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
 714         VM_BUG_ON(!PageCompound(page));
 715         pgtable = pte_alloc_one(mm, haddr);
 716         if (unlikely(!pgtable))
 717                 return VM_FAULT_OOM;
 718
 719         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 720         /*
 721          * The memory barrier inside __SetPageUptodate makes sure that
 722          * clear_huge_page writes become visible before the set_pmd_at()
 723          * write.
 724          */
 725         __SetPageUptodate(page);
 726
 727         spin_lock(&mm->page_table_lock);
 728         if (unlikely(!pmd_none(*pmd))) {
 729                 spin_unlock(&mm->page_table_lock);
 730                 mem_cgroup_uncharge_page(page);
 731                 put_page(page);
 732                 pte_free(mm, pgtable);
 733         } else {
 734                 pmd_t entry;
 735                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 736                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 737                 page_add_new_anon_rmap(page, vma, haddr);
 738                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 739                 set_pmd_at(mm, haddr, pmd, entry);
 740                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 741                 atomic_long_inc(&mm->nr_ptes);
 742                 spin_unlock(&mm->page_table_lock);
 743         }
 744
 745         return 0;
 746 }
 747
 748 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 749 {
 750         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 751 }
 752
 753 static inline struct page *alloc_hugepage_vma(int defrag,
 754                                               struct vm_area_struct *vma,
 755                                               unsigned long haddr, int nd,
 756                                               gfp_t extra_gfp)
 757 {
 758         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 759                                HPAGE_PMD_ORDER, vma, haddr, nd);
 760 }
 761
 762 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 763                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 764                 struct page *zero_page)
 765 {
 766         pmd_t entry;
 767         if (!pmd_none(*pmd))
 768                 return false;
 769         entry = mk_pmd(zero_page, vma->vm_page_prot);
 770         entry = pmd_wrprotect(entry);
 771         entry = pmd_mkhuge(entry);
 772         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 773         set_pmd_at(mm, haddr, pmd, entry);
 774         atomic_long_inc(&mm->nr_ptes);
 775         return true;
 776 }
 777
 778 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 779                                unsigned long address, pmd_t *pmd,
 780                                unsigned int flags)
 781 {
 782         struct page *page;
 783         unsigned long haddr = address & HPAGE_PMD_MASK;
 784
 785         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 786                 return VM_FAULT_FALLBACK;
 787         if (unlikely(anon_vma_prepare(vma)))
 788                 return VM_FAULT_OOM;
 789         if (unlikely(khugepaged_enter(vma)))
 790                 return VM_FAULT_OOM;
 791         if (!(flags & FAULT_FLAG_WRITE) &&
 792                         transparent_hugepage_use_zero_page()) {
 793                 pgtable_t pgtable;
 794                 struct page *zero_page;
 795                 bool set;
 796                 pgtable = pte_alloc_one(mm, haddr);
 797                 if (unlikely(!pgtable))
 798                         return VM_FAULT_OOM;
 799                 zero_page = get_huge_zero_page();
 800                 if (unlikely(!zero_page)) {
 801                         pte_free(mm, pgtable);
 802                         count_vm_event(THP_FAULT_FALLBACK);
 803                         return VM_FAULT_FALLBACK;
 804                 }
 805                 spin_lock(&mm->page_table_lock);
 806                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 807                                 zero_page);
 808                 spin_unlock(&mm->page_table_lock);
 809                 if (!set) {
 810                         pte_free(mm, pgtable);
 811                         put_huge_zero_page();
 812                 }
 813                 return 0;
 814         }
 815         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 816                         vma, haddr, numa_node_id(), 0);
 817         if (unlikely(!page)) {
 818                 count_vm_event(THP_FAULT_FALLBACK);
 819                 return VM_FAULT_FALLBACK;
 820         }
 821         if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 822                 put_page(page);
 823                 count_vm_event(THP_FAULT_FALLBACK);
 824                 return VM_FAULT_FALLBACK;
 825         }
 826         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 827                 mem_cgroup_uncharge_page(page);
 828                 put_page(page);
 829                 count_vm_event(THP_FAULT_FALLBACK);
 830                 return VM_FAULT_FALLBACK;
 831         }
 832
 833         count_vm_event(THP_FAULT_ALLOC);
 834         return 0;
 835 }
 836
 837 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 838                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 839                   struct vm_area_struct *vma)
 840 {
 841         struct page *src_page;
 842         pmd_t pmd;
 843         pgtable_t pgtable;
 844         int ret;
 845
 846         ret = -ENOMEM;
 847         pgtable = pte_alloc_one(dst_mm, addr);
 848         if (unlikely(!pgtable))
 849                 goto out;
 850
 851         spin_lock(&dst_mm->page_table_lock);
 852         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 853
 854         ret = -EAGAIN;
 855         pmd = *src_pmd;
 856         if (unlikely(!pmd_trans_huge(pmd))) {
 857                 pte_free(dst_mm, pgtable);
 858                 goto out_unlock;
 859         }
 860         /*
 861          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 862          * under splitting since we don't split the page itself, only pmd to
 863          * a page table.
 864          */
 865         if (is_huge_zero_pmd(pmd)) {
 866                 struct page *zero_page;
 867                 bool set;
 868                 /*
 869                  * get_huge_zero_page() will never allocate a new page here,
 870                  * since we already have a zero page to copy. It just takes a
 871                  * reference.
 872                  */
 873                 zero_page = get_huge_zero_page();
 874                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 875                                 zero_page);
 876                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 877                 ret = 0;
 878                 goto out_unlock;
 879         }
 880         if (unlikely(pmd_trans_splitting(pmd))) {
 881                 /* split huge page running from under us */
 882                 spin_unlock(&src_mm->page_table_lock);
 883                 spin_unlock(&dst_mm->page_table_lock);
 884                 pte_free(dst_mm, pgtable);
 885
 886                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 887                 goto out;
 888         }
 889         src_page = pmd_page(pmd);
 890         VM_BUG_ON(!PageHead(src_page));
 891         get_page(src_page);
 892         page_dup_rmap(src_page);
 893         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 894
 895         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 896         pmd = pmd_mkold(pmd_wrprotect(pmd));
 897         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 898         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 899         atomic_long_inc(&dst_mm->nr_ptes);
 900
 901         ret = 0;
 902 out_unlock:
 903         spin_unlock(&src_mm->page_table_lock);
 904         spin_unlock(&dst_mm->page_table_lock);
 905 out:
 906         return ret;
 907 }
 908
 909 void huge_pmd_set_accessed(struct mm_struct *mm,
 910                            struct vm_area_struct *vma,
 911                            unsigned long address,
 912                            pmd_t *pmd, pmd_t orig_pmd,
 913                            int dirty)
 914 {
 915         pmd_t entry;
 916         unsigned long haddr;
 917
 918         spin_lock(&mm->page_table_lock);
 919         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 920                 goto unlock;
 921
 922         entry = pmd_mkyoung(orig_pmd);
 923         haddr = address & HPAGE_PMD_MASK;
 924         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 925                 update_mmu_cache_pmd(vma, address, pmd);
 926
 927 unlock:
 928         spin_unlock(&mm->page_table_lock);
 929 }
 930
 931 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 932                 struct vm_area_struct *vma, unsigned long address,
 933                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 934 {
 935         pgtable_t pgtable;
 936         pmd_t _pmd;
 937         struct page *page;
 938         int i, ret = 0;
 939         unsigned long mmun_start;       /* For mmu_notifiers */
 940         unsigned long mmun_end;         /* For mmu_notifiers */
 941
 942         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 943         if (!page) {
 944                 ret |= VM_FAULT_OOM;
 945                 goto out;
 946         }
 947
 948         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 949                 put_page(page);
 950                 ret |= VM_FAULT_OOM;
 951                 goto out;
 952         }
 953
 954         clear_user_highpage(page, address);
 955         __SetPageUptodate(page);
 956
 957         mmun_start = haddr;
 958         mmun_end   = haddr + HPAGE_PMD_SIZE;
 959         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 960
 961         spin_lock(&mm->page_table_lock);
 962         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 963                 goto out_free_page;
 964
 965         pmdp_clear_flush(vma, haddr, pmd);
 966         /* leave pmd empty until pte is filled */
 967
 968         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
 969         pmd_populate(mm, &_pmd, pgtable);
 970
 971         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 972                 pte_t *pte, entry;
 973                 if (haddr == (address & PAGE_MASK)) {
 974                         entry = mk_pte(page, vma->vm_page_prot);
 975                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 976                         page_add_new_anon_rmap(page, vma, haddr);
 977                 } else {
 978                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
 979                         entry = pte_mkspecial(entry);
 980                 }
 981                 pte = pte_offset_map(&_pmd, haddr);
 982                 VM_BUG_ON(!pte_none(*pte));
