mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is disabled in order that avoid
  31  * to risk increase the memory footprint of applications without a guaranteed
  32  * benefit. When transparent hugepage support is enabled, is for all mappings,
  33  * and khugepaged scans all mappings.
  34  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  35  * for all hugepage allocations.
  36  */
  37 unsigned long transparent_hugepage_flags __read_mostly =
  38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  39         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  40 #endif
  41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  42         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  43 #endif
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  46         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  47
  48 /* default scan 8*512 pte (or vmas) every 30 second */
  49 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  50 static unsigned int khugepaged_pages_collapsed;
  51 static unsigned int khugepaged_full_scans;
  52 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  53 /* during fragmentation poll the hugepage allocator once every minute */
  54 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  55 static struct task_struct *khugepaged_thread __read_mostly;
  56 static DEFINE_MUTEX(khugepaged_mutex);
  57 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  58 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  59 /*
  60  * default collapse hugepages if there is at least one pte mapped like
  61  * it would have happened if the vma was large enough during page
  62  * fault.
  63  */
  64 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  65
  66 static int khugepaged(void *none);
  67 static int khugepaged_slab_init(void);
  68
  69 #define MM_SLOTS_HASH_BITS 10
  70 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  71
  72 static struct kmem_cache *mm_slot_cache __read_mostly;
  73
  74 /**
  75  * struct mm_slot - hash lookup from mm to mm_slot
  76  * @hash: hash collision list
  77  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  78  * @mm: the mm that this information is valid for
  79  */
  80 struct mm_slot {
  81         struct hlist_node hash;
  82         struct list_head mm_node;
  83         struct mm_struct *mm;
  84 };
  85
  86 /**
  87  * struct khugepaged_scan - cursor for scanning
  88  * @mm_head: the head of the mm list to scan
  89  * @mm_slot: the current mm_slot we are scanning
  90  * @address: the next address inside that to be scanned
  91  *
  92  * There is only the one khugepaged_scan instance of this cursor structure.
  93  */
  94 struct khugepaged_scan {
  95         struct list_head mm_head;
  96         struct mm_slot *mm_slot;
  97         unsigned long address;
  98 };
  99 static struct khugepaged_scan khugepaged_scan = {
 100         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 101 };
 102
 103
 104 static int set_recommended_min_free_kbytes(void)
 105 {
 106         struct zone *zone;
 107         int nr_zones = 0;
 108         unsigned long recommended_min;
 109
 110         if (!khugepaged_enabled())
 111                 return 0;
 112
 113         for_each_populated_zone(zone)
 114                 nr_zones++;
 115
 116         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 117         recommended_min = pageblock_nr_pages * nr_zones * 2;
 118
 119         /*
 120          * Make sure that on average at least two pageblocks are almost free
 121          * of another type, one for a migratetype to fall back to and a
 122          * second to avoid subsequent fallbacks of other types There are 3
 123          * MIGRATE_TYPES we care about.
 124          */
 125         recommended_min += pageblock_nr_pages * nr_zones *
 126                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 127
 128         /* don't ever allow to reserve more than 5% of the lowmem */
 129         recommended_min = min(recommended_min,
 130                               (unsigned long) nr_free_buffer_pages() / 20);
 131         recommended_min <<= (PAGE_SHIFT-10);
 132
 133         if (recommended_min > min_free_kbytes)
 134                 min_free_kbytes = recommended_min;
 135         setup_per_zone_wmarks();
 136         return 0;
 137 }
 138 late_initcall(set_recommended_min_free_kbytes);
 139
 140 static int start_khugepaged(void)
 141 {
 142         int err = 0;
 143         if (khugepaged_enabled()) {
 144                 if (!khugepaged_thread)
 145                         khugepaged_thread = kthread_run(khugepaged, NULL,
 146                                                         "khugepaged");
 147                 if (unlikely(IS_ERR(khugepaged_thread))) {
 148                         printk(KERN_ERR
 149                                "khugepaged: kthread_run(khugepaged) failed\n");
 150                         err = PTR_ERR(khugepaged_thread);
 151                         khugepaged_thread = NULL;
 152                 }
 153
 154                 if (!list_empty(&khugepaged_scan.mm_head))
 155                         wake_up_interruptible(&khugepaged_wait);
 156
 157                 set_recommended_min_free_kbytes();
 158         } else if (khugepaged_thread) {
 159                 kthread_stop(khugepaged_thread);
 160                 khugepaged_thread = NULL;
 161         }
 162
 163         return err;
 164 }
 165
 166 static atomic_t huge_zero_refcount;
 167 static struct page *huge_zero_page __read_mostly;
 168
 169 static inline bool is_huge_zero_page(struct page *page)
 170 {
 171         return ACCESS_ONCE(huge_zero_page) == page;
 172 }
 173
 174 static inline bool is_huge_zero_pmd(pmd_t pmd)
 175 {
 176         return is_huge_zero_page(pmd_page(pmd));
 177 }
 178
 179 static struct page *get_huge_zero_page(void)
 180 {
 181         struct page *zero_page;
 182 retry:
 183         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 184                 return ACCESS_ONCE(huge_zero_page);
 185
 186         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 187                         HPAGE_PMD_ORDER);
 188         if (!zero_page) {
 189                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 190                 return NULL;
 191         }
 192         count_vm_event(THP_ZERO_PAGE_ALLOC);
 193         preempt_disable();
 194         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 195                 preempt_enable();
 196                 __free_page(zero_page);
 197                 goto retry;
 198         }
 199
 200         /* We take additional reference here. It will be put back by shrinker */
 201         atomic_set(&huge_zero_refcount, 2);
 202         preempt_enable();
 203         return ACCESS_ONCE(huge_zero_page);
 204 }
 205
 206 static void put_huge_zero_page(void)
 207 {
 208         /*
 209          * Counter should never go to zero here. Only shrinker can put
 210          * last reference.
 211          */
 212         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 213 }
 214
 215 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 216                                         struct shrink_control *sc)
 217 {
 218         /* we can free zero page only if last reference remains */
 219         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 220 }
 221
 222 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 223                                        struct shrink_control *sc)
 224 {
 225         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 226                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 227                 BUG_ON(zero_page == NULL);
 228                 __free_page(zero_page);
 229                 return HPAGE_PMD_NR;
 230         }
 231
 232         return 0;
 233 }
 234
 235 static struct shrinker huge_zero_page_shrinker = {
 236         .count_objects = shrink_huge_zero_page_count,
 237         .scan_objects = shrink_huge_zero_page_scan,
 238         .seeks = DEFAULT_SEEKS,
 239 };
 240
 241 #ifdef CONFIG_SYSFS
 242
 243 static ssize_t double_flag_show(struct kobject *kobj,
 244                                 struct kobj_attribute *attr, char *buf,
 245                                 enum transparent_hugepage_flag enabled,
 246                                 enum transparent_hugepage_flag req_madv)
 247 {
 248         if (test_bit(enabled, &transparent_hugepage_flags)) {
 249                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 250                 return sprintf(buf, "[always] madvise never\n");
 251         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 252                 return sprintf(buf, "always [madvise] never\n");
 253         else
 254                 return sprintf(buf, "always madvise [never]\n");
 255 }
 256 static ssize_t double_flag_store(struct kobject *kobj,
 257                                  struct kobj_attribute *attr,
 258                                  const char *buf, size_t count,
 259                                  enum transparent_hugepage_flag enabled,
 260                                  enum transparent_hugepage_flag req_madv)
 261 {
 262         if (!memcmp("always", buf,
 263                     min(sizeof("always")-1, count))) {
 264                 set_bit(enabled, &transparent_hugepage_flags);
 265                 clear_bit(req_madv, &transparent_hugepage_flags);
 266         } else if (!memcmp("madvise", buf,
 267                            min(sizeof("madvise")-1, count))) {
 268                 clear_bit(enabled, &transparent_hugepage_flags);
 269                 set_bit(req_madv, &transparent_hugepage_flags);
 270         } else if (!memcmp("never", buf,
 271                            min(sizeof("never")-1, count))) {
 272                 clear_bit(enabled, &transparent_hugepage_flags);
 273                 clear_bit(req_madv, &transparent_hugepage_flags);
 274         } else
 275                 return -EINVAL;
 276
 277         return count;
 278 }
 279
 280 static ssize_t enabled_show(struct kobject *kobj,
 281                             struct kobj_attribute *attr, char *buf)
 282 {
 283         return double_flag_show(kobj, attr, buf,
 284                                 TRANSPARENT_HUGEPAGE_FLAG,
 285                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 286 }
 287 static ssize_t enabled_store(struct kobject *kobj,
 288                              struct kobj_attribute *attr,
 289                              const char *buf, size_t count)
 290 {
 291         ssize_t ret;
 292
 293         ret = double_flag_store(kobj, attr, buf, count,
 294                                 TRANSPARENT_HUGEPAGE_FLAG,
 295                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 296
 297         if (ret > 0) {
 298                 int err;
 299
 300                 mutex_lock(&khugepaged_mutex);
 301                 err = start_khugepaged();
 302                 mutex_unlock(&khugepaged_mutex);
 303
 304                 if (err)
 305                         ret = err;
 306         }
 307
 308         return ret;
 309 }
 310 static struct kobj_attribute enabled_attr =
 311         __ATTR(enabled, 0644, enabled_show, enabled_store);
 312
 313 static ssize_t single_flag_show(struct kobject *kobj,
 314                                 struct kobj_attribute *attr, char *buf,
 315                                 enum transparent_hugepage_flag flag)
 316 {
 317         return sprintf(buf, "%d\n",
 318                        !!test_bit(flag, &transparent_hugepage_flags));
 319 }
 320
 321 static ssize_t single_flag_store(struct kobject *kobj,
 322                                  struct kobj_attribute *attr,
 323                                  const char *buf, size_t count,
 324                                  enum transparent_hugepage_flag flag)
 325 {
 326         unsigned long value;
 327         int ret;
 328
 329         ret = kstrtoul(buf, 10, &value);
 330         if (ret < 0)
 331                 return ret;
 332         if (value > 1)
 333                 return -EINVAL;
 334
 335         if (value)
 336                 set_bit(flag, &transparent_hugepage_flags);
 337         else
 338                 clear_bit(flag, &transparent_hugepage_flags);
 339
 340         return count;
 341 }
 342
 343 /*
 344  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 345  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 346  * memory just to allocate one more hugepage.
