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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/kthread.h>
  20 #include <linux/khugepaged.h>
  21 #include <linux/freezer.h>
  22 #include <linux/mman.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/migrate.h>
  25 #include <linux/hashtable.h>
  26
  27 #include <asm/tlb.h>
  28 #include <asm/pgalloc.h>
  29 #include "internal.h"
  30
  31 /*
  32  * By default transparent hugepage support is disabled in order that avoid
  33  * to risk increase the memory footprint of applications without a guaranteed
  34  * benefit. When transparent hugepage support is enabled, is for all mappings,
  35  * and khugepaged scans all mappings.
  36  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  37  * for all hugepage allocations.
  38  */
  39 unsigned long transparent_hugepage_flags __read_mostly =
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  41         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  42 #endif
  43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  44         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  45 #endif
  46         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  47         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  48         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  49
  50 /* default scan 8*512 pte (or vmas) every 30 second */
  51 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  52 static unsigned int khugepaged_pages_collapsed;
  53 static unsigned int khugepaged_full_scans;
  54 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  55 /* during fragmentation poll the hugepage allocator once every minute */
  56 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  57 static struct task_struct *khugepaged_thread __read_mostly;
  58 static DEFINE_MUTEX(khugepaged_mutex);
  59 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  60 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  61 /*
  62  * default collapse hugepages if there is at least one pte mapped like
  63  * it would have happened if the vma was large enough during page
  64  * fault.
  65  */
  66 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  67
  68 static int khugepaged(void *none);
  69 static int khugepaged_slab_init(void);
  70
  71 #define MM_SLOTS_HASH_BITS 10
  72 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  73
  74 static struct kmem_cache *mm_slot_cache __read_mostly;
  75
  76 /**
  77  * struct mm_slot - hash lookup from mm to mm_slot
  78  * @hash: hash collision list
  79  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  80  * @mm: the mm that this information is valid for
  81  */
  82 struct mm_slot {
  83         struct hlist_node hash;
  84         struct list_head mm_node;
  85         struct mm_struct *mm;
  86 };
  87
  88 /**
  89  * struct khugepaged_scan - cursor for scanning
  90  * @mm_head: the head of the mm list to scan
  91  * @mm_slot: the current mm_slot we are scanning
  92  * @address: the next address inside that to be scanned
  93  *
  94  * There is only the one khugepaged_scan instance of this cursor structure.
  95  */
  96 struct khugepaged_scan {
  97         struct list_head mm_head;
  98         struct mm_slot *mm_slot;
  99         unsigned long address;
 100 };
 101 static struct khugepaged_scan khugepaged_scan = {
 102         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 103 };
 104
 105
 106 static int set_recommended_min_free_kbytes(void)
 107 {
 108         struct zone *zone;
 109         int nr_zones = 0;
 110         unsigned long recommended_min;
 111
 112         if (!khugepaged_enabled())
 113                 return 0;
 114
 115         for_each_populated_zone(zone)
 116                 nr_zones++;
 117
 118         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 119         recommended_min = pageblock_nr_pages * nr_zones * 2;
 120
 121         /*
 122          * Make sure that on average at least two pageblocks are almost free
 123          * of another type, one for a migratetype to fall back to and a
 124          * second to avoid subsequent fallbacks of other types There are 3
 125          * MIGRATE_TYPES we care about.
 126          */
 127         recommended_min += pageblock_nr_pages * nr_zones *
 128                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 129
 130         /* don't ever allow to reserve more than 5% of the lowmem */
 131         recommended_min = min(recommended_min,
 132                               (unsigned long) nr_free_buffer_pages() / 20);
 133         recommended_min <<= (PAGE_SHIFT-10);
 134
 135         if (recommended_min > min_free_kbytes) {
 136                 if (user_min_free_kbytes >= 0)
 137                         pr_info("raising min_free_kbytes from %d to %lu "
 138                                 "to help transparent hugepage allocations\n",
 139                                 min_free_kbytes, recommended_min);
 140
 141                 min_free_kbytes = recommended_min;
 142         }
 143         setup_per_zone_wmarks();
 144         return 0;
 145 }
 146 late_initcall(set_recommended_min_free_kbytes);
 147
 148 static int start_khugepaged(void)
 149 {
 150         int err = 0;
 151         if (khugepaged_enabled()) {
 152                 if (!khugepaged_thread)
 153                         khugepaged_thread = kthread_run(khugepaged, NULL,
 154                                                         "khugepaged");
 155                 if (unlikely(IS_ERR(khugepaged_thread))) {
 156                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 157                         err = PTR_ERR(khugepaged_thread);
 158                         khugepaged_thread = NULL;
 159                 }
 160
 161                 if (!list_empty(&khugepaged_scan.mm_head))
 162                         wake_up_interruptible(&khugepaged_wait);
 163
 164                 set_recommended_min_free_kbytes();
 165         } else if (khugepaged_thread) {
 166                 kthread_stop(khugepaged_thread);
 167                 khugepaged_thread = NULL;
 168         }
 169
 170         return err;
 171 }
 172
 173 static atomic_t huge_zero_refcount;
 174 static struct page *huge_zero_page __read_mostly;
 175
 176 static inline bool is_huge_zero_page(struct page *page)
 177 {
 178         return ACCESS_ONCE(huge_zero_page) == page;
 179 }
 180
 181 static inline bool is_huge_zero_pmd(pmd_t pmd)
 182 {
 183         return is_huge_zero_page(pmd_page(pmd));
 184 }
 185
 186 static struct page *get_huge_zero_page(void)
 187 {
 188         struct page *zero_page;
 189 retry:
 190         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 191                 return ACCESS_ONCE(huge_zero_page);
 192
 193         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 194                         HPAGE_PMD_ORDER);
 195         if (!zero_page) {
 196                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 197                 return NULL;
 198         }
 199         count_vm_event(THP_ZERO_PAGE_ALLOC);
 200         preempt_disable();
 201         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 202                 preempt_enable();
 203                 __free_page(zero_page);
 204                 goto retry;
 205         }
 206
 207         /* We take additional reference here. It will be put back by shrinker */
 208         atomic_set(&huge_zero_refcount, 2);
 209         preempt_enable();
 210         return ACCESS_ONCE(huge_zero_page);
 211 }
 212
 213 static void put_huge_zero_page(void)
 214 {
 215         /*
 216          * Counter should never go to zero here. Only shrinker can put
 217          * last reference.
