2382
2383 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2384 {
2385         if (!*hpage)
2386                 *hpage = khugepaged_alloc_hugepage(wait);
2387
2388         if (unlikely(!*hpage))
2389                 return false;
2390
2391         return true;
2392 }
2393
2394 static struct page
2395 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2396                        struct vm_area_struct *vma, unsigned long address,
2397                        int node)
2398 {
2399         up_read(&mm->mmap_sem);
2400         VM_BUG_ON(!*hpage);
2401         return  *hpage;
2402 }
2403 #endif
2404
2405 static bool hugepage_vma_check(struct vm_area_struct *vma)
2406 {
2407         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2408             (vma->vm_flags & VM_NOHUGEPAGE))
2409                 return false;
2410
2411         if (!vma->anon_vma || vma->vm_ops)
2412                 return false;
2413         if (is_vma_temporary_stack(vma))
2414                 return false;
2415         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2416         return true;
2417 }
2418
2419 static void collapse_huge_page(struct mm_struct *mm,
2420                                    unsigned long address,
2421                                    struct page **hpage,
2422                                    struct vm_area_struct *vma,
2423                                    int node)
2424 {
2425         pmd_t *pmd, _pmd;
2426         pte_t *pte;
2427         pgtable_t pgtable;
2428         struct page *new_page;
2429         spinlock_t *pmd_ptl, *pte_ptl;
2430         int isolated;
2431         unsigned long hstart, hend;
2432         unsigned long mmun_start;       /* For mmu_notifiers */
2433         unsigned long mmun_end;         /* For mmu_notifiers */
2434
2435         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2436
2437         /* release the mmap_sem read lock. */
2438         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2439         if (!new_page)
2440                 return;
2441
2442         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2443                 return;
2444
2445         /*
2446          * Prevent all access to pagetables with the exception of
2447          * gup_fast later hanlded by the ptep_clear_flush and the VM
2448          * handled by the anon_vma lock + PG_lock.
2449          */
2450         down_write(&mm->mmap_sem);
2451         if (unlikely(khugepaged_test_exit(mm)))
2452                 goto out;
2453
2454         vma = find_vma(mm, address);
2455         if (!vma)
2456                 goto out;
2457         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2458         hend = vma->vm_end & HPAGE_PMD_MASK;
2459         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2460                 goto out;
2461         if (!hugepage_vma_check(vma))
2462                 goto out;
2463         pmd = mm_find_pmd(mm, address);
2464         if (!pmd)
2465                 goto out;
2466         if (pmd_trans_huge(*pmd))
2467                 goto out;
2468
2469         anon_vma_lock_write(vma->anon_vma);
2470
2471         pte = pte_offset_map(pmd, address);
2472         pte_ptl = pte_lockptr(mm, pmd);
2473
2474         mmun_start = address;
2475         mmun_end   = address + HPAGE_PMD_SIZE;
2476         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2477         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2478         /*
2479          * After this gup_fast can't run anymore. This also removes
2480          * any huge TLB entry from the CPU so we won't allow
2481          * huge and small TLB entries for the same virtual address
2482          * to avoid the risk of CPU bugs in that area.
2483          */
2484         _pmd = pmdp_clear_flush(vma, address, pmd);
2485         spin_unlock(pmd_ptl);
2486         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2487
2488         spin_lock(pte_ptl);
2489         isolated = __collapse_huge_page_isolate(vma, address, pte);
2490         spin_unlock(pte_ptl);
2491
2492         if (unlikely(!isolated)) {
2493                 pte_unmap(pte);
2494                 spin_lock(pmd_ptl);
2495                 BUG_ON(!pmd_none(*pmd));
2496                 /*
2497                  * We can only use set_pmd_at when establishing
2498                  * hugepmds and never for establishing regular pmds that
2499                  * points to regular pagetables. Use pmd_populate for that
2500                  */
2501                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2502                 spin_unlock(pmd_ptl);
2503                 anon_vma_unlock_write(vma->anon_vma);
2504                 goto out;
2505         }
2506
2507         /*
2508          * All pages are isolated and locked so anon_vma rmap
2509          * can't run anymore.
2510          */
2511         anon_vma_unlock_write(vma->anon_vma);
2512
2513         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2514         pte_unmap(pte);
2515         __SetPageUptodate(new_page);
2516         pgtable = pmd_pgtable(_pmd);
2517
2518         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2519         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2520
2521         /*
2522          * spin_lock() below is not the equivalent of smp_wmb(), so
2523          * this is needed to avoid the copy_huge_page writes to become
2524          * visible after the set_pmd_at() write.