 983                 set_pte_at(mm, haddr, pte, entry);
 984                 pte_unmap(pte);
 985         }
 986         smp_wmb(); /* make pte visible before pmd */
 987         pmd_populate(mm, pmd, pgtable);
 988         spin_unlock(&mm->page_table_lock);
 989         put_huge_zero_page();
 990         inc_mm_counter(mm, MM_ANONPAGES);
 991
 992         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 993
 994         ret |= VM_FAULT_WRITE;
 995 out:
 996         return ret;
 997 out_free_page:
 998         spin_unlock(&mm->page_table_lock);
 999         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1000         mem_cgroup_uncharge_page(page);
1001         put_page(page);
1002         goto out;
1003 }
1004
1005 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1006                                         struct vm_area_struct *vma,
1007                                         unsigned long address,
1008                                         pmd_t *pmd, pmd_t orig_pmd,
1009                                         struct page *page,
1010                                         unsigned long haddr)
1011 {
1012         pgtable_t pgtable;
1013         pmd_t _pmd;
1014         int ret = 0, i;
1015         struct page **pages;
1016         unsigned long mmun_start;       /* For mmu_notifiers */
1017         unsigned long mmun_end;         /* For mmu_notifiers */
1018
1019         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1020                         GFP_KERNEL);
1021         if (unlikely(!pages)) {
1022                 ret |= VM_FAULT_OOM;
1023                 goto out;
1024         }
1025
1026         for (i = 0; i < HPAGE_PMD_NR; i++) {
1027                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1028                                                __GFP_OTHER_NODE,
1029                                                vma, address, page_to_nid(page));
1030                 if (unlikely(!pages[i] ||
1031                              mem_cgroup_newpage_charge(pages[i], mm,
1032                                                        GFP_KERNEL))) {
1033                         if (pages[i])
1034                                 put_page(pages[i]);
1035                         mem_cgroup_uncharge_start();
1036                         while (--i >= 0) {
1037                                 mem_cgroup_uncharge_page(pages[i]);
1038                                 put_page(pages[i]);
1039                         }
1040                         mem_cgroup_uncharge_end();
1041                         kfree(pages);
1042                         ret |= VM_FAULT_OOM;
1043                         goto out;
1044                 }
1045         }
1046
1047         for (i = 0; i < HPAGE_PMD_NR; i++) {
1048                 copy_user_highpage(pages[i], page + i,
1049                                    haddr + PAGE_SIZE * i, vma);
1050                 __SetPageUptodate(pages[i]);
1051                 cond_resched();
1052         }
1053
1054         mmun_start = haddr;
1055         mmun_end   = haddr + HPAGE_PMD_SIZE;
1056         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1057
1058         spin_lock(&mm->page_table_lock);
1059         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1060                 goto out_free_pages;
1061         VM_BUG_ON(!PageHead(page));
1062
1063         pmdp_clear_flush(vma, haddr, pmd);
1064         /* leave pmd empty until pte is filled */
1065
1066         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1067         pmd_populate(mm, &_pmd, pgtable);
1068
1069         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1070                 pte_t *pte, entry;
1071                 entry = mk_pte(pages[i], vma->vm_page_prot);
1072                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1073                 page_add_new_anon_rmap(pages[i], vma, haddr);
1074                 pte = pte_offset_map(&_pmd, haddr);
1075                 VM_BUG_ON(!pte_none(*pte));
1076                 set_pte_at(mm, haddr, pte, entry);
1077                 pte_unmap(pte);
1078         }
1079         kfree(pages);
1080
1081         smp_wmb(); /* make pte visible before pmd */
1082         pmd_populate(mm, pmd, pgtable);
1083         page_remove_rmap(page);
1084         spin_unlock(&mm->page_table_lock);
1085
1086         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1087
1088         ret |= VM_FAULT_WRITE;
1089         put_page(page);
1090
1091 out:
1092         return ret;
1093
1094 out_free_pages:
1095         spin_unlock(&mm->page_table_lock);
1096         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1097         mem_cgroup_uncharge_start();
1098         for (i = 0; i < HPAGE_PMD_NR; i++) {
1099                 mem_cgroup_uncharge_page(pages[i]);
1100                 put_page(pages[i]);
1101         }
1102         mem_cgroup_uncharge_end();
1103         kfree(pages);
1104         goto out;
1105 }
1106
1107 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1108                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1109 {
1110         int ret = 0;
1111         struct page *page = NULL, *new_page;
1112         unsigned long haddr;
1113         unsigned long mmun_start;       /* For mmu_notifiers */
1114         unsigned long mmun_end;         /* For mmu_notifiers */
1115
1116         VM_BUG_ON(!vma->anon_vma);
1117         haddr = address & HPAGE_PMD_MASK;
1118         if (is_huge_zero_pmd(orig_pmd))
1119                 goto alloc;
1120         spin_lock(&mm->page_table_lock);
1121         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1122                 goto out_unlock;
1123
1124         page = pmd_page(orig_pmd);
1125         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1126         if (page_mapcount(page) == 1) {
1127                 pmd_t entry;
1128                 entry = pmd_mkyoung(orig_pmd);
1129                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1130                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1131                         update_mmu_cache_pmd(vma, address, pmd);
1132                 ret |= VM_FAULT_WRITE;
1133                 goto out_unlock;
1134         }
1135         get_page(page);
1136         spin_unlock(&mm->page_table_lock);
1137 alloc:
1138         if (transparent_hugepage_enabled(vma) &&
1139             !transparent_hugepage_debug_cow())
1140                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1141                                               vma, haddr, numa_node_id(), 0);
1142         else
1143                 new_page = NULL;
1144
1145         if (unlikely(!new_page)) {
1146                 if (is_huge_zero_pmd(orig_pmd)) {
1147                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1148                                         address, pmd, orig_pmd, haddr);
1149                 } else {
1150                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1151                                         pmd, orig_pmd, page, haddr);
1152                         if (ret & VM_FAULT_OOM)
1153                                 split_huge_page(page);
1154                         put_page(page);
1155                 }
1156                 count_vm_event(THP_FAULT_FALLBACK);
1157                 goto out;
1158         }
1159
1160         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1161                 put_page(new_page);
1162                 if (page) {
1163                         split_huge_page(page);
1164                         put_page(page);
1165                 }
1166                 count_vm_event(THP_FAULT_FALLBACK);
1167                 ret |= VM_FAULT_OOM;
1168                 goto out;
1169         }
1170
1171         count_vm_event(THP_FAULT_ALLOC);
1172
1173         if (is_huge_zero_pmd(orig_pmd))
1174                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1175         else
1176                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1177         __SetPageUptodate(new_page);
1178
1179         mmun_start = haddr;
1180         mmun_end   = haddr + HPAGE_PMD_SIZE;
1181         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1182
1183         spin_lock(&mm->page_table_lock);
1184         if (page)
1185                 put_page(page);
1186         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1187                 spin_unlock(&mm->page_table_lock);
1188                 mem_cgroup_uncharge_page(new_page);
1189                 put_page(new_page);
1190                 goto out_mn;
1191         } else {
1192                 pmd_t entry;
1193                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1194                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1195                 pmdp_clear_flush(vma, haddr, pmd);
1196                 page_add_new_anon_rmap(new_page, vma, haddr);
1197                 set_pmd_at(mm, haddr, pmd, entry);
1198                 update_mmu_cache_pmd(vma, address, pmd);
1199                 if (is_huge_zero_pmd(orig_pmd)) {
1200                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1201                         put_huge_zero_page();
1202                 } else {
1203                         VM_BUG_ON(!PageHead(page));
1204                         page_remove_rmap(page);
1205                         put_page(page);
1206                 }
1207                 ret |= VM_FAULT_WRITE;
1208         }
1209         spin_unlock(&mm->page_table_lock);
1210 out_mn:
1211         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1212 out:
1213         return ret;
1214 out_unlock:
1215         spin_unlock(&mm->page_table_lock);
1216         return ret;
1217 }
1218
1219 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1220                                    unsigned long addr,
1221                                    pmd_t *pmd,
1222                                    unsigned int flags)
1223 {
1224         struct mm_struct *mm = vma->vm_mm;
1225         struct page *page = NULL;
1226
1227         assert_spin_locked(&mm->page_table_lock);
1228
1229         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1230                 goto out;
1231
1232         /* Avoid dumping huge zero page */
1233         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1234                 return ERR_PTR(-EFAULT);
1235
1236         page = pmd_page(*pmd);
1237         VM_BUG_ON(!PageHead(page));
1238         if (flags & FOLL_TOUCH) {
1239                 pmd_t _pmd;
1240                 /*
1241                  * We should set the dirty bit only for FOLL_WRITE but
1242                  * for now the dirty bit in the pmd is meaningless.
1243                  * And if the dirty bit will become meaningful and
1244                  * we'll only set it with FOLL_WRITE, an atomic
1245                  * set_bit will be required on the pmd to set the
1246                  * young bit, instead of the current set_pmd_at.
