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         /*
 717          * The memory barrier inside __SetPageUptodate makes sure that
 718          * clear_huge_page writes become visible before the set_pmd_at()
 719          * write.
 720          */
 721         __SetPageUptodate(page);
 722
 723         spin_lock(&mm->page_table_lock);
 724         if (unlikely(!pmd_none(*pmd))) {
 725                 spin_unlock(&mm->page_table_lock);
 726                 mem_cgroup_uncharge_page(page);
 727                 put_page(page);
 728                 pte_free(mm, pgtable);
 729         } else {
 730                 pmd_t entry;
 731                 entry = mk_huge_pmd(page, vma);
 732                 page_add_new_anon_rmap(page, vma, haddr);
 733                 set_pmd_at(mm, haddr, pmd, entry);
 734                 pgtable_trans_huge_deposit(mm, pgtable);
 735                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 736                 mm->nr_ptes++;
 737                 spin_unlock(&mm->page_table_lock);
 738         }
 739
 740         return 0;
 741 }
 742
 743 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 744 {
 745         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 746 }
 747
 748 static inline struct page *alloc_hugepage_vma(int defrag,
 749                                               struct vm_area_struct *vma,
 750                                               unsigned long haddr, int nd,
 751                                               gfp_t extra_gfp)
 752 {
 753         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 754                                HPAGE_PMD_ORDER, vma, haddr, nd);
 755 }
 756
 757 #ifndef CONFIG_NUMA
 758 static inline struct page *alloc_hugepage(int defrag)
 759 {
 760         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 761                            HPAGE_PMD_ORDER);
 762 }
 763 #endif
 764
 765 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 766                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 767                 unsigned long zero_pfn)
 768 {
 769         pmd_t entry;
 770         if (!pmd_none(*pmd))
 771                 return false;
 772         entry = pfn_pmd(zero_pfn, vma->vm_page_prot);
 773         entry = pmd_wrprotect(entry);
 774         entry = pmd_mkhuge(entry);
 775         set_pmd_at(mm, haddr, pmd, entry);
 776         pgtable_trans_huge_deposit(mm, pgtable);
 777         mm->nr_ptes++;
 778         return true;
 779 }
 780
 781 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 782                                unsigned long address, pmd_t *pmd,
 783                                unsigned int flags)
 784 {
 785         struct page *page;
 786         unsigned long haddr = address & HPAGE_PMD_MASK;
 787         pte_t *pte;
 788
 789         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 790                 if (unlikely(anon_vma_prepare(vma)))
 791                         return VM_FAULT_OOM;
 792                 if (unlikely(khugepaged_enter(vma)))
 793                         return VM_FAULT_OOM;
 794                 if (!(flags & FAULT_FLAG_WRITE) &&
 795                                 transparent_hugepage_use_zero_page()) {
 796                         pgtable_t pgtable;
 797                         unsigned long zero_pfn;
 798                         bool set;
 799                         pgtable = pte_alloc_one(mm, haddr);
 800                         if (unlikely(!pgtable))
 801                                 return VM_FAULT_OOM;
 802                         zero_pfn = get_huge_zero_page();
 803                         if (unlikely(!zero_pfn)) {
 804                                 pte_free(mm, pgtable);
 805                                 count_vm_event(THP_FAULT_FALLBACK);
 806                                 goto out;
 807                         }
 808                         spin_lock(&mm->page_table_lock);
 809                         set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 810                                         zero_pfn);
 811                         spin_unlock(&mm->page_table_lock);
 812                         if (!set) {
 813                                 pte_free(mm, pgtable);
 814                                 put_huge_zero_page();
 815                         }
 816                         return 0;
 817                 }
 818                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 819                                           vma, haddr, numa_node_id(), 0);
 820                 if (unlikely(!page)) {
 821                         count_vm_event(THP_FAULT_FALLBACK);
 822                         goto out;
 823                 }
 824                 count_vm_event(THP_FAULT_ALLOC);
 825                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 826                         put_page(page);
 827                         goto out;
 828                 }
 829                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 830                                                           page))) {
 831                         mem_cgroup_uncharge_page(page);
 832                         put_page(page);
 833                         goto out;
 834                 }
 835
 836                 return 0;
 837         }
 838 out:
 839         /*
 840          * Use __pte_alloc instead of pte_alloc_map, because we can't
 841          * run pte_offset_map on the pmd, if an huge pmd could
 842          * materialize from under us from a different thread.
 843          */
 844         if (unlikely(pmd_none(*pmd)) &&
 845             unlikely(__pte_alloc(mm, vma, pmd, address)))
 846                 return VM_FAULT_OOM;
 847         /* if an huge pmd materialized from under us just retry later */
 848         if (unlikely(pmd_trans_huge(*pmd)))
 849                 return 0;
 850         /*
 851          * A regular pmd is established and it can't morph into a huge pmd
 852          * from under us anymore at this point because we hold the mmap_sem
 853          * read mode and khugepaged takes it in write mode. So now it's
 854          * safe to run pte_offset_map().
 855          */
 856         pte = pte_offset_map(pmd, address);
 857         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 858 }
 859
 860 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 861                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 862                   struct vm_area_struct *vma)
 863 {
 864         struct page *src_page;
 865         pmd_t pmd;
 866         pgtable_t pgtable;
 867         int ret;
 868
 869         ret = -ENOMEM;
 870         pgtable = pte_alloc_one(dst_mm, addr);
 871         if (unlikely(!pgtable))
 872                 goto out;
 873
 874         spin_lock(&dst_mm->page_table_lock);
 875         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 876
 877         ret = -EAGAIN;
 878         pmd = *src_pmd;
 879         if (unlikely(!pmd_trans_huge(pmd))) {
 880                 pte_free(dst_mm, pgtable);
 881                 goto out_unlock;
 882         }
 883         /*
 884          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 885          * under splitting since we don't split the page itself, only pmd to
 886          * a page table.
 887          */
 888         if (is_huge_zero_pmd(pmd)) {
 889                 unsigned long zero_pfn;
 890                 bool set;
 891                 /*
 892                  * get_huge_zero_page() will never allocate a new page here,
 893                  * since we already have a zero page to copy. It just takes a
 894                  * reference.