 218          */
 219         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 220 }
 221
 222 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 223                                         struct shrink_control *sc)
 224 {
 225         /* we can free zero page only if last reference remains */
 226         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 227 }
 228
 229 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 230                                        struct shrink_control *sc)
 231 {
 232         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 233                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 234                 BUG_ON(zero_page == NULL);
 235                 __free_page(zero_page);
 236                 return HPAGE_PMD_NR;
 237         }
 238
 239         return 0;
 240 }
 241
 242 static struct shrinker huge_zero_page_shrinker = {
 243         .count_objects = shrink_huge_zero_page_count,
 244         .scan_objects = shrink_huge_zero_page_scan,
 245         .seeks = DEFAULT_SEEKS,
 246 };
 247
 248 #ifdef CONFIG_SYSFS
 249
 250 static ssize_t double_flag_show(struct kobject *kobj,
 251                                 struct kobj_attribute *attr, char *buf,
 252                                 enum transparent_hugepage_flag enabled,
 253                                 enum transparent_hugepage_flag req_madv)
 254 {
 255         if (test_bit(enabled, &transparent_hugepage_flags)) {
 256                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 257                 return sprintf(buf, "[always] madvise never\n");
 258         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 259                 return sprintf(buf, "always [madvise] never\n");
 260         else
 261                 return sprintf(buf, "always madvise [never]\n");
 262 }
 263 static ssize_t double_flag_store(struct kobject *kobj,
 264                                  struct kobj_attribute *attr,
 265                                  const char *buf, size_t count,
 266                                  enum transparent_hugepage_flag enabled,
 267                                  enum transparent_hugepage_flag req_madv)
 268 {
 269         if (!memcmp("always", buf,
 270                     min(sizeof("always")-1, count))) {
 271                 set_bit(enabled, &transparent_hugepage_flags);
 272                 clear_bit(req_madv, &transparent_hugepage_flags);
 273         } else if (!memcmp("madvise", buf,
 274                            min(sizeof("madvise")-1, count))) {
 275                 clear_bit(enabled, &transparent_hugepage_flags);
 276                 set_bit(req_madv, &transparent_hugepage_flags);
 277         } else if (!memcmp("never", buf,
 278                            min(sizeof("never")-1, count))) {
 279                 clear_bit(enabled, &transparent_hugepage_flags);
 280                 clear_bit(req_madv, &transparent_hugepage_flags);
 281         } else
 282                 return -EINVAL;
 283
 284         return count;
 285 }
 286
 287 static ssize_t enabled_show(struct kobject *kobj,
 288                             struct kobj_attribute *attr, char *buf)
 289 {
 290         return double_flag_show(kobj, attr, buf,
 291                                 TRANSPARENT_HUGEPAGE_FLAG,
 292                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 293 }
 294 static ssize_t enabled_store(struct kobject *kobj,
 295                              struct kobj_attribute *attr,
 296                              const char *buf, size_t count)
 297 {
 298         ssize_t ret;
 299
 300         ret = double_flag_store(kobj, attr, buf, count,
 301                                 TRANSPARENT_HUGEPAGE_FLAG,
 302                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 303
 304         if (ret > 0) {
 305                 int err;
 306
 307                 mutex_lock(&khugepaged_mutex);
 308                 err = start_khugepaged();
 309                 mutex_unlock(&khugepaged_mutex);
 310
 311                 if (err)
 312                         ret = err;
 313         }
 314
 315         return ret;
 316 }
 317 static struct kobj_attribute enabled_attr =
 318         __ATTR(enabled, 0644, enabled_show, enabled_store);
 319
 320 static ssize_t single_flag_show(struct kobject *kobj,
 321                                 struct kobj_attribute *attr, char *buf,
 322                                 enum transparent_hugepage_flag flag)
 323 {
 324         return sprintf(buf, "%d\n",
 325                        !!test_bit(flag, &transparent_hugepage_flags));
 326 }
 327
 328 static ssize_t single_flag_store(struct kobject *kobj,
 329                                  struct kobj_attribute *attr,
 330                                  const char *buf, size_t count,
 331                                  enum transparent_hugepage_flag flag)
 332 {
 333         unsigned long value;
 334         int ret;
 335
 336         ret = kstrtoul(buf, 10, &value);
 337         if (ret < 0)
 338                 return ret;
 339         if (value > 1)
 340                 return -EINVAL;
 341
 342         if (value)
 343                 set_bit(flag, &transparent_hugepage_flags);
 344         else
 345                 clear_bit(flag, &transparent_hugepage_flags);
 346
 347         return count;
 348 }
 349
 350 /*
 351  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 352  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 353  * memory just to allocate one more hugepage.