1247                  */
1248                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1249                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1250                                           pmd, _pmd,  1))
1251                         update_mmu_cache_pmd(vma, addr, pmd);
1252         }
1253         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1254                 if (page->mapping && trylock_page(page)) {
1255                         lru_add_drain();
1256                         if (page->mapping)
1257                                 mlock_vma_page(page);
1258                         unlock_page(page);
1259                 }
1260         }
1261         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1262         VM_BUG_ON(!PageCompound(page));
1263         if (flags & FOLL_GET)
1264                 get_page_foll(page);
1265
1266 out:
1267         return page;
1268 }
1269
1270 /* NUMA hinting page fault entry point for trans huge pmds */
1271 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1272                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1273 {
1274         struct anon_vma *anon_vma = NULL;
1275         struct page *page;
1276         unsigned long haddr = addr & HPAGE_PMD_MASK;
1277         int page_nid = -1, this_nid = numa_node_id();
1278         int target_nid, last_cpupid = -1;
1279         bool page_locked;
1280         bool migrated = false;
1281         int flags = 0;
1282
1283         spin_lock(&mm->page_table_lock);
1284         if (unlikely(!pmd_same(pmd, *pmdp)))
1285                 goto out_unlock;
1286
1287         page = pmd_page(pmd);
1288         BUG_ON(is_huge_zero_page(page));
1289         page_nid = page_to_nid(page);
1290         last_cpupid = page_cpupid_last(page);
1291         count_vm_numa_event(NUMA_HINT_FAULTS);
1292         if (page_nid == this_nid) {
1293                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1294                 flags |= TNF_FAULT_LOCAL;
1295         }
1296
1297         /*
1298          * Avoid grouping on DSO/COW pages in specific and RO pages
1299          * in general, RO pages shouldn't hurt as much anyway since
1300          * they can be in shared cache state.
1301          */
1302         if (!pmd_write(pmd))
1303                 flags |= TNF_NO_GROUP;
1304
1305         /*
1306          * Acquire the page lock to serialise THP migrations but avoid dropping
1307          * page_table_lock if at all possible
1308          */
1309         page_locked = trylock_page(page);
1310         target_nid = mpol_misplaced(page, vma, haddr);
1311         if (target_nid == -1) {
1312                 /* If the page was locked, there are no parallel migrations */
1313                 if (page_locked)
1314                         goto clear_pmdnuma;
1315
1316                 /*
1317                  * Otherwise wait for potential migrations and retry. We do
1318                  * relock and check_same as the page may no longer be mapped.
1319                  * As the fault is being retried, do not account for it.
1320                  */
1321                 spin_unlock(&mm->page_table_lock);
1322                 wait_on_page_locked(page);
1323                 page_nid = -1;
1324                 goto out;
1325         }
1326
1327         /* Page is misplaced, serialise migrations and parallel THP splits */
1328         get_page(page);
1329         spin_unlock(&mm->page_table_lock);
1330         if (!page_locked)
1331                 lock_page(page);
1332         anon_vma = page_lock_anon_vma_read(page);
1333
1334         /* Confirm the PMD did not change while page_table_lock was released */
1335         spin_lock(&mm->page_table_lock);
1336         if (unlikely(!pmd_same(pmd, *pmdp))) {
1337                 unlock_page(page);
1338                 put_page(page);
1339                 page_nid = -1;
1340                 goto out_unlock;
1341         }
1342
1343         /*
1344          * Migrate the THP to the requested node, returns with page unlocked
1345          * and pmd_numa cleared.
1346          */
1347         spin_unlock(&mm->page_table_lock);
1348         migrated = migrate_misplaced_transhuge_page(mm, vma,
1349                                 pmdp, pmd, addr, page, target_nid);
1350         if (migrated) {
1351                 flags |= TNF_MIGRATED;
1352                 page_nid = target_nid;
1353         }
1354
1355         goto out;
1356 clear_pmdnuma:
1357         BUG_ON(!PageLocked(page));
1358         pmd = pmd_mknonnuma(pmd);
1359         set_pmd_at(mm, haddr, pmdp, pmd);
1360         VM_BUG_ON(pmd_numa(*pmdp));
1361         update_mmu_cache_pmd(vma, addr, pmdp);
1362         unlock_page(page);
1363 out_unlock:
1364         spin_unlock(&mm->page_table_lock);
1365
1366 out:
1367         if (anon_vma)
1368                 page_unlock_anon_vma_read(anon_vma);
1369
1370         if (page_nid != -1)
1371                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1372
1373         return 0;
1374 }
1375
1376 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1377                  pmd_t *pmd, unsigned long addr)
1378 {
1379         spinlock_t *ptl;
1380         int ret = 0;
1381
1382         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1383                 struct page *page;
1384                 pgtable_t pgtable;
1385                 pmd_t orig_pmd;
1386                 /*
1387                  * For architectures like ppc64 we look at deposited pgtable
1388                  * when calling pmdp_get_and_clear. So do the
1389                  * pgtable_trans_huge_withdraw after finishing pmdp related
1390                  * operations.
1391                  */
1392                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1393                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1394                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1395                 if (is_huge_zero_pmd(orig_pmd)) {
1396                         atomic_long_dec(&tlb->mm->nr_ptes);
1397                         spin_unlock(ptl);
1398                         put_huge_zero_page();
1399                 } else {
1400                         page = pmd_page(orig_pmd);
1401                         page_remove_rmap(page);
1402                         VM_BUG_ON(page_mapcount(page) < 0);
1403                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1404                         VM_BUG_ON(!PageHead(page));
1405                         atomic_long_dec(&tlb->mm->nr_ptes);
1406                         spin_unlock(ptl);
1407                         tlb_remove_page(tlb, page);
1408                 }
1409                 pte_free(tlb->mm, pgtable);
1410                 ret = 1;
1411         }
1412         return ret;
1413 }
1414
1415 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1416                 unsigned long addr, unsigned long end,
1417                 unsigned char *vec)
1418 {
1419         spinlock_t *ptl;
1420         int ret = 0;
1421
1422         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1423                 /*
1424                  * All logical pages in the range are present
1425                  * if backed by a huge page.
1426                  */
1427                 spin_unlock(ptl);
1428                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1429                 ret = 1;
1430         }
1431
1432         return ret;
1433 }
1434
1435 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1436                   unsigned long old_addr,
1437                   unsigned long new_addr, unsigned long old_end,
1438                   pmd_t *old_pmd, pmd_t *new_pmd)
1439 {
1440         spinlock_t *old_ptl, *new_ptl;
1441         int ret = 0;
1442         pmd_t pmd;
1443
1444         struct mm_struct *mm = vma->vm_mm;
1445
1446         if ((old_addr & ~HPAGE_PMD_MASK) ||
1447             (new_addr & ~HPAGE_PMD_MASK) ||
1448             old_end - old_addr < HPAGE_PMD_SIZE ||
1449             (new_vma->vm_flags & VM_NOHUGEPAGE))
1450                 goto out;
1451
1452         /*
1453          * The destination pmd shouldn't be established, free_pgtables()
1454          * should have release it.
1455          */
1456         if (WARN_ON(!pmd_none(*new_pmd))) {
1457                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1458                 goto out;
1459         }
1460
1461         /*
1462          * We don't have to worry about the ordering of src and dst
1463          * ptlocks because exclusive mmap_sem prevents deadlock.
1464          */
1465         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1466         if (ret == 1) {
1467                 new_ptl = pmd_lockptr(mm, new_pmd);
1468                 if (new_ptl != old_ptl)
1469                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1470                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1471                 VM_BUG_ON(!pmd_none(*new_pmd));
1472                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1473                 if (new_ptl != old_ptl)
1474                         spin_unlock(new_ptl);
1475                 spin_unlock(old_ptl);
1476         }
1477 out:
1478         return ret;
1479 }
1480
1481 /*
1482  * Returns
1483  *  - 0 if PMD could not be locked
1484  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1485  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1486  */
1487 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1488                 unsigned long addr, pgprot_t newprot, int prot_numa)
1489 {
1490         struct mm_struct *mm = vma->vm_mm;
1491         spinlock_t *ptl;
1492         int ret = 0;
1493
1494         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1495                 pmd_t entry;
1496                 ret = 1;
1497                 if (!prot_numa) {
1498                         entry = pmdp_get_and_clear(mm, addr, pmd);
1499                         entry = pmd_modify(entry, newprot);
1500                         ret = HPAGE_PMD_NR;
1501                         BUG_ON(pmd_write(entry));
1502                 } else {
1503                         struct page *page = pmd_page(*pmd);
1504
1505                         /*
1506                          * Do not trap faults against the zero page. The
1507                          * read-only data is likely to be read-cached on the
1508                          * local CPU cache and it is less useful to know about
1509                          * local vs remote hits on the zero page.
1510                          */
1511                         if (!is_huge_zero_page(page) &&
1512                             !pmd_numa(*pmd)) {
1513                                 entry = pmdp_get_and_clear(mm, addr, pmd);
1514                                 entry = pmd_mknuma(entry);
1515                                 ret = HPAGE_PMD_NR;
1516                         }
1517                 }
1518
1519                 /* Set PMD if cleared earlier */
1520                 if (ret == HPAGE_PMD_NR)
1521                         set_pmd_at(mm, addr, pmd, entry);
1522
1523                 spin_unlock(ptl);
1524         }
1525
1526         return ret;
1527 }
1528
1529 /*
1530  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1531  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1532  *
1533  * Note that if it returns 1, this routine returns without unlocking page
1534  * table locks. So callers must unlock them.