 347  */
 348 static ssize_t defrag_show(struct kobject *kobj,
 349                            struct kobj_attribute *attr, char *buf)
 350 {
 351         return double_flag_show(kobj, attr, buf,
 352                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 353                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 354 }
 355 static ssize_t defrag_store(struct kobject *kobj,
 356                             struct kobj_attribute *attr,
 357                             const char *buf, size_t count)
 358 {
 359         return double_flag_store(kobj, attr, buf, count,
 360                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 361                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 362 }
 363 static struct kobj_attribute defrag_attr =
 364         __ATTR(defrag, 0644, defrag_show, defrag_store);
 365
 366 static ssize_t use_zero_page_show(struct kobject *kobj,
 367                 struct kobj_attribute *attr, char *buf)
 368 {
 369         return single_flag_show(kobj, attr, buf,
 370                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 371 }
 372 static ssize_t use_zero_page_store(struct kobject *kobj,
 373                 struct kobj_attribute *attr, const char *buf, size_t count)
 374 {
 375         return single_flag_store(kobj, attr, buf, count,
 376                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static struct kobj_attribute use_zero_page_attr =
 379         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 380 #ifdef CONFIG_DEBUG_VM
 381 static ssize_t debug_cow_show(struct kobject *kobj,
 382                                 struct kobj_attribute *attr, char *buf)
 383 {
 384         return single_flag_show(kobj, attr, buf,
 385                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 386 }
 387 static ssize_t debug_cow_store(struct kobject *kobj,
 388                                struct kobj_attribute *attr,
 389                                const char *buf, size_t count)
 390 {
 391         return single_flag_store(kobj, attr, buf, count,
 392                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 393 }
 394 static struct kobj_attribute debug_cow_attr =
 395         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 396 #endif /* CONFIG_DEBUG_VM */
 397
 398 static struct attribute *hugepage_attr[] = {
 399         &enabled_attr.attr,
 400         &defrag_attr.attr,
 401         &use_zero_page_attr.attr,
 402 #ifdef CONFIG_DEBUG_VM
 403         &debug_cow_attr.attr,
 404 #endif
 405         NULL,
 406 };
 407
 408 static struct attribute_group hugepage_attr_group = {
 409         .attrs = hugepage_attr,
 410 };
 411
 412 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 413                                          struct kobj_attribute *attr,
 414                                          char *buf)
 415 {
 416         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 417 }
 418
 419 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 420                                           struct kobj_attribute *attr,
 421                                           const char *buf, size_t count)
 422 {
 423         unsigned long msecs;
 424         int err;
 425
 426         err = kstrtoul(buf, 10, &msecs);
 427         if (err || msecs > UINT_MAX)
 428                 return -EINVAL;
 429
 430         khugepaged_scan_sleep_millisecs = msecs;
 431         wake_up_interruptible(&khugepaged_wait);
 432
 433         return count;
 434 }
 435 static struct kobj_attribute scan_sleep_millisecs_attr =
 436         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 437                scan_sleep_millisecs_store);
 438
 439 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 440                                           struct kobj_attribute *attr,
 441                                           char *buf)
 442 {
 443         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 444 }
 445
 446 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 447                                            struct kobj_attribute *attr,
 448                                            const char *buf, size_t count)
 449 {
 450         unsigned long msecs;
 451         int err;
 452
 453         err = kstrtoul(buf, 10, &msecs);
 454         if (err || msecs > UINT_MAX)
 455                 return -EINVAL;
 456
 457         khugepaged_alloc_sleep_millisecs = msecs;
 458         wake_up_interruptible(&khugepaged_wait);
 459
 460         return count;
 461 }
 462 static struct kobj_attribute alloc_sleep_millisecs_attr =
 463         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 464                alloc_sleep_millisecs_store);
 465
 466 static ssize_t pages_to_scan_show(struct kobject *kobj,
 467                                   struct kobj_attribute *attr,
 468                                   char *buf)
 469 {
 470         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 471 }
 472 static ssize_t pages_to_scan_store(struct kobject *kobj,
 473                                    struct kobj_attribute *attr,
 474                                    const char *buf, size_t count)
 475 {
 476         int err;
 477         unsigned long pages;
 478
 479         err = kstrtoul(buf, 10, &pages);
 480         if (err || !pages || pages > UINT_MAX)
 481                 return -EINVAL;
 482
 483         khugepaged_pages_to_scan = pages;
 484
 485         return count;
 486 }
 487 static struct kobj_attribute pages_to_scan_attr =
 488         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 489                pages_to_scan_store);
 490
 491 static ssize_t pages_collapsed_show(struct kobject *kobj,
 492                                     struct kobj_attribute *attr,
 493                                     char *buf)
 494 {
 495         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 496 }
 497 static struct kobj_attribute pages_collapsed_attr =
 498         __ATTR_RO(pages_collapsed);
 499
 500 static ssize_t full_scans_show(struct kobject *kobj,
 501                                struct kobj_attribute *attr,
 502                                char *buf)
 503 {
 504         return sprintf(buf, "%u\n", khugepaged_full_scans);
 505 }
 506 static struct kobj_attribute full_scans_attr =
 507         __ATTR_RO(full_scans);
 508
 509 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 510                                       struct kobj_attribute *attr, char *buf)
 511 {
 512         return single_flag_show(kobj, attr, buf,
 513                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 514 }
 515 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 516                                        struct kobj_attribute *attr,
 517                                        const char *buf, size_t count)
 518 {
 519         return single_flag_store(kobj, attr, buf, count,
 520                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 521 }
 522 static struct kobj_attribute khugepaged_defrag_attr =
 523         __ATTR(defrag, 0644, khugepaged_defrag_show,
 524                khugepaged_defrag_store);
 525
 526 /*
 527  * max_ptes_none controls if khugepaged should collapse hugepages over
 528  * any unmapped ptes in turn potentially increasing the memory
 529  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 530  * reduce the available free memory in the system as it
 531  * runs. Increasing max_ptes_none will instead potentially reduce the
 532  * free memory in the system during the khugepaged scan.
 533  */
 534 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 535                                              struct kobj_attribute *attr,
 536                                              char *buf)
 537 {
 538         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 539 }
 540 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 541                                               struct kobj_attribute *attr,
 542                                               const char *buf, size_t count)
 543 {
 544         int err;
 545         unsigned long max_ptes_none;
 546
 547         err = kstrtoul(buf, 10, &max_ptes_none);
 548         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 549                 return -EINVAL;
 550
 551         khugepaged_max_ptes_none = max_ptes_none;
 552
 553         return count;
 554 }
 555 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 556         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 557                khugepaged_max_ptes_none_store);
 558
 559 static struct attribute *khugepaged_attr[] = {
 560         &khugepaged_defrag_attr.attr,
 561         &khugepaged_max_ptes_none_attr.attr,
 562         &pages_to_scan_attr.attr,
 563         &pages_collapsed_attr.attr,
 564         &full_scans_attr.attr,
 565         &scan_sleep_millisecs_attr.attr,
 566         &alloc_sleep_millisecs_attr.attr,
 567         NULL,
 568 };
 569
 570 static struct attribute_group khugepaged_attr_group = {
 571         .attrs = khugepaged_attr,
 572         .name = "khugepaged",
 573 };
 574
 575 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 576 {
 577         int err;
 578
 579         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 580         if (unlikely(!*hugepage_kobj)) {
 581                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 582                 return -ENOMEM;
 583         }
 584
 585         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 586         if (err) {
 587                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 588                 goto delete_obj;
 589         }
 590
 591         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 592         if (err) {
 593                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 594                 goto remove_hp_group;
 595         }
 596
 597         return 0;
 598
 599 remove_hp_group:
 600         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 601 delete_obj:
 602         kobject_put(*hugepage_kobj);
 603         return err;
 604 }
 605
 606 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 607 {
 608         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 609         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 610         kobject_put(hugepage_kobj);
 611 }
 612 #else
 613 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 614 {
 615         return 0;
 616 }
 617
 618 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 619 {
 620 }
 621 #endif /* CONFIG_SYSFS */
 622
 623 static int __init hugepage_init(void)
 624 {
 625         int err;
 626         struct kobject *hugepage_kobj;
 627
 628         if (!has_transparent_hugepage()) {
 629                 transparent_hugepage_flags = 0;
 630                 return -EINVAL;
 631         }
 632
 633         err = hugepage_init_sysfs(&hugepage_kobj);
 634         if (err)
 635                 return err;
 636
 637         err = khugepaged_slab_init();
 638         if (err)
 639                 goto out;
 640
 641         register_shrinker(&huge_zero_page_shrinker);
 642
 643         /*
 644          * By default disable transparent hugepages on smaller systems,
 645          * where the extra memory used could hurt more than TLB overhead
 646          * is likely to save.  The admin can still enable it through /sys.
 647          */
 648         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 649                 transparent_hugepage_flags = 0;
 650
 651         start_khugepaged();
 652
 653         return 0;
 654 out:
 655         hugepage_exit_sysfs(hugepage_kobj);
 656         return err;
 657 }
 658 module_init(hugepage_init)
 659
 660 static int __init setup_transparent_hugepage(char *str)
 661 {
 662         int ret = 0;
 663         if (!str)
 664                 goto out;
 665         if (!strcmp(str, "always")) {
 666                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 667                         &transparent_hugepage_flags);
 668                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 669                           &transparent_hugepage_flags);
 670                 ret = 1;
 671         } else if (!strcmp(str, "madvise")) {
 672                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 673                           &transparent_hugepage_flags);
 674                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 675                         &transparent_hugepage_flags);
 676                 ret = 1;
 677         } else if (!strcmp(str, "never")) {
 678                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 679                           &transparent_hugepage_flags);
 680                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 681                           &transparent_hugepage_flags);
 682                 ret = 1;
 683         }
 684 out:
 685         if (!ret)
 686                 printk(KERN_WARNING
 687                        "transparent_hugepage= cannot parse, ignored\n");
 688         return ret;
 689 }
 690 __setup("transparent_hugepage=", setup_transparent_hugepage);
 691
 692 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 693 {
 694         if (likely(vma->vm_flags & VM_WRITE))
 695                 pmd = pmd_mkwrite(pmd);
 696         return pmd;
 697 }
 698
 699 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 700 {
 701         pmd_t entry;
 702         entry = mk_pmd(page, prot);
 703         entry = pmd_mkhuge(entry);
 704         return entry;
 705 }
 706
 707 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 708                                         struct vm_area_struct *vma,
 709                                         unsigned long haddr, pmd_t *pmd,
 710                                         struct page *page)
 711 {
 712         pgtable_t pgtable;
 713         spinlock_t *ptl;
 714
 715         VM_BUG_ON(!PageCompound(page));
 716         pgtable = pte_alloc_one(mm, haddr);
 717         if (unlikely(!pgtable))
 718                 return VM_FAULT_OOM;
 719
 720         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 721         /*
 722          * The memory barrier inside __SetPageUptodate makes sure that
 723          * clear_huge_page writes become visible before the set_pmd_at()
 724          * write.
 725          */
 726         __SetPageUptodate(page);
 727
 728         ptl = pmd_lock(mm, pmd);
 729         if (unlikely(!pmd_none(*pmd))) {
 730                 spin_unlock(ptl);
 731                 mem_cgroup_uncharge_page(page);
 732                 put_page(page);
 733                 pte_free(mm, pgtable);
 734         } else {
 735                 pmd_t entry;
 736                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 737                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 738                 page_add_new_anon_rmap(page, vma, haddr);
 739                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 740                 set_pmd_at(mm, haddr, pmd, entry);
 741                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 742                 atomic_long_inc(&mm->nr_ptes);
 743                 spin_unlock(ptl);
 744         }
 745
 746         return 0;
 747 }
 748
 749 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 750 {
 751         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 752 }
 753
 754 static inline struct page *alloc_hugepage_vma(int defrag,
 755                                               struct vm_area_struct *vma,
 756                                               unsigned long haddr, int nd,
 757                                               gfp_t extra_gfp)
 758 {
 759         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 760                                HPAGE_PMD_ORDER, vma, haddr, nd);
 761 }
 762
 763 /* Caller must hold page table lock. */
 764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 765                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 766                 struct page *zero_page)
 767 {
 768         pmd_t entry;
 769         if (!pmd_none(*pmd))
 770                 return false;
 771         entry = mk_pmd(zero_page, vma->vm_page_prot);
 772         entry = pmd_wrprotect(entry);
 773         entry = pmd_mkhuge(entry);
 774         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 775         set_pmd_at(mm, haddr, pmd, entry);
 776         atomic_long_inc(&mm->nr_ptes);
 777         return true;
 778 }
 779
 780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 781                                unsigned long address, pmd_t *pmd,
 782                                unsigned int flags)
 783 {
 784         struct page *page;
 785         unsigned long haddr = address & HPAGE_PMD_MASK;
 786
 787         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 788                 return VM_FAULT_FALLBACK;
 789         if (unlikely(anon_vma_prepare(vma)))
 790                 return VM_FAULT_OOM;
 791         if (unlikely(khugepaged_enter(vma)))
 792                 return VM_FAULT_OOM;
 793         if (!(flags & FAULT_FLAG_WRITE) &&
 794                         transparent_hugepage_use_zero_page()) {
 795                 spinlock_t *ptl;
 796                 pgtable_t pgtable;
 797                 struct page *zero_page;
 798                 bool set;
 799                 pgtable = pte_alloc_one(mm, haddr);
 800                 if (unlikely(!pgtable))
 801                         return VM_FAULT_OOM;
 802                 zero_page = get_huge_zero_page();
 803                 if (unlikely(!zero_page)) {
 804                         pte_free(mm, pgtable);
 805                         count_vm_event(THP_FAULT_FALLBACK);
 806                         return VM_FAULT_FALLBACK;
 807                 }
 808                 ptl = pmd_lock(mm, pmd);
 809                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 810                                 zero_page);
 811                 spin_unlock(ptl);
 812                 if (!set) {
 813                         pte_free(mm, pgtable);
 814                         put_huge_zero_page();
 815                 }
 816                 return 0;
 817         }
 818         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 819                         vma, haddr, numa_node_id(), 0);
 820         if (unlikely(!page)) {
 821                 count_vm_event(THP_FAULT_FALLBACK);
 822                 return VM_FAULT_FALLBACK;
 823         }
 824         if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 825                 put_page(page);
 826                 count_vm_event(THP_FAULT_FALLBACK);
 827                 return VM_FAULT_FALLBACK;
 828         }
 829         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 830                 mem_cgroup_uncharge_page(page);
 831                 put_page(page);
 832                 count_vm_event(THP_FAULT_FALLBACK);
 833                 return VM_FAULT_FALLBACK;
 834         }
 835
 836         count_vm_event(THP_FAULT_ALLOC);
 837         return 0;
 838 }
 839
 840 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 841                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 842                   struct vm_area_struct *vma)
 843 {
 844         spinlock_t *dst_ptl, *src_ptl;
 845         struct page *src_page;
 846         pmd_t pmd;
 847         pgtable_t pgtable;
 848         int ret;
 849
 850         ret = -ENOMEM;
 851         pgtable = pte_alloc_one(dst_mm, addr);
 852         if (unlikely(!pgtable))
 853                 goto out;
 854
 855         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 856         src_ptl = pmd_lockptr(src_mm, src_pmd);
 857         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 858
 859         ret = -EAGAIN;
 860         pmd = *src_pmd;
 861         if (unlikely(!pmd_trans_huge(pmd))) {
 862                 pte_free(dst_mm, pgtable);
 863                 goto out_unlock;
 864         }
 865         /*
 866          * When page table lock is held, the huge zero pmd should not be
 867          * under splitting since we don't split the page itself, only pmd to
 868          * a page table.