 354  */
 355 static ssize_t defrag_show(struct kobject *kobj,
 356                            struct kobj_attribute *attr, char *buf)
 357 {
 358         return double_flag_show(kobj, attr, buf,
 359                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 360                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 361 }
 362 static ssize_t defrag_store(struct kobject *kobj,
 363                             struct kobj_attribute *attr,
 364                             const char *buf, size_t count)
 365 {
 366         return double_flag_store(kobj, attr, buf, count,
 367                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 368                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 369 }
 370 static struct kobj_attribute defrag_attr =
 371         __ATTR(defrag, 0644, defrag_show, defrag_store);
 372
 373 static ssize_t use_zero_page_show(struct kobject *kobj,
 374                 struct kobj_attribute *attr, char *buf)
 375 {
 376         return single_flag_show(kobj, attr, buf,
 377                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 378 }
 379 static ssize_t use_zero_page_store(struct kobject *kobj,
 380                 struct kobj_attribute *attr, const char *buf, size_t count)
 381 {
 382         return single_flag_store(kobj, attr, buf, count,
 383                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 384 }
 385 static struct kobj_attribute use_zero_page_attr =
 386         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 387 #ifdef CONFIG_DEBUG_VM
 388 static ssize_t debug_cow_show(struct kobject *kobj,
 389                                 struct kobj_attribute *attr, char *buf)
 390 {
 391         return single_flag_show(kobj, attr, buf,
 392                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 393 }
 394 static ssize_t debug_cow_store(struct kobject *kobj,
 395                                struct kobj_attribute *attr,
 396                                const char *buf, size_t count)
 397 {
 398         return single_flag_store(kobj, attr, buf, count,
 399                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 400 }
 401 static struct kobj_attribute debug_cow_attr =
 402         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 403 #endif /* CONFIG_DEBUG_VM */
 404
 405 static struct attribute *hugepage_attr[] = {
 406         &enabled_attr.attr,
 407         &defrag_attr.attr,
 408         &use_zero_page_attr.attr,
 409 #ifdef CONFIG_DEBUG_VM
 410         &debug_cow_attr.attr,
 411 #endif
 412         NULL,
 413 };
 414
 415 static struct attribute_group hugepage_attr_group = {
 416         .attrs = hugepage_attr,
 417 };
 418
 419 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 420                                          struct kobj_attribute *attr,
 421                                          char *buf)
 422 {
 423         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 424 }
 425
 426 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 427                                           struct kobj_attribute *attr,
 428                                           const char *buf, size_t count)
 429 {
 430         unsigned long msecs;
 431         int err;
 432
 433         err = kstrtoul(buf, 10, &msecs);
 434         if (err || msecs > UINT_MAX)
 435                 return -EINVAL;
 436
 437         khugepaged_scan_sleep_millisecs = msecs;
 438         wake_up_interruptible(&khugepaged_wait);
 439
 440         return count;
 441 }
 442 static struct kobj_attribute scan_sleep_millisecs_attr =
 443         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 444                scan_sleep_millisecs_store);
 445
 446 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 447                                           struct kobj_attribute *attr,
 448                                           char *buf)
 449 {
 450         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 451 }
 452
 453 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 454                                            struct kobj_attribute *attr,
 455                                            const char *buf, size_t count)
 456 {
 457         unsigned long msecs;
 458         int err;
 459
 460         err = kstrtoul(buf, 10, &msecs);
 461         if (err || msecs > UINT_MAX)
 462                 return -EINVAL;
 463
 464         khugepaged_alloc_sleep_millisecs = msecs;
 465         wake_up_interruptible(&khugepaged_wait);
 466
 467         return count;
 468 }
 469 static struct kobj_attribute alloc_sleep_millisecs_attr =
 470         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 471                alloc_sleep_millisecs_store);
 472
 473 static ssize_t pages_to_scan_show(struct kobject *kobj,
 474                                   struct kobj_attribute *attr,
 475                                   char *buf)
 476 {
 477         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 478 }
 479 static ssize_t pages_to_scan_store(struct kobject *kobj,
 480                                    struct kobj_attribute *attr,
 481                                    const char *buf, size_t count)
 482 {
 483         int err;
 484         unsigned long pages;
 485
 486         err = kstrtoul(buf, 10, &pages);
 487         if (err || !pages || pages > UINT_MAX)
 488                 return -EINVAL;
 489
 490         khugepaged_pages_to_scan = pages;
 491
 492         return count;
 493 }
 494 static struct kobj_attribute pages_to_scan_attr =
 495         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 496                pages_to_scan_store);
 497
 498 static ssize_t pages_collapsed_show(struct kobject *kobj,
 499                                     struct kobj_attribute *attr,
 500                                     char *buf)
 501 {
 502         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 503 }
 504 static struct kobj_attribute pages_collapsed_attr =
 505         __ATTR_RO(pages_collapsed);
 506
 507 static ssize_t full_scans_show(struct kobject *kobj,
 508                                struct kobj_attribute *attr,
 509                                char *buf)
 510 {
 511         return sprintf(buf, "%u\n", khugepaged_full_scans);
 512 }
 513 static struct kobj_attribute full_scans_attr =
 514         __ATTR_RO(full_scans);
 515
 516 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 517                                       struct kobj_attribute *attr, char *buf)
 518 {
 519         return single_flag_show(kobj, attr, buf,
 520                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 521 }
 522 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 523                                        struct kobj_attribute *attr,
 524                                        const char *buf, size_t count)
 525 {
 526         return single_flag_store(kobj, attr, buf, count,
 527                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 528 }
 529 static struct kobj_attribute khugepaged_defrag_attr =
 530         __ATTR(defrag, 0644, khugepaged_defrag_show,
 531                khugepaged_defrag_store);
 532
 533 /*
 534  * max_ptes_none controls if khugepaged should collapse hugepages over
 535  * any unmapped ptes in turn potentially increasing the memory
 536  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 537  * reduce the available free memory in the system as it
 538  * runs. Increasing max_ptes_none will instead potentially reduce the
 539  * free memory in the system during the khugepaged scan.
