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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/kthread.h>
  20 #include <linux/khugepaged.h>
  21 #include <linux/freezer.h>
  22 #include <linux/mman.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/migrate.h>
  25 #include <linux/hashtable.h>
  26
  27 #include <asm/tlb.h>
  28 #include <asm/pgalloc.h>
  29 #include "internal.h"
  30
  31 /*
  32  * By default transparent hugepage support is disabled in order that avoid
  33  * to risk increase the memory footprint of applications without a guaranteed
  34  * benefit. When transparent hugepage support is enabled, is for all mappings,
  35  * and khugepaged scans all mappings.
  36  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  37  * for all hugepage allocations.
  38  */
  39 unsigned long transparent_hugepage_flags __read_mostly =
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  41         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  42 #endif
  43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  44         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  45 #endif
  46         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  47         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  48         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  49
  50 /* default scan 8*512 pte (or vmas) every 30 second */
  51 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  52 static unsigned int khugepaged_pages_collapsed;
  53 static unsigned int khugepaged_full_scans;
  54 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  55 /* during fragmentation poll the hugepage allocator once every minute */
  56 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  57 static struct task_struct *khugepaged_thread __read_mostly;
  58 static DEFINE_MUTEX(khugepaged_mutex);
  59 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  60 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  61 /*
  62  * default collapse hugepages if there is at least one pte mapped like
  63  * it would have happened if the vma was large enough during page
  64  * fault.
  65  */
  66 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  67
  68 static int khugepaged(void *none);
  69 static int khugepaged_slab_init(void);
  70
  71 #define MM_SLOTS_HASH_BITS 10
  72 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  73
  74 static struct kmem_cache *mm_slot_cache __read_mostly;
  75
  76 /**
  77  * struct mm_slot - hash lookup from mm to mm_slot
  78  * @hash: hash collision list
  79  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  80  * @mm: the mm that this information is valid for
  81  */
  82 struct mm_slot {
  83         struct hlist_node hash;
  84         struct list_head mm_node;
  85         struct mm_struct *mm;
  86 };
  87
  88 /**
  89  * struct khugepaged_scan - cursor for scanning
  90  * @mm_head: the head of the mm list to scan
  91  * @mm_slot: the current mm_slot we are scanning
  92  * @address: the next address inside that to be scanned
  93  *
  94  * There is only the one khugepaged_scan instance of this cursor structure.
  95  */
  96 struct khugepaged_scan {
  97         struct list_head mm_head;
  98         struct mm_slot *mm_slot;
  99         unsigned long address;
 100 };
 101 static struct khugepaged_scan khugepaged_scan = {
 102         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 103 };
 104
 105
 106 static int set_recommended_min_free_kbytes(void)
 107 {
 108         struct zone *zone;
 109         int nr_zones = 0;
 110         unsigned long recommended_min;
 111
 112         if (!khugepaged_enabled())
 113                 return 0;
 114
 115         for_each_populated_zone(zone)
 116                 nr_zones++;
 117
 118         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 119         recommended_min = pageblock_nr_pages * nr_zones * 2;
 120
 121         /*
 122          * Make sure that on average at least two pageblocks are almost free
 123          * of another type, one for a migratetype to fall back to and a
 124          * second to avoid subsequent fallbacks of other types There are 3
 125          * MIGRATE_TYPES we care about.
 126          */
 127         recommended_min += pageblock_nr_pages * nr_zones *
 128                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 129
 130         /* don't ever allow to reserve more than 5% of the lowmem */
 131         recommended_min = min(recommended_min,
 132                               (unsigned long) nr_free_buffer_pages() / 20);
 133         recommended_min <<= (PAGE_SHIFT-10);
 134
 135         if (recommended_min > min_free_kbytes) {
 136                 if (user_min_free_kbytes >= 0)
 137                         pr_info("raising min_free_kbytes from %d to %lu "
 138                                 "to help transparent hugepage allocations\n",
 139                                 min_free_kbytes, recommended_min);
 140
 141                 min_free_kbytes = recommended_min;
 142         }
 143         setup_per_zone_wmarks();
 144         return 0;
 145 }
 146 late_initcall(set_recommended_min_free_kbytes);
 147
 148 static int start_khugepaged(void)
 149 {
 150         int err = 0;
 151         if (khugepaged_enabled()) {
 152                 if (!khugepaged_thread)
 153                         khugepaged_thread = kthread_run(khugepaged, NULL,
 154                                                         "khugepaged");
 155                 if (unlikely(IS_ERR(khugepaged_thread))) {
 156                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 157                         err = PTR_ERR(khugepaged_thread);
 158                         khugepaged_thread = NULL;
 159                 }
 160
 161                 if (!list_empty(&khugepaged_scan.mm_head))
 162                         wake_up_interruptible(&khugepaged_wait);
 163
 164                 set_recommended_min_free_kbytes();
 165         } else if (khugepaged_thread) {
 166                 kthread_stop(khugepaged_thread);
 167                 khugepaged_thread = NULL;
 168         }
 169
 170         return err;
 171 }
 172
 173 static atomic_t huge_zero_refcount;
 174 static struct page *huge_zero_page __read_mostly;
 175
 176 static inline bool is_huge_zero_page(struct page *page)
 177 {
 178         return ACCESS_ONCE(huge_zero_page) == page;
 179 }
 180
 181 static inline bool is_huge_zero_pmd(pmd_t pmd)
 182 {
 183         return is_huge_zero_page(pmd_page(pmd));
 184 }
 185
 186 static struct page *get_huge_zero_page(void)
 187 {
 188         struct page *zero_page;
 189 retry:
 190         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 191                 return ACCESS_ONCE(huge_zero_page);
 192
 193         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 194                         HPAGE_PMD_ORDER);
 195         if (!zero_page) {
 196                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 197                 return NULL;
 198         }
 199         count_vm_event(THP_ZERO_PAGE_ALLOC);
 200         preempt_disable();
 201         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 202                 preempt_enable();
 203                 __free_page(zero_page);
 204                 goto retry;
 205         }
 206
 207         /* We take additional reference here. It will be put back by shrinker */
 208         atomic_set(&huge_zero_refcount, 2);
 209         preempt_enable();
 210         return ACCESS_ONCE(huge_zero_page);
 211 }
 212
 213 static void put_huge_zero_page(void)
 214 {
 215         /*
 216          * Counter should never go to zero here. Only shrinker can put
 217          * last reference.
 218          */
 219         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 220 }
 221
 222 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 223                                         struct shrink_control *sc)
 224 {
 225         /* we can free zero page only if last reference remains */
 226         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 227 }
 228
 229 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 230                                        struct shrink_control *sc)
 231 {
 232         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 233                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 234                 BUG_ON(zero_page == NULL);
 235                 __free_page(zero_page);
 236                 return HPAGE_PMD_NR;
 237         }
 238
 239         return 0;
 240 }
 241
 242 static struct shrinker huge_zero_page_shrinker = {
 243         .count_objects = shrink_huge_zero_page_count,
 244         .scan_objects = shrink_huge_zero_page_scan,
 245         .seeks = DEFAULT_SEEKS,
 246 };
 247
 248 #ifdef CONFIG_SYSFS
 249
 250 static ssize_t double_flag_show(struct kobject *kobj,
 251                                 struct kobj_attribute *attr, char *buf,
 252                                 enum transparent_hugepage_flag enabled,
 253                                 enum transparent_hugepage_flag req_madv)
 254 {
 255         if (test_bit(enabled, &transparent_hugepage_flags)) {
 256                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 257                 return sprintf(buf, "[always] madvise never\n");
 258         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 259                 return sprintf(buf, "always [madvise] never\n");
 260         else
 261                 return sprintf(buf, "always madvise [never]\n");
 262 }
 263 static ssize_t double_flag_store(struct kobject *kobj,
 264                                  struct kobj_attribute *attr,
 265                                  const char *buf, size_t count,
 266                                  enum transparent_hugepage_flag enabled,
 267                                  enum transparent_hugepage_flag req_madv)
 268 {
 269         if (!memcmp("always", buf,
 270                     min(sizeof("always")-1, count))) {
 271                 set_bit(enabled, &transparent_hugepage_flags);
 272                 clear_bit(req_madv, &transparent_hugepage_flags);
 273         } else if (!memcmp("madvise", buf,
 274                            min(sizeof("madvise")-1, count))) {
 275                 clear_bit(enabled, &transparent_hugepage_flags);
 276                 set_bit(req_madv, &transparent_hugepage_flags);
 277         } else if (!memcmp("never", buf,
 278                            min(sizeof("never")-1, count))) {
 279                 clear_bit(enabled, &transparent_hugepage_flags);
 280                 clear_bit(req_madv, &transparent_hugepage_flags);
 281         } else
 282                 return -EINVAL;
 283
 284         return count;
 285 }
 286
 287 static ssize_t enabled_show(struct kobject *kobj,
 288                             struct kobj_attribute *attr, char *buf)
 289 {
 290         return double_flag_show(kobj, attr, buf,
 291                                 TRANSPARENT_HUGEPAGE_FLAG,
 292                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 293 }
 294 static ssize_t enabled_store(struct kobject *kobj,
 295                              struct kobj_attribute *attr,
 296                              const char *buf, size_t count)
 297 {
 298         ssize_t ret;
 299
 300         ret = double_flag_store(kobj, attr, buf, count,
 301                                 TRANSPARENT_HUGEPAGE_FLAG,
 302                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 303
 304         if (ret > 0) {
 305                 int err;
 306
 307                 mutex_lock(&khugepaged_mutex);
 308                 err = start_khugepaged();
 309                 mutex_unlock(&khugepaged_mutex);
 310
 311                 if (err)
 312                         ret = err;
 313         }
 314
 315         return ret;
 316 }
 317 static struct kobj_attribute enabled_attr =
 318         __ATTR(enabled, 0644, enabled_show, enabled_store);
 319
 320 static ssize_t single_flag_show(struct kobject *kobj,
 321                                 struct kobj_attribute *attr, char *buf,
 322                                 enum transparent_hugepage_flag flag)
 323 {
 324         return sprintf(buf, "%d\n",
 325                        !!test_bit(flag, &transparent_hugepage_flags));
 326 }
 327
 328 static ssize_t single_flag_store(struct kobject *kobj,
 329                                  struct kobj_attribute *attr,
 330                                  const char *buf, size_t count,
 331                                  enum transparent_hugepage_flag flag)
 332 {
 333         unsigned long value;
 334         int ret;
 335
 336         ret = kstrtoul(buf, 10, &value);
 337         if (ret < 0)
 338                 return ret;
 339         if (value > 1)
 340                 return -EINVAL;
 341
 342         if (value)
 343                 set_bit(flag, &transparent_hugepage_flags);
 344         else
 345                 clear_bit(flag, &transparent_hugepage_flags);
 346
 347         return count;
 348 }
 349
 350 /*
 351  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 352  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 353  * memory just to allocate one more hugepage.