1535  */
1536 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1537                 spinlock_t **ptl)
1538 {
1539         *ptl = pmd_lock(vma->vm_mm, pmd);
1540         if (likely(pmd_trans_huge(*pmd))) {
1541                 if (unlikely(pmd_trans_splitting(*pmd))) {
1542                         spin_unlock(*ptl);
1543                         wait_split_huge_page(vma->anon_vma, pmd);
1544                         return -1;
1545                 } else {
1546                         /* Thp mapped by 'pmd' is stable, so we can
1547                          * handle it as it is. */
1548                         return 1;
1549                 }
1550         }
1551         spin_unlock(*ptl);
1552         return 0;
1553 }
1554
1555 /*
1556  * This function returns whether a given @page is mapped onto the @address
1557  * in the virtual space of @mm.
1558  *
1559  * When it's true, this function returns *pmd with holding the page table lock
1560  * and passing it back to the caller via @ptl.
1561  * If it's false, returns NULL without holding the page table lock.
1562  */
1563 pmd_t *page_check_address_pmd(struct page *page,
1564                               struct mm_struct *mm,
1565                               unsigned long address,
1566                               enum page_check_address_pmd_flag flag,
1567                               spinlock_t **ptl)
1568 {
1569         pmd_t *pmd;
1570
1571         if (address & ~HPAGE_PMD_MASK)
1572                 return NULL;
1573
1574         pmd = mm_find_pmd(mm, address);
1575         if (!pmd)
1576                 return NULL;
1577         *ptl = pmd_lock(mm, pmd);
1578         if (pmd_none(*pmd))
1579                 goto unlock;
1580         if (pmd_page(*pmd) != page)
1581                 goto unlock;
1582         /*
1583          * split_vma() may create temporary aliased mappings. There is
1584          * no risk as long as all huge pmd are found and have their
1585          * splitting bit set before __split_huge_page_refcount
1586          * runs. Finding the same huge pmd more than once during the
1587          * same rmap walk is not a problem.
1588          */
1589         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1590             pmd_trans_splitting(*pmd))
1591                 goto unlock;
1592         if (pmd_trans_huge(*pmd)) {
1593                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1594                           !pmd_trans_splitting(*pmd));
1595                 return pmd;
1596         }
1597 unlock:
1598         spin_unlock(*ptl);
1599         return NULL;
1600 }
1601
1602 static int __split_huge_page_splitting(struct page *page,
1603                                        struct vm_area_struct *vma,
1604                                        unsigned long address)
1605 {
1606         struct mm_struct *mm = vma->vm_mm;
1607         spinlock_t *ptl;
1608         pmd_t *pmd;
1609         int ret = 0;
1610         /* For mmu_notifiers */
1611         const unsigned long mmun_start = address;
1612         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1613
1614         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1615         pmd = page_check_address_pmd(page, mm, address,
1616                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1617         if (pmd) {
1618                 /*
1619                  * We can't temporarily set the pmd to null in order
1620                  * to split it, the pmd must remain marked huge at all
1621                  * times or the VM won't take the pmd_trans_huge paths
1622                  * and it won't wait on the anon_vma->root->rwsem to
1623                  * serialize against split_huge_page*.
1624                  */
1625                 pmdp_splitting_flush(vma, address, pmd);
1626                 ret = 1;
1627                 spin_unlock(ptl);
1628         }
1629         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1630
1631         return ret;
1632 }
1633
1634 static void __split_huge_page_refcount(struct page *page,
1635                                        struct list_head *list)
1636 {
1637         int i;
1638         struct zone *zone = page_zone(page);
1639         struct lruvec *lruvec;
1640         int tail_count = 0;
1641
1642         /* prevent PageLRU to go away from under us, and freeze lru stats */
1643         spin_lock_irq(&zone->lru_lock);
1644         lruvec = mem_cgroup_page_lruvec(page, zone);
1645
1646         compound_lock(page);
1647         /* complete memcg works before add pages to LRU */
1648         mem_cgroup_split_huge_fixup(page);
1649
1650         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1651                 struct page *page_tail = page + i;
1652
1653                 /* tail_page->_mapcount cannot change */
1654                 BUG_ON(page_mapcount(page_tail) < 0);
1655                 tail_count += page_mapcount(page_tail);
1656                 /* check for overflow */
1657                 BUG_ON(tail_count < 0);
1658                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1659                 /*
1660                  * tail_page->_count is zero and not changing from
1661                  * under us. But get_page_unless_zero() may be running
1662                  * from under us on the tail_page. If we used
1663                  * atomic_set() below instead of atomic_add(), we
1664                  * would then run atomic_set() concurrently with
1665                  * get_page_unless_zero(), and atomic_set() is
1666                  * implemented in C not using locked ops. spin_unlock
1667                  * on x86 sometime uses locked ops because of PPro
1668                  * errata 66, 92, so unless somebody can guarantee
1669                  * atomic_set() here would be safe on all archs (and
1670                  * not only on x86), it's safer to use atomic_add().
1671                  */
1672                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1673                            &page_tail->_count);
1674
1675                 /* after clearing PageTail the gup refcount can be released */
1676                 smp_mb();
1677
1678                 /*
1679                  * retain hwpoison flag of the poisoned tail page:
1680                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1681                  *   by the memory-failure.
1682                  */
1683                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1684                 page_tail->flags |= (page->flags &
1685                                      ((1L << PG_referenced) |
1686                                       (1L << PG_swapbacked) |
1687                                       (1L << PG_mlocked) |
1688                                       (1L << PG_uptodate) |
1689                                       (1L << PG_active) |
1690                                       (1L << PG_unevictable)));
1691                 page_tail->flags |= (1L << PG_dirty);
1692
1693                 /* clear PageTail before overwriting first_page */
1694                 smp_wmb();
1695
1696                 /*
1697                  * __split_huge_page_splitting() already set the
1698                  * splitting bit in all pmd that could map this
1699                  * hugepage, that will ensure no CPU can alter the
1700                  * mapcount on the head page. The mapcount is only
1701                  * accounted in the head page and it has to be
1702                  * transferred to all tail pages in the below code. So
1703                  * for this code to be safe, the split the mapcount
1704                  * can't change. But that doesn't mean userland can't
1705                  * keep changing and reading the page contents while
1706                  * we transfer the mapcount, so the pmd splitting
1707                  * status is achieved setting a reserved bit in the
1708                  * pmd, not by clearing the present bit.
1709                 */
1710                 page_tail->_mapcount = page->_mapcount;
1711
1712                 BUG_ON(page_tail->mapping);
1713                 page_tail->mapping = page->mapping;
1714
1715                 page_tail->index = page->index + i;
1716                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1717
1718                 BUG_ON(!PageAnon(page_tail));
1719                 BUG_ON(!PageUptodate(page_tail));
1720                 BUG_ON(!PageDirty(page_tail));
1721                 BUG_ON(!PageSwapBacked(page_tail));
1722
1723                 lru_add_page_tail(page, page_tail, lruvec, list);
1724         }
1725         atomic_sub(tail_count, &page->_count);
1726         BUG_ON(atomic_read(&page->_count) <= 0);
1727
1728         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1729
1730         ClearPageCompound(page);
1731         compound_unlock(page);
1732         spin_unlock_irq(&zone->lru_lock);
1733
1734         for (i = 1; i < HPAGE_PMD_NR; i++) {
1735                 struct page *page_tail = page + i;
1736                 BUG_ON(page_count(page_tail) <= 0);
1737                 /*
1738                  * Tail pages may be freed if there wasn't any mapping
1739                  * like if add_to_swap() is running on a lru page that
1740                  * had its mapping zapped. And freeing these pages
1741                  * requires taking the lru_lock so we do the put_page
1742                  * of the tail pages after the split is complete.
1743                  */
1744                 put_page(page_tail);
1745         }
1746
1747         /*
1748          * Only the head page (now become a regular page) is required
1749          * to be pinned by the caller.
1750          */
1751         BUG_ON(page_count(page) <= 0);
1752 }
1753
1754 static int __split_huge_page_map(struct page *page,
1755                                  struct vm_area_struct *vma,
1756                                  unsigned long address)
1757 {
1758         struct mm_struct *mm = vma->vm_mm;
1759         spinlock_t *ptl;
1760         pmd_t *pmd, _pmd;
1761         int ret = 0, i;
1762         pgtable_t pgtable;
1763         unsigned long haddr;
1764
1765         pmd = page_check_address_pmd(page, mm, address,
1766                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1767         if (pmd) {
1768                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1769                 pmd_populate(mm, &_pmd, pgtable);
1770
1771                 haddr = address;
1772                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1773                         pte_t *pte, entry;
1774                         BUG_ON(PageCompound(page+i));
1775                         entry = mk_pte(page + i, vma->vm_page_prot);
1776                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1777                         if (!pmd_write(*pmd))
1778                                 entry = pte_wrprotect(entry);
1779                         else
1780                                 BUG_ON(page_mapcount(page) != 1);
1781                         if (!pmd_young(*pmd))
1782                                 entry = pte_mkold(entry);
1783                         if (pmd_numa(*pmd))
1784                                 entry = pte_mknuma(entry);
1785                         pte = pte_offset_map(&_pmd, haddr);
1786                         BUG_ON(!pte_none(*pte));
1787                         set_pte_at(mm, haddr, pte, entry);
1788                         pte_unmap(pte);
1789                 }
1790
1791                 smp_wmb(); /* make pte visible before pmd */
1792                 /*
1793                  * Up to this point the pmd is present and huge and
1794                  * userland has the whole access to the hugepage
1795                  * during the split (which happens in place). If we
1796                  * overwrite the pmd with the not-huge version
1797                  * pointing to the pte here (which of course we could
1798                  * if all CPUs were bug free), userland could trigger
1799                  * a small page size TLB miss on the small sized TLB
1800                  * while the hugepage TLB entry is still established
1801                  * in the huge TLB. Some CPU doesn't like that. See
1802                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1803                  * Erratum 383 on page 93. Intel should be safe but is
1804                  * also warns that it's only safe if the permission
1805                  * and cache attributes of the two entries loaded in
1806                  * the two TLB is identical (which should be the case
1807                  * here). But it is generally safer to never allow
1808                  * small and huge TLB entries for the same virtual
1809                  * address to be loaded simultaneously. So instead of
1810                  * doing "pmd_populate(); flush_tlb_range();" we first
1811                  * mark the current pmd notpresent (atomically because
1812                  * here the pmd_trans_huge and pmd_trans_splitting
1813                  * must remain set at all times on the pmd until the
1814                  * split is complete for this pmd), then we flush the
1815                  * SMP TLB and finally we write the non-huge version
1816                  * of the pmd entry with pmd_populate.