 869          */
 870         if (is_huge_zero_pmd(pmd)) {
 871                 struct page *zero_page;
 872                 bool set;
 873                 /*
 874                  * get_huge_zero_page() will never allocate a new page here,
 875                  * since we already have a zero page to copy. It just takes a
 876                  * reference.
 877                  */
 878                 zero_page = get_huge_zero_page();
 879                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 880                                 zero_page);
 881                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 882                 ret = 0;
 883                 goto out_unlock;
 884         }
 885
 886         /* mmap_sem prevents this happening but warn if that changes */
 887         WARN_ON(pmd_trans_migrating(pmd));
 888
 889         if (unlikely(pmd_trans_splitting(pmd))) {
 890                 /* split huge page running from under us */
 891                 spin_unlock(src_ptl);
 892                 spin_unlock(dst_ptl);
 893                 pte_free(dst_mm, pgtable);
 894
 895                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 896                 goto out;
 897         }
 898         src_page = pmd_page(pmd);
 899         VM_BUG_ON(!PageHead(src_page));
 900         get_page(src_page);
 901         page_dup_rmap(src_page);
 902         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 903
 904         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 905         pmd = pmd_mkold(pmd_wrprotect(pmd));
 906         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 907         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 908         atomic_long_inc(&dst_mm->nr_ptes);
 909
 910         ret = 0;
 911 out_unlock:
 912         spin_unlock(src_ptl);
 913         spin_unlock(dst_ptl);
 914 out:
 915         return ret;
 916 }
 917
 918 void huge_pmd_set_accessed(struct mm_struct *mm,
 919                            struct vm_area_struct *vma,
 920                            unsigned long address,
 921                            pmd_t *pmd, pmd_t orig_pmd,
 922                            int dirty)
 923 {
 924         spinlock_t *ptl;
 925         pmd_t entry;
 926         unsigned long haddr;
 927
 928         ptl = pmd_lock(mm, pmd);
 929         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 930                 goto unlock;
 931
 932         entry = pmd_mkyoung(orig_pmd);
 933         haddr = address & HPAGE_PMD_MASK;
 934         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 935                 update_mmu_cache_pmd(vma, address, pmd);
 936
 937 unlock:
 938         spin_unlock(ptl);
 939 }
 940
 941 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 942                 struct vm_area_struct *vma, unsigned long address,
 943                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 944 {
 945         spinlock_t *ptl;
 946         pgtable_t pgtable;
 947         pmd_t _pmd;
 948         struct page *page;
 949         int i, ret = 0;
 950         unsigned long mmun_start;       /* For mmu_notifiers */
 951         unsigned long mmun_end;         /* For mmu_notifiers */
 952
 953         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 954         if (!page) {
 955                 ret |= VM_FAULT_OOM;
 956                 goto out;
 957         }
 958
 959         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 960                 put_page(page);
 961                 ret |= VM_FAULT_OOM;
 962                 goto out;
 963         }
 964
 965         clear_user_highpage(page, address);
 966         __SetPageUptodate(page);
 967
 968         mmun_start = haddr;
 969         mmun_end   = haddr + HPAGE_PMD_SIZE;
 970         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 971
 972         ptl = pmd_lock(mm, pmd);
 973         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 974                 goto out_free_page;
 975
 976         pmdp_clear_flush(vma, haddr, pmd);
 977         /* leave pmd empty until pte is filled */
 978
 979         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
 980         pmd_populate(mm, &_pmd, pgtable);
 981
 982         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 983                 pte_t *pte, entry;
 984                 if (haddr == (address & PAGE_MASK)) {
 985                         entry = mk_pte(page, vma->vm_page_prot);
 986                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 987                         page_add_new_anon_rmap(page, vma, haddr);
 988                 } else {
 989                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
 990                         entry = pte_mkspecial(entry);
 991                 }
 992                 pte = pte_offset_map(&_pmd, haddr);
 993                 VM_BUG_ON(!pte_none(*pte));
 994                 set_pte_at(mm, haddr, pte, entry);
 995                 pte_unmap(pte);
 996         }
 997         smp_wmb(); /* make pte visible before pmd */
 998         pmd_populate(mm, pmd, pgtable);
 999         spin_unlock(ptl);
1000         put_huge_zero_page();
1001         inc_mm_counter(mm, MM_ANONPAGES);
1002
1003         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1004
1005         ret |= VM_FAULT_WRITE;
1006 out:
1007         return ret;
1008 out_free_page:
1009         spin_unlock(ptl);
1010         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1011         mem_cgroup_uncharge_page(page);
1012         put_page(page);
1013         goto out;
1014 }
1015
1016 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1017                                         struct vm_area_struct *vma,
1018                                         unsigned long address,
1019                                         pmd_t *pmd, pmd_t orig_pmd,
1020                                         struct page *page,
1021                                         unsigned long haddr)
1022 {
1023         spinlock_t *ptl;
1024         pgtable_t pgtable;
1025         pmd_t _pmd;
1026         int ret = 0, i;
1027         struct page **pages;
1028         unsigned long mmun_start;       /* For mmu_notifiers */
1029         unsigned long mmun_end;         /* For mmu_notifiers */
1030
1031         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1032                         GFP_KERNEL);
1033         if (unlikely(!pages)) {
1034                 ret |= VM_FAULT_OOM;
1035                 goto out;
1036         }
1037
1038         for (i = 0; i < HPAGE_PMD_NR; i++) {
1039                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1040                                                __GFP_OTHER_NODE,
1041                                                vma, address, page_to_nid(page));
1042                 if (unlikely(!pages[i] ||
1043                              mem_cgroup_newpage_charge(pages[i], mm,
1044                                                        GFP_KERNEL))) {
1045                         if (pages[i])
1046                                 put_page(pages[i]);
1047                         mem_cgroup_uncharge_start();
1048                         while (--i >= 0) {
1049                                 mem_cgroup_uncharge_page(pages[i]);
1050                                 put_page(pages[i]);
1051                         }
1052                         mem_cgroup_uncharge_end();
1053                         kfree(pages);
1054                         ret |= VM_FAULT_OOM;
1055                         goto out;
1056                 }
1057         }
1058
1059         for (i = 0; i < HPAGE_PMD_NR; i++) {
1060                 copy_user_highpage(pages[i], page + i,
1061                                    haddr + PAGE_SIZE * i, vma);
1062                 __SetPageUptodate(pages[i]);
1063                 cond_resched();
1064         }
1065
1066         mmun_start = haddr;
1067         mmun_end   = haddr + HPAGE_PMD_SIZE;
1068         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1069
1070         ptl = pmd_lock(mm, pmd);
1071         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1072                 goto out_free_pages;
1073         VM_BUG_ON(!PageHead(page));
1074
1075         pmdp_clear_flush(vma, haddr, pmd);
1076         /* leave pmd empty until pte is filled */
1077
1078         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1079         pmd_populate(mm, &_pmd, pgtable);
1080
1081         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1082                 pte_t *pte, entry;
1083                 entry = mk_pte(pages[i], vma->vm_page_prot);
1084                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1085                 page_add_new_anon_rmap(pages[i], vma, haddr);
1086                 pte = pte_offset_map(&_pmd, haddr);
1087                 VM_BUG_ON(!pte_none(*pte));
1088                 set_pte_at(mm, haddr, pte, entry);
1089                 pte_unmap(pte);
1090         }
1091         kfree(pages);
1092
1093         smp_wmb(); /* make pte visible before pmd */
1094         pmd_populate(mm, pmd, pgtable);
1095         page_remove_rmap(page);
1096         spin_unlock(ptl);
1097
1098         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1099
1100         ret |= VM_FAULT_WRITE;
1101         put_page(page);
1102
1103 out:
1104         return ret;
1105
1106 out_free_pages:
1107         spin_unlock(ptl);
1108         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1109         mem_cgroup_uncharge_start();
1110         for (i = 0; i < HPAGE_PMD_NR; i++) {
1111                 mem_cgroup_uncharge_page(pages[i]);
1112                 put_page(pages[i]);
1113         }
1114         mem_cgroup_uncharge_end();
1115         kfree(pages);
1116         goto out;
1117 }
1118
1119 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1120                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1121 {
1122         spinlock_t *ptl;
1123         int ret = 0;
1124         struct page *page = NULL, *new_page;
1125         unsigned long haddr;
1126         unsigned long mmun_start;       /* For mmu_notifiers */
1127         unsigned long mmun_end;         /* For mmu_notifiers */
1128
1129         ptl = pmd_lockptr(mm, pmd);
1130         VM_BUG_ON(!vma->anon_vma);
1131         haddr = address & HPAGE_PMD_MASK;
1132         if (is_huge_zero_pmd(orig_pmd))
1133                 goto alloc;
1134         spin_lock(ptl);
1135         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1136                 goto out_unlock;
1137
1138         page = pmd_page(orig_pmd);
1139         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1140         if (page_mapcount(page) == 1) {
1141                 pmd_t entry;
1142                 entry = pmd_mkyoung(orig_pmd);
1143                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1144                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1145                         update_mmu_cache_pmd(vma, address, pmd);
1146                 ret |= VM_FAULT_WRITE;
1147                 goto out_unlock;
1148         }
1149         get_page(page);
1150         spin_unlock(ptl);
1151 alloc:
1152         if (transparent_hugepage_enabled(vma) &&
1153             !transparent_hugepage_debug_cow())
1154                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1155                                               vma, haddr, numa_node_id(), 0);
1156         else
1157                 new_page = NULL;
1158
1159         if (unlikely(!new_page)) {
1160                 if (is_huge_zero_pmd(orig_pmd)) {
1161                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1162                                         address, pmd, orig_pmd, haddr);
1163                 } else {
1164                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1165                                         pmd, orig_pmd, page, haddr);
1166                         if (ret & VM_FAULT_OOM)
1167                                 split_huge_page(page);
1168                         put_page(page);
1169                 }
1170                 count_vm_event(THP_FAULT_FALLBACK);
1171                 goto out;
1172         }
1173
1174         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1175                 put_page(new_page);
1176                 if (page) {
1177                         split_huge_page(page);
1178                         put_page(page);
1179                 }
1180                 count_vm_event(THP_FAULT_FALLBACK);
1181                 ret |= VM_FAULT_OOM;
1182                 goto out;
1183         }
1184
1185         count_vm_event(THP_FAULT_ALLOC);
1186
1187         if (is_huge_zero_pmd(orig_pmd))
1188                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1189         else
1190                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1191         __SetPageUptodate(new_page);
1192
1193         mmun_start = haddr;
1194         mmun_end   = haddr + HPAGE_PMD_SIZE;
1195         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1196
1197         spin_lock(ptl);
1198         if (page)
1199                 put_page(page);
1200         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1201                 spin_unlock(ptl);
1202                 mem_cgroup_uncharge_page(new_page);
1203                 put_page(new_page);
1204                 goto out_mn;
1205         } else {
1206                 pmd_t entry;
1207                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1208                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1209                 pmdp_clear_flush(vma, haddr, pmd);
1210                 page_add_new_anon_rmap(new_page, vma, haddr);
1211                 set_pmd_at(mm, haddr, pmd, entry);
1212                 update_mmu_cache_pmd(vma, address, pmd);
1213                 if (is_huge_zero_pmd(orig_pmd)) {
1214                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1215                         put_huge_zero_page();
1216                 } else {
1217                         VM_BUG_ON(!PageHead(page));
1218                         page_remove_rmap(page);
1219                         put_page(page);
1220                 }
1221                 ret |= VM_FAULT_WRITE;
1222         }
1223         spin_unlock(ptl);
1224 out_mn:
1225         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1226 out:
1227         return ret;
1228 out_unlock:
1229         spin_unlock(ptl);
1230         return ret;
1231 }
1232
1233 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1234                                    unsigned long addr,
1235                                    pmd_t *pmd,
1236                                    unsigned int flags)
1237 {
1238         struct mm_struct *mm = vma->vm_mm;
1239         struct page *page = NULL;
1240
1241         assert_spin_locked(pmd_lockptr(mm, pmd));
1242
1243         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1244                 goto out;
1245
1246         /* Avoid dumping huge zero page */
1247         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1248                 return ERR_PTR(-EFAULT);
1249
1250         /* Full NUMA hinting faults to serialise migration in fault paths */
1251         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1252                 goto out;
1253
1254         page = pmd_page(*pmd);
1255         VM_BUG_ON(!PageHead(page));
1256         if (flags & FOLL_TOUCH) {
1257                 pmd_t _pmd;
1258                 /*
1259                  * We should set the dirty bit only for FOLL_WRITE but
1260                  * for now the dirty bit in the pmd is meaningless.