 354  */
 355 static ssize_t defrag_show(struct kobject *kobj,
 356                            struct kobj_attribute *attr, char *buf)
 357 {
 358         return double_flag_show(kobj, attr, buf,
 359                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 360                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 361 }
 362 static ssize_t defrag_store(struct kobject *kobj,
 363                             struct kobj_attribute *attr,
 364                             const char *buf, size_t count)
 365 {
 366         return double_flag_store(kobj, attr, buf, count,
 367                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 368                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 369 }
 370 static struct kobj_attribute defrag_attr =
 371         __ATTR(defrag, 0644, defrag_show, defrag_store);
 372
 373 static ssize_t use_zero_page_show(struct kobject *kobj,
 374                 struct kobj_attribute *attr, char *buf)
 375 {
 376         return single_flag_show(kobj, attr, buf,
 377                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 378 }
 379 static ssize_t use_zero_page_store(struct kobject *kobj,
 380                 struct kobj_attribute *attr, const char *buf, size_t count)
 381 {
 382         return single_flag_store(kobj, attr, buf, count,
 383                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 384 }
 385 static struct kobj_attribute use_zero_page_attr =
 386         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 387 #ifdef CONFIG_DEBUG_VM
 388 static ssize_t debug_cow_show(struct kobject *kobj,
 389                                 struct kobj_attribute *attr, char *buf)
 390 {
 391         return single_flag_show(kobj, attr, buf,
 392                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 393 }
 394 static ssize_t debug_cow_store(struct kobject *kobj,
 395                                struct kobj_attribute *attr,
 396                                const char *buf, size_t count)
 397 {
 398         return single_flag_store(kobj, attr, buf, count,
 399                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 400 }
 401 static struct kobj_attribute debug_cow_attr =
 402         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 403 #endif /* CONFIG_DEBUG_VM */
 404
 405 static struct attribute *hugepage_attr[] = {
 406         &enabled_attr.attr,
 407         &defrag_attr.attr,
 408         &use_zero_page_attr.attr,
 409 #ifdef CONFIG_DEBUG_VM
 410         &debug_cow_attr.attr,
 411 #endif
 412         NULL,
 413 };
 414
 415 static struct attribute_group hugepage_attr_group = {
 416         .attrs = hugepage_attr,
 417 };
 418
 419 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 420                                          struct kobj_attribute *attr,
 421                                          char *buf)
 422 {
 423         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 424 }
 425
 426 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 427                                           struct kobj_attribute *attr,
 428                                           const char *buf, size_t count)
 429 {
 430         unsigned long msecs;
 431         int err;
 432
 433         err = kstrtoul(buf, 10, &msecs);
 434         if (err || msecs > UINT_MAX)
 435                 return -EINVAL;
 436
 437         khugepaged_scan_sleep_millisecs = msecs;
 438         wake_up_interruptible(&khugepaged_wait);
 439
 440         return count;
 441 }
 442 static struct kobj_attribute scan_sleep_millisecs_attr =
 443         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 444                scan_sleep_millisecs_store);
 445
 446 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 447                                           struct kobj_attribute *attr,
 448                                           char *buf)
 449 {
 450         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 451 }
 452
 453 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 454                                            struct kobj_attribute *attr,
 455                                            const char *buf, size_t count)
 456 {
 457         unsigned long msecs;
 458         int err;
 459
 460         err = kstrtoul(buf, 10, &msecs);
 461         if (err || msecs > UINT_MAX)
 462                 return -EINVAL;
 463
 464         khugepaged_alloc_sleep_millisecs = msecs;
 465         wake_up_interruptible(&khugepaged_wait);
 466
 467         return count;
 468 }
 469 static struct kobj_attribute alloc_sleep_millisecs_attr =
 470         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 471                alloc_sleep_millisecs_store);
 472
 473 static ssize_t pages_to_scan_show(struct kobject *kobj,
 474                                   struct kobj_attribute *attr,
 475                                   char *buf)
 476 {
 477         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 478 }
 479 static ssize_t pages_to_scan_store(struct kobject *kobj,
 480                                    struct kobj_attribute *attr,
 481                                    const char *buf, size_t count)
 482 {
 483         int err;
 484         unsigned long pages;
 485
 486         err = kstrtoul(buf, 10, &pages);
 487         if (err || !pages || pages > UINT_MAX)
 488                 return -EINVAL;
 489
 490         khugepaged_pages_to_scan = pages;
 491
 492         return count;
 493 }
 494 static struct kobj_attribute pages_to_scan_attr =
 495         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 496                pages_to_scan_store);
 497
 498 static ssize_t pages_collapsed_show(struct kobject *kobj,
 499                                     struct kobj_attribute *attr,
 500                                     char *buf)
 501 {
 502         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 503 }
 504 static struct kobj_attribute pages_collapsed_attr =
 505         __ATTR_RO(pages_collapsed);
 506
 507 static ssize_t full_scans_show(struct kobject *kobj,
 508                                struct kobj_attribute *attr,
 509                                char *buf)
 510 {
 511         return sprintf(buf, "%u\n", khugepaged_full_scans);
 512 }
 513 static struct kobj_attribute full_scans_attr =
 514         __ATTR_RO(full_scans);
 515
 516 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 517                                       struct kobj_attribute *attr, char *buf)
 518 {
 519         return single_flag_show(kobj, attr, buf,
 520                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 521 }
 522 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 523                                        struct kobj_attribute *attr,
 524                                        const char *buf, size_t count)
 525 {
 526         return single_flag_store(kobj, attr, buf, count,
 527                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 528 }
 529 static struct kobj_attribute khugepaged_defrag_attr =
 530         __ATTR(defrag, 0644, khugepaged_defrag_show,
 531                khugepaged_defrag_store);
 532
 533 /*
 534  * max_ptes_none controls if khugepaged should collapse hugepages over
 535  * any unmapped ptes in turn potentially increasing the memory
 536  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 537  * reduce the available free memory in the system as it
 538  * runs. Increasing max_ptes_none will instead potentially reduce the
 539  * free memory in the system during the khugepaged scan.
 540  */
 541 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 542                                              struct kobj_attribute *attr,
 543                                              char *buf)
 544 {
 545         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 546 }
 547 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 548                                               struct kobj_attribute *attr,
 549                                               const char *buf, size_t count)
 550 {
 551         int err;
 552         unsigned long max_ptes_none;
 553
 554         err = kstrtoul(buf, 10, &max_ptes_none);
 555         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 556                 return -EINVAL;
 557
 558         khugepaged_max_ptes_none = max_ptes_none;
 559
 560         return count;
 561 }
 562 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 563         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 564                khugepaged_max_ptes_none_store);
 565
 566 static struct attribute *khugepaged_attr[] = {
 567         &khugepaged_defrag_attr.attr,
 568         &khugepaged_max_ptes_none_attr.attr,
 569         &pages_to_scan_attr.attr,
 570         &pages_collapsed_attr.attr,
 571         &full_scans_attr.attr,
 572         &scan_sleep_millisecs_attr.attr,
 573         &alloc_sleep_millisecs_attr.attr,
 574         NULL,
 575 };
 576
 577 static struct attribute_group khugepaged_attr_group = {
 578         .attrs = khugepaged_attr,
 579         .name = "khugepaged",
 580 };
 581
 582 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 583 {
 584         int err;
 585
 586         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 587         if (unlikely(!*hugepage_kobj)) {
 588                 pr_err("failed to create transparent hugepage kobject\n");
 589                 return -ENOMEM;
 590         }
 591
 592         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 593         if (err) {
 594                 pr_err("failed to register transparent hugepage group\n");
 595                 goto delete_obj;
 596         }
 597
 598         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 599         if (err) {
 600                 pr_err("failed to register transparent hugepage group\n");
 601                 goto remove_hp_group;
 602         }
 603
 604         return 0;
 605
 606 remove_hp_group:
 607         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 608 delete_obj:
 609         kobject_put(*hugepage_kobj);
 610         return err;
 611 }
 612
 613 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 614 {
 615         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 616         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 617         kobject_put(hugepage_kobj);
 618 }
 619 #else
 620 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 621 {
 622         return 0;
 623 }
 624
 625 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 626 {
 627 }
 628 #endif /* CONFIG_SYSFS */
 629
 630 static int __init hugepage_init(void)
 631 {
 632         int err;
 633         struct kobject *hugepage_kobj;
 634
 635         if (!has_transparent_hugepage()) {
 636                 transparent_hugepage_flags = 0;
 637                 return -EINVAL;
 638         }
 639
 640         err = hugepage_init_sysfs(&hugepage_kobj);
 641         if (err)
 642                 return err;
 643
 644         err = khugepaged_slab_init();
 645         if (err)
 646                 goto out;
 647
 648         register_shrinker(&huge_zero_page_shrinker);
 649
 650         /*
 651          * By default disable transparent hugepages on smaller systems,
 652          * where the extra memory used could hurt more than TLB overhead
 653          * is likely to save.  The admin can still enable it through /sys.
