mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is disabled in order that avoid
  31  * to risk increase the memory footprint of applications without a guaranteed
  32  * benefit. When transparent hugepage support is enabled, is for all mappings,
  33  * and khugepaged scans all mappings.
  34  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  35  * for all hugepage allocations.
  36  */
  37 unsigned long transparent_hugepage_flags __read_mostly =
  38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  39         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  40 #endif
  41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  42         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  43 #endif
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  46         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  47
  48 /* default scan 8*512 pte (or vmas) every 30 second */
  49 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  50 static unsigned int khugepaged_pages_collapsed;
  51 static unsigned int khugepaged_full_scans;
  52 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  53 /* during fragmentation poll the hugepage allocator once every minute */
  54 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  55 static struct task_struct *khugepaged_thread __read_mostly;
  56 static DEFINE_MUTEX(khugepaged_mutex);
  57 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  58 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  59 /*
  60  * default collapse hugepages if there is at least one pte mapped like
  61  * it would have happened if the vma was large enough during page
  62  * fault.
  63  */
  64 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  65
  66 static int khugepaged(void *none);
  67 static int khugepaged_slab_init(void);
  68
  69 #define MM_SLOTS_HASH_BITS 10
  70 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  71
  72 static struct kmem_cache *mm_slot_cache __read_mostly;
  73
  74 /**
  75  * struct mm_slot - hash lookup from mm to mm_slot
  76  * @hash: hash collision list
  77  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  78  * @mm: the mm that this information is valid for
  79  */
  80 struct mm_slot {
  81         struct hlist_node hash;
  82         struct list_head mm_node;
  83         struct mm_struct *mm;
  84 };
  85
  86 /**
  87  * struct khugepaged_scan - cursor for scanning
  88  * @mm_head: the head of the mm list to scan
  89  * @mm_slot: the current mm_slot we are scanning
  90  * @address: the next address inside that to be scanned
  91  *
  92  * There is only the one khugepaged_scan instance of this cursor structure.
  93  */
  94 struct khugepaged_scan {
  95         struct list_head mm_head;
  96         struct mm_slot *mm_slot;
  97         unsigned long address;
  98 };
  99 static struct khugepaged_scan khugepaged_scan = {
 100         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 101 };
 102
 103
 104 static int set_recommended_min_free_kbytes(void)
 105 {
 106         struct zone *zone;
 107         int nr_zones = 0;
 108         unsigned long recommended_min;
 109
 110         if (!khugepaged_enabled())
 111                 return 0;
 112
 113         for_each_populated_zone(zone)
 114                 nr_zones++;
 115
 116         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 117         recommended_min = pageblock_nr_pages * nr_zones * 2;
 118
 119         /*
 120          * Make sure that on average at least two pageblocks are almost free
 121          * of another type, one for a migratetype to fall back to and a
 122          * second to avoid subsequent fallbacks of other types There are 3
 123          * MIGRATE_TYPES we care about.
 124          */
 125         recommended_min += pageblock_nr_pages * nr_zones *
 126                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 127
 128         /* don't ever allow to reserve more than 5% of the lowmem */
 129         recommended_min = min(recommended_min,
 130                               (unsigned long) nr_free_buffer_pages() / 20);
 131         recommended_min <<= (PAGE_SHIFT-10);
 132
 133         if (recommended_min > min_free_kbytes)
 134                 min_free_kbytes = recommended_min;
 135         setup_per_zone_wmarks();
 136         return 0;
 137 }
 138 late_initcall(set_recommended_min_free_kbytes);
 139
 140 static int start_khugepaged(void)
 141 {
 142         int err = 0;
 143         if (khugepaged_enabled()) {
 144                 if (!khugepaged_thread)
 145                         khugepaged_thread = kthread_run(khugepaged, NULL,
 146                                                         "khugepaged");
 147                 if (unlikely(IS_ERR(khugepaged_thread))) {
 148                         printk(KERN_ERR
 149                                "khugepaged: kthread_run(khugepaged) failed\n");
 150                         err = PTR_ERR(khugepaged_thread);
 151                         khugepaged_thread = NULL;
 152                 }
 153
 154                 if (!list_empty(&khugepaged_scan.mm_head))
 155                         wake_up_interruptible(&khugepaged_wait);
 156
 157                 set_recommended_min_free_kbytes();
 158         } else if (khugepaged_thread) {
 159                 kthread_stop(khugepaged_thread);
 160                 khugepaged_thread = NULL;
 161         }
 162
 163         return err;
 164 }
 165
 166 static atomic_t huge_zero_refcount;
 167 static struct page *huge_zero_page __read_mostly;
 168
 169 static inline bool is_huge_zero_page(struct page *page)
 170 {
 171         return ACCESS_ONCE(huge_zero_page) == page;
 172 }
 173
 174 static inline bool is_huge_zero_pmd(pmd_t pmd)
 175 {
 176         return is_huge_zero_page(pmd_page(pmd));
 177 }
 178
 179 static struct page *get_huge_zero_page(void)
 180 {
 181         struct page *zero_page;
 182 retry:
 183         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 184                 return ACCESS_ONCE(huge_zero_page);
 185
 186         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 187                         HPAGE_PMD_ORDER);
 188         if (!zero_page) {
 189                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 190                 return NULL;
 191         }
 192         count_vm_event(THP_ZERO_PAGE_ALLOC);
 193         preempt_disable();
 194         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 195                 preempt_enable();
 196                 __free_page(zero_page);
 197                 goto retry;
 198         }
 199
 200         /* We take additional reference here. It will be put back by shrinker */
 201         atomic_set(&huge_zero_refcount, 2);
 202         preempt_enable();
 203         return ACCESS_ONCE(huge_zero_page);
 204 }
 205
 206 static void put_huge_zero_page(void)
 207 {
 208         /*
 209          * Counter should never go to zero here. Only shrinker can put
 210          * last reference.
 211          */
 212         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 213 }
 214
 215 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 216                                         struct shrink_control *sc)
 217 {
 218         /* we can free zero page only if last reference remains */
 219         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 220 }
 221
 222 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 223                                        struct shrink_control *sc)
 224 {
 225         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 226                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 227                 BUG_ON(zero_page == NULL);
 228                 __free_page(zero_page);
 229                 return HPAGE_PMD_NR;
 230         }
 231
 232         return 0;
 233 }
 234
 235 static struct shrinker huge_zero_page_shrinker = {
 236         .count_objects = shrink_huge_zero_page_count,
 237         .scan_objects = shrink_huge_zero_page_scan,
 238         .seeks = DEFAULT_SEEKS,
 239 };
 240
 241 #ifdef CONFIG_SYSFS
 242
 243 static ssize_t double_flag_show(struct kobject *kobj,
 244                                 struct kobj_attribute *attr, char *buf,
 245                                 enum transparent_hugepage_flag enabled,
 246                                 enum transparent_hugepage_flag req_madv)
 247 {
 248         if (test_bit(enabled, &transparent_hugepage_flags)) {
 249                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 250                 return sprintf(buf, "[always] madvise never\n");
 251         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 252                 return sprintf(buf, "always [madvise] never\n");
 253         else
 254                 return sprintf(buf, "always madvise [never]\n");
 255 }
 256 static ssize_t double_flag_store(struct kobject *kobj,
 257                                  struct kobj_attribute *attr,
 258                                  const char *buf, size_t count,
 259                                  enum transparent_hugepage_flag enabled,
 260                                  enum transparent_hugepage_flag req_madv)
 261 {
 262         if (!memcmp("always", buf,
 263                     min(sizeof("always")-1, count))) {
 264                 set_bit(enabled, &transparent_hugepage_flags);
 265                 clear_bit(req_madv, &transparent_hugepage_flags);
 266         } else if (!memcmp("madvise", buf,
 267                            min(sizeof("madvise")-1, count))) {
 268                 clear_bit(enabled, &transparent_hugepage_flags);
 269                 set_bit(req_madv, &transparent_hugepage_flags);
 270         } else if (!memcmp("never", buf,
 271                            min(sizeof("never")-1, count))) {
 272                 clear_bit(enabled, &transparent_hugepage_flags);
 273                 clear_bit(req_madv, &transparent_hugepage_flags);
 274         } else
 275                 return -EINVAL;
 276
 277         return count;
 278 }
 279
 280 static ssize_t enabled_show(struct kobject *kobj,
 281                             struct kobj_attribute *attr, char *buf)
 282 {
 283         return double_flag_show(kobj, attr, buf,
 284                                 TRANSPARENT_HUGEPAGE_FLAG,
 285                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 286 }
 287 static ssize_t enabled_store(struct kobject *kobj,
 288                              struct kobj_attribute *attr,
 289                              const char *buf, size_t count)
 290 {
 291         ssize_t ret;
 292
 293         ret = double_flag_store(kobj, attr, buf, count,
 294                                 TRANSPARENT_HUGEPAGE_FLAG,
 295                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 296
 297         if (ret > 0) {
 298                 int err;
 299
 300                 mutex_lock(&khugepaged_mutex);
 301                 err = start_khugepaged();
 302                 mutex_unlock(&khugepaged_mutex);
 303
 304                 if (err)
 305                         ret = err;
 306         }
 307
 308         return ret;
 309 }
 310 static struct kobj_attribute enabled_attr =
 311         __ATTR(enabled, 0644, enabled_show, enabled_store);
 312
 313 static ssize_t single_flag_show(struct kobject *kobj,
 314                                 struct kobj_attribute *attr, char *buf,
 315                                 enum transparent_hugepage_flag flag)
 316 {
 317         return sprintf(buf, "%d\n",
 318                        !!test_bit(flag, &transparent_hugepage_flags));
 319 }
 320
 321 static ssize_t single_flag_store(struct kobject *kobj,
 322                                  struct kobj_attribute *attr,
 323                                  const char *buf, size_t count,
 324                                  enum transparent_hugepage_flag flag)
 325 {
 326         unsigned long value;
 327         int ret;
 328
 329         ret = kstrtoul(buf, 10, &value);
 330         if (ret < 0)
 331                 return ret;
 332         if (value > 1)
 333                 return -EINVAL;
 334
 335         if (value)
 336                 set_bit(flag, &transparent_hugepage_flags);
 337         else
 338                 clear_bit(flag, &transparent_hugepage_flags);
 339
 340         return count;
 341 }
 342
 343 /*
 344  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 345  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 346  * memory just to allocate one more hugepage.