 895                  */
 896                 zero_pfn = get_huge_zero_page();
 897                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 898                                 zero_pfn);
 899                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 900                 ret = 0;
 901                 goto out_unlock;
 902         }
 903         if (unlikely(pmd_trans_splitting(pmd))) {
 904                 /* split huge page running from under us */
 905                 spin_unlock(&src_mm->page_table_lock);
 906                 spin_unlock(&dst_mm->page_table_lock);
 907                 pte_free(dst_mm, pgtable);
 908
 909                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 910                 goto out;
 911         }
 912         src_page = pmd_page(pmd);
 913         VM_BUG_ON(!PageHead(src_page));
 914         get_page(src_page);
 915         page_dup_rmap(src_page);
 916         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 917
 918         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 919         pmd = pmd_mkold(pmd_wrprotect(pmd));
 920         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 921         pgtable_trans_huge_deposit(dst_mm, pgtable);
 922         dst_mm->nr_ptes++;
 923
 924         ret = 0;
 925 out_unlock:
 926         spin_unlock(&src_mm->page_table_lock);
 927         spin_unlock(&dst_mm->page_table_lock);
 928 out:
 929         return ret;
 930 }
 931
 932 void huge_pmd_set_accessed(struct mm_struct *mm,
 933                            struct vm_area_struct *vma,
 934                            unsigned long address,
 935                            pmd_t *pmd, pmd_t orig_pmd,
 936                            int dirty)
 937 {
 938         pmd_t entry;
 939         unsigned long haddr;
 940
 941         spin_lock(&mm->page_table_lock);
 942         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 943                 goto unlock;
 944
 945         entry = pmd_mkyoung(orig_pmd);
 946         haddr = address & HPAGE_PMD_MASK;
 947         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 948                 update_mmu_cache_pmd(vma, address, pmd);
 949
 950 unlock:
 951         spin_unlock(&mm->page_table_lock);
 952 }
 953
 954 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 955                 struct vm_area_struct *vma, unsigned long address,
 956                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 957 {
 958         pgtable_t pgtable;
 959         pmd_t _pmd;
 960         struct page *page;
 961         int i, ret = 0;
 962         unsigned long mmun_start;       /* For mmu_notifiers */
 963         unsigned long mmun_end;         /* For mmu_notifiers */
 964
 965         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 966         if (!page) {
 967                 ret |= VM_FAULT_OOM;
 968                 goto out;
 969         }
 970
 971         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 972                 put_page(page);
 973                 ret |= VM_FAULT_OOM;
 974                 goto out;
 975         }
 976
 977         clear_user_highpage(page, address);
 978         __SetPageUptodate(page);
 979
 980         mmun_start = haddr;
 981         mmun_end   = haddr + HPAGE_PMD_SIZE;
 982         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 983
 984         spin_lock(&mm->page_table_lock);
 985         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 986                 goto out_free_page;
 987
 988         pmdp_clear_flush(vma, haddr, pmd);
 989         /* leave pmd empty until pte is filled */
 990
 991         pgtable = pgtable_trans_huge_withdraw(mm);
 992         pmd_populate(mm, &_pmd, pgtable);
 993
 994         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 995                 pte_t *pte, entry;
 996                 if (haddr == (address & PAGE_MASK)) {
 997                         entry = mk_pte(page, vma->vm_page_prot);
 998                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 999                         page_add_new_anon_rmap(page, vma, haddr);
1000                 } else {
1001                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1002                         entry = pte_mkspecial(entry);
1003                 }
1004                 pte = pte_offset_map(&_pmd, haddr);
1005                 VM_BUG_ON(!pte_none(*pte));
1006                 set_pte_at(mm, haddr, pte, entry);
1007                 pte_unmap(pte);
1008         }
1009         smp_wmb(); /* make pte visible before pmd */
1010         pmd_populate(mm, pmd, pgtable);
1011         spin_unlock(&mm->page_table_lock);
1012         put_huge_zero_page();
1013         inc_mm_counter(mm, MM_ANONPAGES);
1014
1015         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1016
1017         ret |= VM_FAULT_WRITE;
1018 out:
1019         return ret;
1020 out_free_page:
1021         spin_unlock(&mm->page_table_lock);
1022         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1023         mem_cgroup_uncharge_page(page);
1024         put_page(page);
1025         goto out;
1026 }
1027
1028 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1029                                         struct vm_area_struct *vma,
1030                                         unsigned long address,
1031                                         pmd_t *pmd, pmd_t orig_pmd,
1032                                         struct page *page,
1033                                         unsigned long haddr)
1034 {
1035         pgtable_t pgtable;
1036         pmd_t _pmd;
1037         int ret = 0, i;
1038         struct page **pages;
1039         unsigned long mmun_start;       /* For mmu_notifiers */
1040         unsigned long mmun_end;         /* For mmu_notifiers */
1041
1042         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1043                         GFP_KERNEL);
1044         if (unlikely(!pages)) {
1045                 ret |= VM_FAULT_OOM;
1046                 goto out;
1047         }
1048
1049         for (i = 0; i < HPAGE_PMD_NR; i++) {
1050                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1051                                                __GFP_OTHER_NODE,
1052                                                vma, address, page_to_nid(page));
1053                 if (unlikely(!pages[i] ||
1054                              mem_cgroup_newpage_charge(pages[i], mm,
1055                                                        GFP_KERNEL))) {
1056                         if (pages[i])
1057                                 put_page(pages[i]);
1058                         mem_cgroup_uncharge_start();
1059                         while (--i >= 0) {
1060                                 mem_cgroup_uncharge_page(pages[i]);
1061                                 put_page(pages[i]);
1062                         }
1063                         mem_cgroup_uncharge_end();
1064                         kfree(pages);
1065                         ret |= VM_FAULT_OOM;
1066                         goto out;
1067                 }
1068         }
1069
1070         for (i = 0; i < HPAGE_PMD_NR; i++) {
1071                 copy_user_highpage(pages[i], page + i,
1072                                    haddr + PAGE_SIZE * i, vma);
1073                 __SetPageUptodate(pages[i]);
1074                 cond_resched();
1075         }
1076
1077         mmun_start = haddr;
1078         mmun_end   = haddr + HPAGE_PMD_SIZE;
1079         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1080
1081         spin_lock(&mm->page_table_lock);
1082         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1083                 goto out_free_pages;
1084         VM_BUG_ON(!PageHead(page));
1085
1086         pmdp_clear_flush(vma, haddr, pmd);
1087         /* leave pmd empty until pte is filled */
1088
1089         pgtable = pgtable_trans_huge_withdraw(mm);
1090         pmd_populate(mm, &_pmd, pgtable);
1091
1092         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1093                 pte_t *pte, entry;
1094                 entry = mk_pte(pages[i], vma->vm_page_prot);
1095                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1096                 page_add_new_anon_rmap(pages[i], vma, haddr);
1097                 pte = pte_offset_map(&_pmd, haddr);
1098                 VM_BUG_ON(!pte_none(*pte));
1099                 set_pte_at(mm, haddr, pte, entry);
1100                 pte_unmap(pte);
1101         }
1102         kfree(pages);
1103
1104         smp_wmb(); /* make pte visible before pmd */
1105         pmd_populate(mm, pmd, pgtable);
1106         page_remove_rmap(page);
1107         spin_unlock(&mm->page_table_lock);
1108
1109         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1110
1111         ret |= VM_FAULT_WRITE;
1112         put_page(page);
1113
1114 out:
1115         return ret;
1116
1117 out_free_pages:
1118         spin_unlock(&mm->page_table_lock);
1119         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1120         mem_cgroup_uncharge_start();
1121         for (i = 0; i < HPAGE_PMD_NR; i++) {
1122                 mem_cgroup_uncharge_page(pages[i]);
1123                 put_page(pages[i]);
1124         }
1125         mem_cgroup_uncharge_end();
1126         kfree(pages);
1127         goto out;
1128 }
1129
1130 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1131                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1132 {
1133         int ret = 0;
1134         struct page *page = NULL, *new_page;
1135         unsigned long haddr;
1136         unsigned long mmun_start;       /* For mmu_notifiers */
1137         unsigned long mmun_end;         /* For mmu_notifiers */
1138
1139         VM_BUG_ON(!vma->anon_vma);
1140         haddr = address & HPAGE_PMD_MASK;
1141         if (is_huge_zero_pmd(orig_pmd))
1142                 goto alloc;
1143         spin_lock(&mm->page_table_lock);
1144         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1145                 goto out_unlock;
1146
1147         page = pmd_page(orig_pmd);
1148         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1149         if (page_mapcount(page) == 1) {
1150                 pmd_t entry;
1151                 entry = pmd_mkyoung(orig_pmd);
1152                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1153                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1154                         update_mmu_cache_pmd(vma, address, pmd);
1155                 ret |= VM_FAULT_WRITE;
1156                 goto out_unlock;
1157         }
1158         get_page(page);
1159         spin_unlock(&mm->page_table_lock);
1160 alloc:
1161         if (transparent_hugepage_enabled(vma) &&
1162             !transparent_hugepage_debug_cow())
1163                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1164                                               vma, haddr, numa_node_id(), 0);
1165         else
1166                 new_page = NULL;
1167
1168         if (unlikely(!new_page)) {
1169                 count_vm_event(THP_FAULT_FALLBACK);
1170                 if (is_huge_zero_pmd(orig_pmd)) {
1171                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1172                                         address, pmd, orig_pmd, haddr);
1173                 } else {
1174                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1175                                         pmd, orig_pmd, page, haddr);
1176                         if (ret & VM_FAULT_OOM)
1177                                 split_huge_page(page);
1178                         put_page(page);
1179                 }
1180                 goto out;
1181         }
1182         count_vm_event(THP_FAULT_ALLOC);
1183
1184         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1185                 put_page(new_page);
1186                 if (page) {
1187                         split_huge_page(page);
1188                         put_page(page);
1189                 }
1190                 ret |= VM_FAULT_OOM;
1191                 goto out;
1192         }
1193
1194         if (is_huge_zero_pmd(orig_pmd))
1195                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1196         else
1197                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1198         __SetPageUptodate(new_page);
1199
1200         mmun_start = haddr;
1201         mmun_end   = haddr + HPAGE_PMD_SIZE;
1202         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1203
1204         spin_lock(&mm->page_table_lock);
1205         if (page)
1206                 put_page(page);
1207         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1208                 spin_unlock(&mm->page_table_lock);
1209                 mem_cgroup_uncharge_page(new_page);
1210                 put_page(new_page);
1211                 goto out_mn;
1212         } else {
1213                 pmd_t entry;
1214                 entry = mk_huge_pmd(new_page, vma);
1215                 pmdp_clear_flush(vma, haddr, pmd);
1216                 page_add_new_anon_rmap(new_page, vma, haddr);
1217                 set_pmd_at(mm, haddr, pmd, entry);
1218                 update_mmu_cache_pmd(vma, address, pmd);
1219                 if (is_huge_zero_pmd(orig_pmd)) {
1220                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1221                         put_huge_zero_page();
1222                 } else {
1223                         VM_BUG_ON(!PageHead(page));
1224                         page_remove_rmap(page);
1225                         put_page(page);
1226                 }
1227                 ret |= VM_FAULT_WRITE;
1228         }
1229         spin_unlock(&mm->page_table_lock);
1230 out_mn:
1231         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1232 out:
1233         return ret;
1234 out_unlock:
1235         spin_unlock(&mm->page_table_lock);
1236         return ret;
1237 }
1238
1239 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1240                                    unsigned long addr,
1241                                    pmd_t *pmd,
1242                                    unsigned int flags)
1243 {
1244         struct mm_struct *mm = vma->vm_mm;
1245         struct page *page = NULL;
1246
1247         assert_spin_locked(&mm->page_table_lock);
1248
1249         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1250                 goto out;
1251
1252         /* Avoid dumping huge zero page */
1253         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1254                 return ERR_PTR(-EFAULT);
1255
1256         page = pmd_page(*pmd);
1257         VM_BUG_ON(!PageHead(page));
1258         if (flags & FOLL_TOUCH) {
1259                 pmd_t _pmd;
1260                 /*
1261                  * We should set the dirty bit only for FOLL_WRITE but
1262                  * for now the dirty bit in the pmd is meaningless.