2525          */
2526         smp_wmb();
2527
2528         spin_lock(pmd_ptl);
2529         BUG_ON(!pmd_none(*pmd));
2530         page_add_new_anon_rmap(new_page, vma, address);
2531         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2532         set_pmd_at(mm, address, pmd, _pmd);
2533         update_mmu_cache_pmd(vma, address, pmd);
2534         spin_unlock(pmd_ptl);
2535
2536         *hpage = NULL;
2537
2538         khugepaged_pages_collapsed++;
2539 out_up_write:
2540         up_write(&mm->mmap_sem);
2541         return;
2542
2543 out:
2544         mem_cgroup_uncharge_page(new_page);
2545         goto out_up_write;
2546 }
2547
2548 static int khugepaged_scan_pmd(struct mm_struct *mm,
2549                                struct vm_area_struct *vma,
2550                                unsigned long address,
2551                                struct page **hpage)
2552 {
2553         pmd_t *pmd;
2554         pte_t *pte, *_pte;
2555         int ret = 0, referenced = 0, none = 0;
2556         struct page *page;
2557         unsigned long _address;
2558         spinlock_t *ptl;
2559         int node = NUMA_NO_NODE;
2560
2561         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2562
2563         pmd = mm_find_pmd(mm, address);
2564         if (!pmd)
2565                 goto out;
2566         if (pmd_trans_huge(*pmd))
2567                 goto out;
2568
2569         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2570         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2571         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2572              _pte++, _address += PAGE_SIZE) {
2573                 pte_t pteval = *_pte;
2574                 if (pte_none(pteval)) {
2575                         if (++none <= khugepaged_max_ptes_none)
2576                                 continue;
2577                         else
2578                                 goto out_unmap;
2579                 }
2580                 if (!pte_present(pteval) || !pte_write(pteval))
2581                         goto out_unmap;
2582                 page = vm_normal_page(vma, _address, pteval);
2583                 if (unlikely(!page))
2584                         goto out_unmap;
2585                 /*
2586                  * Record which node the original page is from and save this
2587                  * information to khugepaged_node_load[].
2588                  * Khupaged will allocate hugepage from the node has the max
2589                  * hit record.
2590                  */
2591                 node = page_to_nid(page);
2592                 khugepaged_node_load[node]++;
2593                 VM_BUG_ON_PAGE(PageCompound(page), page);
2594                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2595                         goto out_unmap;
2596                 /* cannot use mapcount: can't collapse if there's a gup pin */
2597                 if (page_count(page) != 1)
2598                         goto out_unmap;
2599                 if (pte_young(pteval) || PageReferenced(page) ||
2600                     mmu_notifier_test_young(vma->vm_mm, address))
2601                         referenced = 1;
2602         }
2603         if (referenced)
2604                 ret = 1;
2605 out_unmap:
2606         pte_unmap_unlock(pte, ptl);
2607         if (ret) {
2608                 node = khugepaged_find_target_node();
2609                 /* collapse_huge_page will return with the mmap_sem released */
2610                 collapse_huge_page(mm, address, hpage, vma, node);
2611         }
2612 out:
2613         return ret;
2614 }
2615
2616 static void collect_mm_slot(struct mm_slot *mm_slot)
2617 {
2618         struct mm_struct *mm = mm_slot->mm;
2619
2620         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2621
2622         if (khugepaged_test_exit(mm)) {
2623                 /* free mm_slot */
2624                 hash_del(&mm_slot->hash);
2625                 list_del(&mm_slot->mm_node);
2626
2627                 /*
2628                  * Not strictly needed because the mm exited already.
2629                  *
2630                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2631                  */
2632
2633                 /* khugepaged_mm_lock actually not necessary for the below */
2634                 free_mm_slot(mm_slot);
2635                 mmdrop(mm);
2636         }
2637 }
2638
2639 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2640                                             struct page **hpage)
2641         __releases(&khugepaged_mm_lock)
2642         __acquires(&khugepaged_mm_lock)
2643 {
2644         struct mm_slot *mm_slot;
2645         struct mm_struct *mm;
2646         struct vm_area_struct *vma;
2647         int progress = 0;
2648
2649         VM_BUG_ON(!pages);
2650         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2651
2652         if (khugepaged_scan.mm_slot)
2653                 mm_slot = khugepaged_scan.mm_slot;
2654         else {
2655                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2656                                      struct mm_slot, mm_node);
2657                 khugepaged_scan.address = 0;
2658                 khugepaged_scan.mm_slot = mm_slot;
2659         }
2660         spin_unlock(&khugepaged_mm_lock);
2661
2662         mm = mm_slot->mm;
2663         down_read(&mm->mmap_sem);
2664         if (unlikely(khugepaged_test_exit(mm)))
2665                 vma = NULL;
2666         else
2667                 vma = find_vma(mm, khugepaged_scan.address);
2668
2669         progress++;
2670         for (; vma; vma = vma->vm_next) {
2671                 unsigned long hstart, hend;
2672
2673                 cond_resched();
2674                 if (unlikely(khugepaged_test_exit(mm))) {
2675                         progress++;
2676                         break;