 654          */
 655         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 656                 transparent_hugepage_flags = 0;
 657
 658         start_khugepaged();
 659
 660         return 0;
 661 out:
 662         hugepage_exit_sysfs(hugepage_kobj);
 663         return err;
 664 }
 665 subsys_initcall(hugepage_init);
 666
 667 static int __init setup_transparent_hugepage(char *str)
 668 {
 669         int ret = 0;
 670         if (!str)
 671                 goto out;
 672         if (!strcmp(str, "always")) {
 673                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 674                         &transparent_hugepage_flags);
 675                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 676                           &transparent_hugepage_flags);
 677                 ret = 1;
 678         } else if (!strcmp(str, "madvise")) {
 679                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 680                           &transparent_hugepage_flags);
 681                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 682                         &transparent_hugepage_flags);
 683                 ret = 1;
 684         } else if (!strcmp(str, "never")) {
 685                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 686                           &transparent_hugepage_flags);
 687                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 688                           &transparent_hugepage_flags);
 689                 ret = 1;
 690         }
 691 out:
 692         if (!ret)
 693                 pr_warn("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_charge_anon(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_page_fallback(struct mm_struct *mm,
 945                                         struct vm_area_struct *vma,
 946                                         unsigned long address,
 947                                         pmd_t *pmd, pmd_t orig_pmd,
 948                                         struct page *page,
 949                                         unsigned long haddr)
 950 {
 951         spinlock_t *ptl;
 952         pgtable_t pgtable;
 953         pmd_t _pmd;
 954         int ret = 0, i;
 955         struct page **pages;
 956         unsigned long mmun_start;       /* For mmu_notifiers */
 957         unsigned long mmun_end;         /* For mmu_notifiers */
 958
 959         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 960                         GFP_KERNEL);
 961         if (unlikely(!pages)) {
 962                 ret |= VM_FAULT_OOM;
 963                 goto out;
 964         }
 965
 966         for (i = 0; i < HPAGE_PMD_NR; i++) {
 967                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 968                                                __GFP_OTHER_NODE,
 969                                                vma, address, page_to_nid(page));
 970                 if (unlikely(!pages[i] ||
 971                              mem_cgroup_charge_anon(pages[i], mm,
 972                                                        GFP_KERNEL))) {
 973                         if (pages[i])
 974                                 put_page(pages[i]);
 975                         mem_cgroup_uncharge_start();
 976                         while (--i >= 0) {
 977                                 mem_cgroup_uncharge_page(pages[i]);
 978                                 put_page(pages[i]);
 979                         }
 980                         mem_cgroup_uncharge_end();
 981                         kfree(pages);
 982                         ret |= VM_FAULT_OOM;
 983                         goto out;
 984                 }
 985         }
 986
 987         for (i = 0; i < HPAGE_PMD_NR; i++) {
 988                 copy_user_highpage(pages[i], page + i,
 989                                    haddr + PAGE_SIZE * i, vma);
 990                 __SetPageUptodate(pages[i]);
 991                 cond_resched();
 992         }
 993
 994         mmun_start = haddr;
 995         mmun_end   = haddr + HPAGE_PMD_SIZE;
 996         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 997
 998         ptl = pmd_lock(mm, pmd);
 999         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1000                 goto out_free_pages;
1001         VM_BUG_ON_PAGE(!PageHead(page), page);
1002
1003         pmdp_clear_flush(vma, haddr, pmd);
1004         /* leave pmd empty until pte is filled */
1005
1006         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1007         pmd_populate(mm, &_pmd, pgtable);
1008
1009         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1010                 pte_t *pte, entry;
1011                 entry = mk_pte(pages[i], vma->vm_page_prot);
1012                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1013                 page_add_new_anon_rmap(pages[i], vma, haddr);
1014                 pte = pte_offset_map(&_pmd, haddr);
1015                 VM_BUG_ON(!pte_none(*pte));
1016                 set_pte_at(mm, haddr, pte, entry);
1017                 pte_unmap(pte);
1018         }
1019         kfree(pages);
1020
1021         smp_wmb(); /* make pte visible before pmd */
1022         pmd_populate(mm, pmd, pgtable);
1023         page_remove_rmap(page);
1024         spin_unlock(ptl);
1025
1026         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1027
1028         ret |= VM_FAULT_WRITE;
1029         put_page(page);
1030
1031 out:
1032         return ret;
1033
1034 out_free_pages:
1035         spin_unlock(ptl);
1036         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1037         mem_cgroup_uncharge_start();
1038         for (i = 0; i < HPAGE_PMD_NR; i++) {
1039                 mem_cgroup_uncharge_page(pages[i]);
1040                 put_page(pages[i]);
1041         }
1042         mem_cgroup_uncharge_end();
1043         kfree(pages);
1044         goto out;
1045 }
1046
1047 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1048                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1049 {
1050         spinlock_t *ptl;
1051         int ret = 0;
1052         struct page *page = NULL, *new_page;
1053         unsigned long haddr;
1054         unsigned long mmun_start;       /* For mmu_notifiers */
1055         unsigned long mmun_end;         /* For mmu_notifiers */
1056
1057         ptl = pmd_lockptr(mm, pmd);
1058         VM_BUG_ON(!vma->anon_vma);
1059         haddr = address & HPAGE_PMD_MASK;
1060         if (is_huge_zero_pmd(orig_pmd))
1061                 goto alloc;
1062         spin_lock(ptl);
1063         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1064                 goto out_unlock;
1065
1066         page = pmd_page(orig_pmd);
1067         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1068         if (page_mapcount(page) == 1) {
1069                 pmd_t entry;
1070                 entry = pmd_mkyoung(orig_pmd);
1071                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1072                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1073                         update_mmu_cache_pmd(vma, address, pmd);
1074                 ret |= VM_FAULT_WRITE;
1075                 goto out_unlock;
1076         }
1077         get_page(page);
1078         spin_unlock(ptl);
1079 alloc:
1080         if (transparent_hugepage_enabled(vma) &&
1081             !transparent_hugepage_debug_cow())
1082                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1083                                               vma, haddr, numa_node_id(), 0);
1084         else
1085                 new_page = NULL;
1086
1087         if (unlikely(!new_page)) {
1088                 if (!page) {
1089                         split_huge_page_pmd(vma, address, pmd);
1090                         ret |= VM_FAULT_FALLBACK;
1091                 } else {
1092                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1093                                         pmd, orig_pmd, page, haddr);
1094                         if (ret & VM_FAULT_OOM) {
1095                                 split_huge_page(page);
1096                                 ret |= VM_FAULT_FALLBACK;
1097                         }
1098                         put_page(page);
1099                 }
1100                 count_vm_event(THP_FAULT_FALLBACK);
1101                 goto out;
1102         }
1103
1104         if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))) {
1105                 put_page(new_page);
1106                 if (page) {
1107                         split_huge_page(page);
1108                         put_page(page);
1109                 } else
1110                         split_huge_page_pmd(vma, address, pmd);
1111                 ret |= VM_FAULT_FALLBACK;
1112                 count_vm_event(THP_FAULT_FALLBACK);
1113                 goto out;
1114         }
1115
1116         count_vm_event(THP_FAULT_ALLOC);
1117
1118         if (!page)
1119                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1120         else
1121                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1122         __SetPageUptodate(new_page);
1123
1124         mmun_start = haddr;
1125         mmun_end   = haddr + HPAGE_PMD_SIZE;
1126         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1127
1128         spin_lock(ptl);
1129         if (page)
1130                 put_page(page);
1131         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1132                 spin_unlock(ptl);
1133                 mem_cgroup_uncharge_page(new_page);
1134                 put_page(new_page);
1135                 goto out_mn;
1136         } else {
1137                 pmd_t entry;
1138                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1139                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1140                 pmdp_clear_flush(vma, haddr, pmd);
1141                 page_add_new_anon_rmap(new_page, vma, haddr);
1142                 set_pmd_at(mm, haddr, pmd, entry);
1143                 update_mmu_cache_pmd(vma, address, pmd);
1144                 if (!page) {
1145                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1146                         put_huge_zero_page();
1147                 } else {
1148                         VM_BUG_ON_PAGE(!PageHead(page), page);
1149                         page_remove_rmap(page);
1150                         put_page(page);
1151                 }
1152                 ret |= VM_FAULT_WRITE;
1153         }
1154         spin_unlock(ptl);
1155 out_mn:
1156         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1157 out:
1158         return ret;
1159 out_unlock:
1160         spin_unlock(ptl);
1161         return ret;
1162 }
1163
1164 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1165                                    unsigned long addr,
1166                                    pmd_t *pmd,
1167                                    unsigned int flags)
1168 {
1169         struct mm_struct *mm = vma->vm_mm;
1170         struct page *page = NULL;
1171
1172         assert_spin_locked(pmd_lockptr(mm, pmd));
1173
1174         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1175                 goto out;
1176
1177         /* Avoid dumping huge zero page */
1178         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1179                 return ERR_PTR(-EFAULT);
1180
1181         /* Full NUMA hinting faults to serialise migration in fault paths */
1182         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1183                 goto out;
1184
1185         page = pmd_page(*pmd);
1186         VM_BUG_ON_PAGE(!PageHead(page), page);
1187         if (flags & FOLL_TOUCH) {
1188                 pmd_t _pmd;
1189                 /*
1190                  * We should set the dirty bit only for FOLL_WRITE but
1191                  * for now the dirty bit in the pmd is meaningless.
1192                  * And if the dirty bit will become meaningful and
1193                  * we'll only set it with FOLL_WRITE, an atomic
1194                  * set_bit will be required on the pmd to set the
1195                  * young bit, instead of the current set_pmd_at.
1196                  */
1197                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1198                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1199                                           pmd, _pmd,  1))
1200                         update_mmu_cache_pmd(vma, addr, pmd);
1201         }
1202         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1203                 if (page->mapping && trylock_page(page)) {
1204                         lru_add_drain();
1205                         if (page->mapping)
1206                                 mlock_vma_page(page);
1207                         unlock_page(page);
1208                 }
1209         }
1210         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1211         VM_BUG_ON_PAGE(!PageCompound(page), page);
1212         if (flags & FOLL_GET)
1213                 get_page_foll(page);
1214
1215 out:
1216         return page;
1217 }
1218
1219 /* NUMA hinting page fault entry point for trans huge pmds */
1220 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1221                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1222 {
1223         spinlock_t *ptl;
1224         struct anon_vma *anon_vma = NULL;
1225         struct page *page;
1226         unsigned long haddr = addr & HPAGE_PMD_MASK;
1227         int page_nid = -1, this_nid = numa_node_id();
1228         int target_nid, last_cpupid = -1;
1229         bool page_locked;
1230         bool migrated = false;
1231         int flags = 0;
1232
1233         ptl = pmd_lock(mm, pmdp);
1234         if (unlikely(!pmd_same(pmd, *pmdp)))
1235                 goto out_unlock;
1236
1237         /*
1238          * If there are potential migrations, wait for completion and retry
1239          * without disrupting NUMA hinting information. Do not relock and
1240          * check_same as the page may no longer be mapped.
1241          */
1242         if (unlikely(pmd_trans_migrating(*pmdp))) {
1243                 spin_unlock(ptl);
1244                 wait_migrate_huge_page(vma->anon_vma, pmdp);
1245                 goto out;
1246         }
1247
1248         page = pmd_page(pmd);
1249         BUG_ON(is_huge_zero_page(page));
1250         page_nid = page_to_nid(page);
1251         last_cpupid = page_cpupid_last(page);
1252         count_vm_numa_event(NUMA_HINT_FAULTS);
1253         if (page_nid == this_nid) {
1254                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1255                 flags |= TNF_FAULT_LOCAL;
1256         }
1257
1258         /*
1259          * Avoid grouping on DSO/COW pages in specific and RO pages
1260          * in general, RO pages shouldn't hurt as much anyway since
1261          * they can be in shared cache state.