1817                  */
1818                 pmdp_invalidate(vma, address, pmd);
1819                 pmd_populate(mm, pmd, pgtable);
1820                 ret = 1;
1821                 spin_unlock(ptl);
1822         }
1823
1824         return ret;
1825 }
1826
1827 /* must be called with anon_vma->root->rwsem held */
1828 static void __split_huge_page(struct page *page,
1829                               struct anon_vma *anon_vma,
1830                               struct list_head *list)
1831 {
1832         int mapcount, mapcount2;
1833         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1834         struct anon_vma_chain *avc;
1835
1836         BUG_ON(!PageHead(page));
1837         BUG_ON(PageTail(page));
1838
1839         mapcount = 0;
1840         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1841                 struct vm_area_struct *vma = avc->vma;
1842                 unsigned long addr = vma_address(page, vma);
1843                 BUG_ON(is_vma_temporary_stack(vma));
1844                 mapcount += __split_huge_page_splitting(page, vma, addr);
1845         }
1846         /*
1847          * It is critical that new vmas are added to the tail of the
1848          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1849          * and establishes a child pmd before
1850          * __split_huge_page_splitting() freezes the parent pmd (so if
1851          * we fail to prevent copy_huge_pmd() from running until the
1852          * whole __split_huge_page() is complete), we will still see
1853          * the newly established pmd of the child later during the
1854          * walk, to be able to set it as pmd_trans_splitting too.
1855          */
1856         if (mapcount != page_mapcount(page))
1857                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1858                        mapcount, page_mapcount(page));
1859         BUG_ON(mapcount != page_mapcount(page));
1860
1861         __split_huge_page_refcount(page, list);
1862
1863         mapcount2 = 0;
1864         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1865                 struct vm_area_struct *vma = avc->vma;
1866                 unsigned long addr = vma_address(page, vma);
1867                 BUG_ON(is_vma_temporary_stack(vma));
1868                 mapcount2 += __split_huge_page_map(page, vma, addr);
1869         }
1870         if (mapcount != mapcount2)
1871                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1872                        mapcount, mapcount2, page_mapcount(page));
1873         BUG_ON(mapcount != mapcount2);
1874 }
1875
1876 /*
1877  * Split a hugepage into normal pages. This doesn't change the position of head
1878  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1879  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1880  * from the hugepage.
1881  * Return 0 if the hugepage is split successfully otherwise return 1.
1882  */
1883 int split_huge_page_to_list(struct page *page, struct list_head *list)
1884 {
1885         struct anon_vma *anon_vma;
1886         int ret = 1;
1887
1888         BUG_ON(is_huge_zero_page(page));
1889         BUG_ON(!PageAnon(page));
1890
1891         /*
1892          * The caller does not necessarily hold an mmap_sem that would prevent
1893          * the anon_vma disappearing so we first we take a reference to it
1894          * and then lock the anon_vma for write. This is similar to
1895          * page_lock_anon_vma_read except the write lock is taken to serialise
1896          * against parallel split or collapse operations.
1897          */
1898         anon_vma = page_get_anon_vma(page);
1899         if (!anon_vma)
1900                 goto out;
1901         anon_vma_lock_write(anon_vma);
1902
1903         ret = 0;
1904         if (!PageCompound(page))
1905                 goto out_unlock;
1906
1907         BUG_ON(!PageSwapBacked(page));
1908         __split_huge_page(page, anon_vma, list);
1909         count_vm_event(THP_SPLIT);
1910
1911         BUG_ON(PageCompound(page));
1912 out_unlock:
1913         anon_vma_unlock_write(anon_vma);
1914         put_anon_vma(anon_vma);
1915 out:
1916         return ret;
1917 }
1918
1919 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1920
1921 int hugepage_madvise(struct vm_area_struct *vma,
1922                      unsigned long *vm_flags, int advice)
1923 {
1924         struct mm_struct *mm = vma->vm_mm;
1925
1926         switch (advice) {
1927         case MADV_HUGEPAGE:
1928                 /*
1929                  * Be somewhat over-protective like KSM for now!
1930                  */
1931                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1932                         return -EINVAL;
1933                 if (mm->def_flags & VM_NOHUGEPAGE)
1934                         return -EINVAL;
1935                 *vm_flags &= ~VM_NOHUGEPAGE;
1936                 *vm_flags |= VM_HUGEPAGE;
1937                 /*
1938                  * If the vma become good for khugepaged to scan,
1939                  * register it here without waiting a page fault that
1940                  * may not happen any time soon.
1941                  */
1942                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1943                         return -ENOMEM;
1944                 break;
1945         case MADV_NOHUGEPAGE:
1946                 /*
1947                  * Be somewhat over-protective like KSM for now!
1948                  */
1949                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1950                         return -EINVAL;
1951                 *vm_flags &= ~VM_HUGEPAGE;
1952                 *vm_flags |= VM_NOHUGEPAGE;
1953                 /*
1954                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1955                  * this vma even if we leave the mm registered in khugepaged if
1956                  * it got registered before VM_NOHUGEPAGE was set.
1957                  */
1958                 break;
1959         }
1960
1961         return 0;
1962 }
1963
1964 static int __init khugepaged_slab_init(void)
1965 {
1966         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1967                                           sizeof(struct mm_slot),
1968                                           __alignof__(struct mm_slot), 0, NULL);
1969         if (!mm_slot_cache)
1970                 return -ENOMEM;
1971
1972         return 0;
1973 }
1974
1975 static inline struct mm_slot *alloc_mm_slot(void)
1976 {
1977         if (!mm_slot_cache)     /* initialization failed */
1978                 return NULL;
1979         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1980 }
1981
1982 static inline void free_mm_slot(struct mm_slot *mm_slot)
1983 {
1984         kmem_cache_free(mm_slot_cache, mm_slot);
1985 }
1986
1987 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1988 {
1989         struct mm_slot *mm_slot;
1990
1991         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1992                 if (mm == mm_slot->mm)
1993                         return mm_slot;
1994
1995         return NULL;
1996 }
1997
1998 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1999                                     struct mm_slot *mm_slot)
2000 {
2001         mm_slot->mm = mm;
2002         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2003 }
2004
2005 static inline int khugepaged_test_exit(struct mm_struct *mm)
2006 {
2007         return atomic_read(&mm->mm_users) == 0;
2008 }
2009
2010 int __khugepaged_enter(struct mm_struct *mm)
2011 {
2012         struct mm_slot *mm_slot;
2013         int wakeup;
2014
2015         mm_slot = alloc_mm_slot();
2016         if (!mm_slot)
2017                 return -ENOMEM;
2018
2019         /* __khugepaged_exit() must not run from under us */
2020         VM_BUG_ON(khugepaged_test_exit(mm));
2021         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2022                 free_mm_slot(mm_slot);
2023                 return 0;
2024         }
2025
2026         spin_lock(&khugepaged_mm_lock);
2027         insert_to_mm_slots_hash(mm, mm_slot);
2028         /*
2029          * Insert just behind the scanning cursor, to let the area settle
2030          * down a little.
2031          */
2032         wakeup = list_empty(&khugepaged_scan.mm_head);
2033         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2034         spin_unlock(&khugepaged_mm_lock);
2035
2036         atomic_inc(&mm->mm_count);
2037         if (wakeup)
2038                 wake_up_interruptible(&khugepaged_wait);
2039
2040         return 0;
2041 }
2042
2043 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2044 {
2045         unsigned long hstart, hend;
2046         if (!vma->anon_vma)
2047                 /*
2048                  * Not yet faulted in so we will register later in the
2049                  * page fault if needed.
