mm/huge_memory.c

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