 347  */
 348 static ssize_t defrag_show(struct kobject *kobj,
 349                            struct kobj_attribute *attr, char *buf)
 350 {
 351         return double_flag_show(kobj, attr, buf,
 352                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 353                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 354 }
 355 static ssize_t defrag_store(struct kobject *kobj,
 356                             struct kobj_attribute *attr,
 357                             const char *buf, size_t count)
 358 {
 359         return double_flag_store(kobj, attr, buf, count,
 360                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 361                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 362 }
 363 static struct kobj_attribute defrag_attr =
 364         __ATTR(defrag, 0644, defrag_show, defrag_store);
 365
 366 static ssize_t use_zero_page_show(struct kobject *kobj,
 367                 struct kobj_attribute *attr, char *buf)
 368 {
 369         return single_flag_show(kobj, attr, buf,
 370                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 371 }
 372 static ssize_t use_zero_page_store(struct kobject *kobj,
 373                 struct kobj_attribute *attr, const char *buf, size_t count)
 374 {
 375         return single_flag_store(kobj, attr, buf, count,
 376                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static struct kobj_attribute use_zero_page_attr =
 379         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 380 #ifdef CONFIG_DEBUG_VM
 381 static ssize_t debug_cow_show(struct kobject *kobj,
 382                                 struct kobj_attribute *attr, char *buf)
 383 {
 384         return single_flag_show(kobj, attr, buf,
 385                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 386 }
 387 static ssize_t debug_cow_store(struct kobject *kobj,
 388                                struct kobj_attribute *attr,
 389                                const char *buf, size_t count)
 390 {
 391         return single_flag_store(kobj, attr, buf, count,
 392                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 393 }
 394 static struct kobj_attribute debug_cow_attr =
 395         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 396 #endif /* CONFIG_DEBUG_VM */
 397
 398 static struct attribute *hugepage_attr[] = {
 399         &enabled_attr.attr,
 400         &defrag_attr.attr,
 401         &use_zero_page_attr.attr,
 402 #ifdef CONFIG_DEBUG_VM
 403         &debug_cow_attr.attr,
 404 #endif
 405         NULL,
 406 };
 407
 408 static struct attribute_group hugepage_attr_group = {
 409         .attrs = hugepage_attr,
 410 };
 411
 412 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 413                                          struct kobj_attribute *attr,
 414                                          char *buf)
 415 {
 416         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 417 }
 418
 419 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 420                                           struct kobj_attribute *attr,
 421                                           const char *buf, size_t count)
 422 {
 423         unsigned long msecs;
 424         int err;
 425
 426         err = kstrtoul(buf, 10, &msecs);
 427         if (err || msecs > UINT_MAX)
 428                 return -EINVAL;
 429
 430         khugepaged_scan_sleep_millisecs = msecs;
 431         wake_up_interruptible(&khugepaged_wait);
 432
 433         return count;
 434 }
 435 static struct kobj_attribute scan_sleep_millisecs_attr =
 436         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 437                scan_sleep_millisecs_store);
 438
 439 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 440                                           struct kobj_attribute *attr,
 441                                           char *buf)
 442 {
 443         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 444 }
 445
 446 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 447                                            struct kobj_attribute *attr,
 448                                            const char *buf, size_t count)
 449 {
 450         unsigned long msecs;
 451         int err;
 452
 453         err = kstrtoul(buf, 10, &msecs);
 454         if (err || msecs > UINT_MAX)
 455                 return -EINVAL;
 456
 457         khugepaged_alloc_sleep_millisecs = msecs;
 458         wake_up_interruptible(&khugepaged_wait);
 459
 460         return count;
 461 }
 462 static struct kobj_attribute alloc_sleep_millisecs_attr =
 463         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 464                alloc_sleep_millisecs_store);
 465
 466 static ssize_t pages_to_scan_show(struct kobject *kobj,
 467                                   struct kobj_attribute *attr,
 468                                   char *buf)
 469 {
 470         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 471 }
 472 static ssize_t pages_to_scan_store(struct kobject *kobj,
 473                                    struct kobj_attribute *attr,
 474                                    const char *buf, size_t count)
 475 {
 476         int err;
 477         unsigned long pages;
 478
 479         err = kstrtoul(buf, 10, &pages);
 480         if (err || !pages || pages > UINT_MAX)
 481                 return -EINVAL;
 482
 483         khugepaged_pages_to_scan = pages;
 484
 485         return count;
 486 }
 487 static struct kobj_attribute pages_to_scan_attr =
 488         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 489                pages_to_scan_store);
 490
 491 static ssize_t pages_collapsed_show(struct kobject *kobj,
 492                                     struct kobj_attribute *attr,
 493                                     char *buf)
 494 {
 495         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 496 }
 497 static struct kobj_attribute pages_collapsed_attr =
 498         __ATTR_RO(pages_collapsed);
 499
 500 static ssize_t full_scans_show(struct kobject *kobj,
 501                                struct kobj_attribute *attr,
 502                                char *buf)
 503 {
 504         return sprintf(buf, "%u\n", khugepaged_full_scans);
 505 }
 506 static struct kobj_attribute full_scans_attr =
 507         __ATTR_RO(full_scans);
 508
 509 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 510                                       struct kobj_attribute *attr, char *buf)
 511 {
 512         return single_flag_show(kobj, attr, buf,
 513                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 514 }
 515 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 516                                        struct kobj_attribute *attr,
 517                                        const char *buf, size_t count)
 518 {
 519         return single_flag_store(kobj, attr, buf, count,
 520                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 521 }
 522 static struct kobj_attribute khugepaged_defrag_attr =
 523         __ATTR(defrag, 0644, khugepaged_defrag_show,
 524                khugepaged_defrag_store);
 525
 526 /*
 527  * max_ptes_none controls if khugepaged should collapse hugepages over
 528  * any unmapped ptes in turn potentially increasing the memory
 529  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 530  * reduce the available free memory in the system as it
 531  * runs. Increasing max_ptes_none will instead potentially reduce the
 532  * free memory in the system during the khugepaged scan.
 533  */
 534 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 535                                              struct kobj_attribute *attr,
 536                                              char *buf)
 537 {
 538         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 539 }
 540 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 541                                               struct kobj_attribute *attr,
 542                                               const char *buf, size_t count)
 543 {
 544         int err;
 545         unsigned long max_ptes_none;
 546
 547         err = kstrtoul(buf, 10, &max_ptes_none);
 548         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 549                 return -EINVAL;
 550
 551         khugepaged_max_ptes_none = max_ptes_none;
 552
 553         return count;
 554 }
 555 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 556         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 557                khugepaged_max_ptes_none_store);
 558
 559 static struct attribute *khugepaged_attr[] = {
 560         &khugepaged_defrag_attr.attr,
 561         &khugepaged_max_ptes_none_attr.attr,
 562         &pages_to_scan_attr.attr,
 563         &pages_collapsed_attr.attr,
 564         &full_scans_attr.attr,
 565         &scan_sleep_millisecs_attr.attr,
 566         &alloc_sleep_millisecs_attr.attr,
 567         NULL,
 568 };
 569
 570 static struct attribute_group khugepaged_attr_group = {
 571         .attrs = khugepaged_attr,
 572         .name = "khugepaged",
 573 };
 574
 575 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 576 {
 577         int err;
 578
 579         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 580         if (unlikely(!*hugepage_kobj)) {
 581                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 582                 return -ENOMEM;
 583         }
 584
 585         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 586         if (err) {
 587                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 588                 goto delete_obj;
 589         }
 590
 591         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 592         if (err) {
 593                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 594                 goto remove_hp_group;
 595         }
 596
 597         return 0;
 598
 599 remove_hp_group:
 600         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 601 delete_obj:
 602         kobject_put(*hugepage_kobj);
 603         return err;
 604 }
 605
 606 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 607 {
 608         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 609         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 610         kobject_put(hugepage_kobj);
 611 }
 612 #else
 613 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 614 {
 615         return 0;
 616 }
 617
 618 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 619 {
 620 }
 621 #endif /* CONFIG_SYSFS */
 622
 623 static int __init hugepage_init(void)
 624 {
 625         int err;
 626         struct kobject *hugepage_kobj;
 627
 628         if (!has_transparent_hugepage()) {
 629                 transparent_hugepage_flags = 0;
 630                 return -EINVAL;
 631         }
 632
 633         err = hugepage_init_sysfs(&hugepage_kobj);
 634         if (err)
 635                 return err;
 636
 637         err = khugepaged_slab_init();
 638         if (err)
 639                 goto out;
 640
 641         register_shrinker(&huge_zero_page_shrinker);
 642
 643         /*
 644          * By default disable transparent hugepages on smaller systems,
 645          * where the extra memory used could hurt more than TLB overhead
 646          * is likely to save.  The admin can still enable it through /sys.
 647          */
 648         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 649                 transparent_hugepage_flags = 0;
 650
 651         start_khugepaged();
 652
 653         return 0;
 654 out:
 655         hugepage_exit_sysfs(hugepage_kobj);
 656         return err;
 657 }
 658 module_init(hugepage_init)
 659
 660 static int __init setup_transparent_hugepage(char *str)
 661 {
 662         int ret = 0;
 663         if (!str)
 664                 goto out;
 665         if (!strcmp(str, "always")) {
 666                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 667                         &transparent_hugepage_flags);
 668                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 669                           &transparent_hugepage_flags);
 670                 ret = 1;
 671         } else if (!strcmp(str, "madvise")) {
 672                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 673                           &transparent_hugepage_flags);
 674                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 675                         &transparent_hugepage_flags);
 676                 ret = 1;
 677         } else if (!strcmp(str, "never")) {
 678                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 679                           &transparent_hugepage_flags);
 680                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 681                           &transparent_hugepage_flags);
 682                 ret = 1;
 683         }
 684 out:
 685         if (!ret)
 686                 printk(KERN_WARNING
 687                        "transparent_hugepage= cannot parse, ignored\n");
 688         return ret;
 689 }
 690 __setup("transparent_hugepage=", setup_transparent_hugepage);
 691
 692 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 693 {
 694         if (likely(vma->vm_flags & VM_WRITE))
 695                 pmd = pmd_mkwrite(pmd);
 696         return pmd;
 697 }
 698
 699 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 700 {
 701         pmd_t entry;
 702         entry = mk_pmd(page, prot);
 703         entry = pmd_mkhuge(entry);
 704         return entry;
 705 }
 706
 707 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 708                                         struct vm_area_struct *vma,
 709                                         unsigned long haddr, pmd_t *pmd,
 710                                         struct page *page)
 711 {
 712         pgtable_t pgtable;
 713         spinlock_t *ptl;
 714
 715         VM_BUG_ON(!PageCompound(page));
 716         pgtable = pte_alloc_one(mm, haddr);
 717         if (unlikely(!pgtable))
 718                 return VM_FAULT_OOM;
 719
 720         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 721         /*
 722          * The memory barrier inside __SetPageUptodate makes sure that
 723          * clear_huge_page writes become visible before the set_pmd_at()
 724          * write.