2050                  */
2051                 return 0;
2052         if (vma->vm_ops)
2053                 /* khugepaged not yet working on file or special mappings */
2054                 return 0;
2055         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2056         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2057         hend = vma->vm_end & HPAGE_PMD_MASK;
2058         if (hstart < hend)
2059                 return khugepaged_enter(vma);
2060         return 0;
2061 }
2062
2063 void __khugepaged_exit(struct mm_struct *mm)
2064 {
2065         struct mm_slot *mm_slot;
2066         int free = 0;
2067
2068         spin_lock(&khugepaged_mm_lock);
2069         mm_slot = get_mm_slot(mm);
2070         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2071                 hash_del(&mm_slot->hash);
2072                 list_del(&mm_slot->mm_node);
2073                 free = 1;
2074         }
2075         spin_unlock(&khugepaged_mm_lock);
2076
2077         if (free) {
2078                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2079                 free_mm_slot(mm_slot);
2080                 mmdrop(mm);
2081         } else if (mm_slot) {
2082                 /*
2083                  * This is required to serialize against
2084                  * khugepaged_test_exit() (which is guaranteed to run
2085                  * under mmap sem read mode). Stop here (after we
2086                  * return all pagetables will be destroyed) until
2087                  * khugepaged has finished working on the pagetables
2088                  * under the mmap_sem.
2089                  */
2090                 down_write(&mm->mmap_sem);
2091                 up_write(&mm->mmap_sem);
2092         }
2093 }
2094
2095 static void release_pte_page(struct page *page)
2096 {
2097         /* 0 stands for page_is_file_cache(page) == false */
2098         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2099         unlock_page(page);
2100         putback_lru_page(page);
2101 }
2102
2103 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2104 {
2105         while (--_pte >= pte) {
2106                 pte_t pteval = *_pte;
2107                 if (!pte_none(pteval))
2108                         release_pte_page(pte_page(pteval));
2109         }
2110 }
2111
2112 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2113                                         unsigned long address,
2114                                         pte_t *pte)
2115 {
2116         struct page *page;
2117         pte_t *_pte;
2118         int referenced = 0, none = 0;
2119         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2120              _pte++, address += PAGE_SIZE) {
2121                 pte_t pteval = *_pte;
2122                 if (pte_none(pteval)) {
2123                         if (++none <= khugepaged_max_ptes_none)
2124                                 continue;
2125                         else
2126                                 goto out;
2127                 }
2128                 if (!pte_present(pteval) || !pte_write(pteval))
2129                         goto out;
2130                 page = vm_normal_page(vma, address, pteval);
2131                 if (unlikely(!page))
2132                         goto out;
2133
2134                 VM_BUG_ON(PageCompound(page));
2135                 BUG_ON(!PageAnon(page));
2136                 VM_BUG_ON(!PageSwapBacked(page));
2137
2138                 /* cannot use mapcount: can't collapse if there's a gup pin */
2139                 if (page_count(page) != 1)
2140                         goto out;
2141                 /*
2142                  * We can do it before isolate_lru_page because the
2143                  * page can't be freed from under us. NOTE: PG_lock
2144                  * is needed to serialize against split_huge_page
2145                  * when invoked from the VM.
2146                  */
2147                 if (!trylock_page(page))
2148                         goto out;
2149                 /*
2150                  * Isolate the page to avoid collapsing an hugepage
2151                  * currently in use by the VM.
2152                  */
2153                 if (isolate_lru_page(page)) {
2154                         unlock_page(page);
2155                         goto out;
2156                 }
2157                 /* 0 stands for page_is_file_cache(page) == false */
2158                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2159                 VM_BUG_ON(!PageLocked(page));
2160                 VM_BUG_ON(PageLRU(page));
2161
2162                 /* If there is no mapped pte young don't collapse the page */
2163                 if (pte_young(pteval) || PageReferenced(page) ||
2164                     mmu_notifier_test_young(vma->vm_mm, address))
2165                         referenced = 1;
2166         }
2167         if (likely(referenced))
2168                 return 1;
2169 out:
2170         release_pte_pages(pte, _pte);
2171         return 0;
2172 }
2173
2174 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2175                                       struct vm_area_struct *vma,
2176                                       unsigned long address,
2177                                       spinlock_t *ptl)
2178 {
2179         pte_t *_pte;
2180         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2181                 pte_t pteval = *_pte;
2182                 struct page *src_page;
2183
2184                 if (pte_none(pteval)) {
2185                         clear_user_highpage(page, address);
2186                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2187                 } else {
2188                         src_page = pte_page(pteval);
2189                         copy_user_highpage(page, src_page, address, vma);
2190                         VM_BUG_ON(page_mapcount(src_page) != 1);
2191                         release_pte_page(src_page);
2192                         /*
2193                          * ptl mostly unnecessary, but preempt has to
2194                          * be disabled to update the per-cpu stats
2195                          * inside page_remove_rmap().
2196                          */
2197                         spin_lock(ptl);
2198                         /*
2199                          * paravirt calls inside pte_clear here are
2200                          * superfluous.
2201                          */
2202                         pte_clear(vma->vm_mm, address, _pte);
2203                         page_remove_rmap(src_page);
2204                         spin_unlock(ptl);
2205                         free_page_and_swap_cache(src_page);
2206                 }
2207
2208                 address += PAGE_SIZE;
2209                 page++;
2210         }
2211 }
2212
2213 static void khugepaged_alloc_sleep(void)
2214 {
2215         wait_event_freezable_timeout(khugepaged_wait, false,
2216                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2217 }
2218
2219 static int khugepaged_node_load[MAX_NUMNODES];
2220
2221 #ifdef CONFIG_NUMA
2222 static int khugepaged_find_target_node(void)
2223 {
2224         static int last_khugepaged_target_node = NUMA_NO_NODE;
2225         int nid, target_node = 0, max_value = 0;
2226
2227         /* find first node with max normal pages hit */
2228         for (nid = 0; nid < MAX_NUMNODES; nid++)
2229                 if (khugepaged_node_load[nid] > max_value) {
2230                         max_value = khugepaged_node_load[nid];
2231                         target_node = nid;
2232                 }
2233
2234         /* do some balance if several nodes have the same hit record */
2235         if (target_node <= last_khugepaged_target_node)
2236                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2237                                 nid++)
2238                         if (max_value == khugepaged_node_load[nid]) {
2239                                 target_node = nid;
2240                                 break;
2241                         }
2242
2243         last_khugepaged_target_node = target_node;
2244         return target_node;
2245 }
2246
2247 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2248 {
2249         if (IS_ERR(*hpage)) {
2250                 if (!*wait)
2251                         return false;
2252
2253                 *wait = false;
2254                 *hpage = NULL;
2255                 khugepaged_alloc_sleep();
2256         } else if (*hpage) {
2257                 put_page(*hpage);
2258                 *hpage = NULL;
2259         }
2260
2261         return true;
2262 }
2263
2264 static struct page
2265 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2266                        struct vm_area_struct *vma, unsigned long address,
2267                        int node)
2268 {
2269         VM_BUG_ON(*hpage);
2270         /*
2271          * Allocate the page while the vma is still valid and under
2272          * the mmap_sem read mode so there is no memory allocation
2273          * later when we take the mmap_sem in write mode. This is more
2274          * friendly behavior (OTOH it may actually hide bugs) to
2275          * filesystems in userland with daemons allocating memory in
2276          * the userland I/O paths.  Allocating memory with the
2277          * mmap_sem in read mode is good idea also to allow greater
2278          * scalability.
2279          */
2280         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2281                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2282         /*
2283          * After allocating the hugepage, release the mmap_sem read lock in
2284          * preparation for taking it in write mode.
2285          */
2286         up_read(&mm->mmap_sem);
2287         if (unlikely(!*hpage)) {
2288                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2289                 *hpage = ERR_PTR(-ENOMEM);
2290                 return NULL;
2291         }
2292
2293         count_vm_event(THP_COLLAPSE_ALLOC);
2294         return *hpage;
2295 }
2296 #else
2297 static int khugepaged_find_target_node(void)
2298 {
2299         return 0;
2300 }
2301
2302 static inline struct page *alloc_hugepage(int defrag)
2303 {
2304         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2305                            HPAGE_PMD_ORDER);
2306 }
2307
2308 static struct page *khugepaged_alloc_hugepage(bool *wait)
2309 {
2310         struct page *hpage;
2311
2312         do {
2313                 hpage = alloc_hugepage(khugepaged_defrag());
2314                 if (!hpage) {
2315                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2316                         if (!*wait)
2317                                 return NULL;
2318
2319                         *wait = false;
2320                         khugepaged_alloc_sleep();
2321                 } else
2322                         count_vm_event(THP_COLLAPSE_ALLOC);
2323         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2324
2325         return hpage;
2326 }
2327
2328 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2329 {
2330         if (!*hpage)
2331                 *hpage = khugepaged_alloc_hugepage(wait);
2332
2333         if (unlikely(!*hpage))
2334                 return false;
2335
2336         return true;
2337 }
2338
2339 static struct page
2340 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2341                        struct vm_area_struct *vma, unsigned long address,
2342                        int node)
2343 {
2344         up_read(&mm->mmap_sem);
2345         VM_BUG_ON(!*hpage);
2346         return  *hpage;
2347 }
2348 #endif
2349
2350 static bool hugepage_vma_check(struct vm_area_struct *vma)
2351 {
2352         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2353             (vma->vm_flags & VM_NOHUGEPAGE))
2354                 return false;
2355
2356         if (!vma->anon_vma || vma->vm_ops)
2357                 return false;
2358         if (is_vma_temporary_stack(vma))
2359                 return false;
2360         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2361         return true;
2362 }
2363
2364 static void collapse_huge_page(struct mm_struct *mm,
2365                                    unsigned long address,
2366                                    struct page **hpage,
2367                                    struct vm_area_struct *vma,
2368                                    int node)
2369 {
2370         pmd_t *pmd, _pmd;
2371         pte_t *pte;
2372         pgtable_t pgtable;
2373         struct page *new_page;
2374         spinlock_t *ptl;
2375         int isolated;
2376         unsigned long hstart, hend;
2377         unsigned long mmun_start;       /* For mmu_notifiers */
2378         unsigned long mmun_end;         /* For mmu_notifiers */
2379
2380         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2381
2382         /* release the mmap_sem read lock. */
2383         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2384         if (!new_page)
2385                 return;
2386
2387         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2388                 return;
2389
2390         /*
2391          * Prevent all access to pagetables with the exception of
2392          * gup_fast later hanlded by the ptep_clear_flush and the VM
2393          * handled by the anon_vma lock + PG_lock.