2677                 }
2678                 if (!hugepage_vma_check(vma)) {
2679 skip:
2680                         progress++;
2681                         continue;
2682                 }
2683                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2684                 hend = vma->vm_end & HPAGE_PMD_MASK;
2685                 if (hstart >= hend)
2686                         goto skip;
2687                 if (khugepaged_scan.address > hend)
2688                         goto skip;
2689                 if (khugepaged_scan.address < hstart)
2690                         khugepaged_scan.address = hstart;
2691                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2692
2693                 while (khugepaged_scan.address < hend) {
2694                         int ret;
2695                         cond_resched();
2696                         if (unlikely(khugepaged_test_exit(mm)))
2697                                 goto breakouterloop;
2698
2699                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2700                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2701                                   hend);
2702                         ret = khugepaged_scan_pmd(mm, vma,
2703                                                   khugepaged_scan.address,
2704                                                   hpage);
2705                         /* move to next address */
2706                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2707                         progress += HPAGE_PMD_NR;
2708                         if (ret)
2709                                 /* we released mmap_sem so break loop */
2710                                 goto breakouterloop_mmap_sem;
2711                         if (progress >= pages)
2712                                 goto breakouterloop;
2713                 }
2714         }
2715 breakouterloop:
2716         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2717 breakouterloop_mmap_sem:
2718
2719         spin_lock(&khugepaged_mm_lock);
2720         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2721         /*
2722          * Release the current mm_slot if this mm is about to die, or
2723          * if we scanned all vmas of this mm.
2724          */
2725         if (khugepaged_test_exit(mm) || !vma) {
2726                 /*
2727                  * Make sure that if mm_users is reaching zero while
2728                  * khugepaged runs here, khugepaged_exit will find
2729                  * mm_slot not pointing to the exiting mm.
2730                  */
2731                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2732                         khugepaged_scan.mm_slot = list_entry(
2733                                 mm_slot->mm_node.next,
2734                                 struct mm_slot, mm_node);
2735                         khugepaged_scan.address = 0;
2736                 } else {
2737                         khugepaged_scan.mm_slot = NULL;
2738                         khugepaged_full_scans++;
2739                 }
2740
2741                 collect_mm_slot(mm_slot);
2742         }
2743
2744         return progress;
2745 }
2746
2747 static int khugepaged_has_work(void)
2748 {
2749         return !list_empty(&khugepaged_scan.mm_head) &&
2750                 khugepaged_enabled();
2751 }
2752
2753 static int khugepaged_wait_event(void)
2754 {
2755         return !list_empty(&khugepaged_scan.mm_head) ||
2756                 kthread_should_stop();
2757 }
2758
2759 static void khugepaged_do_scan(void)
2760 {
2761         struct page *hpage = NULL;
2762         unsigned int progress = 0, pass_through_head = 0;
2763         unsigned int pages = khugepaged_pages_to_scan;
2764         bool wait = true;
2765
2766         barrier(); /* write khugepaged_pages_to_scan to local stack */
2767
2768         while (progress < pages) {
2769                 if (!khugepaged_prealloc_page(&hpage, &wait))
2770                         break;
2771
2772                 cond_resched();
2773
2774                 if (unlikely(kthread_should_stop() || freezing(current)))
2775                         break;
2776
2777                 spin_lock(&khugepaged_mm_lock);
2778                 if (!khugepaged_scan.mm_slot)
2779                         pass_through_head++;
2780                 if (khugepaged_has_work() &&
2781                     pass_through_head < 2)
2782                         progress += khugepaged_scan_mm_slot(pages - progress,
2783                                                             &hpage);
2784                 else
2785                         progress = pages;
2786                 spin_unlock(&khugepaged_mm_lock);
2787         }
2788
2789         if (!IS_ERR_OR_NULL(hpage))
2790                 put_page(hpage);
2791 }
2792
2793 static void khugepaged_wait_work(void)
2794 {
2795         try_to_freeze();
2796
2797         if (khugepaged_has_work()) {
2798                 if (!khugepaged_scan_sleep_millisecs)
2799                         return;
2800
2801                 wait_event_freezable_timeout(khugepaged_wait,
2802                                              kthread_should_stop(),
2803                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2804                 return;
2805         }
2806
2807         if (khugepaged_enabled())
2808                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2809 }
2810
2811 static int khugepaged(void *none)
2812 {
2813         struct mm_slot *mm_slot;
2814
2815         set_freezable();
2816         set_user_nice(current, 19);
2817
2818         while (!kthread_should_stop()) {
2819                 khugepaged_do_scan();
2820                 khugepaged_wait_work();
2821         }
2822
2823         spin_lock(&khugepaged_mm_lock);
2824         mm_slot = khugepaged_scan.mm_slot;
