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