2394          */
2395         down_write(&mm->mmap_sem);
2396         if (unlikely(khugepaged_test_exit(mm)))
2397                 goto out;
2398
2399         vma = find_vma(mm, address);
2400         if (!vma)
2401                 goto out;
2402         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2403         hend = vma->vm_end & HPAGE_PMD_MASK;
2404         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2405                 goto out;
2406         if (!hugepage_vma_check(vma))
2407                 goto out;
2408         pmd = mm_find_pmd(mm, address);
2409         if (!pmd)
2410                 goto out;
2411         if (pmd_trans_huge(*pmd))
2412                 goto out;
2413
2414         anon_vma_lock_write(vma->anon_vma);
2415
2416         pte = pte_offset_map(pmd, address);
2417         ptl = pte_lockptr(mm, pmd);
2418
2419         mmun_start = address;
2420         mmun_end   = address + HPAGE_PMD_SIZE;
2421         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2422         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2423         /*
2424          * After this gup_fast can't run anymore. This also removes
2425          * any huge TLB entry from the CPU so we won't allow
2426          * huge and small TLB entries for the same virtual address
2427          * to avoid the risk of CPU bugs in that area.
2428          */
2429         _pmd = pmdp_clear_flush(vma, address, pmd);
2430         spin_unlock(&mm->page_table_lock);
2431         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2432
2433         spin_lock(ptl);
2434         isolated = __collapse_huge_page_isolate(vma, address, pte);
2435         spin_unlock(ptl);
2436
2437         if (unlikely(!isolated)) {
2438                 pte_unmap(pte);
2439                 spin_lock(&mm->page_table_lock);
2440                 BUG_ON(!pmd_none(*pmd));
2441                 /*
2442                  * We can only use set_pmd_at when establishing
2443                  * hugepmds and never for establishing regular pmds that
2444                  * points to regular pagetables. Use pmd_populate for that
2445                  */
2446                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2447                 spin_unlock(&mm->page_table_lock);
2448                 anon_vma_unlock_write(vma->anon_vma);
2449                 goto out;
2450         }
2451
2452         /*
2453          * All pages are isolated and locked so anon_vma rmap
2454          * can't run anymore.
2455          */
2456         anon_vma_unlock_write(vma->anon_vma);
2457
2458         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2459         pte_unmap(pte);
2460         __SetPageUptodate(new_page);
2461         pgtable = pmd_pgtable(_pmd);
2462
2463         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2464         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2465
2466         /*
2467          * spin_lock() below is not the equivalent of smp_wmb(), so
2468          * this is needed to avoid the copy_huge_page writes to become
2469          * visible after the set_pmd_at() write.
2470          */
2471         smp_wmb();
2472
2473         spin_lock(&mm->page_table_lock);
2474         BUG_ON(!pmd_none(*pmd));
2475         page_add_new_anon_rmap(new_page, vma, address);
2476         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2477         set_pmd_at(mm, address, pmd, _pmd);
2478         update_mmu_cache_pmd(vma, address, pmd);
2479         spin_unlock(&mm->page_table_lock);
2480
2481         *hpage = NULL;
2482
2483         khugepaged_pages_collapsed++;
2484 out_up_write:
2485         up_write(&mm->mmap_sem);
2486         return;
2487
2488 out:
2489         mem_cgroup_uncharge_page(new_page);
2490         goto out_up_write;
2491 }
2492
2493 static int khugepaged_scan_pmd(struct mm_struct *mm,
2494                                struct vm_area_struct *vma,
2495                                unsigned long address,
2496                                struct page **hpage)
2497 {
2498         pmd_t *pmd;
2499         pte_t *pte, *_pte;
2500         int ret = 0, referenced = 0, none = 0;
2501         struct page *page;
2502         unsigned long _address;
2503         spinlock_t *ptl;
2504         int node = NUMA_NO_NODE;
2505
2506         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2507
2508         pmd = mm_find_pmd(mm, address);
2509         if (!pmd)
2510                 goto out;
2511         if (pmd_trans_huge(*pmd))
2512                 goto out;
2513
2514         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2515         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2516         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2517              _pte++, _address += PAGE_SIZE) {
2518                 pte_t pteval = *_pte;
2519                 if (pte_none(pteval)) {
2520                         if (++none <= khugepaged_max_ptes_none)
2521                                 continue;
2522                         else
2523                                 goto out_unmap;
2524                 }
2525                 if (!pte_present(pteval) || !pte_write(pteval))
2526                         goto out_unmap;
2527                 page = vm_normal_page(vma, _address, pteval);
2528                 if (unlikely(!page))
2529                         goto out_unmap;
2530                 /*
2531                  * Record which node the original page is from and save this
2532                  * information to khugepaged_node_load[].
2533                  * Khupaged will allocate hugepage from the node has the max
2534                  * hit record.
2535                  */
2536                 node = page_to_nid(page);
2537                 khugepaged_node_load[node]++;
2538                 VM_BUG_ON(PageCompound(page));
2539                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2540                         goto out_unmap;
2541                 /* cannot use mapcount: can't collapse if there's a gup pin */
2542                 if (page_count(page) != 1)
2543                         goto out_unmap;
2544                 if (pte_young(pteval) || PageReferenced(page) ||
2545                     mmu_notifier_test_young(vma->vm_mm, address))
2546                         referenced = 1;
2547         }
2548         if (referenced)
2549                 ret = 1;
2550 out_unmap:
2551         pte_unmap_unlock(pte, ptl);
2552         if (ret) {
2553                 node = khugepaged_find_target_node();
2554                 /* collapse_huge_page will return with the mmap_sem released */
2555                 collapse_huge_page(mm, address, hpage, vma, node);
2556         }
2557 out:
2558         return ret;
2559 }
2560
2561 static void collect_mm_slot(struct mm_slot *mm_slot)
2562 {
2563         struct mm_struct *mm = mm_slot->mm;
2564
2565         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2566
2567         if (khugepaged_test_exit(mm)) {
2568                 /* free mm_slot */
2569                 hash_del(&mm_slot->hash);
2570                 list_del(&mm_slot->mm_node);
2571
2572                 /*
2573                  * Not strictly needed because the mm exited already.
2574                  *
2575                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2576                  */
2577
2578                 /* khugepaged_mm_lock actually not necessary for the below */
2579                 free_mm_slot(mm_slot);
2580                 mmdrop(mm);
2581         }
2582 }
2583
2584 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2585                                             struct page **hpage)
2586         __releases(&khugepaged_mm_lock)
2587         __acquires(&khugepaged_mm_lock)
2588 {
2589         struct mm_slot *mm_slot;
2590         struct mm_struct *mm;
2591         struct vm_area_struct *vma;
2592         int progress = 0;
2593
2594         VM_BUG_ON(!pages);
2595         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2596
2597         if (khugepaged_scan.mm_slot)
2598                 mm_slot = khugepaged_scan.mm_slot;
2599         else {
2600                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2601                                      struct mm_slot, mm_node);
2602                 khugepaged_scan.address = 0;
2603                 khugepaged_scan.mm_slot = mm_slot;
2604         }
2605         spin_unlock(&khugepaged_mm_lock);
2606
2607         mm = mm_slot->mm;
2608         down_read(&mm->mmap_sem);
2609         if (unlikely(khugepaged_test_exit(mm)))
2610                 vma = NULL;
2611         else
2612                 vma = find_vma(mm, khugepaged_scan.address);
2613
2614         progress++;
2615         for (; vma; vma = vma->vm_next) {
2616                 unsigned long hstart, hend;
2617
2618                 cond_resched();
2619                 if (unlikely(khugepaged_test_exit(mm))) {
2620                         progress++;
2621                         break;
2622                 }
2623                 if (!hugepage_vma_check(vma)) {
2624 skip:
2625                         progress++;
2626                         continue;
2627                 }
2628                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2629                 hend = vma->vm_end & HPAGE_PMD_MASK;
2630                 if (hstart >= hend)
2631                         goto skip;
2632                 if (khugepaged_scan.address > hend)
2633                         goto skip;
2634                 if (khugepaged_scan.address < hstart)
2635                         khugepaged_scan.address = hstart;
2636                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2637
2638                 while (khugepaged_scan.address < hend) {
2639                         int ret;
2640                         cond_resched();
2641                         if (unlikely(khugepaged_test_exit(mm)))
2642                                 goto breakouterloop;
2643
2644                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2645                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2646                                   hend);
2647                         ret = khugepaged_scan_pmd(mm, vma,
2648                                                   khugepaged_scan.address,
2649                                                   hpage);
2650                         /* move to next address */
2651                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2652                         progress += HPAGE_PMD_NR;
2653                         if (ret)
2654                                 /* we released mmap_sem so break loop */
2655                                 goto breakouterloop_mmap_sem;
2656                         if (progress >= pages)
2657                                 goto breakouterloop;
2658                 }
2659         }
2660 breakouterloop:
2661         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2662 breakouterloop_mmap_sem:
2663
2664         spin_lock(&khugepaged_mm_lock);
2665         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2666         /*
2667          * Release the current mm_slot if this mm is about to die, or
2668          * if we scanned all vmas of this mm.
