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