1263                  * And if the dirty bit will become meaningful and
1264                  * we'll only set it with FOLL_WRITE, an atomic
1265                  * set_bit will be required on the pmd to set the
1266                  * young bit, instead of the current set_pmd_at.
1267                  */
1268                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1269                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1270         }
1271         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1272                 if (page->mapping && trylock_page(page)) {
1273                         lru_add_drain();
1274                         if (page->mapping)
1275                                 mlock_vma_page(page);
1276                         unlock_page(page);
1277                 }
1278         }
1279         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1280         VM_BUG_ON(!PageCompound(page));
1281         if (flags & FOLL_GET)
1282                 get_page_foll(page);
1283
1284 out:
1285         return page;
1286 }
1287
1288 /* NUMA hinting page fault entry point for trans huge pmds */
1289 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1290                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1291 {
1292         struct page *page;
1293         unsigned long haddr = addr & HPAGE_PMD_MASK;
1294         int target_nid;
1295         int current_nid = -1;
1296         bool migrated;
1297
1298         spin_lock(&mm->page_table_lock);
1299         if (unlikely(!pmd_same(pmd, *pmdp)))
1300                 goto out_unlock;
1301
1302         page = pmd_page(pmd);
1303         get_page(page);
1304         current_nid = page_to_nid(page);
1305         count_vm_numa_event(NUMA_HINT_FAULTS);
1306         if (current_nid == numa_node_id())
1307                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1308
1309         target_nid = mpol_misplaced(page, vma, haddr);
1310         if (target_nid == -1) {
1311                 put_page(page);
1312                 goto clear_pmdnuma;
1313         }
1314
1315         /* Acquire the page lock to serialise THP migrations */
1316         spin_unlock(&mm->page_table_lock);
1317         lock_page(page);
1318
1319         /* Confirm the PTE did not while locked */
1320         spin_lock(&mm->page_table_lock);
1321         if (unlikely(!pmd_same(pmd, *pmdp))) {
1322                 unlock_page(page);
1323                 put_page(page);
1324                 goto out_unlock;
1325         }
1326         spin_unlock(&mm->page_table_lock);
1327
1328         /* Migrate the THP to the requested node */
1329         migrated = migrate_misplaced_transhuge_page(mm, vma,
1330                                 pmdp, pmd, addr, page, target_nid);
1331         if (!migrated)
1332                 goto check_same;
1333
1334         task_numa_fault(target_nid, HPAGE_PMD_NR, true);
1335         return 0;
1336
1337 check_same:
1338         spin_lock(&mm->page_table_lock);
1339         if (unlikely(!pmd_same(pmd, *pmdp)))
1340                 goto out_unlock;
1341 clear_pmdnuma:
1342         pmd = pmd_mknonnuma(pmd);
1343         set_pmd_at(mm, haddr, pmdp, pmd);
1344         VM_BUG_ON(pmd_numa(*pmdp));
1345         update_mmu_cache_pmd(vma, addr, pmdp);
1346 out_unlock:
1347         spin_unlock(&mm->page_table_lock);
1348         if (current_nid != -1)
1349                 task_numa_fault(current_nid, HPAGE_PMD_NR, false);
1350         return 0;
1351 }
1352
1353 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1354                  pmd_t *pmd, unsigned long addr)
1355 {
1356         int ret = 0;
1357
1358         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1359                 struct page *page;
1360                 pgtable_t pgtable;
1361                 pmd_t orig_pmd;
1362                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1363                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1364                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1365                 if (is_huge_zero_pmd(orig_pmd)) {
1366                         tlb->mm->nr_ptes--;
1367                         spin_unlock(&tlb->mm->page_table_lock);
1368                         put_huge_zero_page();
1369                 } else {
1370                         page = pmd_page(orig_pmd);
1371                         page_remove_rmap(page);
1372                         VM_BUG_ON(page_mapcount(page) < 0);
1373                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1374                         VM_BUG_ON(!PageHead(page));
1375                         tlb->mm->nr_ptes--;
1376                         spin_unlock(&tlb->mm->page_table_lock);
1377                         tlb_remove_page(tlb, page);
1378                 }
1379                 pte_free(tlb->mm, pgtable);
1380                 ret = 1;
1381         }
1382         return ret;
1383 }
1384
1385 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1386                 unsigned long addr, unsigned long end,
1387                 unsigned char *vec)
1388 {
1389         int ret = 0;
1390
1391         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1392                 /*
1393                  * All logical pages in the range are present
1394                  * if backed by a huge page.
1395                  */
1396                 spin_unlock(&vma->vm_mm->page_table_lock);
1397                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1398                 ret = 1;
1399         }
1400
1401         return ret;
1402 }
1403
1404 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1405                   unsigned long old_addr,
1406                   unsigned long new_addr, unsigned long old_end,
1407                   pmd_t *old_pmd, pmd_t *new_pmd)
1408 {
1409         int ret = 0;
1410         pmd_t pmd;
1411
1412         struct mm_struct *mm = vma->vm_mm;
1413
1414         if ((old_addr & ~HPAGE_PMD_MASK) ||
1415             (new_addr & ~HPAGE_PMD_MASK) ||
1416             old_end - old_addr < HPAGE_PMD_SIZE ||
1417             (new_vma->vm_flags & VM_NOHUGEPAGE))
1418                 goto out;
1419
1420         /*
1421          * The destination pmd shouldn't be established, free_pgtables()
1422          * should have release it.
1423          */
1424         if (WARN_ON(!pmd_none(*new_pmd))) {
1425                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1426                 goto out;
1427         }
1428
1429         ret = __pmd_trans_huge_lock(old_pmd, vma);
1430         if (ret == 1) {
1431                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1432                 VM_BUG_ON(!pmd_none(*new_pmd));
1433                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1434                 spin_unlock(&mm->page_table_lock);
1435         }
1436 out:
1437         return ret;
1438 }
1439
1440 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1441                 unsigned long addr, pgprot_t newprot, int prot_numa)
1442 {
1443         struct mm_struct *mm = vma->vm_mm;
1444         int ret = 0;
1445
1446         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1447                 pmd_t entry;
1448                 entry = pmdp_get_and_clear(mm, addr, pmd);
1449                 if (!prot_numa) {
1450                         entry = pmd_modify(entry, newprot);
1451                         BUG_ON(pmd_write(entry));
1452                 } else {
1453                         struct page *page = pmd_page(*pmd);
1454
1455                         /* only check non-shared pages */
1456                         if (page_mapcount(page) == 1 &&
1457                             !pmd_numa(*pmd)) {
1458                                 entry = pmd_mknuma(entry);
1459                         }
1460                 }
1461                 set_pmd_at(mm, addr, pmd, entry);
1462                 spin_unlock(&vma->vm_mm->page_table_lock);
1463                 ret = 1;
1464         }
1465
1466         return ret;
1467 }
1468
1469 /*
1470  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1471  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1472  *
1473  * Note that if it returns 1, this routine returns without unlocking page
1474  * table locks. So callers must unlock them.