 725          */
 726         __SetPageUptodate(page);
 727
 728         ptl = pmd_lock(mm, pmd);
 729         if (unlikely(!pmd_none(*pmd))) {
 730                 spin_unlock(ptl);
 731                 mem_cgroup_uncharge_page(page);
 732                 put_page(page);
 733                 pte_free(mm, pgtable);
 734         } else {
 735                 pmd_t entry;
 736                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 737                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 738                 page_add_new_anon_rmap(page, vma, haddr);
 739                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 740                 set_pmd_at(mm, haddr, pmd, entry);
 741                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 742                 atomic_long_inc(&mm->nr_ptes);
 743                 spin_unlock(ptl);
 744         }
 745
 746         return 0;
 747 }
 748
 749 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 750 {
 751         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 752 }
 753
 754 static inline struct page *alloc_hugepage_vma(int defrag,
 755                                               struct vm_area_struct *vma,
 756                                               unsigned long haddr, int nd,
 757                                               gfp_t extra_gfp)
 758 {
 759         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 760                                HPAGE_PMD_ORDER, vma, haddr, nd);
 761 }
 762
 763 /* Caller must hold page table lock. */
 764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 765                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 766                 struct page *zero_page)
 767 {
 768         pmd_t entry;
 769         if (!pmd_none(*pmd))
 770                 return false;
 771         entry = mk_pmd(zero_page, vma->vm_page_prot);
 772         entry = pmd_wrprotect(entry);
 773         entry = pmd_mkhuge(entry);
 774         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 775         set_pmd_at(mm, haddr, pmd, entry);
 776         atomic_long_inc(&mm->nr_ptes);
 777         return true;
 778 }
 779
 780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 781                                unsigned long address, pmd_t *pmd,
 782                                unsigned int flags)
 783 {
 784         struct page *page;
 785         unsigned long haddr = address & HPAGE_PMD_MASK;
 786
 787         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 788                 return VM_FAULT_FALLBACK;
 789         if (unlikely(anon_vma_prepare(vma)))
 790                 return VM_FAULT_OOM;
 791         if (unlikely(khugepaged_enter(vma)))
 792                 return VM_FAULT_OOM;
 793         if (!(flags & FAULT_FLAG_WRITE) &&
 794                         transparent_hugepage_use_zero_page()) {
 795                 spinlock_t *ptl;
 796                 pgtable_t pgtable;
 797                 struct page *zero_page;
 798                 bool set;
 799                 pgtable = pte_alloc_one(mm, haddr);
 800                 if (unlikely(!pgtable))
 801                         return VM_FAULT_OOM;
 802                 zero_page = get_huge_zero_page();
 803                 if (unlikely(!zero_page)) {
 804                         pte_free(mm, pgtable);
 805                         count_vm_event(THP_FAULT_FALLBACK);
 806                         return VM_FAULT_FALLBACK;
 807                 }
 808                 ptl = pmd_lock(mm, pmd);
 809                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 810                                 zero_page);
 811                 spin_unlock(ptl);
 812                 if (!set) {
 813                         pte_free(mm, pgtable);
 814                         put_huge_zero_page();
 815                 }
 816                 return 0;
 817         }
 818         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 819                         vma, haddr, numa_node_id(), 0);
 820         if (unlikely(!page)) {
 821                 count_vm_event(THP_FAULT_FALLBACK);
 822                 return VM_FAULT_FALLBACK;
 823         }
 824         if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 825                 put_page(page);
 826                 count_vm_event(THP_FAULT_FALLBACK);
 827                 return VM_FAULT_FALLBACK;
 828         }
 829         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 830                 mem_cgroup_uncharge_page(page);
 831                 put_page(page);
 832                 count_vm_event(THP_FAULT_FALLBACK);
 833                 return VM_FAULT_FALLBACK;
 834         }
 835
 836         count_vm_event(THP_FAULT_ALLOC);
 837         return 0;
 838 }
 839
 840 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 841                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 842                   struct vm_area_struct *vma)
 843 {
 844         spinlock_t *dst_ptl, *src_ptl;
 845         struct page *src_page;
 846         pmd_t pmd;
 847         pgtable_t pgtable;
 848         int ret;
 849
 850         ret = -ENOMEM;
 851         pgtable = pte_alloc_one(dst_mm, addr);
 852         if (unlikely(!pgtable))
 853                 goto out;
 854
 855         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 856         src_ptl = pmd_lockptr(src_mm, src_pmd);
 857         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 858
 859         ret = -EAGAIN;
 860         pmd = *src_pmd;
 861         if (unlikely(!pmd_trans_huge(pmd))) {
 862                 pte_free(dst_mm, pgtable);
 863                 goto out_unlock;
 864         }
 865         /*
 866          * When page table lock is held, the huge zero pmd should not be
 867          * under splitting since we don't split the page itself, only pmd to
 868          * a page table.
 869          */
 870         if (is_huge_zero_pmd(pmd)) {
 871                 struct page *zero_page;
 872                 bool set;
 873                 /*
 874                  * get_huge_zero_page() will never allocate a new page here,
 875                  * since we already have a zero page to copy. It just takes a
 876                  * reference.
 877                  */
 878                 zero_page = get_huge_zero_page();
 879                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 880                                 zero_page);
 881                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 882                 ret = 0;
 883                 goto out_unlock;
 884         }
 885         if (unlikely(pmd_trans_splitting(pmd))) {
 886                 /* split huge page running from under us */
 887                 spin_unlock(src_ptl);
 888                 spin_unlock(dst_ptl);
 889                 pte_free(dst_mm, pgtable);
 890
 891                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 892                 goto out;
 893         }
 894         src_page = pmd_page(pmd);
 895         VM_BUG_ON(!PageHead(src_page));
 896         get_page(src_page);
 897         page_dup_rmap(src_page);
 898         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 899
 900         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 901         pmd = pmd_mkold(pmd_wrprotect(pmd));
 902         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 903         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 904         atomic_long_inc(&dst_mm->nr_ptes);
 905
 906         ret = 0;
 907 out_unlock:
 908         spin_unlock(src_ptl);
 909         spin_unlock(dst_ptl);
 910 out:
 911         return ret;
 912 }
 913
 914 void huge_pmd_set_accessed(struct mm_struct *mm,
 915                            struct vm_area_struct *vma,
 916                            unsigned long address,
 917                            pmd_t *pmd, pmd_t orig_pmd,
 918                            int dirty)
 919 {
 920         spinlock_t *ptl;
 921         pmd_t entry;
 922         unsigned long haddr;
 923
 924         ptl = pmd_lock(mm, pmd);
 925         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 926                 goto unlock;
 927
 928         entry = pmd_mkyoung(orig_pmd);
 929         haddr = address & HPAGE_PMD_MASK;
 930         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 931                 update_mmu_cache_pmd(vma, address, pmd);
 932
 933 unlock:
 934         spin_unlock(ptl);
 935 }
 936
 937 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 938                 struct vm_area_struct *vma, unsigned long address,
 939                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 940 {
 941         spinlock_t *ptl;
 942         pgtable_t pgtable;
 943         pmd_t _pmd;
 944         struct page *page;
 945         int i, ret = 0;
 946         unsigned long mmun_start;       /* For mmu_notifiers */
 947         unsigned long mmun_end;         /* For mmu_notifiers */
 948
 949         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 950         if (!page) {
 951                 ret |= VM_FAULT_OOM;
 952                 goto out;
 953         }
 954
 955         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 956                 put_page(page);
 957                 ret |= VM_FAULT_OOM;
 958                 goto out;
 959         }
 960
 961         clear_user_highpage(page, address);
 962         __SetPageUptodate(page);
 963
 964         mmun_start = haddr;
 965         mmun_end   = haddr + HPAGE_PMD_SIZE;
 966         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 967
 968         ptl = pmd_lock(mm, pmd);
 969         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 970                 goto out_free_page;
 971
 972         pmdp_clear_flush(vma, haddr, pmd);
 973         /* leave pmd empty until pte is filled */
 974
 975         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
 976         pmd_populate(mm, &_pmd, pgtable);
 977
 978         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 979                 pte_t *pte, entry;
 980                 if (haddr == (address & PAGE_MASK)) {
 981                         entry = mk_pte(page, vma->vm_page_prot);
 982                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 983                         page_add_new_anon_rmap(page, vma, haddr);
 984                 } else {
 985                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
 986                         entry = pte_mkspecial(entry);
 987                 }
 988                 pte = pte_offset_map(&_pmd, haddr);
 989                 VM_BUG_ON(!pte_none(*pte));
 990                 set_pte_at(mm, haddr, pte, entry);
 991                 pte_unmap(pte);
 992         }
 993         smp_wmb(); /* make pte visible before pmd */
 994         pmd_populate(mm, pmd, pgtable);
 995         spin_unlock(ptl);
 996         put_huge_zero_page();
 997         inc_mm_counter(mm, MM_ANONPAGES);
 998
 999         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1000
1001         ret |= VM_FAULT_WRITE;
1002 out:
1003         return ret;
1004 out_free_page:
1005         spin_unlock(ptl);
1006         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007         mem_cgroup_uncharge_page(page);
1008         put_page(page);
1009         goto out;
1010 }
1011
1012 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1013                                         struct vm_area_struct *vma,
1014                                         unsigned long address,
1015                                         pmd_t *pmd, pmd_t orig_pmd,
1016                                         struct page *page,
1017                                         unsigned long haddr)
1018 {
1019         spinlock_t *ptl;
1020         pgtable_t pgtable;
1021         pmd_t _pmd;
1022         int ret = 0, i;
1023         struct page **pages;
1024         unsigned long mmun_start;       /* For mmu_notifiers */
1025         unsigned long mmun_end;         /* For mmu_notifiers */
1026
1027         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1028                         GFP_KERNEL);
1029         if (unlikely(!pages)) {
1030                 ret |= VM_FAULT_OOM;
1031                 goto out;
1032         }
1033
1034         for (i = 0; i < HPAGE_PMD_NR; i++) {
1035                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1036                                                __GFP_OTHER_NODE,
1037                                                vma, address, page_to_nid(page));
1038                 if (unlikely(!pages[i] ||
1039                              mem_cgroup_newpage_charge(pages[i], mm,
1040                                                        GFP_KERNEL))) {
1041                         if (pages[i])
1042                                 put_page(pages[i]);
1043                         mem_cgroup_uncharge_start();
1044                         while (--i >= 0) {
1045                                 mem_cgroup_uncharge_page(pages[i]);
1046                                 put_page(pages[i]);
1047                         }
1048                         mem_cgroup_uncharge_end();
1049                         kfree(pages);
1050                         ret |= VM_FAULT_OOM;
1051                         goto out;
1052                 }
1053         }
1054
1055         for (i = 0; i < HPAGE_PMD_NR; i++) {
1056                 copy_user_highpage(pages[i], page + i,
1057                                    haddr + PAGE_SIZE * i, vma);
1058                 __SetPageUptodate(pages[i]);
1059                 cond_resched();
1060         }
1061
1062         mmun_start = haddr;
1063         mmun_end   = haddr + HPAGE_PMD_SIZE;
1064         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1065
1066         ptl = pmd_lock(mm, pmd);
1067         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1068                 goto out_free_pages;
1069         VM_BUG_ON(!PageHead(page));
1070
1071         pmdp_clear_flush(vma, haddr, pmd);
1072         /* leave pmd empty until pte is filled */
1073
1074         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1075         pmd_populate(mm, &_pmd, pgtable);
1076
1077         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1078                 pte_t *pte, entry;
1079                 entry = mk_pte(pages[i], vma->vm_page_prot);
1080                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1081                 page_add_new_anon_rmap(pages[i], vma, haddr);
1082                 pte = pte_offset_map(&_pmd, haddr);
1083                 VM_BUG_ON(!pte_none(*pte));
1084                 set_pte_at(mm, haddr, pte, entry);
1085                 pte_unmap(pte);
1086         }
1087         kfree(pages);
1088
1089         smp_wmb(); /* make pte visible before pmd */
1090         pmd_populate(mm, pmd, pgtable);
1091         page_remove_rmap(page);
1092         spin_unlock(ptl);
1093
1094         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1095
1096         ret |= VM_FAULT_WRITE;
1097         put_page(page);
1098
1099 out:
1100         return ret;
1101
1102 out_free_pages:
1103         spin_unlock(ptl);
1104         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1105         mem_cgroup_uncharge_start();
1106         for (i = 0; i < HPAGE_PMD_NR; i++) {
1107                 mem_cgroup_uncharge_page(pages[i]);
1108                 put_page(pages[i]);
1109         }
1110         mem_cgroup_uncharge_end();
1111         kfree(pages);
1112         goto out;
1113 }
1114
1115 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1116                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1117 {
1118         spinlock_t *ptl;
1119         int ret = 0;
1120         struct page *page = NULL, *new_page;
1121         unsigned long haddr;
1122         unsigned long mmun_start;       /* For mmu_notifiers */
1123         unsigned long mmun_end;         /* For mmu_notifiers */
1124
1125         ptl = pmd_lockptr(mm, pmd);
1126         VM_BUG_ON(!vma->anon_vma);
1127         haddr = address & HPAGE_PMD_MASK;
1128         if (is_huge_zero_pmd(orig_pmd))
1129                 goto alloc;
1130         spin_lock(ptl);
1131         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1132                 goto out_unlock;
1133
1134         page = pmd_page(orig_pmd);
1135         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1136         if (page_mapcount(page) == 1) {
1137                 pmd_t entry;
1138                 entry = pmd_mkyoung(orig_pmd);
1139                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1140                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1141                         update_mmu_cache_pmd(vma, address, pmd);
1142                 ret |= VM_FAULT_WRITE;
1143                 goto out_unlock;
1144         }
1145         get_page(page);
1146         spin_unlock(ptl);
1147 alloc:
1148         if (transparent_hugepage_enabled(vma) &&
1149             !transparent_hugepage_debug_cow())
1150                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1151                                               vma, haddr, numa_node_id(), 0);
1152         else
1153                 new_page = NULL;
1154
1155         if (unlikely(!new_page)) {
1156                 if (is_huge_zero_pmd(orig_pmd)) {
1157                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1158                                         address, pmd, orig_pmd, haddr);
1159                 } else {
1160                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1161                                         pmd, orig_pmd, page, haddr);
1162                         if (ret & VM_FAULT_OOM)
1163                                 split_huge_page(page);
1164                         put_page(page);
1165                 }
1166                 count_vm_event(THP_FAULT_FALLBACK);
1167                 goto out;
1168         }
1169
1170         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1171                 put_page(new_page);
1172                 if (page) {
1173                         split_huge_page(page);
1174                         put_page(page);
1175                 }
1176                 count_vm_event(THP_FAULT_FALLBACK);
1177                 ret |= VM_FAULT_OOM;
1178                 goto out;
1179         }
1180
1181         count_vm_event(THP_FAULT_ALLOC);
1182
1183         if (is_huge_zero_pmd(orig_pmd))
1184                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1185         else
1186                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1187         __SetPageUptodate(new_page);
1188
1189         mmun_start = haddr;
1190         mmun_end   = haddr + HPAGE_PMD_SIZE;
1191         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1192
1193         spin_lock(ptl);
1194         if (page)
1195                 put_page(page);
1196         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1197                 spin_unlock(ptl);
1198                 mem_cgroup_uncharge_page(new_page);
1199                 put_page(new_page);
1200                 goto out_mn;
1201         } else {
1202                 pmd_t entry;
1203                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1204                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1205                 pmdp_clear_flush(vma, haddr, pmd);
1206                 page_add_new_anon_rmap(new_page, vma, haddr);
1207                 set_pmd_at(mm, haddr, pmd, entry);
1208                 update_mmu_cache_pmd(vma, address, pmd);
1209                 if (is_huge_zero_pmd(orig_pmd)) {
1210                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1211                         put_huge_zero_page();
1212                 } else {
1213                         VM_BUG_ON(!PageHead(page));
1214                         page_remove_rmap(page);
1215                         put_page(page);
1216                 }
1217                 ret |= VM_FAULT_WRITE;
1218         }
1219         spin_unlock(ptl);
1220 out_mn:
1221         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1222 out:
1223         return ret;
1224 out_unlock:
1225         spin_unlock(ptl);
1226         return ret;
1227 }
1228
1229 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1230                                    unsigned long addr,
1231                                    pmd_t *pmd,
1232                                    unsigned int flags)
1233 {
1234         struct mm_struct *mm = vma->vm_mm;
1235         struct page *page = NULL;
1236
1237         assert_spin_locked(pmd_lockptr(mm, pmd));
1238
1239         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1240                 goto out;
1241
1242         /* Avoid dumping huge zero page */
1243         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1244                 return ERR_PTR(-EFAULT);
1245
1246         page = pmd_page(*pmd);
1247         VM_BUG_ON(!PageHead(page));
1248         if (flags & FOLL_TOUCH) {
1249                 pmd_t _pmd;
1250                 /*
1251                  * We should set the dirty bit only for FOLL_WRITE but
1252                  * for now the dirty bit in the pmd is meaningless.