 540  */
 541 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 542                                              struct kobj_attribute *attr,
 543                                              char *buf)
 544 {
 545         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 546 }
 547 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 548                                               struct kobj_attribute *attr,
 549                                               const char *buf, size_t count)
 550 {
 551         int err;
 552         unsigned long max_ptes_none;
 553
 554         err = kstrtoul(buf, 10, &max_ptes_none);
 555         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 556                 return -EINVAL;
 557
 558         khugepaged_max_ptes_none = max_ptes_none;
 559
 560         return count;
 561 }
 562 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 563         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 564                khugepaged_max_ptes_none_store);
 565
 566 static struct attribute *khugepaged_attr[] = {
 567         &khugepaged_defrag_attr.attr,
 568         &khugepaged_max_ptes_none_attr.attr,
 569         &pages_to_scan_attr.attr,
 570         &pages_collapsed_attr.attr,
 571         &full_scans_attr.attr,
 572         &scan_sleep_millisecs_attr.attr,
 573         &alloc_sleep_millisecs_attr.attr,
 574         NULL,
 575 };
 576
 577 static struct attribute_group khugepaged_attr_group = {
 578         .attrs = khugepaged_attr,
 579         .name = "khugepaged",
 580 };
 581
 582 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 583 {
 584         int err;
 585
 586         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 587         if (unlikely(!*hugepage_kobj)) {
 588                 pr_err("failed to create transparent hugepage kobject\n");
 589                 return -ENOMEM;
 590         }
 591
 592         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 593         if (err) {
 594                 pr_err("failed to register transparent hugepage group\n");
 595                 goto delete_obj;
 596         }
 597
 598         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 599         if (err) {
 600                 pr_err("failed to register transparent hugepage group\n");
 601                 goto remove_hp_group;
 602         }
 603
 604         return 0;
 605
 606 remove_hp_group:
 607         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 608 delete_obj:
 609         kobject_put(*hugepage_kobj);
 610         return err;
 611 }
 612
 613 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 614 {
 615         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 616         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 617         kobject_put(hugepage_kobj);
 618 }
 619 #else
 620 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 621 {
 622         return 0;
 623 }
 624
 625 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 626 {
 627 }
 628 #endif /* CONFIG_SYSFS */
 629
 630 static int __init hugepage_init(void)
 631 {
 632         int err;
 633         struct kobject *hugepage_kobj;
 634
 635         if (!has_transparent_hugepage()) {
 636                 transparent_hugepage_flags = 0;
 637                 return -EINVAL;
 638         }
 639
 640         err = hugepage_init_sysfs(&hugepage_kobj);
 641         if (err)
 642                 return err;
 643
 644         err = khugepaged_slab_init();
 645         if (err)
 646                 goto out;
 647
 648         register_shrinker(&huge_zero_page_shrinker);
 649
 650         /*
 651          * By default disable transparent hugepages on smaller systems,
 652          * where the extra memory used could hurt more than TLB overhead
 653          * is likely to save.  The admin can still enable it through /sys.
 654          */
 655         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 656                 transparent_hugepage_flags = 0;
 657
 658         start_khugepaged();
 659
 660         return 0;
 661 out:
 662         hugepage_exit_sysfs(hugepage_kobj);
 663         return err;
 664 }
 665 subsys_initcall(hugepage_init);
 666
 667 static int __init setup_transparent_hugepage(char *str)
 668 {
 669         int ret = 0;
 670         if (!str)
 671                 goto out;
 672         if (!strcmp(str, "always")) {
 673                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 674                         &transparent_hugepage_flags);
 675                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 676                           &transparent_hugepage_flags);
 677                 ret = 1;
 678         } else if (!strcmp(str, "madvise")) {
 679                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 680                           &transparent_hugepage_flags);
 681                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 682                         &transparent_hugepage_flags);
 683                 ret = 1;
 684         } else if (!strcmp(str, "never")) {
 685                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 686                           &transparent_hugepage_flags);
 687                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 688                           &transparent_hugepage_flags);
 689                 ret = 1;
 690         }
 691 out:
 692         if (!ret)
 693                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 694         return ret;
 695 }
 696 __setup("transparent_hugepage=", setup_transparent_hugepage);
 697
 698 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 699 {
 700         if (likely(vma->vm_flags & VM_WRITE))
 701                 pmd = pmd_mkwrite(pmd);
 702         return pmd;
 703 }
 704
 705 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 706 {
 707         pmd_t entry;
 708         entry = mk_pmd(page, prot);
 709         entry = pmd_mkhuge(entry);
 710         return entry;
 711 }
 712
 713 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 714                                         struct vm_area_struct *vma,
 715                                         unsigned long haddr, pmd_t *pmd,
 716                                         struct page *page)
 717 {
 718         struct mem_cgroup *memcg;
 719         pgtable_t pgtable;
 720         spinlock_t *ptl;
 721
 722         VM_BUG_ON_PAGE(!PageCompound(page), page);
 723
 724         if (mem_cgroup_try_charge(page, mm, GFP_TRANSHUGE, &memcg))
 725                 return VM_FAULT_OOM;
 726
 727         pgtable = pte_alloc_one(mm, haddr);
 728         if (unlikely(!pgtable)) {
 729                 mem_cgroup_cancel_charge(page, memcg);
 730                 return VM_FAULT_OOM;
 731         }
 732
 733         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 734         /*
 735          * The memory barrier inside __SetPageUptodate makes sure that
 736          * clear_huge_page writes become visible before the set_pmd_at()
 737          * write.
