mm/huge_memory.c

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