1262          */
1263         if (!pmd_write(pmd))
1264                 flags |= TNF_NO_GROUP;
1265
1266         /*
1267          * Acquire the page lock to serialise THP migrations but avoid dropping
1268          * page_table_lock if at all possible
1269          */
1270         page_locked = trylock_page(page);
1271         target_nid = mpol_misplaced(page, vma, haddr);
1272         if (target_nid == -1) {
1273                 /* If the page was locked, there are no parallel migrations */
1274                 if (page_locked)
1275                         goto clear_pmdnuma;
1276         }
1277
1278         /* Migration could have started since the pmd_trans_migrating check */
1279         if (!page_locked) {
1280                 spin_unlock(ptl);
1281                 wait_on_page_locked(page);
1282                 page_nid = -1;
1283                 goto out;
1284         }
1285
1286         /*
1287          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1288          * to serialises splits
1289          */
1290         get_page(page);
1291         spin_unlock(ptl);
1292         anon_vma = page_lock_anon_vma_read(page);
1293
1294         /* Confirm the PMD did not change while page_table_lock was released */
1295         spin_lock(ptl);
1296         if (unlikely(!pmd_same(pmd, *pmdp))) {
1297                 unlock_page(page);
1298                 put_page(page);
1299                 page_nid = -1;
1300                 goto out_unlock;
1301         }
1302
1303         /* Bail if we fail to protect against THP splits for any reason */
1304         if (unlikely(!anon_vma)) {
1305                 put_page(page);
1306                 page_nid = -1;
1307                 goto clear_pmdnuma;
1308         }
1309
1310         /*
1311          * Migrate the THP to the requested node, returns with page unlocked
1312          * and pmd_numa cleared.
1313          */
1314         spin_unlock(ptl);
1315         migrated = migrate_misplaced_transhuge_page(mm, vma,
1316                                 pmdp, pmd, addr, page, target_nid);
1317         if (migrated) {
1318                 flags |= TNF_MIGRATED;
1319                 page_nid = target_nid;
1320         }
1321
1322         goto out;
1323 clear_pmdnuma:
1324         BUG_ON(!PageLocked(page));
1325         pmd = pmd_mknonnuma(pmd);
1326         set_pmd_at(mm, haddr, pmdp, pmd);
1327         VM_BUG_ON(pmd_numa(*pmdp));
1328         update_mmu_cache_pmd(vma, addr, pmdp);
1329         unlock_page(page);
1330 out_unlock:
1331         spin_unlock(ptl);
1332
1333 out:
1334         if (anon_vma)
1335                 page_unlock_anon_vma_read(anon_vma);
1336
1337         if (page_nid != -1)
1338                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1339
1340         return 0;
1341 }
1342
1343 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1344                  pmd_t *pmd, unsigned long addr)
1345 {
1346         spinlock_t *ptl;
1347         int ret = 0;
1348
1349         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1350                 struct page *page;
1351                 pgtable_t pgtable;
1352                 pmd_t orig_pmd;
1353                 /*
1354                  * For architectures like ppc64 we look at deposited pgtable
1355                  * when calling pmdp_get_and_clear. So do the
1356                  * pgtable_trans_huge_withdraw after finishing pmdp related
1357                  * operations.
1358                  */
1359                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1360                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1361                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1362                 if (is_huge_zero_pmd(orig_pmd)) {
1363                         atomic_long_dec(&tlb->mm->nr_ptes);
1364                         spin_unlock(ptl);
1365                         put_huge_zero_page();
1366                 } else {
1367                         page = pmd_page(orig_pmd);
1368                         page_remove_rmap(page);
1369                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1370                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1371                         VM_BUG_ON_PAGE(!PageHead(page), page);
1372                         atomic_long_dec(&tlb->mm->nr_ptes);
1373                         spin_unlock(ptl);
1374                         tlb_remove_page(tlb, page);
1375                 }
1376                 pte_free(tlb->mm, pgtable);
1377                 ret = 1;
1378         }
1379         return ret;
1380 }
1381
1382 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1383                 unsigned long addr, unsigned long end,
1384                 unsigned char *vec)
1385 {
1386         spinlock_t *ptl;
1387         int ret = 0;
1388
1389         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1390                 /*
1391                  * All logical pages in the range are present
1392                  * if backed by a huge page.
1393                  */
1394                 spin_unlock(ptl);
1395                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1396                 ret = 1;
1397         }
1398
1399         return ret;
1400 }
1401
1402 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1403                   unsigned long old_addr,
1404                   unsigned long new_addr, unsigned long old_end,
1405                   pmd_t *old_pmd, pmd_t *new_pmd)
1406 {
1407         spinlock_t *old_ptl, *new_ptl;
1408         int ret = 0;
1409         pmd_t pmd;
1410
1411         struct mm_struct *mm = vma->vm_mm;
1412
1413         if ((old_addr & ~HPAGE_PMD_MASK) ||
1414             (new_addr & ~HPAGE_PMD_MASK) ||
1415             old_end - old_addr < HPAGE_PMD_SIZE ||
1416             (new_vma->vm_flags & VM_NOHUGEPAGE))
1417                 goto out;
1418
1419         /*
1420          * The destination pmd shouldn't be established, free_pgtables()
1421          * should have release it.
1422          */
1423         if (WARN_ON(!pmd_none(*new_pmd))) {
1424                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1425                 goto out;
1426         }
1427
1428         /*
1429          * We don't have to worry about the ordering of src and dst
1430          * ptlocks because exclusive mmap_sem prevents deadlock.
1431          */
1432         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1433         if (ret == 1) {
1434                 new_ptl = pmd_lockptr(mm, new_pmd);
1435                 if (new_ptl != old_ptl)
1436                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1437                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1438                 VM_BUG_ON(!pmd_none(*new_pmd));
1439
1440                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1441                         pgtable_t pgtable;
1442                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1443                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1444                 }
1445                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1446                 if (new_ptl != old_ptl)
1447                         spin_unlock(new_ptl);
1448                 spin_unlock(old_ptl);
1449         }
1450 out:
1451         return ret;
1452 }
1453
1454 /*
1455  * Returns
1456  *  - 0 if PMD could not be locked
1457  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1458  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1459  */
1460 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1461                 unsigned long addr, pgprot_t newprot, int prot_numa)
1462 {
1463         struct mm_struct *mm = vma->vm_mm;
1464         spinlock_t *ptl;
1465         int ret = 0;
1466
1467         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1468                 pmd_t entry;
1469                 ret = 1;
1470                 if (!prot_numa) {
1471                         entry = pmdp_get_and_clear(mm, addr, pmd);
1472                         if (pmd_numa(entry))
1473                                 entry = pmd_mknonnuma(entry);
1474                         entry = pmd_modify(entry, newprot);
1475                         ret = HPAGE_PMD_NR;
1476                         set_pmd_at(mm, addr, pmd, entry);
1477                         BUG_ON(pmd_write(entry));
1478                 } else {
1479                         struct page *page = pmd_page(*pmd);
1480
1481                         /*
1482                          * Do not trap faults against the zero page. The
1483                          * read-only data is likely to be read-cached on the
1484                          * local CPU cache and it is less useful to know about
1485                          * local vs remote hits on the zero page.
1486                          */
1487                         if (!is_huge_zero_page(page) &&
1488                             !pmd_numa(*pmd)) {
1489                                 pmdp_set_numa(mm, addr, pmd);
1490                                 ret = HPAGE_PMD_NR;
1491                         }
1492                 }
1493                 spin_unlock(ptl);
1494         }
1495
1496         return ret;
1497 }
1498
1499 /*
1500  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1501  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1502  *
1503  * Note that if it returns 1, this routine returns without unlocking page
1504  * table locks. So callers must unlock them.
1505  */
1506 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1507                 spinlock_t **ptl)
1508 {
1509         *ptl = pmd_lock(vma->vm_mm, pmd);
1510         if (likely(pmd_trans_huge(*pmd))) {
1511                 if (unlikely(pmd_trans_splitting(*pmd))) {
1512                         spin_unlock(*ptl);
1513                         wait_split_huge_page(vma->anon_vma, pmd);
1514                         return -1;
1515                 } else {
1516                         /* Thp mapped by 'pmd' is stable, so we can
1517                          * handle it as it is. */
1518                         return 1;
1519                 }
1520         }
1521         spin_unlock(*ptl);
1522         return 0;
1523 }
1524
1525 /*
1526  * This function returns whether a given @page is mapped onto the @address
1527  * in the virtual space of @mm.
1528  *
1529  * When it's true, this function returns *pmd with holding the page table lock
1530  * and passing it back to the caller via @ptl.
1531  * If it's false, returns NULL without holding the page table lock.
1532  */
1533 pmd_t *page_check_address_pmd(struct page *page,
1534                               struct mm_struct *mm,
1535                               unsigned long address,
1536                               enum page_check_address_pmd_flag flag,
1537                               spinlock_t **ptl)
1538 {
1539         pgd_t *pgd;
1540         pud_t *pud;
1541         pmd_t *pmd;
1542
1543         if (address & ~HPAGE_PMD_MASK)
1544                 return NULL;
1545
1546         pgd = pgd_offset(mm, address);
1547         if (!pgd_present(*pgd))
1548                 return NULL;
1549         pud = pud_offset(pgd, address);
1550         if (!pud_present(*pud))
1551                 return NULL;
1552         pmd = pmd_offset(pud, address);
1553
1554         *ptl = pmd_lock(mm, pmd);
1555         if (!pmd_present(*pmd))
1556                 goto unlock;
1557         if (pmd_page(*pmd) != page)
1558                 goto unlock;
1559         /*
1560          * split_vma() may create temporary aliased mappings. There is
1561          * no risk as long as all huge pmd are found and have their
1562          * splitting bit set before __split_huge_page_refcount
1563          * runs. Finding the same huge pmd more than once during the
1564          * same rmap walk is not a problem.
1565          */
1566         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1567             pmd_trans_splitting(*pmd))
1568                 goto unlock;
1569         if (pmd_trans_huge(*pmd)) {
1570                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1571                           !pmd_trans_splitting(*pmd));
1572                 return pmd;
1573         }
1574 unlock:
1575         spin_unlock(*ptl);
1576         return NULL;
1577 }
1578
1579 static int __split_huge_page_splitting(struct page *page,
1580                                        struct vm_area_struct *vma,
1581                                        unsigned long address)
1582 {
1583         struct mm_struct *mm = vma->vm_mm;
1584         spinlock_t *ptl;
1585         pmd_t *pmd;
1586         int ret = 0;
1587         /* For mmu_notifiers */
1588         const unsigned long mmun_start = address;
1589         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1590
1591         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1592         pmd = page_check_address_pmd(page, mm, address,
1593                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1594         if (pmd) {
1595                 /*
1596                  * We can't temporarily set the pmd to null in order
1597                  * to split it, the pmd must remain marked huge at all
1598                  * times or the VM won't take the pmd_trans_huge paths
1599                  * and it won't wait on the anon_vma->root->rwsem to
1600                  * serialize against split_huge_page*.