2669          */
2670         if (khugepaged_test_exit(mm) || !vma) {
2671                 /*
2672                  * Make sure that if mm_users is reaching zero while
2673                  * khugepaged runs here, khugepaged_exit will find
2674                  * mm_slot not pointing to the exiting mm.
2675                  */
2676                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2677                         khugepaged_scan.mm_slot = list_entry(
2678                                 mm_slot->mm_node.next,
2679                                 struct mm_slot, mm_node);
2680                         khugepaged_scan.address = 0;
2681                 } else {
2682                         khugepaged_scan.mm_slot = NULL;
2683                         khugepaged_full_scans++;
2684                 }
2685
2686                 collect_mm_slot(mm_slot);
2687         }
2688
2689         return progress;
2690 }
2691
2692 static int khugepaged_has_work(void)
2693 {
2694         return !list_empty(&khugepaged_scan.mm_head) &&
2695                 khugepaged_enabled();
2696 }
2697
2698 static int khugepaged_wait_event(void)
2699 {
2700         return !list_empty(&khugepaged_scan.mm_head) ||
2701                 kthread_should_stop();
2702 }
2703
2704 static void khugepaged_do_scan(void)
2705 {
2706         struct page *hpage = NULL;
2707         unsigned int progress = 0, pass_through_head = 0;
2708         unsigned int pages = khugepaged_pages_to_scan;
2709         bool wait = true;
2710
2711         barrier(); /* write khugepaged_pages_to_scan to local stack */
2712
2713         while (progress < pages) {
2714                 if (!khugepaged_prealloc_page(&hpage, &wait))
2715                         break;
2716
2717                 cond_resched();
2718
2719                 if (unlikely(kthread_should_stop() || freezing(current)))
2720                         break;
2721
2722                 spin_lock(&khugepaged_mm_lock);
2723                 if (!khugepaged_scan.mm_slot)
2724                         pass_through_head++;
2725                 if (khugepaged_has_work() &&
2726                     pass_through_head < 2)
2727                         progress += khugepaged_scan_mm_slot(pages - progress,
2728                                                             &hpage);
2729                 else
2730                         progress = pages;
2731                 spin_unlock(&khugepaged_mm_lock);
2732         }
2733
2734         if (!IS_ERR_OR_NULL(hpage))
2735                 put_page(hpage);
2736 }
2737
2738 static void khugepaged_wait_work(void)
2739 {
2740         try_to_freeze();
2741
2742         if (khugepaged_has_work()) {
2743                 if (!khugepaged_scan_sleep_millisecs)
2744                         return;
2745
2746                 wait_event_freezable_timeout(khugepaged_wait,
2747                                              kthread_should_stop(),
2748                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2749                 return;
2750         }
2751
2752         if (khugepaged_enabled())
2753                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2754 }
2755
2756 static int khugepaged(void *none)
2757 {
2758         struct mm_slot *mm_slot;
2759
2760         set_freezable();
2761         set_user_nice(current, 19);
2762
2763         while (!kthread_should_stop()) {
2764                 khugepaged_do_scan();
2765                 khugepaged_wait_work();
2766         }
2767
2768         spin_lock(&khugepaged_mm_lock);
2769         mm_slot = khugepaged_scan.mm_slot;
2770         khugepaged_scan.mm_slot = NULL;
2771         if (mm_slot)
2772                 collect_mm_slot(mm_slot);
2773         spin_unlock(&khugepaged_mm_lock);
2774         return 0;
2775 }
2776
2777 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2778                 unsigned long haddr, pmd_t *pmd)
2779 {
2780         struct mm_struct *mm = vma->vm_mm;
2781         pgtable_t pgtable;
2782         pmd_t _pmd;
2783         int i;
2784
2785         pmdp_clear_flush(vma, haddr, pmd);
2786         /* leave pmd empty until pte is filled */
2787
2788         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2789         pmd_populate(mm, &_pmd, pgtable);
2790
2791         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2792                 pte_t *pte, entry;
2793                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2794                 entry = pte_mkspecial(entry);
2795                 pte = pte_offset_map(&_pmd, haddr);
2796                 VM_BUG_ON(!pte_none(*pte));
2797                 set_pte_at(mm, haddr, pte, entry);
2798                 pte_unmap(pte);
2799         }
2800         smp_wmb(); /* make pte visible before pmd */
2801         pmd_populate(mm, pmd, pgtable);
2802         put_huge_zero_page();
2803 }
2804
2805 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2806                 pmd_t *pmd)
2807 {
2808         struct page *page;
2809         struct mm_struct *mm = vma->vm_mm;
2810         unsigned long haddr = address & HPAGE_PMD_MASK;
2811         unsigned long mmun_start;       /* For mmu_notifiers */
2812         unsigned long mmun_end;         /* For mmu_notifiers */
2813
2814         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2815
2816         mmun_start = haddr;
2817         mmun_end   = haddr + HPAGE_PMD_SIZE;
2818 again:
2819         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2820         spin_lock(&mm->page_table_lock);
2821         if (unlikely(!pmd_trans_huge(*pmd))) {
2822                 spin_unlock(&mm->page_table_lock);
2823                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2824                 return;
2825         }
2826         if (is_huge_zero_pmd(*pmd)) {
2827                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2828                 spin_unlock(&mm->page_table_lock);
2829                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2830                 return;
2831         }
2832         page = pmd_page(*pmd);
2833         VM_BUG_ON(!page_count(page));
2834         get_page(page);
2835         spin_unlock(&mm->page_table_lock);
2836         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2837
2838         split_huge_page(page);
2839
2840         put_page(page);
2841
2842         /*
2843          * We don't always have down_write of mmap_sem here: a racing
2844          * do_huge_pmd_wp_page() might have copied-on-write to another
2845          * huge page before our split_huge_page() got the anon_vma lock.
2846          */
2847         if (unlikely(pmd_trans_huge(*pmd)))
2848                 goto again;
2849 }
2850
2851 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2852                 pmd_t *pmd)
2853 {
2854         struct vm_area_struct *vma;
2855
2856         vma = find_vma(mm, address);
2857         BUG_ON(vma == NULL);
2858         split_huge_page_pmd(vma, address, pmd);
2859 }
2860
2861 static void split_huge_page_address(struct mm_struct *mm,
2862                                     unsigned long address)
2863 {
2864         pmd_t *pmd;
2865
2866         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2867
2868         pmd = mm_find_pmd(mm, address);
2869         if (!pmd)
2870                 return;
2871         /*
2872          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2873          * materialize from under us.
2874          */
2875         split_huge_page_pmd_mm(mm, address, pmd);
2876 }
2877
2878 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2879                              unsigned long start,
2880                              unsigned long end,
2881                              long adjust_next)
2882 {
2883         /*
2884          * If the new start address isn't hpage aligned and it could
2885          * previously contain an hugepage: check if we need to split
2886          * an huge pmd.
2887          */
2888         if (start & ~HPAGE_PMD_MASK &&
2889             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2890             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2891                 split_huge_page_address(vma->vm_mm, start);
2892
2893         /*
2894          * If the new end address isn't hpage aligned and it could
2895          * previously contain an hugepage: check if we need to split
2896          * an huge pmd.
2897          */
2898         if (end & ~HPAGE_PMD_MASK &&
2899             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2900             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2901                 split_huge_page_address(vma->vm_mm, end);
2902
2903         /*
2904          * If we're also updating the vma->vm_next->vm_start, if the new
2905          * vm_next->vm_start isn't page aligned and it could previously
2906          * contain an hugepage: check if we need to split an huge pmd.
2907          */
2908         if (adjust_next > 0) {
2909                 struct vm_area_struct *next = vma->vm_next;
2910                 unsigned long nstart = next->vm_start;
2911                 nstart += adjust_next << PAGE_SHIFT;
2912                 if (nstart & ~HPAGE_PMD_MASK &&
2913                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2914                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2915                         split_huge_page_address(next->vm_mm, nstart);
2916         }
2917 }