 738          */
 739         __SetPageUptodate(page);
 740
 741         ptl = pmd_lock(mm, pmd);
 742         if (unlikely(!pmd_none(*pmd))) {
 743                 spin_unlock(ptl);
 744                 mem_cgroup_cancel_charge(page, memcg);
 745                 put_page(page);
 746                 pte_free(mm, pgtable);
 747         } else {
 748                 pmd_t entry;
 749                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 750                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 751                 page_add_new_anon_rmap(page, vma, haddr);
 752                 mem_cgroup_commit_charge(page, memcg, false);
 753                 lru_cache_add_active_or_unevictable(page, vma);
 754                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 755                 set_pmd_at(mm, haddr, pmd, entry);
 756                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 757                 atomic_long_inc(&mm->nr_ptes);
 758                 spin_unlock(ptl);
 759         }
 760
 761         return 0;
 762 }
 763
 764 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 765 {
 766         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 767 }
 768
 769 static inline struct page *alloc_hugepage_vma(int defrag,
 770                                               struct vm_area_struct *vma,
 771                                               unsigned long haddr, int nd,
 772                                               gfp_t extra_gfp)
 773 {
 774         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 775                                HPAGE_PMD_ORDER, vma, haddr, nd);
 776 }
 777
 778 /* Caller must hold page table lock. */
 779 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 780                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 781                 struct page *zero_page)
 782 {
 783         pmd_t entry;
 784         if (!pmd_none(*pmd))
 785                 return false;
 786         entry = mk_pmd(zero_page, vma->vm_page_prot);
 787         entry = pmd_wrprotect(entry);
 788         entry = pmd_mkhuge(entry);
 789         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 790         set_pmd_at(mm, haddr, pmd, entry);
 791         atomic_long_inc(&mm->nr_ptes);
 792         return true;
 793 }
 794
 795 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 796                                unsigned long address, pmd_t *pmd,
 797                                unsigned int flags)
 798 {
 799         struct page *page;
 800         unsigned long haddr = address & HPAGE_PMD_MASK;
 801
 802         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 803                 return VM_FAULT_FALLBACK;
 804         if (unlikely(anon_vma_prepare(vma)))
 805                 return VM_FAULT_OOM;
 806         if (unlikely(khugepaged_enter(vma)))
 807                 return VM_FAULT_OOM;
 808         if (!(flags & FAULT_FLAG_WRITE) &&
 809                         transparent_hugepage_use_zero_page()) {
 810                 spinlock_t *ptl;
 811                 pgtable_t pgtable;
 812                 struct page *zero_page;
 813                 bool set;
 814                 pgtable = pte_alloc_one(mm, haddr);
 815                 if (unlikely(!pgtable))
 816                         return VM_FAULT_OOM;
 817                 zero_page = get_huge_zero_page();
 818                 if (unlikely(!zero_page)) {
 819                         pte_free(mm, pgtable);
 820                         count_vm_event(THP_FAULT_FALLBACK);
 821                         return VM_FAULT_FALLBACK;
 822                 }
 823                 ptl = pmd_lock(mm, pmd);
 824                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 825                                 zero_page);
 826                 spin_unlock(ptl);
 827                 if (!set) {
 828                         pte_free(mm, pgtable);
 829                         put_huge_zero_page();
 830                 }
 831                 return 0;
 832         }
 833         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 834                         vma, haddr, numa_node_id(), 0);
 835         if (unlikely(!page)) {
 836                 count_vm_event(THP_FAULT_FALLBACK);
 837                 return VM_FAULT_FALLBACK;
 838         }
 839         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 840                 put_page(page);
 841                 count_vm_event(THP_FAULT_FALLBACK);
 842                 return VM_FAULT_FALLBACK;
 843         }
 844
 845         count_vm_event(THP_FAULT_ALLOC);
 846         return 0;
 847 }
 848
 849 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 850                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 851                   struct vm_area_struct *vma)
 852 {
 853         spinlock_t *dst_ptl, *src_ptl;
 854         struct page *src_page;
 855         pmd_t pmd;
 856         pgtable_t pgtable;
 857         int ret;
 858
 859         ret = -ENOMEM;
 860         pgtable = pte_alloc_one(dst_mm, addr);
 861         if (unlikely(!pgtable))
 862                 goto out;
 863
 864         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 865         src_ptl = pmd_lockptr(src_mm, src_pmd);
 866         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 867
 868         ret = -EAGAIN;
 869         pmd = *src_pmd;
 870         if (unlikely(!pmd_trans_huge(pmd))) {
 871                 pte_free(dst_mm, pgtable);
 872                 goto out_unlock;
 873         }
 874         /*
 875          * When page table lock is held, the huge zero pmd should not be
 876          * under splitting since we don't split the page itself, only pmd to
 877          * a page table.
 878          */
 879         if (is_huge_zero_pmd(pmd)) {
 880                 struct page *zero_page;
 881                 bool set;
 882                 /*
 883                  * get_huge_zero_page() will never allocate a new page here,
 884                  * since we already have a zero page to copy. It just takes a
 885                  * reference.
