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