1601                  */
1602                 pmdp_splitting_flush(vma, address, pmd);
1603                 ret = 1;
1604                 spin_unlock(ptl);
1605         }
1606         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1607
1608         return ret;
1609 }
1610
1611 static void __split_huge_page_refcount(struct page *page,
1612                                        struct list_head *list)
1613 {
1614         int i;
1615         struct zone *zone = page_zone(page);
1616         struct lruvec *lruvec;
1617         int tail_count = 0;
1618
1619         /* prevent PageLRU to go away from under us, and freeze lru stats */
1620         spin_lock_irq(&zone->lru_lock);
1621         lruvec = mem_cgroup_page_lruvec(page, zone);
1622
1623         compound_lock(page);
1624         /* complete memcg works before add pages to LRU */
1625         mem_cgroup_split_huge_fixup(page);
1626
1627         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1628                 struct page *page_tail = page + i;
1629
1630                 /* tail_page->_mapcount cannot change */
1631                 BUG_ON(page_mapcount(page_tail) < 0);
1632                 tail_count += page_mapcount(page_tail);
1633                 /* check for overflow */
1634                 BUG_ON(tail_count < 0);
1635                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1636                 /*
1637                  * tail_page->_count is zero and not changing from
1638                  * under us. But get_page_unless_zero() may be running
1639                  * from under us on the tail_page. If we used
1640                  * atomic_set() below instead of atomic_add(), we
1641                  * would then run atomic_set() concurrently with
1642                  * get_page_unless_zero(), and atomic_set() is
1643                  * implemented in C not using locked ops. spin_unlock
1644                  * on x86 sometime uses locked ops because of PPro
1645                  * errata 66, 92, so unless somebody can guarantee
1646                  * atomic_set() here would be safe on all archs (and
1647                  * not only on x86), it's safer to use atomic_add().
1648                  */
1649                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1650                            &page_tail->_count);
1651
1652                 /* after clearing PageTail the gup refcount can be released */
1653                 smp_mb();
1654
1655                 /*
1656                  * retain hwpoison flag of the poisoned tail page:
1657                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1658                  *   by the memory-failure.
1659                  */
1660                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1661                 page_tail->flags |= (page->flags &
1662                                      ((1L << PG_referenced) |
1663                                       (1L << PG_swapbacked) |
1664                                       (1L << PG_mlocked) |
1665                                       (1L << PG_uptodate) |
1666                                       (1L << PG_active) |
1667                                       (1L << PG_unevictable)));
1668                 page_tail->flags |= (1L << PG_dirty);
1669
1670                 /* clear PageTail before overwriting first_page */
1671                 smp_wmb();
1672
1673                 /*
1674                  * __split_huge_page_splitting() already set the
1675                  * splitting bit in all pmd that could map this
1676                  * hugepage, that will ensure no CPU can alter the
1677                  * mapcount on the head page. The mapcount is only
1678                  * accounted in the head page and it has to be
1679                  * transferred to all tail pages in the below code. So
1680                  * for this code to be safe, the split the mapcount
1681                  * can't change. But that doesn't mean userland can't
1682                  * keep changing and reading the page contents while
1683                  * we transfer the mapcount, so the pmd splitting
1684                  * status is achieved setting a reserved bit in the
1685                  * pmd, not by clearing the present bit.
1686                 */
1687                 page_tail->_mapcount = page->_mapcount;
1688
1689                 BUG_ON(page_tail->mapping);
1690                 page_tail->mapping = page->mapping;
1691
1692                 page_tail->index = page->index + i;
1693                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1694
1695                 BUG_ON(!PageAnon(page_tail));
1696                 BUG_ON(!PageUptodate(page_tail));
1697                 BUG_ON(!PageDirty(page_tail));
1698                 BUG_ON(!PageSwapBacked(page_tail));
1699
1700                 lru_add_page_tail(page, page_tail, lruvec, list);
1701         }
1702         atomic_sub(tail_count, &page->_count);
1703         BUG_ON(atomic_read(&page->_count) <= 0);
1704
1705         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1706
1707         ClearPageCompound(page);
1708         compound_unlock(page);
1709         spin_unlock_irq(&zone->lru_lock);
1710
1711         for (i = 1; i < HPAGE_PMD_NR; i++) {
1712                 struct page *page_tail = page + i;
1713                 BUG_ON(page_count(page_tail) <= 0);
1714                 /*
1715                  * Tail pages may be freed if there wasn't any mapping
1716                  * like if add_to_swap() is running on a lru page that
1717                  * had its mapping zapped. And freeing these pages
1718                  * requires taking the lru_lock so we do the put_page
1719                  * of the tail pages after the split is complete.
1720                  */
1721                 put_page(page_tail);
1722         }
1723
1724         /*
1725          * Only the head page (now become a regular page) is required
1726          * to be pinned by the caller.
1727          */
1728         BUG_ON(page_count(page) <= 0);
1729 }
1730
1731 static int __split_huge_page_map(struct page *page,
1732                                  struct vm_area_struct *vma,
1733                                  unsigned long address)
1734 {
1735         struct mm_struct *mm = vma->vm_mm;
1736         spinlock_t *ptl;
1737         pmd_t *pmd, _pmd;
1738         int ret = 0, i;
1739         pgtable_t pgtable;
1740         unsigned long haddr;
1741
1742         pmd = page_check_address_pmd(page, mm, address,
1743                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1744         if (pmd) {
1745                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1746                 pmd_populate(mm, &_pmd, pgtable);
1747
1748                 haddr = address;
1749                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1750                         pte_t *pte, entry;
1751                         BUG_ON(PageCompound(page+i));
1752                         entry = mk_pte(page + i, vma->vm_page_prot);
1753                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1754                         if (!pmd_write(*pmd))
1755                                 entry = pte_wrprotect(entry);
1756                         else
1757                                 BUG_ON(page_mapcount(page) != 1);
1758                         if (!pmd_young(*pmd))
1759                                 entry = pte_mkold(entry);
1760                         if (pmd_numa(*pmd))
1761                                 entry = pte_mknuma(entry);
1762                         pte = pte_offset_map(&_pmd, haddr);
1763                         BUG_ON(!pte_none(*pte));
1764                         set_pte_at(mm, haddr, pte, entry);
1765                         pte_unmap(pte);
1766                 }
1767
1768                 smp_wmb(); /* make pte visible before pmd */
1769                 /*
1770                  * Up to this point the pmd is present and huge and
1771                  * userland has the whole access to the hugepage
1772                  * during the split (which happens in place). If we
1773                  * overwrite the pmd with the not-huge version
1774                  * pointing to the pte here (which of course we could
1775                  * if all CPUs were bug free), userland could trigger
1776                  * a small page size TLB miss on the small sized TLB
1777                  * while the hugepage TLB entry is still established
1778                  * in the huge TLB. Some CPU doesn't like that. See
1779                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1780                  * Erratum 383 on page 93. Intel should be safe but is
1781                  * also warns that it's only safe if the permission
1782                  * and cache attributes of the two entries loaded in
1783                  * the two TLB is identical (which should be the case
1784                  * here). But it is generally safer to never allow
1785                  * small and huge TLB entries for the same virtual
1786                  * address to be loaded simultaneously. So instead of
1787                  * doing "pmd_populate(); flush_tlb_range();" we first
1788                  * mark the current pmd notpresent (atomically because
1789                  * here the pmd_trans_huge and pmd_trans_splitting
1790                  * must remain set at all times on the pmd until the
1791                  * split is complete for this pmd), then we flush the
1792                  * SMP TLB and finally we write the non-huge version
1793                  * of the pmd entry with pmd_populate.
1794                  */
1795                 pmdp_invalidate(vma, address, pmd);
1796                 pmd_populate(mm, pmd, pgtable);
1797                 ret = 1;
1798                 spin_unlock(ptl);
1799         }
1800
1801         return ret;
1802 }
1803
1804 /* must be called with anon_vma->root->rwsem held */
1805 static void __split_huge_page(struct page *page,
1806                               struct anon_vma *anon_vma,
1807                               struct list_head *list)
1808 {
1809         int mapcount, mapcount2;
1810         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1811         struct anon_vma_chain *avc;
1812
1813         BUG_ON(!PageHead(page));
1814         BUG_ON(PageTail(page));
1815
1816         mapcount = 0;
1817         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1818                 struct vm_area_struct *vma = avc->vma;
1819                 unsigned long addr = vma_address(page, vma);
1820                 BUG_ON(is_vma_temporary_stack(vma));
1821                 mapcount += __split_huge_page_splitting(page, vma, addr);
1822         }
1823         /*
1824          * It is critical that new vmas are added to the tail of the
1825          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1826          * and establishes a child pmd before
1827          * __split_huge_page_splitting() freezes the parent pmd (so if
1828          * we fail to prevent copy_huge_pmd() from running until the
1829          * whole __split_huge_page() is complete), we will still see
1830          * the newly established pmd of the child later during the
1831          * walk, to be able to set it as pmd_trans_splitting too.
1832          */
1833         if (mapcount != page_mapcount(page)) {
1834                 pr_err("mapcount %d page_mapcount %d\n",
1835                         mapcount, page_mapcount(page));
1836                 BUG();
1837         }
1838
1839         __split_huge_page_refcount(page, list);
1840
1841         mapcount2 = 0;
1842         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1843                 struct vm_area_struct *vma = avc->vma;
1844                 unsigned long addr = vma_address(page, vma);
1845                 BUG_ON(is_vma_temporary_stack(vma));
1846                 mapcount2 += __split_huge_page_map(page, vma, addr);
1847         }
1848         if (mapcount != mapcount2) {
1849                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1850                         mapcount, mapcount2, page_mapcount(page));
1851                 BUG();
1852         }
1853 }
1854
1855 /*
1856  * Split a hugepage into normal pages. This doesn't change the position of head
1857  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1858  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1859  * from the hugepage.
1860  * Return 0 if the hugepage is split successfully otherwise return 1.
1861  */
1862 int split_huge_page_to_list(struct page *page, struct list_head *list)
1863 {
1864         struct anon_vma *anon_vma;
1865         int ret = 1;
1866
1867         BUG_ON(is_huge_zero_page(page));
1868         BUG_ON(!PageAnon(page));
1869
1870         /*
1871          * The caller does not necessarily hold an mmap_sem that would prevent
1872          * the anon_vma disappearing so we first we take a reference to it
1873          * and then lock the anon_vma for write. This is similar to
1874          * page_lock_anon_vma_read except the write lock is taken to serialise
1875          * against parallel split or collapse operations.
