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