1475  */
1476 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1477 {
1478         spin_lock(&vma->vm_mm->page_table_lock);
1479         if (likely(pmd_trans_huge(*pmd))) {
1480                 if (unlikely(pmd_trans_splitting(*pmd))) {
1481                         spin_unlock(&vma->vm_mm->page_table_lock);
1482                         wait_split_huge_page(vma->anon_vma, pmd);
1483                         return -1;
1484                 } else {
1485                         /* Thp mapped by 'pmd' is stable, so we can
1486                          * handle it as it is. */
1487                         return 1;
1488                 }
1489         }
1490         spin_unlock(&vma->vm_mm->page_table_lock);
1491         return 0;
1492 }
1493
1494 pmd_t *page_check_address_pmd(struct page *page,
1495                               struct mm_struct *mm,
1496                               unsigned long address,
1497                               enum page_check_address_pmd_flag flag)
1498 {
1499         pmd_t *pmd, *ret = NULL;
1500
1501         if (address & ~HPAGE_PMD_MASK)
1502                 goto out;
1503
1504         pmd = mm_find_pmd(mm, address);
1505         if (!pmd)
1506                 goto out;
1507         if (pmd_none(*pmd))
1508                 goto out;
1509         if (pmd_page(*pmd) != page)
1510                 goto out;
1511         /*
1512          * split_vma() may create temporary aliased mappings. There is
1513          * no risk as long as all huge pmd are found and have their
1514          * splitting bit set before __split_huge_page_refcount
1515          * runs. Finding the same huge pmd more than once during the
1516          * same rmap walk is not a problem.
1517          */
1518         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1519             pmd_trans_splitting(*pmd))
1520                 goto out;
1521         if (pmd_trans_huge(*pmd)) {
1522                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1523                           !pmd_trans_splitting(*pmd));
1524                 ret = pmd;
1525         }
1526 out:
1527         return ret;
1528 }
1529
1530 static int __split_huge_page_splitting(struct page *page,
1531                                        struct vm_area_struct *vma,
1532                                        unsigned long address)
1533 {
1534         struct mm_struct *mm = vma->vm_mm;
1535         pmd_t *pmd;
1536         int ret = 0;
1537         /* For mmu_notifiers */
1538         const unsigned long mmun_start = address;
1539         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1540
1541         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1542         spin_lock(&mm->page_table_lock);
1543         pmd = page_check_address_pmd(page, mm, address,
1544                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1545         if (pmd) {
1546                 /*
1547                  * We can't temporarily set the pmd to null in order
1548                  * to split it, the pmd must remain marked huge at all
1549                  * times or the VM won't take the pmd_trans_huge paths
1550                  * and it won't wait on the anon_vma->root->rwsem to
1551                  * serialize against split_huge_page*.
1552                  */
1553                 pmdp_splitting_flush(vma, address, pmd);
1554                 ret = 1;
1555         }
1556         spin_unlock(&mm->page_table_lock);
1557         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1558
1559         return ret;
1560 }
1561
1562 static void __split_huge_page_refcount(struct page *page)
1563 {
1564         int i;
1565         struct zone *zone = page_zone(page);
1566         struct lruvec *lruvec;
1567         int tail_count = 0;
1568
1569         /* prevent PageLRU to go away from under us, and freeze lru stats */
1570         spin_lock_irq(&zone->lru_lock);
1571         lruvec = mem_cgroup_page_lruvec(page, zone);
1572
1573         compound_lock(page);
1574         /* complete memcg works before add pages to LRU */
1575         mem_cgroup_split_huge_fixup(page);
1576
1577         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1578                 struct page *page_tail = page + i;
1579
1580                 /* tail_page->_mapcount cannot change */
1581                 BUG_ON(page_mapcount(page_tail) < 0);
1582                 tail_count += page_mapcount(page_tail);
1583                 /* check for overflow */
1584                 BUG_ON(tail_count < 0);
1585                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1586                 /*
1587                  * tail_page->_count is zero and not changing from
1588                  * under us. But get_page_unless_zero() may be running
1589                  * from under us on the tail_page. If we used
1590                  * atomic_set() below instead of atomic_add(), we
1591                  * would then run atomic_set() concurrently with
1592                  * get_page_unless_zero(), and atomic_set() is
1593                  * implemented in C not using locked ops. spin_unlock
1594                  * on x86 sometime uses locked ops because of PPro
1595                  * errata 66, 92, so unless somebody can guarantee
1596                  * atomic_set() here would be safe on all archs (and
1597                  * not only on x86), it's safer to use atomic_add().
1598                  */
1599                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1600                            &page_tail->_count);
1601
1602                 /* after clearing PageTail the gup refcount can be released */
1603                 smp_mb();
1604
1605                 /*
1606                  * retain hwpoison flag of the poisoned tail page:
1607                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1608                  *   by the memory-failure.
1609                  */
1610                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1611                 page_tail->flags |= (page->flags &
1612                                      ((1L << PG_referenced) |
1613                                       (1L << PG_swapbacked) |
1614                                       (1L << PG_mlocked) |
1615                                       (1L << PG_uptodate)));
1616                 page_tail->flags |= (1L << PG_dirty);
1617
1618                 /* clear PageTail before overwriting first_page */
1619                 smp_wmb();
1620
1621                 /*
1622                  * __split_huge_page_splitting() already set the
1623                  * splitting bit in all pmd that could map this
1624                  * hugepage, that will ensure no CPU can alter the
1625                  * mapcount on the head page. The mapcount is only
1626                  * accounted in the head page and it has to be
1627                  * transferred to all tail pages in the below code. So
1628                  * for this code to be safe, the split the mapcount
1629                  * can't change. But that doesn't mean userland can't
1630                  * keep changing and reading the page contents while
1631                  * we transfer the mapcount, so the pmd splitting
1632                  * status is achieved setting a reserved bit in the
1633                  * pmd, not by clearing the present bit.
1634                 */
1635                 page_tail->_mapcount = page->_mapcount;
1636
1637                 BUG_ON(page_tail->mapping);
1638                 page_tail->mapping = page->mapping;
1639
1640                 page_tail->index = page->index + i;
1641                 page_nid_xchg_last(page_tail, page_nid_last(page));
1642
1643                 BUG_ON(!PageAnon(page_tail));
1644                 BUG_ON(!PageUptodate(page_tail));
1645                 BUG_ON(!PageDirty(page_tail));
1646                 BUG_ON(!PageSwapBacked(page_tail));
1647
1648                 lru_add_page_tail(page, page_tail, lruvec);
1649         }
1650         atomic_sub(tail_count, &page->_count);
1651         BUG_ON(atomic_read(&page->_count) <= 0);
1652
1653         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1654         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1655
1656         ClearPageCompound(page);
1657         compound_unlock(page);
1658         spin_unlock_irq(&zone->lru_lock);
1659
1660         for (i = 1; i < HPAGE_PMD_NR; i++) {
1661                 struct page *page_tail = page + i;
1662                 BUG_ON(page_count(page_tail) <= 0);
1663                 /*
1664                  * Tail pages may be freed if there wasn't any mapping
1665                  * like if add_to_swap() is running on a lru page that
1666                  * had its mapping zapped. And freeing these pages
1667                  * requires taking the lru_lock so we do the put_page
1668                  * of the tail pages after the split is complete.
1669                  */
1670                 put_page(page_tail);
1671         }
1672
1673         /*
1674          * Only the head page (now become a regular page) is required
1675          * to be pinned by the caller.
1676          */
1677         BUG_ON(page_count(page) <= 0);
1678 }
1679
1680 static int __split_huge_page_map(struct page *page,
1681                                  struct vm_area_struct *vma,
1682                                  unsigned long address)
1683 {
1684         struct mm_struct *mm = vma->vm_mm;
1685         pmd_t *pmd, _pmd;
1686         int ret = 0, i;
1687         pgtable_t pgtable;
1688         unsigned long haddr;
1689
1690         spin_lock(&mm->page_table_lock);
1691         pmd = page_check_address_pmd(page, mm, address,
1692                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1693         if (pmd) {
1694                 pgtable = pgtable_trans_huge_withdraw(mm);
1695                 pmd_populate(mm, &_pmd, pgtable);
1696
1697                 haddr = address;
1698                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1699                         pte_t *pte, entry;
1700                         BUG_ON(PageCompound(page+i));
1701                         entry = mk_pte(page + i, vma->vm_page_prot);
1702                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1703                         if (!pmd_write(*pmd))
1704                                 entry = pte_wrprotect(entry);
1705                         else
1706                                 BUG_ON(page_mapcount(page) != 1);
1707                         if (!pmd_young(*pmd))
1708                                 entry = pte_mkold(entry);
1709                         if (pmd_numa(*pmd))
1710                                 entry = pte_mknuma(entry);
1711                         pte = pte_offset_map(&_pmd, haddr);
1712                         BUG_ON(!pte_none(*pte));
1713                         set_pte_at(mm, haddr, pte, entry);
1714                         pte_unmap(pte);
1715                 }
1716
1717                 smp_wmb(); /* make pte visible before pmd */
1718                 /*
1719                  * Up to this point the pmd is present and huge and
1720                  * userland has the whole access to the hugepage
1721                  * during the split (which happens in place). If we
1722                  * overwrite the pmd with the not-huge version
1723                  * pointing to the pte here (which of course we could
1724                  * if all CPUs were bug free), userland could trigger
1725                  * a small page size TLB miss on the small sized TLB
1726                  * while the hugepage TLB entry is still established
1727                  * in the huge TLB. Some CPU doesn't like that. See
1728                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1729                  * Erratum 383 on page 93. Intel should be safe but is
1730                  * also warns that it's only safe if the permission
1731                  * and cache attributes of the two entries loaded in
1732                  * the two TLB is identical (which should be the case
1733                  * here). But it is generally safer to never allow
1734                  * small and huge TLB entries for the same virtual
1735                  * address to be loaded simultaneously. So instead of
1736                  * doing "pmd_populate(); flush_tlb_range();" we first
1737                  * mark the current pmd notpresent (atomically because
1738                  * here the pmd_trans_huge and pmd_trans_splitting
1739                  * must remain set at all times on the pmd until the
1740                  * split is complete for this pmd), then we flush the
1741                  * SMP TLB and finally we write the non-huge version
1742                  * of the pmd entry with pmd_populate.