1253                  * And if the dirty bit will become meaningful and
1254                  * we'll only set it with FOLL_WRITE, an atomic
1255                  * set_bit will be required on the pmd to set the
1256                  * young bit, instead of the current set_pmd_at.
1257                  */
1258                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1259                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1260                                           pmd, _pmd,  1))
1261                         update_mmu_cache_pmd(vma, addr, pmd);
1262         }
1263         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1264                 if (page->mapping && trylock_page(page)) {
1265                         lru_add_drain();
1266                         if (page->mapping)
1267                                 mlock_vma_page(page);
1268                         unlock_page(page);
1269                 }
1270         }
1271         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1272         VM_BUG_ON(!PageCompound(page));
1273         if (flags & FOLL_GET)
1274                 get_page_foll(page);
1275
1276 out:
1277         return page;
1278 }
1279
1280 /* NUMA hinting page fault entry point for trans huge pmds */
1281 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1282                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1283 {
1284         spinlock_t *ptl;
1285         struct anon_vma *anon_vma = NULL;
1286         struct page *page;
1287         unsigned long haddr = addr & HPAGE_PMD_MASK;
1288         int page_nid = -1, this_nid = numa_node_id();
1289         int target_nid, last_cpupid = -1;
1290         bool page_locked;
1291         bool migrated = false;
1292         int flags = 0;
1293
1294         ptl = pmd_lock(mm, pmdp);
1295         if (unlikely(!pmd_same(pmd, *pmdp)))
1296                 goto out_unlock;
1297
1298         page = pmd_page(pmd);
1299         BUG_ON(is_huge_zero_page(page));
1300         page_nid = page_to_nid(page);
1301         last_cpupid = page_cpupid_last(page);
1302         count_vm_numa_event(NUMA_HINT_FAULTS);
1303         if (page_nid == this_nid) {
1304                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1305                 flags |= TNF_FAULT_LOCAL;
1306         }
1307
1308         /*
1309          * Avoid grouping on DSO/COW pages in specific and RO pages
1310          * in general, RO pages shouldn't hurt as much anyway since
1311          * they can be in shared cache state.
1312          */
1313         if (!pmd_write(pmd))
1314                 flags |= TNF_NO_GROUP;
1315
1316         /*
1317          * Acquire the page lock to serialise THP migrations but avoid dropping
1318          * page_table_lock if at all possible
1319          */
1320         page_locked = trylock_page(page);
1321         target_nid = mpol_misplaced(page, vma, haddr);
1322         if (target_nid == -1) {
1323                 /* If the page was locked, there are no parallel migrations */
1324                 if (page_locked)
1325                         goto clear_pmdnuma;
1326
1327                 /*
1328                  * Otherwise wait for potential migrations and retry. We do
1329                  * relock and check_same as the page may no longer be mapped.
1330                  * As the fault is being retried, do not account for it.
1331                  */
1332                 spin_unlock(ptl);
1333                 wait_on_page_locked(page);
1334                 page_nid = -1;
1335                 goto out;
1336         }
1337
1338         /* Page is misplaced, serialise migrations and parallel THP splits */
1339         get_page(page);
1340         spin_unlock(ptl);
1341         if (!page_locked)
1342                 lock_page(page);
1343         anon_vma = page_lock_anon_vma_read(page);
1344
1345         /* Confirm the PMD did not change while page_table_lock was released */
1346         spin_lock(ptl);
1347         if (unlikely(!pmd_same(pmd, *pmdp))) {
1348                 unlock_page(page);
1349                 put_page(page);
1350                 page_nid = -1;
1351                 goto out_unlock;
1352         }
1353
1354         /*
1355          * Migrate the THP to the requested node, returns with page unlocked
1356          * and pmd_numa cleared.
1357          */
1358         spin_unlock(ptl);
1359         migrated = migrate_misplaced_transhuge_page(mm, vma,
1360                                 pmdp, pmd, addr, page, target_nid);
1361         if (migrated) {
1362                 flags |= TNF_MIGRATED;
1363                 page_nid = target_nid;
1364         }
1365
1366         goto out;
1367 clear_pmdnuma:
1368         BUG_ON(!PageLocked(page));
1369         pmd = pmd_mknonnuma(pmd);
1370         set_pmd_at(mm, haddr, pmdp, pmd);
1371         VM_BUG_ON(pmd_numa(*pmdp));
1372         update_mmu_cache_pmd(vma, addr, pmdp);
1373         unlock_page(page);
1374 out_unlock:
1375         spin_unlock(ptl);
1376
1377 out:
1378         if (anon_vma)
1379                 page_unlock_anon_vma_read(anon_vma);
1380
1381         if (page_nid != -1)
1382                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1383
1384         return 0;
1385 }
1386
1387 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1388                  pmd_t *pmd, unsigned long addr)
1389 {
1390         spinlock_t *ptl;
1391         int ret = 0;
1392
1393         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1394                 struct page *page;
1395                 pgtable_t pgtable;
1396                 pmd_t orig_pmd;
1397                 /*
1398                  * For architectures like ppc64 we look at deposited pgtable
1399                  * when calling pmdp_get_and_clear. So do the
1400                  * pgtable_trans_huge_withdraw after finishing pmdp related
1401                  * operations.
1402                  */
1403                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1404                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1405                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1406                 if (is_huge_zero_pmd(orig_pmd)) {
1407                         atomic_long_dec(&tlb->mm->nr_ptes);
1408                         spin_unlock(ptl);
1409                         put_huge_zero_page();
1410                 } else {
1411                         page = pmd_page(orig_pmd);
1412                         page_remove_rmap(page);
1413                         VM_BUG_ON(page_mapcount(page) < 0);
1414                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1415                         VM_BUG_ON(!PageHead(page));
1416                         atomic_long_dec(&tlb->mm->nr_ptes);
1417                         spin_unlock(ptl);
1418                         tlb_remove_page(tlb, page);
1419                 }
1420                 pte_free(tlb->mm, pgtable);
1421                 ret = 1;
1422         }
1423         return ret;
1424 }
1425
1426 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1427                 unsigned long addr, unsigned long end,
1428                 unsigned char *vec)
1429 {
1430         spinlock_t *ptl;
1431         int ret = 0;
1432
1433         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1434                 /*
1435                  * All logical pages in the range are present
1436                  * if backed by a huge page.
1437                  */
1438                 spin_unlock(ptl);
1439                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1440                 ret = 1;
1441         }
1442
1443         return ret;
1444 }
1445
1446 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1447                   unsigned long old_addr,
1448                   unsigned long new_addr, unsigned long old_end,
1449                   pmd_t *old_pmd, pmd_t *new_pmd)
1450 {
1451         spinlock_t *old_ptl, *new_ptl;
1452         int ret = 0;
1453         pmd_t pmd;
1454
1455         struct mm_struct *mm = vma->vm_mm;
1456
1457         if ((old_addr & ~HPAGE_PMD_MASK) ||
1458             (new_addr & ~HPAGE_PMD_MASK) ||
1459             old_end - old_addr < HPAGE_PMD_SIZE ||
1460             (new_vma->vm_flags & VM_NOHUGEPAGE))
1461                 goto out;
1462
1463         /*
1464          * The destination pmd shouldn't be established, free_pgtables()
1465          * should have release it.
1466          */
1467         if (WARN_ON(!pmd_none(*new_pmd))) {
1468                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1469                 goto out;
1470         }
1471
1472         /*
1473          * We don't have to worry about the ordering of src and dst
1474          * ptlocks because exclusive mmap_sem prevents deadlock.
1475          */
1476         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1477         if (ret == 1) {
1478                 new_ptl = pmd_lockptr(mm, new_pmd);
1479                 if (new_ptl != old_ptl)
1480                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1481                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1482                 VM_BUG_ON(!pmd_none(*new_pmd));
1483                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1484                 if (new_ptl != old_ptl)
1485                         spin_unlock(new_ptl);
1486                 spin_unlock(old_ptl);
1487         }
1488 out:
1489         return ret;
1490 }
1491
1492 /*
1493  * Returns
1494  *  - 0 if PMD could not be locked
1495  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1496  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1497  */
1498 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1499                 unsigned long addr, pgprot_t newprot, int prot_numa)
1500 {
1501         struct mm_struct *mm = vma->vm_mm;
1502         spinlock_t *ptl;
1503         int ret = 0;
1504
1505         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1506                 pmd_t entry;
1507                 ret = 1;
1508                 if (!prot_numa) {
1509                         entry = pmdp_get_and_clear(mm, addr, pmd);
1510                         entry = pmd_modify(entry, newprot);
1511                         ret = HPAGE_PMD_NR;
1512                         BUG_ON(pmd_write(entry));
1513                 } else {
1514                         struct page *page = pmd_page(*pmd);
1515
1516                         /*
1517                          * Do not trap faults against the zero page. The
1518                          * read-only data is likely to be read-cached on the
1519                          * local CPU cache and it is less useful to know about
1520                          * local vs remote hits on the zero page.