1261                  * And if the dirty bit will become meaningful and
1262                  * we'll only set it with FOLL_WRITE, an atomic
1263                  * set_bit will be required on the pmd to set the
1264                  * young bit, instead of the current set_pmd_at.
1265                  */
1266                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1267                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1268                                           pmd, _pmd,  1))
1269                         update_mmu_cache_pmd(vma, addr, pmd);
1270         }
1271         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1272                 if (page->mapping && trylock_page(page)) {
1273                         lru_add_drain();
1274                         if (page->mapping)
1275                                 mlock_vma_page(page);
1276                         unlock_page(page);
1277                 }
1278         }
1279         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1280         VM_BUG_ON(!PageCompound(page));
1281         if (flags & FOLL_GET)
1282                 get_page_foll(page);
1283
1284 out:
1285         return page;
1286 }
1287
1288 /* NUMA hinting page fault entry point for trans huge pmds */
1289 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1290                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1291 {
1292         spinlock_t *ptl;
1293         struct anon_vma *anon_vma = NULL;
1294         struct page *page;
1295         unsigned long haddr = addr & HPAGE_PMD_MASK;
1296         int page_nid = -1, this_nid = numa_node_id();
1297         int target_nid, last_cpupid = -1;
1298         bool page_locked;
1299         bool migrated = false;
1300         int flags = 0;
1301
1302         ptl = pmd_lock(mm, pmdp);
1303         if (unlikely(!pmd_same(pmd, *pmdp)))
1304                 goto out_unlock;
1305
1306         /*
1307          * If there are potential migrations, wait for completion and retry
1308          * without disrupting NUMA hinting information. Do not relock and
1309          * check_same as the page may no longer be mapped.
1310          */
1311         if (unlikely(pmd_trans_migrating(*pmdp))) {
1312                 spin_unlock(ptl);
1313                 wait_migrate_huge_page(vma->anon_vma, pmdp);
1314                 goto out;
1315         }
1316
1317         page = pmd_page(pmd);
1318         BUG_ON(is_huge_zero_page(page));
1319         page_nid = page_to_nid(page);
1320         last_cpupid = page_cpupid_last(page);
1321         count_vm_numa_event(NUMA_HINT_FAULTS);
1322         if (page_nid == this_nid) {
1323                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1324                 flags |= TNF_FAULT_LOCAL;
1325         }
1326
1327         /*
1328          * Avoid grouping on DSO/COW pages in specific and RO pages
1329          * in general, RO pages shouldn't hurt as much anyway since
1330          * they can be in shared cache state.
1331          */
1332         if (!pmd_write(pmd))
1333                 flags |= TNF_NO_GROUP;
1334
1335         /*
1336          * Acquire the page lock to serialise THP migrations but avoid dropping
1337          * page_table_lock if at all possible
1338          */
1339         page_locked = trylock_page(page);
1340         target_nid = mpol_misplaced(page, vma, haddr);
1341         if (target_nid == -1) {
1342                 /* If the page was locked, there are no parallel migrations */
1343                 if (page_locked)
1344                         goto clear_pmdnuma;
1345         }
1346
1347         /* Migration could have started since the pmd_trans_migrating check */
1348         if (!page_locked) {
1349                 spin_unlock(ptl);
1350                 wait_on_page_locked(page);
1351                 page_nid = -1;
1352                 goto out;
1353         }
1354
1355         /*
1356          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1357          * to serialises splits
1358          */
1359         get_page(page);
1360         spin_unlock(ptl);
1361         anon_vma = page_lock_anon_vma_read(page);
1362
1363         /* Confirm the PMD did not change while page_table_lock was released */
1364         spin_lock(ptl);
1365         if (unlikely(!pmd_same(pmd, *pmdp))) {
1366                 unlock_page(page);
1367                 put_page(page);
1368                 page_nid = -1;
1369                 goto out_unlock;
1370         }
1371
1372         /* Bail if we fail to protect against THP splits for any reason */
1373         if (unlikely(!anon_vma)) {
1374                 put_page(page);
1375                 page_nid = -1;
1376                 goto clear_pmdnuma;
1377         }
1378
1379         /*
1380          * The page_table_lock above provides a memory barrier
1381          * with change_protection_range.
1382          */
1383         if (mm_tlb_flush_pending(mm))
1384                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1385
1386         /*
1387          * Migrate the THP to the requested node, returns with page unlocked
1388          * and pmd_numa cleared.
1389          */
1390         spin_unlock(ptl);
1391         migrated = migrate_misplaced_transhuge_page(mm, vma,
1392                                 pmdp, pmd, addr, page, target_nid);
1393         if (migrated) {
1394                 flags |= TNF_MIGRATED;
1395                 page_nid = target_nid;
1396         }
1397
1398         goto out;
1399 clear_pmdnuma:
1400         BUG_ON(!PageLocked(page));
1401         pmd = pmd_mknonnuma(pmd);
1402         set_pmd_at(mm, haddr, pmdp, pmd);
1403         VM_BUG_ON(pmd_numa(*pmdp));
1404         update_mmu_cache_pmd(vma, addr, pmdp);
1405         unlock_page(page);
1406 out_unlock:
1407         spin_unlock(ptl);
1408
1409 out:
1410         if (anon_vma)
1411                 page_unlock_anon_vma_read(anon_vma);
1412
1413         if (page_nid != -1)
1414                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1415
1416         return 0;
1417 }
1418
1419 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1420                  pmd_t *pmd, unsigned long addr)
1421 {
1422         spinlock_t *ptl;
1423         int ret = 0;
1424
1425         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1426                 struct page *page;
1427                 pgtable_t pgtable;
1428                 pmd_t orig_pmd;
1429                 /*
1430                  * For architectures like ppc64 we look at deposited pgtable
1431                  * when calling pmdp_get_and_clear. So do the
1432                  * pgtable_trans_huge_withdraw after finishing pmdp related
1433                  * operations.
1434                  */
1435                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1436                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1437                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1438                 if (is_huge_zero_pmd(orig_pmd)) {
1439                         atomic_long_dec(&tlb->mm->nr_ptes);
1440                         spin_unlock(ptl);
1441                         put_huge_zero_page();
1442                 } else {
1443                         page = pmd_page(orig_pmd);
1444                         page_remove_rmap(page);
1445                         VM_BUG_ON(page_mapcount(page) < 0);
1446                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1447                         VM_BUG_ON(!PageHead(page));
1448                         atomic_long_dec(&tlb->mm->nr_ptes);
1449                         spin_unlock(ptl);
1450                         tlb_remove_page(tlb, page);
1451                 }
1452                 pte_free(tlb->mm, pgtable);
1453                 ret = 1;
1454         }
1455         return ret;
1456 }
1457
1458 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1459                 unsigned long addr, unsigned long end,
1460                 unsigned char *vec)
1461 {
1462         spinlock_t *ptl;
1463         int ret = 0;
1464
1465         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1466                 /*
1467                  * All logical pages in the range are present
1468                  * if backed by a huge page.
1469                  */
1470                 spin_unlock(ptl);
1471                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1472                 ret = 1;
1473         }
1474
1475         return ret;
1476 }
1477
1478 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1479                   unsigned long old_addr,
1480                   unsigned long new_addr, unsigned long old_end,
1481                   pmd_t *old_pmd, pmd_t *new_pmd)
1482 {
1483         spinlock_t *old_ptl, *new_ptl;
1484         int ret = 0;
1485         pmd_t pmd;
1486
1487         struct mm_struct *mm = vma->vm_mm;
1488
1489         if ((old_addr & ~HPAGE_PMD_MASK) ||
1490             (new_addr & ~HPAGE_PMD_MASK) ||
1491             old_end - old_addr < HPAGE_PMD_SIZE ||
1492             (new_vma->vm_flags & VM_NOHUGEPAGE))
1493                 goto out;
1494
1495         /*
1496          * The destination pmd shouldn't be established, free_pgtables()
1497          * should have release it.
1498          */
1499         if (WARN_ON(!pmd_none(*new_pmd))) {
1500                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1501                 goto out;
1502         }
1503
1504         /*
1505          * We don't have to worry about the ordering of src and dst
1506          * ptlocks because exclusive mmap_sem prevents deadlock.
1507          */
1508         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1509         if (ret == 1) {
1510                 new_ptl = pmd_lockptr(mm, new_pmd);
1511                 if (new_ptl != old_ptl)
1512                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1513                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1514                 VM_BUG_ON(!pmd_none(*new_pmd));
1515                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1516                 if (new_ptl != old_ptl) {
1517                         pgtable_t pgtable;
1518
1519                         /*
1520                          * Move preallocated PTE page table if new_pmd is on
1521                          * different PMD page table.