1743                  */
1744                 pmdp_invalidate(vma, address, pmd);
1745                 pmd_populate(mm, pmd, pgtable);
1746                 ret = 1;
1747         }
1748         spin_unlock(&mm->page_table_lock);
1749
1750         return ret;
1751 }
1752
1753 /* must be called with anon_vma->root->rwsem held */
1754 static void __split_huge_page(struct page *page,
1755                               struct anon_vma *anon_vma)
1756 {
1757         int mapcount, mapcount2;
1758         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1759         struct anon_vma_chain *avc;
1760
1761         BUG_ON(!PageHead(page));
1762         BUG_ON(PageTail(page));
1763
1764         mapcount = 0;
1765         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1766                 struct vm_area_struct *vma = avc->vma;
1767                 unsigned long addr = vma_address(page, vma);
1768                 BUG_ON(is_vma_temporary_stack(vma));
1769                 mapcount += __split_huge_page_splitting(page, vma, addr);
1770         }
1771         /*
1772          * It is critical that new vmas are added to the tail of the
1773          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1774          * and establishes a child pmd before
1775          * __split_huge_page_splitting() freezes the parent pmd (so if
1776          * we fail to prevent copy_huge_pmd() from running until the
1777          * whole __split_huge_page() is complete), we will still see
1778          * the newly established pmd of the child later during the
1779          * walk, to be able to set it as pmd_trans_splitting too.
1780          */
1781         if (mapcount != page_mapcount(page))
1782                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1783                        mapcount, page_mapcount(page));
1784         BUG_ON(mapcount != page_mapcount(page));
1785
1786         __split_huge_page_refcount(page);
1787
1788         mapcount2 = 0;
1789         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1790                 struct vm_area_struct *vma = avc->vma;
1791                 unsigned long addr = vma_address(page, vma);
1792                 BUG_ON(is_vma_temporary_stack(vma));
1793                 mapcount2 += __split_huge_page_map(page, vma, addr);
1794         }
1795         if (mapcount != mapcount2)
1796                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1797                        mapcount, mapcount2, page_mapcount(page));
1798         BUG_ON(mapcount != mapcount2);
1799 }
1800
1801 int split_huge_page(struct page *page)
1802 {
1803         struct anon_vma *anon_vma;
1804         int ret = 1;
1805
1806         BUG_ON(is_huge_zero_pfn(page_to_pfn(page)));
1807         BUG_ON(!PageAnon(page));
1808
1809         /*
1810          * The caller does not necessarily hold an mmap_sem that would prevent
1811          * the anon_vma disappearing so we first we take a reference to it
1812          * and then lock the anon_vma for write. This is similar to
1813          * page_lock_anon_vma_read except the write lock is taken to serialise
1814          * against parallel split or collapse operations.
1815          */
1816         anon_vma = page_get_anon_vma(page);
1817         if (!anon_vma)
1818                 goto out;
1819         anon_vma_lock_write(anon_vma);
1820
1821         ret = 0;
1822         if (!PageCompound(page))
1823                 goto out_unlock;
1824
1825         BUG_ON(!PageSwapBacked(page));
1826         __split_huge_page(page, anon_vma);
1827         count_vm_event(THP_SPLIT);
1828
1829         BUG_ON(PageCompound(page));
1830 out_unlock:
1831         anon_vma_unlock_write(anon_vma);
1832         put_anon_vma(anon_vma);
1833 out:
1834         return ret;
1835 }
1836
1837 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1838
1839 int hugepage_madvise(struct vm_area_struct *vma,
1840                      unsigned long *vm_flags, int advice)
1841 {
1842         struct mm_struct *mm = vma->vm_mm;
1843
1844         switch (advice) {
1845         case MADV_HUGEPAGE:
1846                 /*
1847                  * Be somewhat over-protective like KSM for now!
1848                  */
1849                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1850                         return -EINVAL;
1851                 if (mm->def_flags & VM_NOHUGEPAGE)
1852                         return -EINVAL;
1853                 *vm_flags &= ~VM_NOHUGEPAGE;
1854                 *vm_flags |= VM_HUGEPAGE;
1855                 /*
1856                  * If the vma become good for khugepaged to scan,
1857                  * register it here without waiting a page fault that
1858                  * may not happen any time soon.
1859                  */
1860                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1861                         return -ENOMEM;
1862                 break;
1863         case MADV_NOHUGEPAGE:
1864                 /*
1865                  * Be somewhat over-protective like KSM for now!
1866                  */
1867                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1868                         return -EINVAL;
1869                 *vm_flags &= ~VM_HUGEPAGE;
1870                 *vm_flags |= VM_NOHUGEPAGE;
1871                 /*
1872                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1873                  * this vma even if we leave the mm registered in khugepaged if
1874                  * it got registered before VM_NOHUGEPAGE was set.
1875                  */
1876                 break;
1877         }
1878
1879         return 0;
1880 }
1881
1882 static int __init khugepaged_slab_init(void)
1883 {
1884         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1885                                           sizeof(struct mm_slot),
1886                                           __alignof__(struct mm_slot), 0, NULL);
1887         if (!mm_slot_cache)
1888                 return -ENOMEM;
1889
1890         return 0;
1891 }
1892
1893 static inline struct mm_slot *alloc_mm_slot(void)
1894 {
1895         if (!mm_slot_cache)     /* initialization failed */
1896                 return NULL;
1897         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1898 }
1899
1900 static inline void free_mm_slot(struct mm_slot *mm_slot)
1901 {
1902         kmem_cache_free(mm_slot_cache, mm_slot);
1903 }
1904
1905 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1906 {
1907         struct mm_slot *mm_slot;
1908
1909         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1910                 if (mm == mm_slot->mm)
1911                         return mm_slot;
1912
1913         return NULL;
1914 }
1915
1916 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1917                                     struct mm_slot *mm_slot)
1918 {
1919         mm_slot->mm = mm;
1920         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1921 }
1922
1923 static inline int khugepaged_test_exit(struct mm_struct *mm)
1924 {
1925         return atomic_read(&mm->mm_users) == 0;
1926 }
1927
1928 int __khugepaged_enter(struct mm_struct *mm)
1929 {
1930         struct mm_slot *mm_slot;
1931         int wakeup;
1932
1933         mm_slot = alloc_mm_slot();
1934         if (!mm_slot)
1935                 return -ENOMEM;
1936
1937         /* __khugepaged_exit() must not run from under us */
1938         VM_BUG_ON(khugepaged_test_exit(mm));
1939         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1940                 free_mm_slot(mm_slot);
1941                 return 0;
1942         }
1943
1944         spin_lock(&khugepaged_mm_lock);
1945         insert_to_mm_slots_hash(mm, mm_slot);
1946         /*
1947          * Insert just behind the scanning cursor, to let the area settle
1948          * down a little.
1949          */
1950         wakeup = list_empty(&khugepaged_scan.mm_head);
1951         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1952         spin_unlock(&khugepaged_mm_lock);
1953
1954         atomic_inc(&mm->mm_count);
1955         if (wakeup)
1956                 wake_up_interruptible(&khugepaged_wait);
1957
1958         return 0;
1959 }
1960
1961 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1962 {
1963         unsigned long hstart, hend;
1964         if (!vma->anon_vma)
1965                 /*
1966                  * Not yet faulted in so we will register later in the
1967                  * page fault if needed.