1876          */
1877         anon_vma = page_get_anon_vma(page);
1878         if (!anon_vma)
1879                 goto out;
1880         anon_vma_lock_write(anon_vma);
1881
1882         ret = 0;
1883         if (!PageCompound(page))
1884                 goto out_unlock;
1885
1886         BUG_ON(!PageSwapBacked(page));
1887         __split_huge_page(page, anon_vma, list);
1888         count_vm_event(THP_SPLIT);
1889
1890         BUG_ON(PageCompound(page));
1891 out_unlock:
1892         anon_vma_unlock_write(anon_vma);
1893         put_anon_vma(anon_vma);
1894 out:
1895         return ret;
1896 }
1897
1898 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1899
1900 int hugepage_madvise(struct vm_area_struct *vma,
1901                      unsigned long *vm_flags, int advice)
1902 {
1903         switch (advice) {
1904         case MADV_HUGEPAGE:
1905 #ifdef CONFIG_S390
1906                 /*
1907                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1908                  * can't handle this properly after s390_enable_sie, so we simply
1909                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1910                  */
1911                 if (mm_has_pgste(vma->vm_mm))
1912                         return 0;
1913 #endif
1914                 /*
1915                  * Be somewhat over-protective like KSM for now!
1916                  */
1917                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1918                         return -EINVAL;
1919                 *vm_flags &= ~VM_NOHUGEPAGE;
1920                 *vm_flags |= VM_HUGEPAGE;
1921                 /*
1922                  * If the vma become good for khugepaged to scan,
1923                  * register it here without waiting a page fault that
1924                  * may not happen any time soon.
1925                  */
1926                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1927                         return -ENOMEM;
1928                 break;
1929         case MADV_NOHUGEPAGE:
1930                 /*
1931                  * Be somewhat over-protective like KSM for now!
1932                  */
1933                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1934                         return -EINVAL;
1935                 *vm_flags &= ~VM_HUGEPAGE;
1936                 *vm_flags |= VM_NOHUGEPAGE;
1937                 /*
1938                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1939                  * this vma even if we leave the mm registered in khugepaged if
1940                  * it got registered before VM_NOHUGEPAGE was set.
1941                  */
1942                 break;
1943         }
1944
1945         return 0;
1946 }
1947
1948 static int __init khugepaged_slab_init(void)
1949 {
1950         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1951                                           sizeof(struct mm_slot),
1952                                           __alignof__(struct mm_slot), 0, NULL);
1953         if (!mm_slot_cache)
1954                 return -ENOMEM;
1955
1956         return 0;
1957 }
1958
1959 static inline struct mm_slot *alloc_mm_slot(void)
1960 {
1961         if (!mm_slot_cache)     /* initialization failed */
1962                 return NULL;
1963         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1964 }
1965
1966 static inline void free_mm_slot(struct mm_slot *mm_slot)
1967 {
1968         kmem_cache_free(mm_slot_cache, mm_slot);
1969 }
1970
1971 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1972 {
1973         struct mm_slot *mm_slot;
1974
1975         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1976                 if (mm == mm_slot->mm)
1977                         return mm_slot;
1978
1979         return NULL;
1980 }
1981
1982 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1983                                     struct mm_slot *mm_slot)
1984 {
1985         mm_slot->mm = mm;
1986         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1987 }
1988
1989 static inline int khugepaged_test_exit(struct mm_struct *mm)
1990 {
1991         return atomic_read(&mm->mm_users) == 0;
1992 }
1993
1994 int __khugepaged_enter(struct mm_struct *mm)
1995 {
1996         struct mm_slot *mm_slot;
1997         int wakeup;
1998
1999         mm_slot = alloc_mm_slot();
2000         if (!mm_slot)
2001                 return -ENOMEM;
2002
2003         /* __khugepaged_exit() must not run from under us */
2004         VM_BUG_ON(khugepaged_test_exit(mm));
2005         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2006                 free_mm_slot(mm_slot);
2007                 return 0;
2008         }
2009
2010         spin_lock(&khugepaged_mm_lock);
2011         insert_to_mm_slots_hash(mm, mm_slot);
2012         /*
2013          * Insert just behind the scanning cursor, to let the area settle
2014          * down a little.
2015          */
2016         wakeup = list_empty(&khugepaged_scan.mm_head);
2017         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2018         spin_unlock(&khugepaged_mm_lock);
2019
2020         atomic_inc(&mm->mm_count);
2021         if (wakeup)
2022                 wake_up_interruptible(&khugepaged_wait);
2023
2024         return 0;
2025 }
2026
2027 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2028 {
2029         unsigned long hstart, hend;
2030         if (!vma->anon_vma)
2031                 /*
2032                  * Not yet faulted in so we will register later in the
2033                  * page fault if needed.
2034                  */
2035                 return 0;
2036         if (vma->vm_ops)
2037                 /* khugepaged not yet working on file or special mappings */
2038                 return 0;
2039         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2040         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2041         hend = vma->vm_end & HPAGE_PMD_MASK;
2042         if (hstart < hend)
2043                 return khugepaged_enter(vma);
2044         return 0;
2045 }
2046
2047 void __khugepaged_exit(struct mm_struct *mm)
2048 {
2049         struct mm_slot *mm_slot;
2050         int free = 0;
2051
2052         spin_lock(&khugepaged_mm_lock);
2053         mm_slot = get_mm_slot(mm);
2054         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2055                 hash_del(&mm_slot->hash);
2056                 list_del(&mm_slot->mm_node);
2057                 free = 1;
2058         }
2059         spin_unlock(&khugepaged_mm_lock);
2060
2061         if (free) {
2062                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2063                 free_mm_slot(mm_slot);
2064                 mmdrop(mm);
2065         } else if (mm_slot) {
2066                 /*
2067                  * This is required to serialize against
2068                  * khugepaged_test_exit() (which is guaranteed to run
2069                  * under mmap sem read mode). Stop here (after we
2070                  * return all pagetables will be destroyed) until
2071                  * khugepaged has finished working on the pagetables
2072                  * under the mmap_sem.
2073                  */
2074                 down_write(&mm->mmap_sem);
2075                 up_write(&mm->mmap_sem);
2076         }
2077 }
2078
2079 static void release_pte_page(struct page *page)
2080 {
2081         /* 0 stands for page_is_file_cache(page) == false */
2082         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2083         unlock_page(page);
2084         putback_lru_page(page);
2085 }
2086
2087 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2088 {
2089         while (--_pte >= pte) {
2090                 pte_t pteval = *_pte;
2091                 if (!pte_none(pteval))
2092                         release_pte_page(pte_page(pteval));
2093         }
2094 }
2095
2096 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2097                                         unsigned long address,
2098                                         pte_t *pte)
2099 {
2100         struct page *page;
2101         pte_t *_pte;
2102         int referenced = 0, none = 0;
2103         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2104              _pte++, address += PAGE_SIZE) {
2105                 pte_t pteval = *_pte;
2106                 if (pte_none(pteval)) {
2107                         if (++none <= khugepaged_max_ptes_none)
2108                                 continue;
2109                         else
2110                                 goto out;
2111                 }
2112                 if (!pte_present(pteval) || !pte_write(pteval))
2113                         goto out;
2114                 page = vm_normal_page(vma, address, pteval);
2115                 if (unlikely(!page))
2116                         goto out;
2117
2118                 VM_BUG_ON_PAGE(PageCompound(page), page);
2119                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2120                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2121
2122                 /* cannot use mapcount: can't collapse if there's a gup pin */
2123                 if (page_count(page) != 1)
2124                         goto out;
2125                 /*
2126                  * We can do it before isolate_lru_page because the
2127                  * page can't be freed from under us. NOTE: PG_lock
2128                  * is needed to serialize against split_huge_page
2129                  * when invoked from the VM.
2130                  */
2131                 if (!trylock_page(page))
2132                         goto out;
2133                 /*
2134                  * Isolate the page to avoid collapsing an hugepage
2135                  * currently in use by the VM.
2136                  */
2137                 if (isolate_lru_page(page)) {
2138                         unlock_page(page);
2139                         goto out;
2140                 }
2141                 /* 0 stands for page_is_file_cache(page) == false */
2142                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2143                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2144                 VM_BUG_ON_PAGE(PageLRU(page), page);
2145
2146                 /* If there is no mapped pte young don't collapse the page */
2147                 if (pte_young(pteval) || PageReferenced(page) ||
2148                     mmu_notifier_test_young(vma->vm_mm, address))
2149                         referenced = 1;
2150         }
2151         if (likely(referenced))
2152                 return 1;
2153 out:
2154         release_pte_pages(pte, _pte);
2155         return 0;
2156 }
2157
2158 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2159                                       struct vm_area_struct *vma,
2160                                       unsigned long address,
2161                                       spinlock_t *ptl)
2162 {
2163         pte_t *_pte;
2164         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2165                 pte_t pteval = *_pte;
2166                 struct page *src_page;
2167
2168                 if (pte_none(pteval)) {
2169                         clear_user_highpage(page, address);
2170                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2171                 } else {
2172                         src_page = pte_page(pteval);
2173                         copy_user_highpage(page, src_page, address, vma);
2174                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2175                         release_pte_page(src_page);
2176                         /*
2177                          * ptl mostly unnecessary, but preempt has to
2178                          * be disabled to update the per-cpu stats
2179                          * inside page_remove_rmap().
2180                          */
2181                         spin_lock(ptl);
2182                         /*
2183                          * paravirt calls inside pte_clear here are
2184                          * superfluous.
2185                          */
2186                         pte_clear(vma->vm_mm, address, _pte);
2187                         page_remove_rmap(src_page);
2188                         spin_unlock(ptl);
2189                         free_page_and_swap_cache(src_page);
2190                 }
2191
2192                 address += PAGE_SIZE;
2193                 page++;
2194         }
2195 }
2196
2197 static void khugepaged_alloc_sleep(void)
2198 {
2199         wait_event_freezable_timeout(khugepaged_wait, false,
2200                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2201 }
2202
2203 static int khugepaged_node_load[MAX_NUMNODES];
2204
2205 #ifdef CONFIG_NUMA
2206 static int khugepaged_find_target_node(void)
2207 {
2208         static int last_khugepaged_target_node = NUMA_NO_NODE;
2209         int nid, target_node = 0, max_value = 0;
2210
2211         /* find first node with max normal pages hit */
2212         for (nid = 0; nid < MAX_NUMNODES; nid++)
2213                 if (khugepaged_node_load[nid] > max_value) {
2214                         max_value = khugepaged_node_load[nid];
2215                         target_node = nid;
2216                 }
2217
2218         /* do some balance if several nodes have the same hit record */
2219         if (target_node <= last_khugepaged_target_node)
2220                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2221                                 nid++)
2222                         if (max_value == khugepaged_node_load[nid]) {
2223                                 target_node = nid;
2224                                 break;
2225                         }
2226
2227         last_khugepaged_target_node = target_node;
2228         return target_node;
2229 }
2230
2231 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2232 {
2233         if (IS_ERR(*hpage)) {
2234                 if (!*wait)
2235                         return false;
2236
2237                 *wait = false;
2238                 *hpage = NULL;
2239                 khugepaged_alloc_sleep();
2240         } else if (*hpage) {
2241                 put_page(*hpage);
2242                 *hpage = NULL;
2243         }
2244
2245         return true;
2246 }
2247
2248 static struct page
2249 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2250                        struct vm_area_struct *vma, unsigned long address,
2251                        int node)
2252 {
2253         VM_BUG_ON_PAGE(*hpage, *hpage);
2254         /*
2255          * Allocate the page while the vma is still valid and under
2256          * the mmap_sem read mode so there is no memory allocation
2257          * later when we take the mmap_sem in write mode. This is more
2258          * friendly behavior (OTOH it may actually hide bugs) to
2259          * filesystems in userland with daemons allocating memory in
2260          * the userland I/O paths.  Allocating memory with the
2261          * mmap_sem in read mode is good idea also to allow greater
2262          * scalability.