1522                          */
1523                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1524                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1525
1526                         spin_unlock(new_ptl);
1527                 }
1528                 spin_unlock(old_ptl);
1529         }
1530 out:
1531         return ret;
1532 }
1533
1534 /*
1535  * Returns
1536  *  - 0 if PMD could not be locked
1537  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1538  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1539  */
1540 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1541                 unsigned long addr, pgprot_t newprot, int prot_numa)
1542 {
1543         struct mm_struct *mm = vma->vm_mm;
1544         spinlock_t *ptl;
1545         int ret = 0;
1546
1547         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1548                 pmd_t entry;
1549                 ret = 1;
1550                 if (!prot_numa) {
1551                         entry = pmdp_get_and_clear(mm, addr, pmd);
1552                         if (pmd_numa(entry))
1553                                 entry = pmd_mknonnuma(entry);
1554                         entry = pmd_modify(entry, newprot);
1555                         ret = HPAGE_PMD_NR;
1556                         BUG_ON(pmd_write(entry));
1557                 } else {
1558                         struct page *page = pmd_page(*pmd);
1559
1560                         /*
1561                          * Do not trap faults against the zero page. The
1562                          * read-only data is likely to be read-cached on the
1563                          * local CPU cache and it is less useful to know about
1564                          * local vs remote hits on the zero page.
1565                          */
1566                         if (!is_huge_zero_page(page) &&
1567                             !pmd_numa(*pmd)) {
1568                                 entry = *pmd;
1569                                 entry = pmd_mknuma(entry);
1570                                 ret = HPAGE_PMD_NR;
1571                         }
1572                 }
1573
1574                 /* Set PMD if cleared earlier */
1575                 if (ret == HPAGE_PMD_NR)
1576                         set_pmd_at(mm, addr, pmd, entry);
1577
1578                 spin_unlock(ptl);
1579         }
1580
1581         return ret;
1582 }
1583
1584 /*
1585  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1586  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1587  *
1588  * Note that if it returns 1, this routine returns without unlocking page
1589  * table locks. So callers must unlock them.
1590  */
1591 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1592                 spinlock_t **ptl)
1593 {
1594         *ptl = pmd_lock(vma->vm_mm, pmd);
1595         if (likely(pmd_trans_huge(*pmd))) {
1596                 if (unlikely(pmd_trans_splitting(*pmd))) {
1597                         spin_unlock(*ptl);
1598                         wait_split_huge_page(vma->anon_vma, pmd);
1599                         return -1;
1600                 } else {
1601                         /* Thp mapped by 'pmd' is stable, so we can
1602                          * handle it as it is. */
1603                         return 1;
1604                 }
1605         }
1606         spin_unlock(*ptl);
1607         return 0;
1608 }
1609
1610 /*
1611  * This function returns whether a given @page is mapped onto the @address
1612  * in the virtual space of @mm.
1613  *
1614  * When it's true, this function returns *pmd with holding the page table lock
1615  * and passing it back to the caller via @ptl.
1616  * If it's false, returns NULL without holding the page table lock.
1617  */
1618 pmd_t *page_check_address_pmd(struct page *page,
1619                               struct mm_struct *mm,
1620                               unsigned long address,
1621                               enum page_check_address_pmd_flag flag,
1622                               spinlock_t **ptl)
1623 {
1624         pmd_t *pmd;
1625
1626         if (address & ~HPAGE_PMD_MASK)
1627                 return NULL;
1628
1629         pmd = mm_find_pmd(mm, address);
1630         if (!pmd)
1631                 return NULL;
1632         *ptl = pmd_lock(mm, pmd);
1633         if (pmd_none(*pmd))
1634                 goto unlock;
1635         if (pmd_page(*pmd) != page)
1636                 goto unlock;
1637         /*
1638          * split_vma() may create temporary aliased mappings. There is
1639          * no risk as long as all huge pmd are found and have their
1640          * splitting bit set before __split_huge_page_refcount
1641          * runs. Finding the same huge pmd more than once during the
1642          * same rmap walk is not a problem.
1643          */
1644         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1645             pmd_trans_splitting(*pmd))
1646                 goto unlock;
1647         if (pmd_trans_huge(*pmd)) {
1648                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1649                           !pmd_trans_splitting(*pmd));
1650                 return pmd;
1651         }
1652 unlock:
1653         spin_unlock(*ptl);
1654         return NULL;
1655 }
1656
1657 static int __split_huge_page_splitting(struct page *page,
1658                                        struct vm_area_struct *vma,
1659                                        unsigned long address)
1660 {
1661         struct mm_struct *mm = vma->vm_mm;
1662         spinlock_t *ptl;
1663         pmd_t *pmd;
1664         int ret = 0;
1665         /* For mmu_notifiers */
1666         const unsigned long mmun_start = address;
1667         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1668
1669         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1670         pmd = page_check_address_pmd(page, mm, address,
1671                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1672         if (pmd) {
1673                 /*
1674                  * We can't temporarily set the pmd to null in order
1675                  * to split it, the pmd must remain marked huge at all
1676                  * times or the VM won't take the pmd_trans_huge paths
1677                  * and it won't wait on the anon_vma->root->rwsem to
1678                  * serialize against split_huge_page*.
1679                  */
1680                 pmdp_splitting_flush(vma, address, pmd);
1681                 ret = 1;
1682                 spin_unlock(ptl);
1683         }
1684         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1685
1686         return ret;
1687 }
1688
1689 static void __split_huge_page_refcount(struct page *page,
1690                                        struct list_head *list)
1691 {
1692         int i;
1693         struct zone *zone = page_zone(page);
1694         struct lruvec *lruvec;
1695         int tail_count = 0;
1696
1697         /* prevent PageLRU to go away from under us, and freeze lru stats */
1698         spin_lock_irq(&zone->lru_lock);
1699         lruvec = mem_cgroup_page_lruvec(page, zone);
1700
1701         compound_lock(page);
1702         /* complete memcg works before add pages to LRU */
1703         mem_cgroup_split_huge_fixup(page);
1704
1705         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1706                 struct page *page_tail = page + i;
1707
1708                 /* tail_page->_mapcount cannot change */
1709                 BUG_ON(page_mapcount(page_tail) < 0);
1710                 tail_count += page_mapcount(page_tail);
1711                 /* check for overflow */
1712                 BUG_ON(tail_count < 0);
1713                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1714                 /*
1715                  * tail_page->_count is zero and not changing from
1716                  * under us. But get_page_unless_zero() may be running
1717                  * from under us on the tail_page. If we used
1718                  * atomic_set() below instead of atomic_add(), we
1719                  * would then run atomic_set() concurrently with
1720                  * get_page_unless_zero(), and atomic_set() is
1721                  * implemented in C not using locked ops. spin_unlock
1722                  * on x86 sometime uses locked ops because of PPro
1723                  * errata 66, 92, so unless somebody can guarantee
1724                  * atomic_set() here would be safe on all archs (and
1725                  * not only on x86), it's safer to use atomic_add().
1726                  */
1727                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1728                            &page_tail->_count);
1729
1730                 /* after clearing PageTail the gup refcount can be released */
1731                 smp_mb();
1732
1733                 /*
1734                  * retain hwpoison flag of the poisoned tail page:
1735                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1736                  *   by the memory-failure.
1737                  */
1738                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1739                 page_tail->flags |= (page->flags &
1740                                      ((1L << PG_referenced) |
1741                                       (1L << PG_swapbacked) |
1742                                       (1L << PG_mlocked) |
1743                                       (1L << PG_uptodate) |
1744                                       (1L << PG_active) |
1745                                       (1L << PG_unevictable)));
1746                 page_tail->flags |= (1L << PG_dirty);
1747
1748                 /* clear PageTail before overwriting first_page */
1749                 smp_wmb();
1750
1751                 /*
1752                  * __split_huge_page_splitting() already set the
1753                  * splitting bit in all pmd that could map this
1754                  * hugepage, that will ensure no CPU can alter the
1755                  * mapcount on the head page. The mapcount is only
1756                  * accounted in the head page and it has to be
1757                  * transferred to all tail pages in the below code. So
1758                  * for this code to be safe, the split the mapcount
1759                  * can't change. But that doesn't mean userland can't
1760                  * keep changing and reading the page contents while
1761                  * we transfer the mapcount, so the pmd splitting
1762                  * status is achieved setting a reserved bit in the
1763                  * pmd, not by clearing the present bit.
1764                 */
1765                 page_tail->_mapcount = page->_mapcount;
1766
1767                 BUG_ON(page_tail->mapping);
1768                 page_tail->mapping = page->mapping;
1769
1770                 page_tail->index = page->index + i;
1771                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1772
1773                 BUG_ON(!PageAnon(page_tail));
1774                 BUG_ON(!PageUptodate(page_tail));
1775                 BUG_ON(!PageDirty(page_tail));
1776                 BUG_ON(!PageSwapBacked(page_tail));
1777
1778                 lru_add_page_tail(page, page_tail, lruvec, list);
1779         }
1780         atomic_sub(tail_count, &page->_count);
1781         BUG_ON(atomic_read(&page->_count) <= 0);
1782
1783         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1784
1785         ClearPageCompound(page);
1786         compound_unlock(page);
1787         spin_unlock_irq(&zone->lru_lock);
1788
1789         for (i = 1; i < HPAGE_PMD_NR; i++) {
1790                 struct page *page_tail = page + i;
1791                 BUG_ON(page_count(page_tail) <= 0);
1792                 /*
1793                  * Tail pages may be freed if there wasn't any mapping
1794                  * like if add_to_swap() is running on a lru page that
1795                  * had its mapping zapped. And freeing these pages
1796                  * requires taking the lru_lock so we do the put_page
1797                  * of the tail pages after the split is complete.
1798                  */
1799                 put_page(page_tail);
1800         }
1801
1802         /*
1803          * Only the head page (now become a regular page) is required
1804          * to be pinned by the caller.
1805          */
1806         BUG_ON(page_count(page) <= 0);
1807 }
1808
1809 static int __split_huge_page_map(struct page *page,
1810                                  struct vm_area_struct *vma,
1811                                  unsigned long address)
1812 {
1813         struct mm_struct *mm = vma->vm_mm;
1814         spinlock_t *ptl;
1815         pmd_t *pmd, _pmd;
1816         int ret = 0, i;
1817         pgtable_t pgtable;
1818         unsigned long haddr;
1819
1820         pmd = page_check_address_pmd(page, mm, address,
1821                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1822         if (pmd) {
1823                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1824                 pmd_populate(mm, &_pmd, pgtable);
1825
1826                 haddr = address;
1827                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1828                         pte_t *pte, entry;
1829                         BUG_ON(PageCompound(page+i));
1830                         entry = mk_pte(page + i, vma->vm_page_prot);
1831                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1832                         if (!pmd_write(*pmd))
1833                                 entry = pte_wrprotect(entry);
1834                         else
1835                                 BUG_ON(page_mapcount(page) != 1);
1836                         if (!pmd_young(*pmd))
1837                                 entry = pte_mkold(entry);
1838                         if (pmd_numa(*pmd))
1839                                 entry = pte_mknuma(entry);
1840                         pte = pte_offset_map(&_pmd, haddr);
1841                         BUG_ON(!pte_none(*pte));
1842                         set_pte_at(mm, haddr, pte, entry);
1843                         pte_unmap(pte);
1844                 }
1845
1846                 smp_wmb(); /* make pte visible before pmd */
1847                 /*
1848                  * Up to this point the pmd is present and huge and
1849                  * userland has the whole access to the hugepage
1850                  * during the split (which happens in place). If we
1851                  * overwrite the pmd with the not-huge version
1852                  * pointing to the pte here (which of course we could
1853                  * if all CPUs were bug free), userland could trigger
1854                  * a small page size TLB miss on the small sized TLB
1855                  * while the hugepage TLB entry is still established
1856                  * in the huge TLB. Some CPU doesn't like that. See
1857                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1858                  * Erratum 383 on page 93. Intel should be safe but is
1859                  * also warns that it's only safe if the permission
1860                  * and cache attributes of the two entries loaded in
1861                  * the two TLB is identical (which should be the case
1862                  * here). But it is generally safer to never allow
1863                  * small and huge TLB entries for the same virtual
1864                  * address to be loaded simultaneously. So instead of
1865                  * doing "pmd_populate(); flush_tlb_range();" we first
1866                  * mark the current pmd notpresent (atomically because
1867                  * here the pmd_trans_huge and pmd_trans_splitting
1868                  * must remain set at all times on the pmd until the
1869                  * split is complete for this pmd), then we flush the
1870                  * SMP TLB and finally we write the non-huge version
1871                  * of the pmd entry with pmd_populate.