1521                          */
1522                         if (!is_huge_zero_page(page) &&
1523                             !pmd_numa(*pmd)) {
1524                                 entry = pmdp_get_and_clear(mm, addr, pmd);
1525                                 entry = pmd_mknuma(entry);
1526                                 ret = HPAGE_PMD_NR;
1527                         }
1528                 }
1529
1530                 /* Set PMD if cleared earlier */
1531                 if (ret == HPAGE_PMD_NR)
1532                         set_pmd_at(mm, addr, pmd, entry);
1533
1534                 spin_unlock(ptl);
1535         }
1536
1537         return ret;
1538 }
1539
1540 /*
1541  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1542  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1543  *
1544  * Note that if it returns 1, this routine returns without unlocking page
1545  * table locks. So callers must unlock them.
1546  */
1547 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1548                 spinlock_t **ptl)
1549 {
1550         *ptl = pmd_lock(vma->vm_mm, pmd);
1551         if (likely(pmd_trans_huge(*pmd))) {
1552                 if (unlikely(pmd_trans_splitting(*pmd))) {
1553                         spin_unlock(*ptl);
1554                         wait_split_huge_page(vma->anon_vma, pmd);
1555                         return -1;
1556                 } else {
1557                         /* Thp mapped by 'pmd' is stable, so we can
1558                          * handle it as it is. */
1559                         return 1;
1560                 }
1561         }
1562         spin_unlock(*ptl);
1563         return 0;
1564 }
1565
1566 /*
1567  * This function returns whether a given @page is mapped onto the @address
1568  * in the virtual space of @mm.
1569  *
1570  * When it's true, this function returns *pmd with holding the page table lock
1571  * and passing it back to the caller via @ptl.
1572  * If it's false, returns NULL without holding the page table lock.
1573  */
1574 pmd_t *page_check_address_pmd(struct page *page,
1575                               struct mm_struct *mm,
1576                               unsigned long address,
1577                               enum page_check_address_pmd_flag flag,
1578                               spinlock_t **ptl)
1579 {
1580         pmd_t *pmd;
1581
1582         if (address & ~HPAGE_PMD_MASK)
1583                 return NULL;
1584
1585         pmd = mm_find_pmd(mm, address);
1586         if (!pmd)
1587                 return NULL;
1588         *ptl = pmd_lock(mm, pmd);
1589         if (pmd_none(*pmd))
1590                 goto unlock;
1591         if (pmd_page(*pmd) != page)
1592                 goto unlock;
1593         /*
1594          * split_vma() may create temporary aliased mappings. There is
1595          * no risk as long as all huge pmd are found and have their
1596          * splitting bit set before __split_huge_page_refcount
1597          * runs. Finding the same huge pmd more than once during the
1598          * same rmap walk is not a problem.
1599          */
1600         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1601             pmd_trans_splitting(*pmd))
1602                 goto unlock;
1603         if (pmd_trans_huge(*pmd)) {
1604                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1605                           !pmd_trans_splitting(*pmd));
1606                 return pmd;
1607         }
1608 unlock:
1609         spin_unlock(*ptl);
1610         return NULL;
1611 }
1612
1613 static int __split_huge_page_splitting(struct page *page,
1614                                        struct vm_area_struct *vma,
1615                                        unsigned long address)
1616 {
1617         struct mm_struct *mm = vma->vm_mm;
1618         spinlock_t *ptl;
1619         pmd_t *pmd;
1620         int ret = 0;
1621         /* For mmu_notifiers */
1622         const unsigned long mmun_start = address;
1623         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1624
1625         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1626         pmd = page_check_address_pmd(page, mm, address,
1627                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1628         if (pmd) {
1629                 /*
1630                  * We can't temporarily set the pmd to null in order
1631                  * to split it, the pmd must remain marked huge at all
1632                  * times or the VM won't take the pmd_trans_huge paths
1633                  * and it won't wait on the anon_vma->root->rwsem to
1634                  * serialize against split_huge_page*.
1635                  */
1636                 pmdp_splitting_flush(vma, address, pmd);
1637                 ret = 1;
1638                 spin_unlock(ptl);
1639         }
1640         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1641
1642         return ret;
1643 }
1644
1645 static void __split_huge_page_refcount(struct page *page,
1646                                        struct list_head *list)
1647 {
1648         int i;
1649         struct zone *zone = page_zone(page);
1650         struct lruvec *lruvec;
1651         int tail_count = 0;
1652
1653         /* prevent PageLRU to go away from under us, and freeze lru stats */
1654         spin_lock_irq(&zone->lru_lock);
1655         lruvec = mem_cgroup_page_lruvec(page, zone);
1656
1657         compound_lock(page);
1658         /* complete memcg works before add pages to LRU */
1659         mem_cgroup_split_huge_fixup(page);
1660
1661         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1662                 struct page *page_tail = page + i;
1663
1664                 /* tail_page->_mapcount cannot change */
1665                 BUG_ON(page_mapcount(page_tail) < 0);
1666                 tail_count += page_mapcount(page_tail);
1667                 /* check for overflow */
1668                 BUG_ON(tail_count < 0);
1669                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1670                 /*
1671                  * tail_page->_count is zero and not changing from
1672                  * under us. But get_page_unless_zero() may be running
1673                  * from under us on the tail_page. If we used
1674                  * atomic_set() below instead of atomic_add(), we
1675                  * would then run atomic_set() concurrently with
1676                  * get_page_unless_zero(), and atomic_set() is
1677                  * implemented in C not using locked ops. spin_unlock
1678                  * on x86 sometime uses locked ops because of PPro
1679                  * errata 66, 92, so unless somebody can guarantee
1680                  * atomic_set() here would be safe on all archs (and
1681                  * not only on x86), it's safer to use atomic_add().
1682                  */
1683                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1684                            &page_tail->_count);
1685
1686                 /* after clearing PageTail the gup refcount can be released */
1687                 smp_mb();
1688
1689                 /*
1690                  * retain hwpoison flag of the poisoned tail page:
1691                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1692                  *   by the memory-failure.
1693                  */
1694                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1695                 page_tail->flags |= (page->flags &
1696                                      ((1L << PG_referenced) |
1697                                       (1L << PG_swapbacked) |
1698                                       (1L << PG_mlocked) |
1699                                       (1L << PG_uptodate) |
1700                                       (1L << PG_active) |
1701                                       (1L << PG_unevictable)));
1702                 page_tail->flags |= (1L << PG_dirty);
1703
1704                 /* clear PageTail before overwriting first_page */
1705                 smp_wmb();
1706
1707                 /*
1708                  * __split_huge_page_splitting() already set the
1709                  * splitting bit in all pmd that could map this
1710                  * hugepage, that will ensure no CPU can alter the
1711                  * mapcount on the head page. The mapcount is only
1712                  * accounted in the head page and it has to be
1713                  * transferred to all tail pages in the below code. So
1714                  * for this code to be safe, the split the mapcount
1715                  * can't change. But that doesn't mean userland can't
1716                  * keep changing and reading the page contents while
1717                  * we transfer the mapcount, so the pmd splitting
1718                  * status is achieved setting a reserved bit in the
1719                  * pmd, not by clearing the present bit.
1720                 */
1721                 page_tail->_mapcount = page->_mapcount;
1722
1723                 BUG_ON(page_tail->mapping);
1724                 page_tail->mapping = page->mapping;
1725
1726                 page_tail->index = page->index + i;
1727                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1728
1729                 BUG_ON(!PageAnon(page_tail));
1730                 BUG_ON(!PageUptodate(page_tail));
1731                 BUG_ON(!PageDirty(page_tail));
1732                 BUG_ON(!PageSwapBacked(page_tail));
1733
1734                 lru_add_page_tail(page, page_tail, lruvec, list);
1735         }
1736         atomic_sub(tail_count, &page->_count);
1737         BUG_ON(atomic_read(&page->_count) <= 0);
1738
1739         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1740
1741         ClearPageCompound(page);
1742         compound_unlock(page);
1743         spin_unlock_irq(&zone->lru_lock);
1744
1745         for (i = 1; i < HPAGE_PMD_NR; i++) {
1746                 struct page *page_tail = page + i;
1747                 BUG_ON(page_count(page_tail) <= 0);
1748                 /*
1749                  * Tail pages may be freed if there wasn't any mapping
1750                  * like if add_to_swap() is running on a lru page that
1751                  * had its mapping zapped. And freeing these pages
1752                  * requires taking the lru_lock so we do the put_page
1753                  * of the tail pages after the split is complete.
1754                  */
1755                 put_page(page_tail);
1756         }
1757
1758         /*
1759          * Only the head page (now become a regular page) is required
1760          * to be pinned by the caller.
1761          */
1762         BUG_ON(page_count(page) <= 0);
1763 }
1764
1765 static int __split_huge_page_map(struct page *page,
1766                                  struct vm_area_struct *vma,
1767                                  unsigned long address)
1768 {
1769         struct mm_struct *mm = vma->vm_mm;
1770         spinlock_t *ptl;
1771         pmd_t *pmd, _pmd;
1772         int ret = 0, i;
1773         pgtable_t pgtable;
1774         unsigned long haddr;
1775
1776         pmd = page_check_address_pmd(page, mm, address,
1777                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1778         if (pmd) {
1779                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1780                 pmd_populate(mm, &_pmd, pgtable);
1781
1782                 haddr = address;
1783                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1784                         pte_t *pte, entry;
1785                         BUG_ON(PageCompound(page+i));
1786                         entry = mk_pte(page + i, vma->vm_page_prot);
1787                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1788                         if (!pmd_write(*pmd))
1789                                 entry = pte_wrprotect(entry);
1790                         else
1791                                 BUG_ON(page_mapcount(page) != 1);
1792                         if (!pmd_young(*pmd))
1793                                 entry = pte_mkold(entry);
1794                         if (pmd_numa(*pmd))
1795                                 entry = pte_mknuma(entry);
1796                         pte = pte_offset_map(&_pmd, haddr);
1797                         BUG_ON(!pte_none(*pte));
1798                         set_pte_at(mm, haddr, pte, entry);
1799                         pte_unmap(pte);
1800                 }
1801
1802                 smp_wmb(); /* make pte visible before pmd */
1803                 /*
1804                  * Up to this point the pmd is present and huge and
1805                  * userland has the whole access to the hugepage
1806                  * during the split (which happens in place). If we
1807                  * overwrite the pmd with the not-huge version
1808                  * pointing to the pte here (which of course we could
1809                  * if all CPUs were bug free), userland could trigger
1810                  * a small page size TLB miss on the small sized TLB
1811                  * while the hugepage TLB entry is still established
1812                  * in the huge TLB. Some CPU doesn't like that. See
1813                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1814                  * Erratum 383 on page 93. Intel should be safe but is
1815                  * also warns that it's only safe if the permission
1816                  * and cache attributes of the two entries loaded in
1817                  * the two TLB is identical (which should be the case
1818                  * here). But it is generally safer to never allow
1819                  * small and huge TLB entries for the same virtual
1820                  * address to be loaded simultaneously. So instead of
1821                  * doing "pmd_populate(); flush_tlb_range();" we first
1822                  * mark the current pmd notpresent (atomically because
1823                  * here the pmd_trans_huge and pmd_trans_splitting
1824                  * must remain set at all times on the pmd until the
1825                  * split is complete for this pmd), then we flush the
1826                  * SMP TLB and finally we write the non-huge version
1827                  * of the pmd entry with pmd_populate.