2263          */
2264         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2265                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2266         /*
2267          * After allocating the hugepage, release the mmap_sem read lock in
2268          * preparation for taking it in write mode.
2269          */
2270         up_read(&mm->mmap_sem);
2271         if (unlikely(!*hpage)) {
2272                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2273                 *hpage = ERR_PTR(-ENOMEM);
2274                 return NULL;
2275         }
2276
2277         count_vm_event(THP_COLLAPSE_ALLOC);
2278         return *hpage;
2279 }
2280 #else
2281 static int khugepaged_find_target_node(void)
2282 {
2283         return 0;
2284 }
2285
2286 static inline struct page *alloc_hugepage(int defrag)
2287 {
2288         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2289                            HPAGE_PMD_ORDER);
2290 }
2291
2292 static struct page *khugepaged_alloc_hugepage(bool *wait)
2293 {
2294         struct page *hpage;
2295
2296         do {
2297                 hpage = alloc_hugepage(khugepaged_defrag());
2298                 if (!hpage) {
2299                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2300                         if (!*wait)
2301                                 return NULL;
2302
2303                         *wait = false;
2304                         khugepaged_alloc_sleep();
2305                 } else
2306                         count_vm_event(THP_COLLAPSE_ALLOC);
2307         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2308
2309         return hpage;
2310 }
2311
2312 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2313 {
2314         if (!*hpage)
2315                 *hpage = khugepaged_alloc_hugepage(wait);
2316
2317         if (unlikely(!*hpage))
2318                 return false;
2319
2320         return true;
2321 }
2322
2323 static struct page
2324 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2325                        struct vm_area_struct *vma, unsigned long address,
2326                        int node)
2327 {
2328         up_read(&mm->mmap_sem);
2329         VM_BUG_ON(!*hpage);
2330         return  *hpage;
2331 }
2332 #endif
2333
2334 static bool hugepage_vma_check(struct vm_area_struct *vma)
2335 {
2336         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2337             (vma->vm_flags & VM_NOHUGEPAGE))
2338                 return false;
2339
2340         if (!vma->anon_vma || vma->vm_ops)
2341                 return false;
2342         if (is_vma_temporary_stack(vma))
2343                 return false;
2344         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2345         return true;
2346 }
2347
2348 static void collapse_huge_page(struct mm_struct *mm,
2349                                    unsigned long address,
2350                                    struct page **hpage,
2351                                    struct vm_area_struct *vma,
2352                                    int node)
2353 {
2354         pmd_t *pmd, _pmd;
2355         pte_t *pte;
2356         pgtable_t pgtable;
2357         struct page *new_page;
2358         spinlock_t *pmd_ptl, *pte_ptl;
2359         int isolated;
2360         unsigned long hstart, hend;
2361         unsigned long mmun_start;       /* For mmu_notifiers */
2362         unsigned long mmun_end;         /* For mmu_notifiers */
2363
2364         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2365
2366         /* release the mmap_sem read lock. */
2367         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2368         if (!new_page)
2369                 return;
2370
2371         if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)))
2372                 return;
2373
2374         /*
2375          * Prevent all access to pagetables with the exception of
2376          * gup_fast later hanlded by the ptep_clear_flush and the VM
2377          * handled by the anon_vma lock + PG_lock.
2378          */
2379         down_write(&mm->mmap_sem);
2380         if (unlikely(khugepaged_test_exit(mm)))
2381                 goto out;
2382
2383         vma = find_vma(mm, address);
2384         if (!vma)
2385                 goto out;
2386         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2387         hend = vma->vm_end & HPAGE_PMD_MASK;
2388         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2389                 goto out;
2390         if (!hugepage_vma_check(vma))
2391                 goto out;
2392         pmd = mm_find_pmd(mm, address);
2393         if (!pmd)
2394                 goto out;
2395         if (pmd_trans_huge(*pmd))
2396                 goto out;
2397
2398         anon_vma_lock_write(vma->anon_vma);
2399
2400         pte = pte_offset_map(pmd, address);
2401         pte_ptl = pte_lockptr(mm, pmd);
2402
2403         mmun_start = address;
2404         mmun_end   = address + HPAGE_PMD_SIZE;
2405         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2406         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2407         /*
2408          * After this gup_fast can't run anymore. This also removes
2409          * any huge TLB entry from the CPU so we won't allow
2410          * huge and small TLB entries for the same virtual address
2411          * to avoid the risk of CPU bugs in that area.
2412          */
2413         _pmd = pmdp_clear_flush(vma, address, pmd);
2414         spin_unlock(pmd_ptl);
2415         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2416
2417         spin_lock(pte_ptl);
2418         isolated = __collapse_huge_page_isolate(vma, address, pte);
2419         spin_unlock(pte_ptl);
2420
2421         if (unlikely(!isolated)) {
2422                 pte_unmap(pte);
2423                 spin_lock(pmd_ptl);
2424                 BUG_ON(!pmd_none(*pmd));
2425                 /*
2426                  * We can only use set_pmd_at when establishing
2427                  * hugepmds and never for establishing regular pmds that
2428                  * points to regular pagetables. Use pmd_populate for that
2429                  */
2430                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2431                 spin_unlock(pmd_ptl);
2432                 anon_vma_unlock_write(vma->anon_vma);
2433                 goto out;
2434         }
2435
2436         /*
2437          * All pages are isolated and locked so anon_vma rmap
2438          * can't run anymore.
2439          */
2440         anon_vma_unlock_write(vma->anon_vma);
2441
2442         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2443         pte_unmap(pte);
2444         __SetPageUptodate(new_page);
2445         pgtable = pmd_pgtable(_pmd);
2446
2447         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2448         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2449
2450         /*
2451          * spin_lock() below is not the equivalent of smp_wmb(), so
2452          * this is needed to avoid the copy_huge_page writes to become
2453          * visible after the set_pmd_at() write.
2454          */
2455         smp_wmb();
2456
2457         spin_lock(pmd_ptl);
2458         BUG_ON(!pmd_none(*pmd));
2459         page_add_new_anon_rmap(new_page, vma, address);
2460         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2461         set_pmd_at(mm, address, pmd, _pmd);
2462         update_mmu_cache_pmd(vma, address, pmd);
2463         spin_unlock(pmd_ptl);
2464
2465         *hpage = NULL;
2466
2467         khugepaged_pages_collapsed++;
2468 out_up_write:
2469         up_write(&mm->mmap_sem);
2470         return;
2471
2472 out:
2473         mem_cgroup_uncharge_page(new_page);
2474         goto out_up_write;
2475 }
2476
2477 static int khugepaged_scan_pmd(struct mm_struct *mm,
2478                                struct vm_area_struct *vma,
2479                                unsigned long address,
2480                                struct page **hpage)
2481 {
2482         pmd_t *pmd;
2483         pte_t *pte, *_pte;
2484         int ret = 0, referenced = 0, none = 0;
2485         struct page *page;
2486         unsigned long _address;
2487         spinlock_t *ptl;
2488         int node = NUMA_NO_NODE;
2489
2490         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2491
2492         pmd = mm_find_pmd(mm, address);
2493         if (!pmd)
2494                 goto out;
2495         if (pmd_trans_huge(*pmd))
2496                 goto out;
2497
2498         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2499         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2500         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2501              _pte++, _address += PAGE_SIZE) {
2502                 pte_t pteval = *_pte;
2503                 if (pte_none(pteval)) {
2504                         if (++none <= khugepaged_max_ptes_none)
2505                                 continue;
2506                         else
2507                                 goto out_unmap;
2508                 }
2509                 if (!pte_present(pteval) || !pte_write(pteval))
2510                         goto out_unmap;
2511                 page = vm_normal_page(vma, _address, pteval);
2512                 if (unlikely(!page))
2513                         goto out_unmap;
2514                 /*
2515                  * Record which node the original page is from and save this
2516                  * information to khugepaged_node_load[].
2517                  * Khupaged will allocate hugepage from the node has the max
2518                  * hit record.
2519                  */
2520                 node = page_to_nid(page);
2521                 khugepaged_node_load[node]++;
2522                 VM_BUG_ON_PAGE(PageCompound(page), page);
2523                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2524                         goto out_unmap;
2525                 /* cannot use mapcount: can't collapse if there's a gup pin */
2526                 if (page_count(page) != 1)
2527                         goto out_unmap;
2528                 if (pte_young(pteval) || PageReferenced(page) ||
2529                     mmu_notifier_test_young(vma->vm_mm, address))
2530                         referenced = 1;
2531         }
2532         if (referenced)
2533                 ret = 1;
2534 out_unmap:
2535         pte_unmap_unlock(pte, ptl);
2536         if (ret) {
2537                 node = khugepaged_find_target_node();
2538                 /* collapse_huge_page will return with the mmap_sem released */
2539                 collapse_huge_page(mm, address, hpage, vma, node);
2540         }
2541 out:
2542         return ret;
2543 }
2544
2545 static void collect_mm_slot(struct mm_slot *mm_slot)
2546 {
2547         struct mm_struct *mm = mm_slot->mm;
2548
2549         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2550
2551         if (khugepaged_test_exit(mm)) {
2552                 /* free mm_slot */
2553                 hash_del(&mm_slot->hash);
2554                 list_del(&mm_slot->mm_node);
2555
2556                 /*
2557                  * Not strictly needed because the mm exited already.