1872                  */
1873                 pmdp_invalidate(vma, address, pmd);
1874                 pmd_populate(mm, pmd, pgtable);
1875                 ret = 1;
1876                 spin_unlock(ptl);
1877         }
1878
1879         return ret;
1880 }
1881
1882 /* must be called with anon_vma->root->rwsem held */
1883 static void __split_huge_page(struct page *page,
1884                               struct anon_vma *anon_vma,
1885                               struct list_head *list)
1886 {
1887         int mapcount, mapcount2;
1888         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1889         struct anon_vma_chain *avc;
1890
1891         BUG_ON(!PageHead(page));
1892         BUG_ON(PageTail(page));
1893
1894         mapcount = 0;
1895         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1896                 struct vm_area_struct *vma = avc->vma;
1897                 unsigned long addr = vma_address(page, vma);
1898                 BUG_ON(is_vma_temporary_stack(vma));
1899                 mapcount += __split_huge_page_splitting(page, vma, addr);
1900         }
1901         /*
1902          * It is critical that new vmas are added to the tail of the
1903          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1904          * and establishes a child pmd before
1905          * __split_huge_page_splitting() freezes the parent pmd (so if
1906          * we fail to prevent copy_huge_pmd() from running until the
1907          * whole __split_huge_page() is complete), we will still see
1908          * the newly established pmd of the child later during the
1909          * walk, to be able to set it as pmd_trans_splitting too.
1910          */
1911         if (mapcount != page_mapcount(page))
1912                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1913                        mapcount, page_mapcount(page));
1914         BUG_ON(mapcount != page_mapcount(page));
1915
1916         __split_huge_page_refcount(page, list);
1917
1918         mapcount2 = 0;
1919         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1920                 struct vm_area_struct *vma = avc->vma;
1921                 unsigned long addr = vma_address(page, vma);
1922                 BUG_ON(is_vma_temporary_stack(vma));
1923                 mapcount2 += __split_huge_page_map(page, vma, addr);
1924         }
1925         if (mapcount != mapcount2)
1926                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1927                        mapcount, mapcount2, page_mapcount(page));
1928         BUG_ON(mapcount != mapcount2);
1929 }
1930
1931 /*
1932  * Split a hugepage into normal pages. This doesn't change the position of head
1933  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1934  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1935  * from the hugepage.
1936  * Return 0 if the hugepage is split successfully otherwise return 1.
1937  */
1938 int split_huge_page_to_list(struct page *page, struct list_head *list)
1939 {
1940         struct anon_vma *anon_vma;
1941         int ret = 1;
1942
1943         BUG_ON(is_huge_zero_page(page));
1944         BUG_ON(!PageAnon(page));
1945
1946         /*
1947          * The caller does not necessarily hold an mmap_sem that would prevent
1948          * the anon_vma disappearing so we first we take a reference to it
1949          * and then lock the anon_vma for write. This is similar to
1950          * page_lock_anon_vma_read except the write lock is taken to serialise
1951          * against parallel split or collapse operations.
1952          */
1953         anon_vma = page_get_anon_vma(page);
1954         if (!anon_vma)
1955                 goto out;
1956         anon_vma_lock_write(anon_vma);
1957
1958         ret = 0;
1959         if (!PageCompound(page))
1960                 goto out_unlock;
1961
1962         BUG_ON(!PageSwapBacked(page));
1963         __split_huge_page(page, anon_vma, list);
1964         count_vm_event(THP_SPLIT);
1965
1966         BUG_ON(PageCompound(page));
1967 out_unlock:
1968         anon_vma_unlock_write(anon_vma);
1969         put_anon_vma(anon_vma);
1970 out:
1971         return ret;
1972 }
1973
1974 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1975
1976 int hugepage_madvise(struct vm_area_struct *vma,
1977                      unsigned long *vm_flags, int advice)
1978 {
1979         struct mm_struct *mm = vma->vm_mm;
1980
1981         switch (advice) {
1982         case MADV_HUGEPAGE:
1983                 /*
1984                  * Be somewhat over-protective like KSM for now!
1985                  */
1986                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1987                         return -EINVAL;
1988                 if (mm->def_flags & VM_NOHUGEPAGE)
1989                         return -EINVAL;
1990                 *vm_flags &= ~VM_NOHUGEPAGE;
1991                 *vm_flags |= VM_HUGEPAGE;
1992                 /*
1993                  * If the vma become good for khugepaged to scan,
1994                  * register it here without waiting a page fault that
1995                  * may not happen any time soon.
1996                  */
1997                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1998                         return -ENOMEM;
1999                 break;
2000         case MADV_NOHUGEPAGE:
2001                 /*
2002                  * Be somewhat over-protective like KSM for now!
2003                  */
2004                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
2005                         return -EINVAL;
2006                 *vm_flags &= ~VM_HUGEPAGE;
2007                 *vm_flags |= VM_NOHUGEPAGE;
2008                 /*
2009                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2010                  * this vma even if we leave the mm registered in khugepaged if
2011                  * it got registered before VM_NOHUGEPAGE was set.
2012                  */
2013                 break;
2014         }
2015
2016         return 0;
2017 }
2018
2019 static int __init khugepaged_slab_init(void)
2020 {
2021         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2022                                           sizeof(struct mm_slot),
2023                                           __alignof__(struct mm_slot), 0, NULL);
2024         if (!mm_slot_cache)
2025                 return -ENOMEM;
2026
2027         return 0;
2028 }
2029
2030 static inline struct mm_slot *alloc_mm_slot(void)
2031 {
2032         if (!mm_slot_cache)     /* initialization failed */
2033                 return NULL;
2034         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2035 }
2036
2037 static inline void free_mm_slot(struct mm_slot *mm_slot)
2038 {
2039         kmem_cache_free(mm_slot_cache, mm_slot);
2040 }
2041
2042 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2043 {
2044         struct mm_slot *mm_slot;
2045
2046         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2047                 if (mm == mm_slot->mm)
2048                         return mm_slot;
2049
2050         return NULL;
2051 }
2052
2053 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2054                                     struct mm_slot *mm_slot)
2055 {
2056         mm_slot->mm = mm;
2057         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2058 }
2059
2060 static inline int khugepaged_test_exit(struct mm_struct *mm)
2061 {
2062         return atomic_read(&mm->mm_users) == 0;
2063 }
2064
2065 int __khugepaged_enter(struct mm_struct *mm)
2066 {
2067         struct mm_slot *mm_slot;
2068         int wakeup;
2069
2070         mm_slot = alloc_mm_slot();
2071         if (!mm_slot)
2072                 return -ENOMEM;
2073
2074         /* __khugepaged_exit() must not run from under us */
2075         VM_BUG_ON(khugepaged_test_exit(mm));
2076         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2077                 free_mm_slot(mm_slot);
2078                 return 0;
2079         }
2080
2081         spin_lock(&khugepaged_mm_lock);
2082         insert_to_mm_slots_hash(mm, mm_slot);
2083         /*
2084          * Insert just behind the scanning cursor, to let the area settle
2085          * down a little.
2086          */
2087         wakeup = list_empty(&khugepaged_scan.mm_head);
2088         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2089         spin_unlock(&khugepaged_mm_lock);
2090
2091         atomic_inc(&mm->mm_count);
2092         if (wakeup)
2093                 wake_up_interruptible(&khugepaged_wait);
2094
2095         return 0;
2096 }
2097
2098 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2099 {
2100         unsigned long hstart, hend;
2101         if (!vma->anon_vma)
2102                 /*
2103                  * Not yet faulted in so we will register later in the
2104                  * page fault if needed.
2105                  */
2106                 return 0;
2107         if (vma->vm_ops)
2108                 /* khugepaged not yet working on file or special mappings */
2109                 return 0;
2110         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2111         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2112         hend = vma->vm_end & HPAGE_PMD_MASK;
2113         if (hstart < hend)
2114                 return khugepaged_enter(vma);
2115         return 0;
2116 }
2117
2118 void __khugepaged_exit(struct mm_struct *mm)
2119 {
2120         struct mm_slot *mm_slot;
2121         int free = 0;
2122
2123         spin_lock(&khugepaged_mm_lock);
2124         mm_slot = get_mm_slot(mm);
2125         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2126                 hash_del(&mm_slot->hash);
2127                 list_del(&mm_slot->mm_node);
2128                 free = 1;
2129         }
2130         spin_unlock(&khugepaged_mm_lock);
2131
2132         if (free) {
2133                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2134                 free_mm_slot(mm_slot);
2135                 mmdrop(mm);
2136         } else if (mm_slot) {
2137                 /*
2138                  * This is required to serialize against
2139                  * khugepaged_test_exit() (which is guaranteed to run
2140                  * under mmap sem read mode). Stop here (after we
2141                  * return all pagetables will be destroyed) until
2142                  * khugepaged has finished working on the pagetables
2143                  * under the mmap_sem.
2144                  */
2145                 down_write(&mm->mmap_sem);
2146                 up_write(&mm->mmap_sem);
2147         }
2148 }
2149
2150 static void release_pte_page(struct page *page)
2151 {
2152         /* 0 stands for page_is_file_cache(page) == false */
2153         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2154         unlock_page(page);
2155         putback_lru_page(page);
2156 }
2157
2158 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2159 {
2160         while (--_pte >= pte) {
2161                 pte_t pteval = *_pte;
2162                 if (!pte_none(pteval))
2163                         release_pte_page(pte_page(pteval));
2164         }
2165 }
2166
2167 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2168                                         unsigned long address,
2169                                         pte_t *pte)
2170 {
2171         struct page *page;
2172         pte_t *_pte;
2173         int referenced = 0, none = 0;
2174         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2175              _pte++, address += PAGE_SIZE) {
2176                 pte_t pteval = *_pte;
2177                 if (pte_none(pteval)) {
2178                         if (++none <= khugepaged_max_ptes_none)
2179                                 continue;
2180                         else
2181                                 goto out;
2182                 }
2183                 if (!pte_present(pteval) || !pte_write(pteval))
2184                         goto out;
2185                 page = vm_normal_page(vma, address, pteval);
2186                 if (unlikely(!page))
2187                         goto out;
2188
2189                 VM_BUG_ON(PageCompound(page));
2190                 BUG_ON(!PageAnon(page));
2191                 VM_BUG_ON(!PageSwapBacked(page));
2192
2193                 /* cannot use mapcount: can't collapse if there's a gup pin */
2194                 if (page_count(page) != 1)
2195                         goto out;
2196                 /*
2197                  * We can do it before isolate_lru_page because the
2198                  * page can't be freed from under us. NOTE: PG_lock
2199                  * is needed to serialize against split_huge_page
2200                  * when invoked from the VM.
2201                  */
2202                 if (!trylock_page(page))
2203                         goto out;
2204                 /*
2205                  * Isolate the page to avoid collapsing an hugepage
2206                  * currently in use by the VM.