1828                  */
1829                 pmdp_invalidate(vma, address, pmd);
1830                 pmd_populate(mm, pmd, pgtable);
1831                 ret = 1;
1832                 spin_unlock(ptl);
1833         }
1834
1835         return ret;
1836 }
1837
1838 /* must be called with anon_vma->root->rwsem held */
1839 static void __split_huge_page(struct page *page,
1840                               struct anon_vma *anon_vma,
1841                               struct list_head *list)
1842 {
1843         int mapcount, mapcount2;
1844         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1845         struct anon_vma_chain *avc;
1846
1847         BUG_ON(!PageHead(page));
1848         BUG_ON(PageTail(page));
1849
1850         mapcount = 0;
1851         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1852                 struct vm_area_struct *vma = avc->vma;
1853                 unsigned long addr = vma_address(page, vma);
1854                 BUG_ON(is_vma_temporary_stack(vma));
1855                 mapcount += __split_huge_page_splitting(page, vma, addr);
1856         }
1857         /*
1858          * It is critical that new vmas are added to the tail of the
1859          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1860          * and establishes a child pmd before
1861          * __split_huge_page_splitting() freezes the parent pmd (so if
1862          * we fail to prevent copy_huge_pmd() from running until the
1863          * whole __split_huge_page() is complete), we will still see
1864          * the newly established pmd of the child later during the
1865          * walk, to be able to set it as pmd_trans_splitting too.
1866          */
1867         if (mapcount != page_mapcount(page))
1868                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1869                        mapcount, page_mapcount(page));
1870         BUG_ON(mapcount != page_mapcount(page));
1871
1872         __split_huge_page_refcount(page, list);
1873
1874         mapcount2 = 0;
1875         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1876                 struct vm_area_struct *vma = avc->vma;
1877                 unsigned long addr = vma_address(page, vma);
1878                 BUG_ON(is_vma_temporary_stack(vma));
1879                 mapcount2 += __split_huge_page_map(page, vma, addr);
1880         }
1881         if (mapcount != mapcount2)
1882                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1883                        mapcount, mapcount2, page_mapcount(page));
1884         BUG_ON(mapcount != mapcount2);
1885 }
1886
1887 /*
1888  * Split a hugepage into normal pages. This doesn't change the position of head
1889  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1890  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1891  * from the hugepage.
1892  * Return 0 if the hugepage is split successfully otherwise return 1.
1893  */
1894 int split_huge_page_to_list(struct page *page, struct list_head *list)
1895 {
1896         struct anon_vma *anon_vma;
1897         int ret = 1;
1898
1899         BUG_ON(is_huge_zero_page(page));
1900         BUG_ON(!PageAnon(page));
1901
1902         /*
1903          * The caller does not necessarily hold an mmap_sem that would prevent
1904          * the anon_vma disappearing so we first we take a reference to it
1905          * and then lock the anon_vma for write. This is similar to
1906          * page_lock_anon_vma_read except the write lock is taken to serialise
1907          * against parallel split or collapse operations.
1908          */
1909         anon_vma = page_get_anon_vma(page);
1910         if (!anon_vma)
1911                 goto out;
1912         anon_vma_lock_write(anon_vma);
1913
1914         ret = 0;
1915         if (!PageCompound(page))
1916                 goto out_unlock;
1917
1918         BUG_ON(!PageSwapBacked(page));
1919         __split_huge_page(page, anon_vma, list);
1920         count_vm_event(THP_SPLIT);
1921
1922         BUG_ON(PageCompound(page));
1923 out_unlock:
1924         anon_vma_unlock_write(anon_vma);
1925         put_anon_vma(anon_vma);
1926 out:
1927         return ret;
1928 }
1929
1930 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1931
1932 int hugepage_madvise(struct vm_area_struct *vma,
1933                      unsigned long *vm_flags, int advice)
1934 {
1935         struct mm_struct *mm = vma->vm_mm;
1936
1937         switch (advice) {
1938         case MADV_HUGEPAGE:
1939                 /*
1940                  * Be somewhat over-protective like KSM for now!
1941                  */
1942                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1943                         return -EINVAL;
1944                 if (mm->def_flags & VM_NOHUGEPAGE)
1945                         return -EINVAL;
1946                 *vm_flags &= ~VM_NOHUGEPAGE;
1947                 *vm_flags |= VM_HUGEPAGE;
1948                 /*
1949                  * If the vma become good for khugepaged to scan,
1950                  * register it here without waiting a page fault that
1951                  * may not happen any time soon.
1952                  */
1953                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1954                         return -ENOMEM;
1955                 break;
1956         case MADV_NOHUGEPAGE:
1957                 /*
1958                  * Be somewhat over-protective like KSM for now!
1959                  */
1960                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1961                         return -EINVAL;
1962                 *vm_flags &= ~VM_HUGEPAGE;
1963                 *vm_flags |= VM_NOHUGEPAGE;
1964                 /*
1965                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1966                  * this vma even if we leave the mm registered in khugepaged if
1967                  * it got registered before VM_NOHUGEPAGE was set.
1968                  */
1969                 break;
1970         }
1971
1972         return 0;
1973 }
1974
1975 static int __init khugepaged_slab_init(void)
1976 {
1977         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1978                                           sizeof(struct mm_slot),
1979                                           __alignof__(struct mm_slot), 0, NULL);
1980         if (!mm_slot_cache)
1981                 return -ENOMEM;
1982
1983         return 0;
1984 }
1985
1986 static inline struct mm_slot *alloc_mm_slot(void)
1987 {
1988         if (!mm_slot_cache)     /* initialization failed */
1989                 return NULL;
1990         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1991 }
1992
1993 static inline void free_mm_slot(struct mm_slot *mm_slot)
1994 {
1995         kmem_cache_free(mm_slot_cache, mm_slot);
1996 }
1997
1998 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1999 {
2000         struct mm_slot *mm_slot;
2001
2002         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2003                 if (mm == mm_slot->mm)
2004                         return mm_slot;
2005
2006         return NULL;
2007 }
2008
2009 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2010                                     struct mm_slot *mm_slot)
2011 {
2012         mm_slot->mm = mm;
2013         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2014 }
2015
2016 static inline int khugepaged_test_exit(struct mm_struct *mm)
2017 {
2018         return atomic_read(&mm->mm_users) == 0;
2019 }
2020
2021 int __khugepaged_enter(struct mm_struct *mm)
2022 {
2023         struct mm_slot *mm_slot;
2024         int wakeup;
2025
2026         mm_slot = alloc_mm_slot();
2027         if (!mm_slot)
2028                 return -ENOMEM;
2029
2030         /* __khugepaged_exit() must not run from under us */
2031         VM_BUG_ON(khugepaged_test_exit(mm));
2032         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2033                 free_mm_slot(mm_slot);
2034                 return 0;
2035         }
2036
2037         spin_lock(&khugepaged_mm_lock);
2038         insert_to_mm_slots_hash(mm, mm_slot);
2039         /*
2040          * Insert just behind the scanning cursor, to let the area settle
2041          * down a little.
2042          */
2043         wakeup = list_empty(&khugepaged_scan.mm_head);
2044         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2045         spin_unlock(&khugepaged_mm_lock);
2046
2047         atomic_inc(&mm->mm_count);
2048         if (wakeup)
2049                 wake_up_interruptible(&khugepaged_wait);
2050
2051         return 0;
2052 }
2053
2054 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2055 {
2056         unsigned long hstart, hend;
2057         if (!vma->anon_vma)
2058                 /*
2059                  * Not yet faulted in so we will register later in the
2060                  * page fault if needed.
2061                  */
2062                 return 0;
2063         if (vma->vm_ops)
2064                 /* khugepaged not yet working on file or special mappings */
2065                 return 0;
2066         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2067         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2068         hend = vma->vm_end & HPAGE_PMD_MASK;
2069         if (hstart < hend)
2070                 return khugepaged_enter(vma);
2071         return 0;
2072 }
2073
2074 void __khugepaged_exit(struct mm_struct *mm)
2075 {
2076         struct mm_slot *mm_slot;
2077         int free = 0;
2078
2079         spin_lock(&khugepaged_mm_lock);
2080         mm_slot = get_mm_slot(mm);
2081         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2082                 hash_del(&mm_slot->hash);
2083                 list_del(&mm_slot->mm_node);
2084                 free = 1;
2085         }
2086         spin_unlock(&khugepaged_mm_lock);
2087
2088         if (free) {
2089                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2090                 free_mm_slot(mm_slot);
2091                 mmdrop(mm);
2092         } else if (mm_slot) {
2093                 /*
2094                  * This is required to serialize against
2095                  * khugepaged_test_exit() (which is guaranteed to run
2096                  * under mmap sem read mode). Stop here (after we
2097                  * return all pagetables will be destroyed) until
2098                  * khugepaged has finished working on the pagetables
2099                  * under the mmap_sem.
2100                  */
2101                 down_write(&mm->mmap_sem);
2102                 up_write(&mm->mmap_sem);
2103         }
2104 }
2105
2106 static void release_pte_page(struct page *page)
2107 {
2108         /* 0 stands for page_is_file_cache(page) == false */
2109         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2110         unlock_page(page);
2111         putback_lru_page(page);
2112 }
2113
2114 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2115 {
2116         while (--_pte >= pte) {
2117                 pte_t pteval = *_pte;
2118                 if (!pte_none(pteval))
2119                         release_pte_page(pte_page(pteval));
2120         }
2121 }
2122
2123 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2124                                         unsigned long address,
2125                                         pte_t *pte)
2126 {
2127         struct page *page;
2128         pte_t *_pte;
2129         int referenced = 0, none = 0;
2130         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2131              _pte++, address += PAGE_SIZE) {
2132                 pte_t pteval = *_pte;
2133                 if (pte_none(pteval)) {
2134                         if (++none <= khugepaged_max_ptes_none)
2135                                 continue;
2136                         else
2137                                 goto out;
2138                 }
2139                 if (!pte_present(pteval) || !pte_write(pteval))
2140                         goto out;
2141                 page = vm_normal_page(vma, address, pteval);
2142                 if (unlikely(!page))
2143                         goto out;
2144
2145                 VM_BUG_ON(PageCompound(page));
2146                 BUG_ON(!PageAnon(page));
2147                 VM_BUG_ON(!PageSwapBacked(page));
2148
2149                 /* cannot use mapcount: can't collapse if there's a gup pin */
2150                 if (page_count(page) != 1)
2151                         goto out;
2152                 /*
2153                  * We can do it before isolate_lru_page because the
2154                  * page can't be freed from under us. NOTE: PG_lock
2155                  * is needed to serialize against split_huge_page
2156                  * when invoked from the VM.
2157                  */
2158                 if (!trylock_page(page))
2159                         goto out;
2160                 /*
2161                  * Isolate the page to avoid collapsing an hugepage
2162                  * currently in use by the VM.
2163                  */
2164                 if (isolate_lru_page(page)) {
2165                         unlock_page(page);
2166                         goto out;
2167                 }
2168                 /* 0 stands for page_is_file_cache(page) == false */
2169                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2170                 VM_BUG_ON(!PageLocked(page));
2171                 VM_BUG_ON(PageLRU(page));
2172
2173                 /* If there is no mapped pte young don't collapse the page */
2174                 if (pte_young(pteval) || PageReferenced(page) ||
2175                     mmu_notifier_test_young(vma->vm_mm, address))
2176                         referenced = 1;
2177         }
2178         if (likely(referenced))
2179                 return 1;
2180 out:
2181         release_pte_pages(pte, _pte);
2182         return 0;
2183 }
2184
2185 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2186                                       struct vm_area_struct *vma,
2187                                       unsigned long address,
2188                                       spinlock_t *ptl)
2189 {
2190         pte_t *_pte;
2191         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2192                 pte_t pteval = *_pte;
2193                 struct page *src_page;
2194
2195                 if (pte_none(pteval)) {
2196                         clear_user_highpage(page, address);
2197                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2198                 } else {
2199                         src_page = pte_page(pteval);
2200                         copy_user_highpage(page, src_page, address, vma);
2201                         VM_BUG_ON(page_mapcount(src_page) != 1);
2202                         release_pte_page(src_page);
2203                         /*
2204                          * ptl mostly unnecessary, but preempt has to
2205                          * be disabled to update the per-cpu stats
2206                          * inside page_remove_rmap().