1968                  */
1969                 return 0;
1970         if (vma->vm_ops)
1971                 /* khugepaged not yet working on file or special mappings */
1972                 return 0;
1973         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1974         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1975         hend = vma->vm_end & HPAGE_PMD_MASK;
1976         if (hstart < hend)
1977                 return khugepaged_enter(vma);
1978         return 0;
1979 }
1980
1981 void __khugepaged_exit(struct mm_struct *mm)
1982 {
1983         struct mm_slot *mm_slot;
1984         int free = 0;
1985
1986         spin_lock(&khugepaged_mm_lock);
1987         mm_slot = get_mm_slot(mm);
1988         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1989                 hash_del(&mm_slot->hash);
1990                 list_del(&mm_slot->mm_node);
1991                 free = 1;
1992         }
1993         spin_unlock(&khugepaged_mm_lock);
1994
1995         if (free) {
1996                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1997                 free_mm_slot(mm_slot);
1998                 mmdrop(mm);
1999         } else if (mm_slot) {
2000                 /*
2001                  * This is required to serialize against
2002                  * khugepaged_test_exit() (which is guaranteed to run
2003                  * under mmap sem read mode). Stop here (after we
2004                  * return all pagetables will be destroyed) until
2005                  * khugepaged has finished working on the pagetables
2006                  * under the mmap_sem.
2007                  */
2008                 down_write(&mm->mmap_sem);
2009                 up_write(&mm->mmap_sem);
2010         }
2011 }
2012
2013 static void release_pte_page(struct page *page)
2014 {
2015         /* 0 stands for page_is_file_cache(page) == false */
2016         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2017         unlock_page(page);
2018         putback_lru_page(page);
2019 }
2020
2021 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2022 {
2023         while (--_pte >= pte) {
2024                 pte_t pteval = *_pte;
2025                 if (!pte_none(pteval))
2026                         release_pte_page(pte_page(pteval));
2027         }
2028 }
2029
2030 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2031                                         unsigned long address,
2032                                         pte_t *pte)
2033 {
2034         struct page *page;
2035         pte_t *_pte;
2036         int referenced = 0, none = 0;
2037         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2038              _pte++, address += PAGE_SIZE) {
2039                 pte_t pteval = *_pte;
2040                 if (pte_none(pteval)) {
2041                         if (++none <= khugepaged_max_ptes_none)
2042                                 continue;
2043                         else
2044                                 goto out;
2045                 }
2046                 if (!pte_present(pteval) || !pte_write(pteval))
2047                         goto out;
2048                 page = vm_normal_page(vma, address, pteval);
2049                 if (unlikely(!page))
2050                         goto out;
2051
2052                 VM_BUG_ON(PageCompound(page));
2053                 BUG_ON(!PageAnon(page));
2054                 VM_BUG_ON(!PageSwapBacked(page));
2055
2056                 /* cannot use mapcount: can't collapse if there's a gup pin */
2057                 if (page_count(page) != 1)
2058                         goto out;
2059                 /*
2060                  * We can do it before isolate_lru_page because the
2061                  * page can't be freed from under us. NOTE: PG_lock
2062                  * is needed to serialize against split_huge_page
2063                  * when invoked from the VM.
2064                  */
2065                 if (!trylock_page(page))
2066                         goto out;
2067                 /*
2068                  * Isolate the page to avoid collapsing an hugepage
2069                  * currently in use by the VM.
2070                  */
2071                 if (isolate_lru_page(page)) {
2072                         unlock_page(page);
2073                         goto out;
2074                 }
2075                 /* 0 stands for page_is_file_cache(page) == false */
2076                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2077                 VM_BUG_ON(!PageLocked(page));
2078                 VM_BUG_ON(PageLRU(page));
2079
2080                 /* If there is no mapped pte young don't collapse the page */
2081                 if (pte_young(pteval) || PageReferenced(page) ||
2082                     mmu_notifier_test_young(vma->vm_mm, address))
2083                         referenced = 1;
2084         }
2085         if (likely(referenced))
2086                 return 1;
2087 out:
2088         release_pte_pages(pte, _pte);
2089         return 0;
2090 }
2091
2092 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2093                                       struct vm_area_struct *vma,
2094                                       unsigned long address,
2095                                       spinlock_t *ptl)
2096 {
2097         pte_t *_pte;
2098         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2099                 pte_t pteval = *_pte;
2100                 struct page *src_page;
2101
2102                 if (pte_none(pteval)) {
2103                         clear_user_highpage(page, address);
2104                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2105                 } else {
2106                         src_page = pte_page(pteval);
2107                         copy_user_highpage(page, src_page, address, vma);
2108                         VM_BUG_ON(page_mapcount(src_page) != 1);
2109                         release_pte_page(src_page);
2110                         /*
2111                          * ptl mostly unnecessary, but preempt has to
2112                          * be disabled to update the per-cpu stats
2113                          * inside page_remove_rmap().
2114                          */
2115                         spin_lock(ptl);
2116                         /*
2117                          * paravirt calls inside pte_clear here are
2118                          * superfluous.
2119                          */
2120                         pte_clear(vma->vm_mm, address, _pte);
2121                         page_remove_rmap(src_page);
2122                         spin_unlock(ptl);
2123                         free_page_and_swap_cache(src_page);
2124                 }
2125
2126                 address += PAGE_SIZE;
2127                 page++;
2128         }
2129 }
2130
2131 static void khugepaged_alloc_sleep(void)
2132 {
2133         wait_event_freezable_timeout(khugepaged_wait, false,
2134                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2135 }
2136
2137 #ifdef CONFIG_NUMA
2138 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2139 {
2140         if (IS_ERR(*hpage)) {
2141                 if (!*wait)
2142                         return false;
2143
2144                 *wait = false;
2145                 *hpage = NULL;
2146                 khugepaged_alloc_sleep();
2147         } else if (*hpage) {
2148                 put_page(*hpage);
2149                 *hpage = NULL;
2150         }
2151
2152         return true;
2153 }
2154
2155 static struct page
2156 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2157                        struct vm_area_struct *vma, unsigned long address,
2158                        int node)
2159 {
2160         VM_BUG_ON(*hpage);
2161         /*
2162          * Allocate the page while the vma is still valid and under
2163          * the mmap_sem read mode so there is no memory allocation
2164          * later when we take the mmap_sem in write mode. This is more
2165          * friendly behavior (OTOH it may actually hide bugs) to
2166          * filesystems in userland with daemons allocating memory in
2167          * the userland I/O paths.  Allocating memory with the
2168          * mmap_sem in read mode is good idea also to allow greater
2169          * scalability.
2170          */
2171         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2172                                       node, __GFP_OTHER_NODE);
2173
2174         /*
2175          * After allocating the hugepage, release the mmap_sem read lock in
2176          * preparation for taking it in write mode.
2177          */
2178         up_read(&mm->mmap_sem);
2179         if (unlikely(!*hpage)) {
2180                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2181                 *hpage = ERR_PTR(-ENOMEM);
2182                 return NULL;
2183         }
2184
2185         count_vm_event(THP_COLLAPSE_ALLOC);
2186         return *hpage;
2187 }
2188 #else
2189 static struct page *khugepaged_alloc_hugepage(bool *wait)
2190 {
2191         struct page *hpage;
2192
2193         do {
2194                 hpage = alloc_hugepage(khugepaged_defrag());
2195                 if (!hpage) {
2196                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2197                         if (!*wait)
2198                                 return NULL;
2199
2200                         *wait = false;
2201                         khugepaged_alloc_sleep();
2202                 } else
2203                         count_vm_event(THP_COLLAPSE_ALLOC);
2204         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2205
2206         return hpage;
2207 }
2208
2209 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2210 {
2211         if (!*hpage)
2212                 *hpage = khugepaged_alloc_hugepage(wait);
2213
2214         if (unlikely(!*hpage))
2215                 return false;
2216
2217         return true;
2218 }
2219
2220 static struct page
2221 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2222                        struct vm_area_struct *vma, unsigned long address,
2223                        int node)
2224 {
2225         up_read(&mm->mmap_sem);
2226         VM_BUG_ON(!*hpage);
2227         return  *hpage;
2228 }
2229 #endif
2230
2231 static bool hugepage_vma_check(struct vm_area_struct *vma)
2232 {
2233         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2234             (vma->vm_flags & VM_NOHUGEPAGE))
2235                 return false;
2236
2237         if (!vma->anon_vma || vma->vm_ops)
2238                 return false;
2239         if (is_vma_temporary_stack(vma))
2240                 return false;
2241         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2242         return true;
2243 }
2244
2245 static void collapse_huge_page(struct mm_struct *mm,
2246                                    unsigned long address,
2247                                    struct page **hpage,
2248                                    struct vm_area_struct *vma,
2249                                    int node)
2250 {
2251         pmd_t *pmd, _pmd;
2252         pte_t *pte;
2253         pgtable_t pgtable;
2254         struct page *new_page;
2255         spinlock_t *ptl;
2256         int isolated;
2257         unsigned long hstart, hend;
2258         unsigned long mmun_start;       /* For mmu_notifiers */
2259         unsigned long mmun_end;         /* For mmu_notifiers */
2260
2261         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2262
2263         /* release the mmap_sem read lock. */
2264         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2265         if (!new_page)
2266                 return;
2267
2268         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2269                 return;
2270
2271         /*
2272          * Prevent all access to pagetables with the exception of
2273          * gup_fast later hanlded by the ptep_clear_flush and the VM
2274          * handled by the anon_vma lock + PG_lock.
