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