 886                  */
 887                 zero_page = get_huge_zero_page();
 888                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 889                                 zero_page);
 890                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 891                 ret = 0;
 892                 goto out_unlock;
 893         }
 894
 895         if (unlikely(pmd_trans_splitting(pmd))) {
 896                 /* split huge page running from under us */
 897                 spin_unlock(src_ptl);
 898                 spin_unlock(dst_ptl);
 899                 pte_free(dst_mm, pgtable);
 900
 901                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 902                 goto out;
 903         }
 904         src_page = pmd_page(pmd);
 905         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 906         get_page(src_page);
 907         page_dup_rmap(src_page);
 908         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 909
 910         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 911         pmd = pmd_mkold(pmd_wrprotect(pmd));
 912         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 913         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 914         atomic_long_inc(&dst_mm->nr_ptes);
 915
 916         ret = 0;
 917 out_unlock:
 918         spin_unlock(src_ptl);
 919         spin_unlock(dst_ptl);
 920 out:
 921         return ret;
 922 }
 923
 924 void huge_pmd_set_accessed(struct mm_struct *mm,
 925                            struct vm_area_struct *vma,
 926                            unsigned long address,
 927                            pmd_t *pmd, pmd_t orig_pmd,
 928                            int dirty)
 929 {
 930         spinlock_t *ptl;
 931         pmd_t entry;
 932         unsigned long haddr;
 933
 934         ptl = pmd_lock(mm, pmd);
 935         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 936                 goto unlock;
 937
 938         entry = pmd_mkyoung(orig_pmd);
 939         haddr = address & HPAGE_PMD_MASK;
 940         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 941                 update_mmu_cache_pmd(vma, address, pmd);
 942
 943 unlock:
 944         spin_unlock(ptl);
 945 }
 946
 947 /*
 948  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 949  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 950  * the source page gets split and a tail freed before copy completes.
 951  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 952  */
 953 static void get_user_huge_page(struct page *page)
 954 {
 955         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 956                 struct page *endpage = page + HPAGE_PMD_NR;
 957
 958                 atomic_add(HPAGE_PMD_NR, &page->_count);
 959                 while (++page < endpage)
 960                         get_huge_page_tail(page);
 961         } else {
 962                 get_page(page);
 963         }
 964 }
 965
 966 static void put_user_huge_page(struct page *page)
 967 {
 968         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 969                 struct page *endpage = page + HPAGE_PMD_NR;
 970
 971                 while (page < endpage)
 972                         put_page(page++);
 973         } else {
 974                 put_page(page);
 975         }
 976 }
 977
 978 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 979                                         struct vm_area_struct *vma,
 980                                         unsigned long address,
 981                                         pmd_t *pmd, pmd_t orig_pmd,
 982                                         struct page *page,
 983                                         unsigned long haddr)
 984 {
 985         struct mem_cgroup *memcg;
 986         spinlock_t *ptl;
 987         pgtable_t pgtable;
 988         pmd_t _pmd;
 989         int ret = 0, i;
 990         struct page **pages;
 991         unsigned long mmun_start;       /* For mmu_notifiers */
 992         unsigned long mmun_end;         /* For mmu_notifiers */
 993
 994         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 995                         GFP_KERNEL);
 996         if (unlikely(!pages)) {
 997                 ret |= VM_FAULT_OOM;
 998                 goto out;
 999         }
1000
1001         for (i = 0; i < HPAGE_PMD_NR; i++) {
1002                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1003                                                __GFP_OTHER_NODE,
1004                                                vma, address, page_to_nid(page));
1005                 if (unlikely(!pages[i] ||
1006                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1007                                                    &memcg))) {
1008                         if (pages[i])
1009                                 put_page(pages[i]);
1010                         while (--i >= 0) {
1011                                 memcg = (void *)page_private(pages[i]);
1012                                 set_page_private(pages[i], 0);
1013                                 mem_cgroup_cancel_charge(pages[i], memcg);
1014                                 put_page(pages[i]);
1015                         }
1016                         kfree(pages);
1017                         ret |= VM_FAULT_OOM;
1018                         goto out;
1019                 }
1020                 set_page_private(pages[i], (unsigned long)memcg);
1021         }
1022
1023         for (i = 0; i < HPAGE_PMD_NR; i++) {
1024                 copy_user_highpage(pages[i], page + i,
1025                                    haddr + PAGE_SIZE * i, vma);
1026                 __SetPageUptodate(pages[i]);
1027                 cond_resched();
1028         }
1029
1030         mmun_start = haddr;
1031         mmun_end   = haddr + HPAGE_PMD_SIZE;
1032         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1033
1034         ptl = pmd_lock(mm, pmd);
1035         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1036                 goto out_free_pages;
1037         VM_BUG_ON_PAGE(!PageHead(page), page);
1038
1039         pmdp_clear_flush(vma, haddr, pmd);
1040         /* leave pmd empty until pte is filled */
1041
1042         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1043         pmd_populate(mm, &_pmd, pgtable);
1044
1045         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1046                 pte_t *pte, entry;
1047                 entry = mk_pte(pages[i], vma->vm_page_prot);
1048                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1049                 memcg = (void *)page_private(pages[i]);
1050                 set_page_private(pages[i], 0);
1051                 page_add_new_anon_rmap(pages[i], vma, haddr);
1052                 mem_cgroup_commit_charge(pages[i], memcg, false);
1053                 lru_cache_add_active_or_unevictable(pages[i], vma);
1054                 pte = pte_offset_map(&_pmd, haddr);
1055                 VM_BUG_ON(!pte_none(*pte));
1056                 set_pte_at(mm, haddr, pte, entry);
1057                 pte_unmap(pte);
1058         }
1059         kfree(pages);