2275          */
2276         down_write(&mm->mmap_sem);
2277         if (unlikely(khugepaged_test_exit(mm)))
2278                 goto out;
2279
2280         vma = find_vma(mm, address);
2281         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2282         hend = vma->vm_end & HPAGE_PMD_MASK;
2283         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2284                 goto out;
2285         if (!hugepage_vma_check(vma))
2286                 goto out;
2287         pmd = mm_find_pmd(mm, address);
2288         if (!pmd)
2289                 goto out;
2290         if (pmd_trans_huge(*pmd))
2291                 goto out;
2292
2293         anon_vma_lock_write(vma->anon_vma);
2294
2295         pte = pte_offset_map(pmd, address);
2296         ptl = pte_lockptr(mm, pmd);
2297
2298         mmun_start = address;
2299         mmun_end   = address + HPAGE_PMD_SIZE;
2300         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2301         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2302         /*
2303          * After this gup_fast can't run anymore. This also removes
2304          * any huge TLB entry from the CPU so we won't allow
2305          * huge and small TLB entries for the same virtual address
2306          * to avoid the risk of CPU bugs in that area.
2307          */
2308         _pmd = pmdp_clear_flush(vma, address, pmd);
2309         spin_unlock(&mm->page_table_lock);
2310         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2311
2312         spin_lock(ptl);
2313         isolated = __collapse_huge_page_isolate(vma, address, pte);
2314         spin_unlock(ptl);
2315
2316         if (unlikely(!isolated)) {
2317                 pte_unmap(pte);
2318                 spin_lock(&mm->page_table_lock);
2319                 BUG_ON(!pmd_none(*pmd));
2320                 set_pmd_at(mm, address, pmd, _pmd);
2321                 spin_unlock(&mm->page_table_lock);
2322                 anon_vma_unlock_write(vma->anon_vma);
2323                 goto out;
2324         }
2325
2326         /*
2327          * All pages are isolated and locked so anon_vma rmap
2328          * can't run anymore.
2329          */
2330         anon_vma_unlock_write(vma->anon_vma);
2331
2332         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2333         pte_unmap(pte);
2334         __SetPageUptodate(new_page);
2335         pgtable = pmd_pgtable(_pmd);
2336
2337         _pmd = mk_huge_pmd(new_page, vma);
2338
2339         /*
2340          * spin_lock() below is not the equivalent of smp_wmb(), so
2341          * this is needed to avoid the copy_huge_page writes to become
2342          * visible after the set_pmd_at() write.
2343          */
2344         smp_wmb();
2345
2346         spin_lock(&mm->page_table_lock);
2347         BUG_ON(!pmd_none(*pmd));
2348         page_add_new_anon_rmap(new_page, vma, address);
2349         set_pmd_at(mm, address, pmd, _pmd);
2350         update_mmu_cache_pmd(vma, address, pmd);
2351         pgtable_trans_huge_deposit(mm, pgtable);
2352         spin_unlock(&mm->page_table_lock);
2353
2354         *hpage = NULL;
2355
2356         khugepaged_pages_collapsed++;
2357 out_up_write:
2358         up_write(&mm->mmap_sem);
2359         return;
2360
2361 out:
2362         mem_cgroup_uncharge_page(new_page);
2363         goto out_up_write;
2364 }
2365
2366 static int khugepaged_scan_pmd(struct mm_struct *mm,
2367                                struct vm_area_struct *vma,
2368                                unsigned long address,
2369                                struct page **hpage)
2370 {
2371         pmd_t *pmd;
2372         pte_t *pte, *_pte;
2373         int ret = 0, referenced = 0, none = 0;
2374         struct page *page;
2375         unsigned long _address;
2376         spinlock_t *ptl;
2377         int node = NUMA_NO_NODE;
2378
2379         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2380
2381         pmd = mm_find_pmd(mm, address);
2382         if (!pmd)
2383                 goto out;
2384         if (pmd_trans_huge(*pmd))
2385                 goto out;
2386
2387         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2388         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2389              _pte++, _address += PAGE_SIZE) {
2390                 pte_t pteval = *_pte;
2391                 if (pte_none(pteval)) {
2392                         if (++none <= khugepaged_max_ptes_none)
2393                                 continue;
2394                         else
2395                                 goto out_unmap;
2396                 }
2397                 if (!pte_present(pteval) || !pte_write(pteval))
2398                         goto out_unmap;
2399                 page = vm_normal_page(vma, _address, pteval);
2400                 if (unlikely(!page))
2401                         goto out_unmap;
2402                 /*
2403                  * Chose the node of the first page. This could
2404                  * be more sophisticated and look at more pages,
2405                  * but isn't for now.
2406                  */
2407                 if (node == NUMA_NO_NODE)
2408                         node = page_to_nid(page);
2409                 VM_BUG_ON(PageCompound(page));
2410                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2411                         goto out_unmap;
2412                 /* cannot use mapcount: can't collapse if there's a gup pin */
2413                 if (page_count(page) != 1)
2414                         goto out_unmap;
2415                 if (pte_young(pteval) || PageReferenced(page) ||
2416                     mmu_notifier_test_young(vma->vm_mm, address))
2417                         referenced = 1;
2418         }
2419         if (referenced)
2420                 ret = 1;
2421 out_unmap:
2422         pte_unmap_unlock(pte, ptl);
2423         if (ret)
2424                 /* collapse_huge_page will return with the mmap_sem released */
2425                 collapse_huge_page(mm, address, hpage, vma, node);
2426 out:
2427         return ret;
2428 }
2429
2430 static void collect_mm_slot(struct mm_slot *mm_slot)
2431 {
2432         struct mm_struct *mm = mm_slot->mm;
2433
2434         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2435
2436         if (khugepaged_test_exit(mm)) {
2437                 /* free mm_slot */
2438                 hash_del(&mm_slot->hash);
2439                 list_del(&mm_slot->mm_node);
2440
2441                 /*
2442                  * Not strictly needed because the mm exited already.
2443                  *
2444                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2445                  */
2446
2447                 /* khugepaged_mm_lock actually not necessary for the below */
2448                 free_mm_slot(mm_slot);
2449                 mmdrop(mm);
2450         }
2451 }
2452
2453 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2454                                             struct page **hpage)
2455         __releases(&khugepaged_mm_lock)
2456         __acquires(&khugepaged_mm_lock)
2457 {
2458         struct mm_slot *mm_slot;
2459         struct mm_struct *mm;
2460         struct vm_area_struct *vma;
2461         int progress = 0;
2462
2463         VM_BUG_ON(!pages);
2464         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2465
2466         if (khugepaged_scan.mm_slot)
2467                 mm_slot = khugepaged_scan.mm_slot;
2468         else {
2469                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2470                                      struct mm_slot, mm_node);
2471                 khugepaged_scan.address = 0;
2472                 khugepaged_scan.mm_slot = mm_slot;
2473         }
2474         spin_unlock(&khugepaged_mm_lock);
2475
2476         mm = mm_slot->mm;
2477         down_read(&mm->mmap_sem);
2478         if (unlikely(khugepaged_test_exit(mm)))
2479                 vma = NULL;
2480         else
2481                 vma = find_vma(mm, khugepaged_scan.address);
2482
2483         progress++;
2484         for (; vma; vma = vma->vm_next) {
2485                 unsigned long hstart, hend;
2486
2487                 cond_resched();
2488                 if (unlikely(khugepaged_test_exit(mm))) {
2489                         progress++;
2490                         break;
2491                 }
2492                 if (!hugepage_vma_check(vma)) {
2493 skip:
2494                         progress++;
2495                         continue;
2496                 }
2497                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2498                 hend = vma->vm_end & HPAGE_PMD_MASK;
2499                 if (hstart >= hend)
2500                         goto skip;
2501                 if (khugepaged_scan.address > hend)
2502                         goto skip;
2503                 if (khugepaged_scan.address < hstart)
2504                         khugepaged_scan.address = hstart;
2505                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2506
2507                 while (khugepaged_scan.address < hend) {
2508                         int ret;
2509                         cond_resched();
2510                         if (unlikely(khugepaged_test_exit(mm)))
2511                                 goto breakouterloop;
2512
2513                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2514                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2515                                   hend);
2516                         ret = khugepaged_scan_pmd(mm, vma,
2517                                                   khugepaged_scan.address,
2518                                                   hpage);
2519                         /* move to next address */
2520                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2521                         progress += HPAGE_PMD_NR;
2522                         if (ret)
2523                                 /* we released mmap_sem so break loop */
2524                                 goto breakouterloop_mmap_sem;
2525                         if (progress >= pages)
2526                                 goto breakouterloop;
2527                 }
2528         }
2529 breakouterloop:
2530         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2531 breakouterloop_mmap_sem:
2532
2533         spin_lock(&khugepaged_mm_lock);
2534         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2535         /*
2536          * Release the current mm_slot if this mm is about to die, or
2537          * if we scanned all vmas of this mm.
