mm/huge_memory.c

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