1060
1061         smp_wmb(); /* make pte visible before pmd */
1062         pmd_populate(mm, pmd, pgtable);
1063         page_remove_rmap(page);
1064         spin_unlock(ptl);
1065
1066         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1067
1068         ret |= VM_FAULT_WRITE;
1069         put_page(page);
1070
1071 out:
1072         return ret;
1073
1074 out_free_pages:
1075         spin_unlock(ptl);
1076         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1077         for (i = 0; i < HPAGE_PMD_NR; i++) {
1078                 memcg = (void *)page_private(pages[i]);
1079                 set_page_private(pages[i], 0);
1080                 mem_cgroup_cancel_charge(pages[i], memcg);
1081                 put_page(pages[i]);
1082         }
1083         kfree(pages);
1084         goto out;
1085 }
1086
1087 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1088                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1089 {
1090         spinlock_t *ptl;
1091         int ret = 0;
1092         struct page *page = NULL, *new_page;
1093         struct mem_cgroup *memcg;
1094         unsigned long haddr;
1095         unsigned long mmun_start;       /* For mmu_notifiers */
1096         unsigned long mmun_end;         /* For mmu_notifiers */
1097
1098         ptl = pmd_lockptr(mm, pmd);
1099         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1100         haddr = address & HPAGE_PMD_MASK;
1101         if (is_huge_zero_pmd(orig_pmd))
1102                 goto alloc;
1103         spin_lock(ptl);
1104         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1105                 goto out_unlock;
1106
1107         page = pmd_page(orig_pmd);
1108         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1109         if (page_mapcount(page) == 1) {
1110                 pmd_t entry;
1111                 entry = pmd_mkyoung(orig_pmd);
1112                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1113                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1114                         update_mmu_cache_pmd(vma, address, pmd);
1115                 ret |= VM_FAULT_WRITE;
1116                 goto out_unlock;
1117         }
1118         get_user_huge_page(page);
1119         spin_unlock(ptl);
1120 alloc:
1121         if (transparent_hugepage_enabled(vma) &&
1122             !transparent_hugepage_debug_cow())
1123                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1124                                               vma, haddr, numa_node_id(), 0);
1125         else
1126                 new_page = NULL;
1127
1128         if (unlikely(!new_page)) {
1129                 if (!page) {
1130                         split_huge_page_pmd(vma, address, pmd);
1131                         ret |= VM_FAULT_FALLBACK;
1132                 } else {
1133                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1134                                         pmd, orig_pmd, page, haddr);
1135                         if (ret & VM_FAULT_OOM) {
1136                                 split_huge_page(page);
1137                                 ret |= VM_FAULT_FALLBACK;
1138                         }
1139                         put_user_huge_page(page);
1140                 }
1141                 count_vm_event(THP_FAULT_FALLBACK);
1142                 goto out;
1143         }
1144
1145         if (unlikely(mem_cgroup_try_charge(new_page, mm,
1146                                            GFP_TRANSHUGE, &memcg))) {
1147                 put_page(new_page);
1148                 if (page) {
1149                         split_huge_page(page);
1150                         put_user_huge_page(page);
1151                 } else
1152                         split_huge_page_pmd(vma, address, pmd);
1153                 ret |= VM_FAULT_FALLBACK;
1154                 count_vm_event(THP_FAULT_FALLBACK);
1155                 goto out;
1156         }
1157
1158         count_vm_event(THP_FAULT_ALLOC);
1159
1160         if (!page)
1161                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1162         else
1163                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1164         __SetPageUptodate(new_page);
1165
1166         mmun_start = haddr;
1167         mmun_end   = haddr + HPAGE_PMD_SIZE;
1168         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1169
1170         spin_lock(ptl);
1171         if (page)
1172                 put_user_huge_page(page);
1173         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1174                 spin_unlock(ptl);
1175                 mem_cgroup_cancel_charge(new_page, memcg);
1176                 put_page(new_page);
1177                 goto out_mn;
1178         } else {
1179                 pmd_t entry;
1180                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1181                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1182                 pmdp_clear_flush(vma, haddr, pmd);
1183                 page_add_new_anon_rmap(new_page, vma, haddr);
1184                 mem_cgroup_commit_charge(new_page, memcg, false);
1185                 lru_cache_add_active_or_unevictable(new_page, vma);
1186                 set_pmd_at(mm, haddr, pmd, entry);
1187                 update_mmu_cache_pmd(vma, address, pmd);
1188                 if (!page) {
1189                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1190                         put_huge_zero_page();
1191                 } else {
1192                         VM_BUG_ON_PAGE(!PageHead(page), page);
1193                         page_remove_rmap(page);
1194                         put_page(page);
1195                 }
1196                 ret |= VM_FAULT_WRITE;
1197         }
1198         spin_unlock(ptl);
1199 out_mn:
1200         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1201 out:
1202         return ret;
1203 out_unlock:
1204         spin_unlock(ptl);
1205         return ret;
1206 }
1207
1208 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1209                                    unsigned long addr,
1210                                    pmd_t *pmd,
1211                                    unsigned int flags)
1212 {
1213         struct mm_struct *mm = vma->vm_mm;
1214         struct page *page = NULL;
1215
1216         assert_spin_locked(pmd_lockptr(mm, pmd));
1217
1218         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1219                 goto out;
1220
1221         /* Avoid dumping huge zero page */
1222         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1223                 return ERR_PTR(-EFAULT);
1224
1225         /* Full NUMA hinting faults to serialise migration in fault paths */
1226         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1227                 goto out;
1228
1229         page = pmd_page(*pmd);
1230         VM_BUG_ON_PAGE(!PageHead(page), page);
1231         if (flags & FOLL_TOUCH) {
1232                 pmd_t _pmd;
1233                 /*
1234                  * We should set the dirty bit only for FOLL_WRITE but
1235                  * for now the dirty bit in the pmd is meaningless.