2825         khugepaged_scan.mm_slot = NULL;
2826         if (mm_slot)
2827                 collect_mm_slot(mm_slot);
2828         spin_unlock(&khugepaged_mm_lock);
2829         return 0;
2830 }
2831
2832 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2833                 unsigned long haddr, pmd_t *pmd)
2834 {
2835         struct mm_struct *mm = vma->vm_mm;
2836         pgtable_t pgtable;
2837         pmd_t _pmd;
2838         int i;
2839
2840         pmdp_clear_flush(vma, haddr, pmd);
2841         /* leave pmd empty until pte is filled */
2842
2843         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2844         pmd_populate(mm, &_pmd, pgtable);
2845
2846         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2847                 pte_t *pte, entry;
2848                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2849                 entry = pte_mkspecial(entry);
2850                 pte = pte_offset_map(&_pmd, haddr);
2851                 VM_BUG_ON(!pte_none(*pte));
2852                 set_pte_at(mm, haddr, pte, entry);
2853                 pte_unmap(pte);
2854         }
2855         smp_wmb(); /* make pte visible before pmd */
2856         pmd_populate(mm, pmd, pgtable);
2857         put_huge_zero_page();
2858 }
2859
2860 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2861                 pmd_t *pmd)
2862 {
2863         spinlock_t *ptl;
2864         struct page *page;
2865         struct mm_struct *mm = vma->vm_mm;
2866         unsigned long haddr = address & HPAGE_PMD_MASK;
2867         unsigned long mmun_start;       /* For mmu_notifiers */
2868         unsigned long mmun_end;         /* For mmu_notifiers */
2869
2870         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2871
2872         mmun_start = haddr;
2873         mmun_end   = haddr + HPAGE_PMD_SIZE;
2874 again:
2875         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2876         ptl = pmd_lock(mm, pmd);
2877         if (unlikely(!pmd_trans_huge(*pmd))) {
2878                 spin_unlock(ptl);
2879                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2880                 return;
2881         }
2882         if (is_huge_zero_pmd(*pmd)) {
2883                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2884                 spin_unlock(ptl);
2885                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2886                 return;
2887         }
2888         page = pmd_page(*pmd);
2889         VM_BUG_ON_PAGE(!page_count(page), page);
2890         get_page(page);
2891         spin_unlock(ptl);
2892         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2893
2894         split_huge_page(page);
2895
2896         put_page(page);
2897
2898         /*
2899          * We don't always have down_write of mmap_sem here: a racing
2900          * do_huge_pmd_wp_page() might have copied-on-write to another
2901          * huge page before our split_huge_page() got the anon_vma lock.
2902          */
2903         if (unlikely(pmd_trans_huge(*pmd)))
2904                 goto again;
2905 }
2906
2907 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2908                 pmd_t *pmd)
2909 {
2910         struct vm_area_struct *vma;
2911
2912         vma = find_vma(mm, address);
2913         BUG_ON(vma == NULL);
2914         split_huge_page_pmd(vma, address, pmd);
2915 }
2916
2917 static void split_huge_page_address(struct mm_struct *mm,
2918                                     unsigned long address)
2919 {
2920         pmd_t *pmd;
2921
2922         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2923
2924         pmd = mm_find_pmd(mm, address);
2925         if (!pmd)
2926                 return;
2927         /*
2928          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2929          * materialize from under us.
2930          */
2931         split_huge_page_pmd_mm(mm, address, pmd);
2932 }
2933
2934 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2935                              unsigned long start,
2936                              unsigned long end,
2937                              long adjust_next)
2938 {
2939         /*
2940          * If the new start address isn't hpage aligned and it could
2941          * previously contain an hugepage: check if we need to split
2942          * an huge pmd.
2943          */
2944         if (start & ~HPAGE_PMD_MASK &&
2945             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2946             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2947                 split_huge_page_address(vma->vm_mm, start);
2948
2949         /*
2950          * If the new end address isn't hpage aligned and it could
2951          * previously contain an hugepage: check if we need to split
2952          * an huge pmd.
2953          */
2954         if (end & ~HPAGE_PMD_MASK &&
2955             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2956             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2957                 split_huge_page_address(vma->vm_mm, end);
2958
2959         /*
2960          * If we're also updating the vma->vm_next->vm_start, if the new
2961          * vm_next->vm_start isn't page aligned and it could previously
2962          * contain an hugepage: check if we need to split an huge pmd.
2963          */
2964         if (adjust_next > 0) {
2965                 struct vm_area_struct *next = vma->vm_next;
2966                 unsigned long nstart = next->vm_start;
2967                 nstart += adjust_next << PAGE_SHIFT;
2968                 if (nstart & ~HPAGE_PMD_MASK &&
2969                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2970                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2971                         split_huge_page_address(next->vm_mm, nstart);
2972         }
2973 }