2558                  *
2559                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2560                  */
2561
2562                 /* khugepaged_mm_lock actually not necessary for the below */
2563                 free_mm_slot(mm_slot);
2564                 mmdrop(mm);
2565         }
2566 }
2567
2568 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2569                                             struct page **hpage)
2570         __releases(&khugepaged_mm_lock)
2571         __acquires(&khugepaged_mm_lock)
2572 {
2573         struct mm_slot *mm_slot;
2574         struct mm_struct *mm;
2575         struct vm_area_struct *vma;
2576         int progress = 0;
2577
2578         VM_BUG_ON(!pages);
2579         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2580
2581         if (khugepaged_scan.mm_slot)
2582                 mm_slot = khugepaged_scan.mm_slot;
2583         else {
2584                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2585                                      struct mm_slot, mm_node);
2586                 khugepaged_scan.address = 0;
2587                 khugepaged_scan.mm_slot = mm_slot;
2588         }
2589         spin_unlock(&khugepaged_mm_lock);
2590
2591         mm = mm_slot->mm;
2592         down_read(&mm->mmap_sem);
2593         if (unlikely(khugepaged_test_exit(mm)))
2594                 vma = NULL;
2595         else
2596                 vma = find_vma(mm, khugepaged_scan.address);
2597
2598         progress++;
2599         for (; vma; vma = vma->vm_next) {
2600                 unsigned long hstart, hend;
2601
2602                 cond_resched();
2603                 if (unlikely(khugepaged_test_exit(mm))) {
2604                         progress++;
2605                         break;
2606                 }
2607                 if (!hugepage_vma_check(vma)) {
2608 skip:
2609                         progress++;
2610                         continue;
2611                 }
2612                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2613                 hend = vma->vm_end & HPAGE_PMD_MASK;
2614                 if (hstart >= hend)
2615                         goto skip;
2616                 if (khugepaged_scan.address > hend)
2617                         goto skip;
2618                 if (khugepaged_scan.address < hstart)
2619                         khugepaged_scan.address = hstart;
2620                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2621
2622                 while (khugepaged_scan.address < hend) {
2623                         int ret;
2624                         cond_resched();
2625                         if (unlikely(khugepaged_test_exit(mm)))
2626                                 goto breakouterloop;
2627
2628                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2629                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2630                                   hend);
2631                         ret = khugepaged_scan_pmd(mm, vma,
2632                                                   khugepaged_scan.address,
2633                                                   hpage);
2634                         /* move to next address */
2635                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2636                         progress += HPAGE_PMD_NR;
2637                         if (ret)
2638                                 /* we released mmap_sem so break loop */
2639                                 goto breakouterloop_mmap_sem;
2640                         if (progress >= pages)
2641                                 goto breakouterloop;
2642                 }
2643         }
2644 breakouterloop:
2645         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2646 breakouterloop_mmap_sem:
2647
2648         spin_lock(&khugepaged_mm_lock);
2649         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2650         /*
2651          * Release the current mm_slot if this mm is about to die, or
2652          * if we scanned all vmas of this mm.
2653          */
2654         if (khugepaged_test_exit(mm) || !vma) {
2655                 /*
2656                  * Make sure that if mm_users is reaching zero while
2657                  * khugepaged runs here, khugepaged_exit will find
2658                  * mm_slot not pointing to the exiting mm.
2659                  */
2660                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2661                         khugepaged_scan.mm_slot = list_entry(
2662                                 mm_slot->mm_node.next,
2663                                 struct mm_slot, mm_node);
2664                         khugepaged_scan.address = 0;
2665                 } else {
2666                         khugepaged_scan.mm_slot = NULL;
2667                         khugepaged_full_scans++;
2668                 }
2669
2670                 collect_mm_slot(mm_slot);
2671         }
2672
2673         return progress;
2674 }
2675
2676 static int khugepaged_has_work(void)
2677 {
2678         return !list_empty(&khugepaged_scan.mm_head) &&
2679                 khugepaged_enabled();
2680 }
2681
2682 static int khugepaged_wait_event(void)
2683 {
2684         return !list_empty(&khugepaged_scan.mm_head) ||
2685                 kthread_should_stop();
2686 }
2687
2688 static void khugepaged_do_scan(void)
2689 {
2690         struct page *hpage = NULL;
2691         unsigned int progress = 0, pass_through_head = 0;
2692         unsigned int pages = khugepaged_pages_to_scan;
2693         bool wait = true;
2694
2695         barrier(); /* write khugepaged_pages_to_scan to local stack */
2696
2697         while (progress < pages) {
2698                 if (!khugepaged_prealloc_page(&hpage, &wait))
2699                         break;
2700
2701                 cond_resched();
2702
2703                 if (unlikely(kthread_should_stop() || freezing(current)))
2704                         break;
2705
2706                 spin_lock(&khugepaged_mm_lock);
2707                 if (!khugepaged_scan.mm_slot)
2708                         pass_through_head++;
2709                 if (khugepaged_has_work() &&
2710                     pass_through_head < 2)
2711                         progress += khugepaged_scan_mm_slot(pages - progress,
2712                                                             &hpage);
2713                 else
2714                         progress = pages;
2715                 spin_unlock(&khugepaged_mm_lock);
2716         }
2717
2718         if (!IS_ERR_OR_NULL(hpage))
2719                 put_page(hpage);
2720 }
2721
2722 static void khugepaged_wait_work(void)
2723 {
2724         try_to_freeze();
2725
2726         if (khugepaged_has_work()) {
2727                 if (!khugepaged_scan_sleep_millisecs)
2728                         return;
2729
2730                 wait_event_freezable_timeout(khugepaged_wait,
2731                                              kthread_should_stop(),
2732                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2733                 return;
2734         }
2735
2736         if (khugepaged_enabled())
2737                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2738 }
2739
2740 static int khugepaged(void *none)
2741 {
2742         struct mm_slot *mm_slot;
2743
2744         set_freezable();
2745         set_user_nice(current, MAX_NICE);
2746
2747         while (!kthread_should_stop()) {
2748                 khugepaged_do_scan();
2749                 khugepaged_wait_work();
2750         }
2751
2752         spin_lock(&khugepaged_mm_lock);
2753         mm_slot = khugepaged_scan.mm_slot;
2754         khugepaged_scan.mm_slot = NULL;
2755         if (mm_slot)
2756                 collect_mm_slot(mm_slot);
2757         spin_unlock(&khugepaged_mm_lock);
2758         return 0;
2759 }
2760
2761 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2762                 unsigned long haddr, pmd_t *pmd)
2763 {
2764         struct mm_struct *mm = vma->vm_mm;
2765         pgtable_t pgtable;
2766         pmd_t _pmd;
2767         int i;
2768
2769         pmdp_clear_flush(vma, haddr, pmd);
2770         /* leave pmd empty until pte is filled */
2771
2772         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2773         pmd_populate(mm, &_pmd, pgtable);
2774
2775         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2776                 pte_t *pte, entry;
2777                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2778                 entry = pte_mkspecial(entry);
2779                 pte = pte_offset_map(&_pmd, haddr);
2780                 VM_BUG_ON(!pte_none(*pte));
2781                 set_pte_at(mm, haddr, pte, entry);
2782                 pte_unmap(pte);
2783         }
2784         smp_wmb(); /* make pte visible before pmd */
2785         pmd_populate(mm, pmd, pgtable);
2786         put_huge_zero_page();
2787 }
2788
2789 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2790                 pmd_t *pmd)
2791 {
2792         spinlock_t *ptl;
2793         struct page *page;
2794         struct mm_struct *mm = vma->vm_mm;
2795         unsigned long haddr = address & HPAGE_PMD_MASK;
2796         unsigned long mmun_start;       /* For mmu_notifiers */
2797         unsigned long mmun_end;         /* For mmu_notifiers */
2798
2799         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2800
2801         mmun_start = haddr;
2802         mmun_end   = haddr + HPAGE_PMD_SIZE;
2803 again:
2804         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2805         ptl = pmd_lock(mm, pmd);
2806         if (unlikely(!pmd_trans_huge(*pmd))) {
2807                 spin_unlock(ptl);
2808                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2809                 return;
2810         }
2811         if (is_huge_zero_pmd(*pmd)) {
2812                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2813                 spin_unlock(ptl);
2814                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2815                 return;
2816         }
2817         page = pmd_page(*pmd);
2818         VM_BUG_ON_PAGE(!page_count(page), page);
2819         get_page(page);
2820         spin_unlock(ptl);
2821         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2822
2823         split_huge_page(page);
2824
2825         put_page(page);
2826
2827         /*
2828          * We don't always have down_write of mmap_sem here: a racing
2829          * do_huge_pmd_wp_page() might have copied-on-write to another
2830          * huge page before our split_huge_page() got the anon_vma lock.
2831          */
2832         if (unlikely(pmd_trans_huge(*pmd)))
2833                 goto again;
2834 }
2835
2836 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2837                 pmd_t *pmd)
2838 {
2839         struct vm_area_struct *vma;
2840
2841         vma = find_vma(mm, address);
2842         BUG_ON(vma == NULL);
2843         split_huge_page_pmd(vma, address, pmd);
2844 }
2845
2846 static void split_huge_page_address(struct mm_struct *mm,
2847                                     unsigned long address)
2848 {
2849         pmd_t *pmd;
2850
2851         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2852
2853         pmd = mm_find_pmd(mm, address);
2854         if (!pmd)
2855                 return;
2856         /*
2857          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2858          * materialize from under us.
2859          */
2860         split_huge_page_pmd_mm(mm, address, pmd);
2861 }
2862
2863 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2864                              unsigned long start,
2865                              unsigned long end,
2866                              long adjust_next)
2867 {
2868         /*
2869          * If the new start address isn't hpage aligned and it could
2870          * previously contain an hugepage: check if we need to split
2871          * an huge pmd.
2872          */
2873         if (start & ~HPAGE_PMD_MASK &&
2874             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2875             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2876                 split_huge_page_address(vma->vm_mm, start);
2877
2878         /*
2879          * If the new end address isn't hpage aligned and it could
2880          * previously contain an hugepage: check if we need to split
2881          * an huge pmd.
2882          */
2883         if (end & ~HPAGE_PMD_MASK &&
2884             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2885             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2886                 split_huge_page_address(vma->vm_mm, end);
2887
2888         /*
2889          * If we're also updating the vma->vm_next->vm_start, if the new
2890          * vm_next->vm_start isn't page aligned and it could previously
2891          * contain an hugepage: check if we need to split an huge pmd.
2892          */
2893         if (adjust_next > 0) {
2894                 struct vm_area_struct *next = vma->vm_next;
2895                 unsigned long nstart = next->vm_start;
2896                 nstart += adjust_next << PAGE_SHIFT;
2897                 if (nstart & ~HPAGE_PMD_MASK &&
2898                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2899                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2900                         split_huge_page_address(next->vm_mm, nstart);
2901         }
2902 }