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