2207                  */
2208                 if (isolate_lru_page(page)) {
2209                         unlock_page(page);
2210                         goto out;
2211                 }
2212                 /* 0 stands for page_is_file_cache(page) == false */
2213                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2214                 VM_BUG_ON(!PageLocked(page));
2215                 VM_BUG_ON(PageLRU(page));
2216
2217                 /* If there is no mapped pte young don't collapse the page */
2218                 if (pte_young(pteval) || PageReferenced(page) ||
2219                     mmu_notifier_test_young(vma->vm_mm, address))
2220                         referenced = 1;
2221         }
2222         if (likely(referenced))
2223                 return 1;
2224 out:
2225         release_pte_pages(pte, _pte);
2226         return 0;
2227 }
2228
2229 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2230                                       struct vm_area_struct *vma,
2231                                       unsigned long address,
2232                                       spinlock_t *ptl)
2233 {
2234         pte_t *_pte;
2235         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2236                 pte_t pteval = *_pte;
2237                 struct page *src_page;
2238
2239                 if (pte_none(pteval)) {
2240                         clear_user_highpage(page, address);
2241                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2242                 } else {
2243                         src_page = pte_page(pteval);
2244                         copy_user_highpage(page, src_page, address, vma);
2245                         VM_BUG_ON(page_mapcount(src_page) != 1);
2246                         release_pte_page(src_page);
2247                         /*
2248                          * ptl mostly unnecessary, but preempt has to
2249                          * be disabled to update the per-cpu stats
2250                          * inside page_remove_rmap().
2251                          */
2252                         spin_lock(ptl);
2253                         /*
2254                          * paravirt calls inside pte_clear here are
2255                          * superfluous.
2256                          */
2257                         pte_clear(vma->vm_mm, address, _pte);
2258                         page_remove_rmap(src_page);
2259                         spin_unlock(ptl);
2260                         free_page_and_swap_cache(src_page);
2261                 }
2262
2263                 address += PAGE_SIZE;
2264                 page++;
2265         }
2266 }
2267
2268 static void khugepaged_alloc_sleep(void)
2269 {
2270         wait_event_freezable_timeout(khugepaged_wait, false,
2271                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2272 }
2273
2274 static int khugepaged_node_load[MAX_NUMNODES];
2275
2276 #ifdef CONFIG_NUMA
2277 static int khugepaged_find_target_node(void)
2278 {
2279         static int last_khugepaged_target_node = NUMA_NO_NODE;
2280         int nid, target_node = 0, max_value = 0;
2281
2282         /* find first node with max normal pages hit */
2283         for (nid = 0; nid < MAX_NUMNODES; nid++)
2284                 if (khugepaged_node_load[nid] > max_value) {
2285                         max_value = khugepaged_node_load[nid];
2286                         target_node = nid;
2287                 }
2288
2289         /* do some balance if several nodes have the same hit record */
2290         if (target_node <= last_khugepaged_target_node)
2291                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2292                                 nid++)
2293                         if (max_value == khugepaged_node_load[nid]) {
2294                                 target_node = nid;
2295                                 break;
2296                         }
2297
2298         last_khugepaged_target_node = target_node;
2299         return target_node;
2300 }
2301
2302 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2303 {
2304         if (IS_ERR(*hpage)) {
2305                 if (!*wait)
2306                         return false;
2307
2308                 *wait = false;
2309                 *hpage = NULL;
2310                 khugepaged_alloc_sleep();
2311         } else if (*hpage) {
2312                 put_page(*hpage);
2313                 *hpage = NULL;
2314         }
2315
2316         return true;
2317 }
2318
2319 static struct page
2320 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2321                        struct vm_area_struct *vma, unsigned long address,
2322                        int node)
2323 {
2324         VM_BUG_ON(*hpage);
2325         /*
2326          * Allocate the page while the vma is still valid and under
2327          * the mmap_sem read mode so there is no memory allocation
2328          * later when we take the mmap_sem in write mode. This is more
2329          * friendly behavior (OTOH it may actually hide bugs) to
2330          * filesystems in userland with daemons allocating memory in
2331          * the userland I/O paths.  Allocating memory with the
2332          * mmap_sem in read mode is good idea also to allow greater
2333          * scalability.
2334          */
2335         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2336                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2337         /*
2338          * After allocating the hugepage, release the mmap_sem read lock in
2339          * preparation for taking it in write mode.
2340          */
2341         up_read(&mm->mmap_sem);
2342         if (unlikely(!*hpage)) {
2343                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2344                 *hpage = ERR_PTR(-ENOMEM);
2345                 return NULL;
2346         }
2347
2348         count_vm_event(THP_COLLAPSE_ALLOC);
2349         return *hpage;
2350 }
2351 #else
2352 static int khugepaged_find_target_node(void)
2353 {
2354         return 0;
2355 }
2356
2357 static inline struct page *alloc_hugepage(int defrag)
2358 {
2359         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2360                            HPAGE_PMD_ORDER);
2361 }
2362
2363 static struct page *khugepaged_alloc_hugepage(bool *wait)
2364 {
2365         struct page *hpage;
2366
2367         do {
2368                 hpage = alloc_hugepage(khugepaged_defrag());
2369                 if (!hpage) {
2370                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2371                         if (!*wait)
2372                                 return NULL;
2373
2374                         *wait = false;
2375                         khugepaged_alloc_sleep();
2376                 } else
2377                         count_vm_event(THP_COLLAPSE_ALLOC);
2378         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2379
2380         return hpage;
2381 }
2382
2383 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2384 {
2385         if (!*hpage)
2386                 *hpage = khugepaged_alloc_hugepage(wait);
2387
2388         if (unlikely(!*hpage))
2389                 return false;
2390
2391         return true;
2392 }
2393
2394 static struct page
2395 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2396                        struct vm_area_struct *vma, unsigned long address,
2397                        int node)
2398 {
2399         up_read(&mm->mmap_sem);
2400         VM_BUG_ON(!*hpage);
2401         return  *hpage;
2402 }
2403 #endif
2404
2405 static bool hugepage_vma_check(struct vm_area_struct *vma)
2406 {
2407         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2408             (vma->vm_flags & VM_NOHUGEPAGE))
2409                 return false;
2410
2411         if (!vma->anon_vma || vma->vm_ops)
2412                 return false;
2413         if (is_vma_temporary_stack(vma))
2414                 return false;
2415         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2416         return true;
2417 }
2418
2419 static void collapse_huge_page(struct mm_struct *mm,
2420                                    unsigned long address,
2421                                    struct page **hpage,
2422                                    struct vm_area_struct *vma,
2423                                    int node)
2424 {
2425         pmd_t *pmd, _pmd;
2426         pte_t *pte;
2427         pgtable_t pgtable;
2428         struct page *new_page;
2429         spinlock_t *pmd_ptl, *pte_ptl;
2430         int isolated;
2431         unsigned long hstart, hend;
2432         unsigned long mmun_start;       /* For mmu_notifiers */
2433         unsigned long mmun_end;         /* For mmu_notifiers */
2434
2435         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2436
2437         /* release the mmap_sem read lock. */
2438         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2439         if (!new_page)
2440                 return;
2441
2442         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2443                 return;
2444
2445         /*
2446          * Prevent all access to pagetables with the exception of
2447          * gup_fast later hanlded by the ptep_clear_flush and the VM
2448          * handled by the anon_vma lock + PG_lock.
2449          */
2450         down_write(&mm->mmap_sem);
2451         if (unlikely(khugepaged_test_exit(mm)))
2452                 goto out;
2453
2454         vma = find_vma(mm, address);
2455         if (!vma)
2456                 goto out;
2457         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2458         hend = vma->vm_end & HPAGE_PMD_MASK;
2459         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2460                 goto out;
2461         if (!hugepage_vma_check(vma))
2462                 goto out;
2463         pmd = mm_find_pmd(mm, address);
2464         if (!pmd)
2465                 goto out;
2466         if (pmd_trans_huge(*pmd))
2467                 goto out;
2468
2469         anon_vma_lock_write(vma->anon_vma);
2470
2471         pte = pte_offset_map(pmd, address);
2472         pte_ptl = pte_lockptr(mm, pmd);
2473
2474         mmun_start = address;
2475         mmun_end   = address + HPAGE_PMD_SIZE;
2476         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2477         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2478         /*
2479          * After this gup_fast can't run anymore. This also removes
2480          * any huge TLB entry from the CPU so we won't allow
2481          * huge and small TLB entries for the same virtual address
2482          * to avoid the risk of CPU bugs in that area.
2483          */
2484         _pmd = pmdp_clear_flush(vma, address, pmd);
2485         spin_unlock(pmd_ptl);
2486         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2487
2488         spin_lock(pte_ptl);
2489         isolated = __collapse_huge_page_isolate(vma, address, pte);
2490         spin_unlock(pte_ptl);
2491
2492         if (unlikely(!isolated)) {
2493                 pte_unmap(pte);
2494                 spin_lock(pmd_ptl);
2495                 BUG_ON(!pmd_none(*pmd));
2496                 /*
2497                  * We can only use set_pmd_at when establishing
2498                  * hugepmds and never for establishing regular pmds that
2499                  * points to regular pagetables. Use pmd_populate for that
2500                  */
2501                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2502                 spin_unlock(pmd_ptl);
2503                 anon_vma_unlock_write(vma->anon_vma);
2504                 goto out;
2505         }
2506
2507         /*
2508          * All pages are isolated and locked so anon_vma rmap
2509          * can't run anymore.
2510          */
2511         anon_vma_unlock_write(vma->anon_vma);
2512
2513         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2514         pte_unmap(pte);
2515         __SetPageUptodate(new_page);
2516         pgtable = pmd_pgtable(_pmd);
2517
2518         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2519         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2520
2521         /*
2522          * spin_lock() below is not the equivalent of smp_wmb(), so
2523          * this is needed to avoid the copy_huge_page writes to become
2524          * visible after the set_pmd_at() write.
2525          */
2526         smp_wmb();
2527
2528         spin_lock(pmd_ptl);
2529         BUG_ON(!pmd_none(*pmd));
2530         page_add_new_anon_rmap(new_page, vma, address);
2531         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2532         set_pmd_at(mm, address, pmd, _pmd);
2533         update_mmu_cache_pmd(vma, address, pmd);
2534         spin_unlock(pmd_ptl);
2535
2536         *hpage = NULL;
2537
2538         khugepaged_pages_collapsed++;
2539 out_up_write:
2540         up_write(&mm->mmap_sem);
2541         return;
2542
2543 out:
2544         mem_cgroup_uncharge_page(new_page);
2545         goto out_up_write;
2546 }
2547
2548 static int khugepaged_scan_pmd(struct mm_struct *mm,
2549                                struct vm_area_struct *vma,
2550                                unsigned long address,
2551                                struct page **hpage)
2552 {
2553         pmd_t *pmd;
2554         pte_t *pte, *_pte;
2555         int ret = 0, referenced = 0, none = 0;
2556         struct page *page;
2557         unsigned long _address;
2558         spinlock_t *ptl;
2559         int node = NUMA_NO_NODE;
2560
2561         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2562
2563         pmd = mm_find_pmd(mm, address);
2564         if (!pmd)
2565                 goto out;
2566         if (pmd_trans_huge(*pmd))
2567                 goto out;
2568
2569         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2570         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2571         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2572              _pte++, _address += PAGE_SIZE) {
2573                 pte_t pteval = *_pte;
2574                 if (pte_none(pteval)) {
2575                         if (++none <= khugepaged_max_ptes_none)
2576                                 continue;
2577                         else
2578                                 goto out_unmap;
2579                 }
2580                 if (!pte_present(pteval) || !pte_write(pteval))
2581                         goto out_unmap;
2582                 page = vm_normal_page(vma, _address, pteval);
2583                 if (unlikely(!page))
2584                         goto out_unmap;
2585                 /*
2586                  * Record which node the original page is from and save this
2587                  * information to khugepaged_node_load[].
2588                  * Khupaged will allocate hugepage from the node has the max
2589                  * hit record.