2207                          */
2208                         spin_lock(ptl);
2209                         /*
2210                          * paravirt calls inside pte_clear here are
2211                          * superfluous.
2212                          */
2213                         pte_clear(vma->vm_mm, address, _pte);
2214                         page_remove_rmap(src_page);
2215                         spin_unlock(ptl);
2216                         free_page_and_swap_cache(src_page);
2217                 }
2218
2219                 address += PAGE_SIZE;
2220                 page++;
2221         }
2222 }
2223
2224 static void khugepaged_alloc_sleep(void)
2225 {
2226         wait_event_freezable_timeout(khugepaged_wait, false,
2227                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2228 }
2229
2230 static int khugepaged_node_load[MAX_NUMNODES];
2231
2232 #ifdef CONFIG_NUMA
2233 static int khugepaged_find_target_node(void)
2234 {
2235         static int last_khugepaged_target_node = NUMA_NO_NODE;
2236         int nid, target_node = 0, max_value = 0;
2237
2238         /* find first node with max normal pages hit */
2239         for (nid = 0; nid < MAX_NUMNODES; nid++)
2240                 if (khugepaged_node_load[nid] > max_value) {
2241                         max_value = khugepaged_node_load[nid];
2242                         target_node = nid;
2243                 }
2244
2245         /* do some balance if several nodes have the same hit record */
2246         if (target_node <= last_khugepaged_target_node)
2247                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2248                                 nid++)
2249                         if (max_value == khugepaged_node_load[nid]) {
2250                                 target_node = nid;
2251                                 break;
2252                         }
2253
2254         last_khugepaged_target_node = target_node;
2255         return target_node;
2256 }
2257
2258 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2259 {
2260         if (IS_ERR(*hpage)) {
2261                 if (!*wait)
2262                         return false;
2263
2264                 *wait = false;
2265                 *hpage = NULL;
2266                 khugepaged_alloc_sleep();
2267         } else if (*hpage) {
2268                 put_page(*hpage);
2269                 *hpage = NULL;
2270         }
2271
2272         return true;
2273 }
2274
2275 static struct page
2276 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2277                        struct vm_area_struct *vma, unsigned long address,
2278                        int node)
2279 {
2280         VM_BUG_ON(*hpage);
2281         /*
2282          * Allocate the page while the vma is still valid and under
2283          * the mmap_sem read mode so there is no memory allocation
2284          * later when we take the mmap_sem in write mode. This is more
2285          * friendly behavior (OTOH it may actually hide bugs) to
2286          * filesystems in userland with daemons allocating memory in
2287          * the userland I/O paths.  Allocating memory with the
2288          * mmap_sem in read mode is good idea also to allow greater
2289          * scalability.
2290          */
2291         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2292                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2293         /*
2294          * After allocating the hugepage, release the mmap_sem read lock in
2295          * preparation for taking it in write mode.
2296          */
2297         up_read(&mm->mmap_sem);
2298         if (unlikely(!*hpage)) {
2299                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2300                 *hpage = ERR_PTR(-ENOMEM);
2301                 return NULL;
2302         }
2303
2304         count_vm_event(THP_COLLAPSE_ALLOC);
2305         return *hpage;
2306 }
2307 #else
2308 static int khugepaged_find_target_node(void)
2309 {
2310         return 0;
2311 }
2312
2313 static inline struct page *alloc_hugepage(int defrag)
2314 {
2315         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2316                            HPAGE_PMD_ORDER);
2317 }
2318
2319 static struct page *khugepaged_alloc_hugepage(bool *wait)
2320 {
2321         struct page *hpage;
2322
2323         do {
2324                 hpage = alloc_hugepage(khugepaged_defrag());
2325                 if (!hpage) {
2326                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2327                         if (!*wait)
2328                                 return NULL;
2329
2330                         *wait = false;
2331                         khugepaged_alloc_sleep();
2332                 } else
2333                         count_vm_event(THP_COLLAPSE_ALLOC);
2334         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2335
2336         return hpage;
2337 }
2338
2339 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2340 {
2341         if (!*hpage)
2342                 *hpage = khugepaged_alloc_hugepage(wait);
2343
2344         if (unlikely(!*hpage))
2345                 return false;
2346
2347         return true;
2348 }
2349
2350 static struct page
2351 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2352                        struct vm_area_struct *vma, unsigned long address,
2353                        int node)
2354 {
2355         up_read(&mm->mmap_sem);
2356         VM_BUG_ON(!*hpage);
2357         return  *hpage;
2358 }
2359 #endif
2360
2361 static bool hugepage_vma_check(struct vm_area_struct *vma)
2362 {
2363         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2364             (vma->vm_flags & VM_NOHUGEPAGE))
2365                 return false;
2366
2367         if (!vma->anon_vma || vma->vm_ops)
2368                 return false;
2369         if (is_vma_temporary_stack(vma))
2370                 return false;
2371         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2372         return true;
2373 }
2374
2375 static void collapse_huge_page(struct mm_struct *mm,
2376                                    unsigned long address,
2377                                    struct page **hpage,
2378                                    struct vm_area_struct *vma,
2379                                    int node)
2380 {
2381         pmd_t *pmd, _pmd;
2382         pte_t *pte;
2383         pgtable_t pgtable;
2384         struct page *new_page;
2385         spinlock_t *pmd_ptl, *pte_ptl;
2386         int isolated;
2387         unsigned long hstart, hend;
2388         unsigned long mmun_start;       /* For mmu_notifiers */
2389         unsigned long mmun_end;         /* For mmu_notifiers */
2390
2391         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2392
2393         /* release the mmap_sem read lock. */
2394         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2395         if (!new_page)
2396                 return;
2397
2398         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2399                 return;
2400
2401         /*
2402          * Prevent all access to pagetables with the exception of
2403          * gup_fast later hanlded by the ptep_clear_flush and the VM
2404          * handled by the anon_vma lock + PG_lock.
2405          */
2406         down_write(&mm->mmap_sem);
2407         if (unlikely(khugepaged_test_exit(mm)))
2408                 goto out;
2409
2410         vma = find_vma(mm, address);
2411         if (!vma)
2412                 goto out;
2413         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2414         hend = vma->vm_end & HPAGE_PMD_MASK;
2415         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2416                 goto out;
2417         if (!hugepage_vma_check(vma))
2418                 goto out;
2419         pmd = mm_find_pmd(mm, address);
2420         if (!pmd)
2421                 goto out;
2422         if (pmd_trans_huge(*pmd))
2423                 goto out;
2424
2425         anon_vma_lock_write(vma->anon_vma);
2426
2427         pte = pte_offset_map(pmd, address);
2428         pte_ptl = pte_lockptr(mm, pmd);
2429
2430         mmun_start = address;
2431         mmun_end   = address + HPAGE_PMD_SIZE;
2432         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2433         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2434         /*
2435          * After this gup_fast can't run anymore. This also removes
2436          * any huge TLB entry from the CPU so we won't allow
2437          * huge and small TLB entries for the same virtual address
2438          * to avoid the risk of CPU bugs in that area.
2439          */
2440         _pmd = pmdp_clear_flush(vma, address, pmd);
2441         spin_unlock(pmd_ptl);
2442         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2443
2444         spin_lock(pte_ptl);
2445         isolated = __collapse_huge_page_isolate(vma, address, pte);
2446         spin_unlock(pte_ptl);
2447
2448         if (unlikely(!isolated)) {
2449                 pte_unmap(pte);
2450                 spin_lock(pmd_ptl);
2451                 BUG_ON(!pmd_none(*pmd));
2452                 /*
2453                  * We can only use set_pmd_at when establishing
2454                  * hugepmds and never for establishing regular pmds that
2455                  * points to regular pagetables. Use pmd_populate for that
2456                  */
2457                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2458                 spin_unlock(pmd_ptl);
2459                 anon_vma_unlock_write(vma->anon_vma);
2460                 goto out;
2461         }
2462
2463         /*
2464          * All pages are isolated and locked so anon_vma rmap
2465          * can't run anymore.
2466          */
2467         anon_vma_unlock_write(vma->anon_vma);
2468
2469         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2470         pte_unmap(pte);
2471         __SetPageUptodate(new_page);
2472         pgtable = pmd_pgtable(_pmd);
2473
2474         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2475         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2476
2477         /*
2478          * spin_lock() below is not the equivalent of smp_wmb(), so
2479          * this is needed to avoid the copy_huge_page writes to become
2480          * visible after the set_pmd_at() write.
2481          */
2482         smp_wmb();
2483
2484         spin_lock(pmd_ptl);
2485         BUG_ON(!pmd_none(*pmd));
2486         page_add_new_anon_rmap(new_page, vma, address);
2487         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2488         set_pmd_at(mm, address, pmd, _pmd);
2489         update_mmu_cache_pmd(vma, address, pmd);
2490         spin_unlock(pmd_ptl);
2491
2492         *hpage = NULL;
2493
2494         khugepaged_pages_collapsed++;
2495 out_up_write:
2496         up_write(&mm->mmap_sem);
2497         return;
2498
2499 out:
2500         mem_cgroup_uncharge_page(new_page);
2501         goto out_up_write;
2502 }
2503
2504 static int khugepaged_scan_pmd(struct mm_struct *mm,
2505                                struct vm_area_struct *vma,
2506                                unsigned long address,
2507                                struct page **hpage)
2508 {
2509         pmd_t *pmd;
2510         pte_t *pte, *_pte;
2511         int ret = 0, referenced = 0, none = 0;
2512         struct page *page;
2513         unsigned long _address;
2514         spinlock_t *ptl;
2515         int node = NUMA_NO_NODE;
2516
2517         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2518
2519         pmd = mm_find_pmd(mm, address);
2520         if (!pmd)
2521                 goto out;
2522         if (pmd_trans_huge(*pmd))
2523                 goto out;
2524
2525         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2526         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2527         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2528              _pte++, _address += PAGE_SIZE) {
2529                 pte_t pteval = *_pte;
2530                 if (pte_none(pteval)) {
2531                         if (++none <= khugepaged_max_ptes_none)
2532                                 continue;
2533                         else
2534                                 goto out_unmap;
2535                 }
2536                 if (!pte_present(pteval) || !pte_write(pteval))
2537                         goto out_unmap;
2538                 page = vm_normal_page(vma, _address, pteval);
2539                 if (unlikely(!page))
2540                         goto out_unmap;
2541                 /*
2542                  * Record which node the original page is from and save this
2543                  * information to khugepaged_node_load[].
2544                  * Khupaged will allocate hugepage from the node has the max
2545                  * hit record.