2538          */
2539         if (khugepaged_test_exit(mm) || !vma) {
2540                 /*
2541                  * Make sure that if mm_users is reaching zero while
2542                  * khugepaged runs here, khugepaged_exit will find
2543                  * mm_slot not pointing to the exiting mm.
2544                  */
2545                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2546                         khugepaged_scan.mm_slot = list_entry(
2547                                 mm_slot->mm_node.next,
2548                                 struct mm_slot, mm_node);
2549                         khugepaged_scan.address = 0;
2550                 } else {
2551                         khugepaged_scan.mm_slot = NULL;
2552                         khugepaged_full_scans++;
2553                 }
2554
2555                 collect_mm_slot(mm_slot);
2556         }
2557
2558         return progress;
2559 }
2560
2561 static int khugepaged_has_work(void)
2562 {
2563         return !list_empty(&khugepaged_scan.mm_head) &&
2564                 khugepaged_enabled();
2565 }
2566
2567 static int khugepaged_wait_event(void)
2568 {
2569         return !list_empty(&khugepaged_scan.mm_head) ||
2570                 kthread_should_stop();
2571 }
2572
2573 static void khugepaged_do_scan(void)
2574 {
2575         struct page *hpage = NULL;
2576         unsigned int progress = 0, pass_through_head = 0;
2577         unsigned int pages = khugepaged_pages_to_scan;
2578         bool wait = true;
2579
2580         barrier(); /* write khugepaged_pages_to_scan to local stack */
2581
2582         while (progress < pages) {
2583                 if (!khugepaged_prealloc_page(&hpage, &wait))
2584                         break;
2585
2586                 cond_resched();
2587
2588                 if (unlikely(kthread_should_stop() || freezing(current)))
2589                         break;
2590
2591                 spin_lock(&khugepaged_mm_lock);
2592                 if (!khugepaged_scan.mm_slot)
2593                         pass_through_head++;
2594                 if (khugepaged_has_work() &&
2595                     pass_through_head < 2)
2596                         progress += khugepaged_scan_mm_slot(pages - progress,
2597                                                             &hpage);
2598                 else
2599                         progress = pages;
2600                 spin_unlock(&khugepaged_mm_lock);
2601         }
2602
2603         if (!IS_ERR_OR_NULL(hpage))
2604                 put_page(hpage);
2605 }
2606
2607 static void khugepaged_wait_work(void)
2608 {
2609         try_to_freeze();
2610
2611         if (khugepaged_has_work()) {
2612                 if (!khugepaged_scan_sleep_millisecs)
2613                         return;
2614
2615                 wait_event_freezable_timeout(khugepaged_wait,
2616                                              kthread_should_stop(),
2617                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2618                 return;
2619         }
2620
2621         if (khugepaged_enabled())
2622                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2623 }
2624
2625 static int khugepaged(void *none)
2626 {
2627         struct mm_slot *mm_slot;
2628
2629         set_freezable();
2630         set_user_nice(current, 19);
2631
2632         while (!kthread_should_stop()) {
2633                 khugepaged_do_scan();
2634                 khugepaged_wait_work();
2635         }
2636
2637         spin_lock(&khugepaged_mm_lock);
2638         mm_slot = khugepaged_scan.mm_slot;
2639         khugepaged_scan.mm_slot = NULL;
2640         if (mm_slot)
2641                 collect_mm_slot(mm_slot);
2642         spin_unlock(&khugepaged_mm_lock);
2643         return 0;
2644 }
2645
2646 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2647                 unsigned long haddr, pmd_t *pmd)
2648 {
2649         struct mm_struct *mm = vma->vm_mm;
2650         pgtable_t pgtable;
2651         pmd_t _pmd;
2652         int i;
2653
2654         pmdp_clear_flush(vma, haddr, pmd);
2655         /* leave pmd empty until pte is filled */
2656
2657         pgtable = pgtable_trans_huge_withdraw(mm);
2658         pmd_populate(mm, &_pmd, pgtable);
2659
2660         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2661                 pte_t *pte, entry;
2662                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2663                 entry = pte_mkspecial(entry);
2664                 pte = pte_offset_map(&_pmd, haddr);
2665                 VM_BUG_ON(!pte_none(*pte));
2666                 set_pte_at(mm, haddr, pte, entry);
2667                 pte_unmap(pte);
2668         }
2669         smp_wmb(); /* make pte visible before pmd */
2670         pmd_populate(mm, pmd, pgtable);
2671         put_huge_zero_page();
2672 }
2673
2674 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2675                 pmd_t *pmd)
2676 {
2677         struct page *page;
2678         struct mm_struct *mm = vma->vm_mm;
2679         unsigned long haddr = address & HPAGE_PMD_MASK;
2680         unsigned long mmun_start;       /* For mmu_notifiers */
2681         unsigned long mmun_end;         /* For mmu_notifiers */
2682
2683         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2684
2685         mmun_start = haddr;
2686         mmun_end   = haddr + HPAGE_PMD_SIZE;
2687         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2688         spin_lock(&mm->page_table_lock);
2689         if (unlikely(!pmd_trans_huge(*pmd))) {
2690                 spin_unlock(&mm->page_table_lock);
2691                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2692                 return;
2693         }
2694         if (is_huge_zero_pmd(*pmd)) {
2695                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2696                 spin_unlock(&mm->page_table_lock);
2697                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2698                 return;
2699         }
2700         page = pmd_page(*pmd);
2701         VM_BUG_ON(!page_count(page));
2702         get_page(page);
2703         spin_unlock(&mm->page_table_lock);
2704         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2705
2706         split_huge_page(page);
2707
2708         put_page(page);
2709         BUG_ON(pmd_trans_huge(*pmd));
2710 }
2711
2712 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2713                 pmd_t *pmd)
2714 {
2715         struct vm_area_struct *vma;
2716
2717         vma = find_vma(mm, address);
2718         BUG_ON(vma == NULL);
2719         split_huge_page_pmd(vma, address, pmd);
2720 }
2721
2722 static void split_huge_page_address(struct mm_struct *mm,
2723                                     unsigned long address)
2724 {
2725         pmd_t *pmd;
2726
2727         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2728
2729         pmd = mm_find_pmd(mm, address);
2730         if (!pmd)
2731                 return;
2732         /*
2733          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2734          * materialize from under us.
2735          */
2736         split_huge_page_pmd_mm(mm, address, pmd);
2737 }
2738
2739 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2740                              unsigned long start,
2741                              unsigned long end,
2742                              long adjust_next)
2743 {
2744         /*
2745          * If the new start address isn't hpage aligned and it could
2746          * previously contain an hugepage: check if we need to split
2747          * an huge pmd.
2748          */
2749         if (start & ~HPAGE_PMD_MASK &&
2750             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2751             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2752                 split_huge_page_address(vma->vm_mm, start);
2753
2754         /*
2755          * If the new end address isn't hpage aligned and it could
2756          * previously contain an hugepage: check if we need to split
2757          * an huge pmd.
2758          */
2759         if (end & ~HPAGE_PMD_MASK &&
2760             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2761             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2762                 split_huge_page_address(vma->vm_mm, end);
2763
2764         /*
2765          * If we're also updating the vma->vm_next->vm_start, if the new
2766          * vm_next->vm_start isn't page aligned and it could previously
2767          * contain an hugepage: check if we need to split an huge pmd.
2768          */
2769         if (adjust_next > 0) {
2770                 struct vm_area_struct *next = vma->vm_next;
2771                 unsigned long nstart = next->vm_start;
2772                 nstart += adjust_next << PAGE_SHIFT;
2773                 if (nstart & ~HPAGE_PMD_MASK &&
2774                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2775                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2776                         split_huge_page_address(next->vm_mm, nstart);
2777         }
2778 }