2590                  */
2591                 node = page_to_nid(page);
2592                 khugepaged_node_load[node]++;
2593                 VM_BUG_ON(PageCompound(page));
2594                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2595                         goto out_unmap;
2596                 /* cannot use mapcount: can't collapse if there's a gup pin */
2597                 if (page_count(page) != 1)
2598                         goto out_unmap;
2599                 if (pte_young(pteval) || PageReferenced(page) ||
2600                     mmu_notifier_test_young(vma->vm_mm, address))
2601                         referenced = 1;
2602         }
2603         if (referenced)
2604                 ret = 1;
2605 out_unmap:
2606         pte_unmap_unlock(pte, ptl);
2607         if (ret) {
2608                 node = khugepaged_find_target_node();
2609                 /* collapse_huge_page will return with the mmap_sem released */
2610                 collapse_huge_page(mm, address, hpage, vma, node);
2611         }
2612 out:
2613         return ret;
2614 }
2615
2616 static void collect_mm_slot(struct mm_slot *mm_slot)
2617 {
2618         struct mm_struct *mm = mm_slot->mm;
2619
2620         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2621
2622         if (khugepaged_test_exit(mm)) {
2623                 /* free mm_slot */
2624                 hash_del(&mm_slot->hash);
2625                 list_del(&mm_slot->mm_node);
2626
2627                 /*
2628                  * Not strictly needed because the mm exited already.
2629                  *
2630                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2631                  */
2632
2633                 /* khugepaged_mm_lock actually not necessary for the below */
2634                 free_mm_slot(mm_slot);
2635                 mmdrop(mm);
2636         }
2637 }
2638
2639 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2640                                             struct page **hpage)
2641         __releases(&khugepaged_mm_lock)
2642         __acquires(&khugepaged_mm_lock)
2643 {
2644         struct mm_slot *mm_slot;
2645         struct mm_struct *mm;
2646         struct vm_area_struct *vma;
2647         int progress = 0;
2648
2649         VM_BUG_ON(!pages);
2650         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2651
2652         if (khugepaged_scan.mm_slot)
2653                 mm_slot = khugepaged_scan.mm_slot;
2654         else {
2655                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2656                                      struct mm_slot, mm_node);
2657                 khugepaged_scan.address = 0;
2658                 khugepaged_scan.mm_slot = mm_slot;
2659         }
2660         spin_unlock(&khugepaged_mm_lock);
2661
2662         mm = mm_slot->mm;
2663         down_read(&mm->mmap_sem);
2664         if (unlikely(khugepaged_test_exit(mm)))
2665                 vma = NULL;
2666         else
2667                 vma = find_vma(mm, khugepaged_scan.address);
2668
2669         progress++;
2670         for (; vma; vma = vma->vm_next) {
2671                 unsigned long hstart, hend;
2672
2673                 cond_resched();
2674                 if (unlikely(khugepaged_test_exit(mm))) {
2675                         progress++;
2676                         break;
2677                 }
2678                 if (!hugepage_vma_check(vma)) {
2679 skip:
2680                         progress++;
2681                         continue;
2682                 }
2683                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2684                 hend = vma->vm_end & HPAGE_PMD_MASK;
2685                 if (hstart >= hend)
2686                         goto skip;
2687                 if (khugepaged_scan.address > hend)
2688                         goto skip;
2689                 if (khugepaged_scan.address < hstart)
2690                         khugepaged_scan.address = hstart;
2691                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2692
2693                 while (khugepaged_scan.address < hend) {
2694                         int ret;
2695                         cond_resched();
2696                         if (unlikely(khugepaged_test_exit(mm)))
2697                                 goto breakouterloop;
2698
2699                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2700                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2701                                   hend);
2702                         ret = khugepaged_scan_pmd(mm, vma,
2703                                                   khugepaged_scan.address,
2704                                                   hpage);
2705                         /* move to next address */
2706                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2707                         progress += HPAGE_PMD_NR;
2708                         if (ret)
2709                                 /* we released mmap_sem so break loop */
2710                                 goto breakouterloop_mmap_sem;
2711                         if (progress >= pages)
2712                                 goto breakouterloop;
2713                 }
2714         }
2715 breakouterloop:
2716         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2717 breakouterloop_mmap_sem:
2718
2719         spin_lock(&khugepaged_mm_lock);
2720         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2721         /*
2722          * Release the current mm_slot if this mm is about to die, or
2723          * if we scanned all vmas of this mm.
2724          */
2725         if (khugepaged_test_exit(mm) || !vma) {
2726                 /*
2727                  * Make sure that if mm_users is reaching zero while
2728                  * khugepaged runs here, khugepaged_exit will find
2729                  * mm_slot not pointing to the exiting mm.
2730                  */
2731                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2732                         khugepaged_scan.mm_slot = list_entry(
2733                                 mm_slot->mm_node.next,
2734                                 struct mm_slot, mm_node);
2735                         khugepaged_scan.address = 0;
2736                 } else {
2737                         khugepaged_scan.mm_slot = NULL;
2738                         khugepaged_full_scans++;
2739                 }
2740
2741                 collect_mm_slot(mm_slot);
2742         }
2743
2744         return progress;
2745 }
2746
2747 static int khugepaged_has_work(void)
2748 {
2749         return !list_empty(&khugepaged_scan.mm_head) &&
2750                 khugepaged_enabled();
2751 }
2752
2753 static int khugepaged_wait_event(void)
2754 {
2755         return !list_empty(&khugepaged_scan.mm_head) ||
2756                 kthread_should_stop();
2757 }
2758
2759 static void khugepaged_do_scan(void)
2760 {
2761         struct page *hpage = NULL;
2762         unsigned int progress = 0, pass_through_head = 0;
2763         unsigned int pages = khugepaged_pages_to_scan;
2764         bool wait = true;
2765
2766         barrier(); /* write khugepaged_pages_to_scan to local stack */
2767
2768         while (progress < pages) {
2769                 if (!khugepaged_prealloc_page(&hpage, &wait))
2770                         break;
2771
2772                 cond_resched();
2773
2774                 if (unlikely(kthread_should_stop() || freezing(current)))
2775                         break;
2776
2777                 spin_lock(&khugepaged_mm_lock);
2778                 if (!khugepaged_scan.mm_slot)
2779                         pass_through_head++;
2780                 if (khugepaged_has_work() &&
2781                     pass_through_head < 2)
2782                         progress += khugepaged_scan_mm_slot(pages - progress,
2783                                                             &hpage);
2784                 else
2785                         progress = pages;
2786                 spin_unlock(&khugepaged_mm_lock);
2787         }
2788
2789         if (!IS_ERR_OR_NULL(hpage))
2790                 put_page(hpage);
2791 }
2792
2793 static void khugepaged_wait_work(void)
2794 {
2795         try_to_freeze();
2796
2797         if (khugepaged_has_work()) {
2798                 if (!khugepaged_scan_sleep_millisecs)
2799                         return;
2800
2801                 wait_event_freezable_timeout(khugepaged_wait,
2802                                              kthread_should_stop(),
2803                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2804                 return;
2805         }
2806
2807         if (khugepaged_enabled())
2808                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2809 }
2810
2811 static int khugepaged(void *none)
2812 {
2813         struct mm_slot *mm_slot;
2814
2815         set_freezable();
2816         set_user_nice(current, 19);
2817
2818         while (!kthread_should_stop()) {
2819                 khugepaged_do_scan();
2820                 khugepaged_wait_work();
2821         }
2822
2823         spin_lock(&khugepaged_mm_lock);
2824         mm_slot = khugepaged_scan.mm_slot;
2825         khugepaged_scan.mm_slot = NULL;
2826         if (mm_slot)
2827                 collect_mm_slot(mm_slot);
2828         spin_unlock(&khugepaged_mm_lock);
2829         return 0;
2830 }
2831
2832 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2833                 unsigned long haddr, pmd_t *pmd)
2834 {
2835         struct mm_struct *mm = vma->vm_mm;
2836         pgtable_t pgtable;
2837         pmd_t _pmd;
2838         int i;
2839
2840         pmdp_clear_flush(vma, haddr, pmd);
2841         /* leave pmd empty until pte is filled */
2842
2843         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2844         pmd_populate(mm, &_pmd, pgtable);
2845
2846         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2847                 pte_t *pte, entry;
2848                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2849                 entry = pte_mkspecial(entry);
2850                 pte = pte_offset_map(&_pmd, haddr);
2851                 VM_BUG_ON(!pte_none(*pte));
2852                 set_pte_at(mm, haddr, pte, entry);
2853                 pte_unmap(pte);
2854         }
2855         smp_wmb(); /* make pte visible before pmd */
2856         pmd_populate(mm, pmd, pgtable);
2857         put_huge_zero_page();
2858 }
2859
2860 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2861                 pmd_t *pmd)
2862 {
2863         spinlock_t *ptl;
2864         struct page *page;
2865         struct mm_struct *mm = vma->vm_mm;
2866         unsigned long haddr = address & HPAGE_PMD_MASK;
2867         unsigned long mmun_start;       /* For mmu_notifiers */
2868         unsigned long mmun_end;         /* For mmu_notifiers */
2869
2870         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2871
2872         mmun_start = haddr;
2873         mmun_end   = haddr + HPAGE_PMD_SIZE;
2874 again:
2875         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2876         ptl = pmd_lock(mm, pmd);
2877         if (unlikely(!pmd_trans_huge(*pmd))) {
2878                 spin_unlock(ptl);
2879                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2880                 return;
2881         }
2882         if (is_huge_zero_pmd(*pmd)) {
2883                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2884                 spin_unlock(ptl);
2885                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2886                 return;
2887         }
2888         page = pmd_page(*pmd);
2889         VM_BUG_ON(!page_count(page));
2890         get_page(page);
2891         spin_unlock(ptl);
2892         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2893
2894         split_huge_page(page);
2895
2896         put_page(page);
2897
2898         /*
2899          * We don't always have down_write of mmap_sem here: a racing
2900          * do_huge_pmd_wp_page() might have copied-on-write to another
2901          * huge page before our split_huge_page() got the anon_vma lock.
2902          */
2903         if (unlikely(pmd_trans_huge(*pmd)))
2904                 goto again;
2905 }
2906
2907 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2908                 pmd_t *pmd)
2909 {
2910         struct vm_area_struct *vma;
2911
2912         vma = find_vma(mm, address);
2913         BUG_ON(vma == NULL);
2914         split_huge_page_pmd(vma, address, pmd);
2915 }
2916
2917 static void split_huge_page_address(struct mm_struct *mm,
2918                                     unsigned long address)
2919 {
2920         pmd_t *pmd;
2921
2922         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2923
2924         pmd = mm_find_pmd(mm, address);
2925         if (!pmd)
2926                 return;
2927         /*
2928          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2929          * materialize from under us.
2930          */
2931         split_huge_page_pmd_mm(mm, address, pmd);
2932 }
2933
2934 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2935                              unsigned long start,
2936                              unsigned long end,
2937                              long adjust_next)
2938 {
2939         /*
2940          * If the new start address isn't hpage aligned and it could
2941          * previously contain an hugepage: check if we need to split
2942          * an huge pmd.
2943          */
2944         if (start & ~HPAGE_PMD_MASK &&
2945             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2946             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2947                 split_huge_page_address(vma->vm_mm, start);
2948
2949         /*
2950          * If the new end address isn't hpage aligned and it could
2951          * previously contain an hugepage: check if we need to split
2952          * an huge pmd.
2953          */
2954         if (end & ~HPAGE_PMD_MASK &&
2955             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2956             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2957                 split_huge_page_address(vma->vm_mm, end);
2958
2959         /*
2960          * If we're also updating the vma->vm_next->vm_start, if the new
2961          * vm_next->vm_start isn't page aligned and it could previously
2962          * contain an hugepage: check if we need to split an huge pmd.
2963          */
2964         if (adjust_next > 0) {
2965                 struct vm_area_struct *next = vma->vm_next;
2966                 unsigned long nstart = next->vm_start;
2967                 nstart += adjust_next << PAGE_SHIFT;
2968                 if (nstart & ~HPAGE_PMD_MASK &&
2969                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2970                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2971                         split_huge_page_address(next->vm_mm, nstart);
2972         }
2973 }