2546                  */
2547                 node = page_to_nid(page);
2548                 khugepaged_node_load[node]++;
2549                 VM_BUG_ON(PageCompound(page));
2550                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2551                         goto out_unmap;
2552                 /* cannot use mapcount: can't collapse if there's a gup pin */
2553                 if (page_count(page) != 1)
2554                         goto out_unmap;
2555                 if (pte_young(pteval) || PageReferenced(page) ||
2556                     mmu_notifier_test_young(vma->vm_mm, address))
2557                         referenced = 1;
2558         }
2559         if (referenced)
2560                 ret = 1;
2561 out_unmap:
2562         pte_unmap_unlock(pte, ptl);
2563         if (ret) {
2564                 node = khugepaged_find_target_node();
2565                 /* collapse_huge_page will return with the mmap_sem released */
2566                 collapse_huge_page(mm, address, hpage, vma, node);
2567         }
2568 out:
2569         return ret;
2570 }
2571
2572 static void collect_mm_slot(struct mm_slot *mm_slot)
2573 {
2574         struct mm_struct *mm = mm_slot->mm;
2575
2576         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2577
2578         if (khugepaged_test_exit(mm)) {
2579                 /* free mm_slot */
2580                 hash_del(&mm_slot->hash);
2581                 list_del(&mm_slot->mm_node);
2582
2583                 /*
2584                  * Not strictly needed because the mm exited already.
2585                  *
2586                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2587                  */
2588
2589                 /* khugepaged_mm_lock actually not necessary for the below */
2590                 free_mm_slot(mm_slot);
2591                 mmdrop(mm);
2592         }
2593 }
2594
2595 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2596                                             struct page **hpage)
2597         __releases(&khugepaged_mm_lock)
2598         __acquires(&khugepaged_mm_lock)
2599 {
2600         struct mm_slot *mm_slot;
2601         struct mm_struct *mm;
2602         struct vm_area_struct *vma;
2603         int progress = 0;
2604
2605         VM_BUG_ON(!pages);
2606         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2607
2608         if (khugepaged_scan.mm_slot)
2609                 mm_slot = khugepaged_scan.mm_slot;
2610         else {
2611                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2612                                      struct mm_slot, mm_node);
2613                 khugepaged_scan.address = 0;
2614                 khugepaged_scan.mm_slot = mm_slot;
2615         }
2616         spin_unlock(&khugepaged_mm_lock);
2617
2618         mm = mm_slot->mm;
2619         down_read(&mm->mmap_sem);
2620         if (unlikely(khugepaged_test_exit(mm)))
2621                 vma = NULL;
2622         else
2623                 vma = find_vma(mm, khugepaged_scan.address);
2624
2625         progress++;
2626         for (; vma; vma = vma->vm_next) {
2627                 unsigned long hstart, hend;
2628
2629                 cond_resched();
2630                 if (unlikely(khugepaged_test_exit(mm))) {
2631                         progress++;
2632                         break;
2633                 }
2634                 if (!hugepage_vma_check(vma)) {
2635 skip:
2636                         progress++;
2637                         continue;
2638                 }
2639                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2640                 hend = vma->vm_end & HPAGE_PMD_MASK;
2641                 if (hstart >= hend)
2642                         goto skip;
2643                 if (khugepaged_scan.address > hend)
2644                         goto skip;
2645                 if (khugepaged_scan.address < hstart)
2646                         khugepaged_scan.address = hstart;
2647                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2648
2649                 while (khugepaged_scan.address < hend) {
2650                         int ret;
2651                         cond_resched();
2652                         if (unlikely(khugepaged_test_exit(mm)))
2653                                 goto breakouterloop;
2654
2655                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2656                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2657                                   hend);
2658                         ret = khugepaged_scan_pmd(mm, vma,
2659                                                   khugepaged_scan.address,
2660                                                   hpage);
2661                         /* move to next address */
2662                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2663                         progress += HPAGE_PMD_NR;
2664                         if (ret)
2665                                 /* we released mmap_sem so break loop */
2666                                 goto breakouterloop_mmap_sem;
2667                         if (progress >= pages)
2668                                 goto breakouterloop;
2669                 }
2670         }
2671 breakouterloop:
2672         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2673 breakouterloop_mmap_sem:
2674
2675         spin_lock(&khugepaged_mm_lock);
2676         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2677         /*
2678          * Release the current mm_slot if this mm is about to die, or
2679          * if we scanned all vmas of this mm.
2680          */
2681         if (khugepaged_test_exit(mm) || !vma) {
2682                 /*
2683                  * Make sure that if mm_users is reaching zero while
2684                  * khugepaged runs here, khugepaged_exit will find
2685                  * mm_slot not pointing to the exiting mm.
2686                  */
2687                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2688                         khugepaged_scan.mm_slot = list_entry(
2689                                 mm_slot->mm_node.next,
2690                                 struct mm_slot, mm_node);
2691                         khugepaged_scan.address = 0;
2692                 } else {
2693                         khugepaged_scan.mm_slot = NULL;
2694                         khugepaged_full_scans++;
2695                 }
2696
2697                 collect_mm_slot(mm_slot);
2698         }
2699
2700         return progress;
2701 }
2702
2703 static int khugepaged_has_work(void)
2704 {
2705         return !list_empty(&khugepaged_scan.mm_head) &&
2706                 khugepaged_enabled();
2707 }
2708
2709 static int khugepaged_wait_event(void)
2710 {
2711         return !list_empty(&khugepaged_scan.mm_head) ||
2712                 kthread_should_stop();
2713 }
2714
2715 static void khugepaged_do_scan(void)
2716 {
2717         struct page *hpage = NULL;
2718         unsigned int progress = 0, pass_through_head = 0;
2719         unsigned int pages = khugepaged_pages_to_scan;
2720         bool wait = true;
2721
2722         barrier(); /* write khugepaged_pages_to_scan to local stack */
2723
2724         while (progress < pages) {
2725                 if (!khugepaged_prealloc_page(&hpage, &wait))
2726                         break;
2727
2728                 cond_resched();
2729
2730                 if (unlikely(kthread_should_stop() || freezing(current)))
2731                         break;
2732
2733                 spin_lock(&khugepaged_mm_lock);
2734                 if (!khugepaged_scan.mm_slot)
2735                         pass_through_head++;
2736                 if (khugepaged_has_work() &&
2737                     pass_through_head < 2)
2738                         progress += khugepaged_scan_mm_slot(pages - progress,
2739                                                             &hpage);
2740                 else
2741                         progress = pages;
2742                 spin_unlock(&khugepaged_mm_lock);
2743         }
2744
2745         if (!IS_ERR_OR_NULL(hpage))
2746                 put_page(hpage);
2747 }
2748
2749 static void khugepaged_wait_work(void)
2750 {
2751         try_to_freeze();
2752
2753         if (khugepaged_has_work()) {
2754                 if (!khugepaged_scan_sleep_millisecs)
2755                         return;
2756
2757                 wait_event_freezable_timeout(khugepaged_wait,
2758                                              kthread_should_stop(),
2759                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2760                 return;
2761         }
2762
2763         if (khugepaged_enabled())
2764                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2765 }
2766
2767 static int khugepaged(void *none)
2768 {
2769         struct mm_slot *mm_slot;
2770
2771         set_freezable();
2772         set_user_nice(current, 19);
2773
2774         while (!kthread_should_stop()) {
2775                 khugepaged_do_scan();
2776                 khugepaged_wait_work();
2777         }
2778
2779         spin_lock(&khugepaged_mm_lock);
2780         mm_slot = khugepaged_scan.mm_slot;
2781         khugepaged_scan.mm_slot = NULL;
2782         if (mm_slot)
2783                 collect_mm_slot(mm_slot);
2784         spin_unlock(&khugepaged_mm_lock);
2785         return 0;
2786 }
2787
2788 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2789                 unsigned long haddr, pmd_t *pmd)
2790 {
2791         struct mm_struct *mm = vma->vm_mm;
2792         pgtable_t pgtable;
2793         pmd_t _pmd;
2794         int i;
2795
2796         pmdp_clear_flush(vma, haddr, pmd);
2797         /* leave pmd empty until pte is filled */
2798
2799         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2800         pmd_populate(mm, &_pmd, pgtable);
2801
2802         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2803                 pte_t *pte, entry;
2804                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2805                 entry = pte_mkspecial(entry);
2806                 pte = pte_offset_map(&_pmd, haddr);
2807                 VM_BUG_ON(!pte_none(*pte));
2808                 set_pte_at(mm, haddr, pte, entry);
2809                 pte_unmap(pte);
2810         }
2811         smp_wmb(); /* make pte visible before pmd */
2812         pmd_populate(mm, pmd, pgtable);
2813         put_huge_zero_page();
2814 }
2815
2816 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2817                 pmd_t *pmd)
2818 {
2819         spinlock_t *ptl;
2820         struct page *page;
2821         struct mm_struct *mm = vma->vm_mm;
2822         unsigned long haddr = address & HPAGE_PMD_MASK;
2823         unsigned long mmun_start;       /* For mmu_notifiers */
2824         unsigned long mmun_end;         /* For mmu_notifiers */
2825
2826         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2827
2828         mmun_start = haddr;
2829         mmun_end   = haddr + HPAGE_PMD_SIZE;
2830 again:
2831         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2832         ptl = pmd_lock(mm, pmd);
2833         if (unlikely(!pmd_trans_huge(*pmd))) {
2834                 spin_unlock(ptl);
2835                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2836                 return;
2837         }
2838         if (is_huge_zero_pmd(*pmd)) {
2839                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2840                 spin_unlock(ptl);
2841                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2842                 return;
2843         }
2844         page = pmd_page(*pmd);
2845         VM_BUG_ON(!page_count(page));
2846         get_page(page);
2847         spin_unlock(ptl);
2848         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2849
2850         split_huge_page(page);
2851
2852         put_page(page);
2853
2854         /*
2855          * We don't always have down_write of mmap_sem here: a racing
2856          * do_huge_pmd_wp_page() might have copied-on-write to another
2857          * huge page before our split_huge_page() got the anon_vma lock.
2858          */
2859         if (unlikely(pmd_trans_huge(*pmd)))
2860                 goto again;
2861 }
2862
2863 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2864                 pmd_t *pmd)
2865 {
2866         struct vm_area_struct *vma;
2867
2868         vma = find_vma(mm, address);
2869         BUG_ON(vma == NULL);
2870         split_huge_page_pmd(vma, address, pmd);
2871 }
2872
2873 static void split_huge_page_address(struct mm_struct *mm,
2874                                     unsigned long address)
2875 {
2876         pmd_t *pmd;
2877
2878         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2879
2880         pmd = mm_find_pmd(mm, address);
2881         if (!pmd)
2882                 return;
2883         /*
2884          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2885          * materialize from under us.
2886          */
2887         split_huge_page_pmd_mm(mm, address, pmd);
2888 }
2889
2890 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2891                              unsigned long start,
2892                              unsigned long end,
2893                              long adjust_next)
2894 {
2895         /*
2896          * If the new start address isn't hpage aligned and it could
2897          * previously contain an hugepage: check if we need to split
2898          * an huge pmd.
2899          */
2900         if (start & ~HPAGE_PMD_MASK &&
2901             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2902             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2903                 split_huge_page_address(vma->vm_mm, start);
2904
2905         /*
2906          * If the new end address isn't hpage aligned and it could
2907          * previously contain an hugepage: check if we need to split
2908          * an huge pmd.
2909          */
2910         if (end & ~HPAGE_PMD_MASK &&
2911             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2912             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2913                 split_huge_page_address(vma->vm_mm, end);
2914
2915         /*
2916          * If we're also updating the vma->vm_next->vm_start, if the new
2917          * vm_next->vm_start isn't page aligned and it could previously
2918          * contain an hugepage: check if we need to split an huge pmd.
2919          */
2920         if (adjust_next > 0) {
2921                 struct vm_area_struct *next = vma->vm_next;
2922                 unsigned long nstart = next->vm_start;
2923                 nstart += adjust_next << PAGE_SHIFT;
2924                 if (nstart & ~HPAGE_PMD_MASK &&
2925                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2926                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2927                         split_huge_page_address(next->vm_mm, nstart);
2928         }
2929 }