1236                  * And if the dirty bit will become meaningful and
1237                  * we'll only set it with FOLL_WRITE, an atomic
1238                  * set_bit will be required on the pmd to set the
1239                  * young bit, instead of the current set_pmd_at.
1240                  */
1241                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1242                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1243                                           pmd, _pmd,  1))
1244                         update_mmu_cache_pmd(vma, addr, pmd);
1245         }
1246         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1247                 if (page->mapping && trylock_page(page)) {
1248                         lru_add_drain();
1249                         if (page->mapping)
1250                                 mlock_vma_page(page);
1251                         unlock_page(page);
1252                 }
1253         }
1254         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1255         VM_BUG_ON_PAGE(!PageCompound(page), page);
1256         if (flags & FOLL_GET)
1257                 get_page_foll(page);
1258
1259 out:
1260         return page;
1261 }
1262
1263 /* NUMA hinting page fault entry point for trans huge pmds */
1264 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1265                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1266 {
1267         spinlock_t *ptl;
1268         struct anon_vma *anon_vma = NULL;
1269         struct page *page;
1270         unsigned long haddr = addr & HPAGE_PMD_MASK;
1271         int page_nid = -1, this_nid = numa_node_id();
1272         int target_nid, last_cpupid = -1;
1273         bool page_locked;
1274         bool migrated = false;
1275         int flags = 0;
1276
1277         ptl = pmd_lock(mm, pmdp);
1278         if (unlikely(!pmd_same(pmd, *pmdp)))
1279                 goto out_unlock;
1280
1281         /*
1282          * If there are potential migrations, wait for completion and retry
1283          * without disrupting NUMA hinting information. Do not relock and
1284          * check_same as the page may no longer be mapped.
1285          */
1286         if (unlikely(pmd_trans_migrating(*pmdp))) {
1287                 spin_unlock(ptl);
1288                 wait_migrate_huge_page(vma->anon_vma, pmdp);
1289                 goto out;
1290         }
1291
1292         page = pmd_page(pmd);
1293         BUG_ON(is_huge_zero_page(page));
1294         page_nid = page_to_nid(page);
1295         last_cpupid = page_cpupid_last(page);
1296         count_vm_numa_event(NUMA_HINT_FAULTS);
1297         if (page_nid == this_nid) {
1298                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1299                 flags |= TNF_FAULT_LOCAL;
1300         }
1301
1302         /*
1303          * Avoid grouping on DSO/COW pages in specific and RO pages
1304          * in general, RO pages shouldn't hurt as much anyway since
1305          * they can be in shared cache state.
1306          */
1307         if (!pmd_write(pmd))
1308                 flags |= TNF_NO_GROUP;
1309
1310         /*
1311          * Acquire the page lock to serialise THP migrations but avoid dropping
1312          * page_table_lock if at all possible
1313          */
1314         page_locked = trylock_page(page);
1315         target_nid = mpol_misplaced(page, vma, haddr);
1316         if (target_nid == -1) {
1317                 /* If the page was locked, there are no parallel migrations */
1318                 if (page_locked)
1319                         goto clear_pmdnuma;
1320         }
1321
1322         /* Migration could have started since the pmd_trans_migrating check */
1323         if (!page_locked) {
1324                 spin_unlock(ptl);
1325                 wait_on_page_locked(page);
1326                 page_nid = -1;
1327                 goto out;
1328         }
1329
1330         /*
1331          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1332          * to serialises splits
1333          */
1334         get_page(page);
1335         spin_unlock(ptl);
1336         anon_vma = page_lock_anon_vma_read(page);
1337
1338         /* Confirm the PMD did not change while page_table_lock was released */
1339         spin_lock(ptl);
1340         if (unlikely(!pmd_same(pmd, *pmdp))) {
1341                 unlock_page(page);
1342                 put_page(page);
1343                 page_nid = -1;
1344                 goto out_unlock;
1345         }
1346
1347         /* Bail if we fail to protect against THP splits for any reason */
1348         if (unlikely(!anon_vma)) {
1349                 put_page(page);
1350                 page_nid = -1;
1351                 goto clear_pmdnuma;
1352         }
1353
1354         /*
1355          * Migrate the THP to the requested node, returns with page unlocked
1356          * and pmd_numa cleared.
1357          */
1358         spin_unlock(ptl);
1359         migrated = migrate_misplaced_transhuge_page(mm, vma,
1360                                 pmdp, pmd, addr, page, target_nid);
1361         if (migrated) {
1362                 flags |= TNF_MIGRATED;
1363                 page_nid = target_nid;
1364         }
1365
1366         goto out;
1367 clear_pmdnuma:
1368         BUG_ON(!PageLocked(page));
1369         pmd = pmd_mknonnuma(pmd);
1370         set_pmd_at(mm, haddr, pmdp, pmd);
1371         VM_BUG_ON(pmd_numa(*pmdp));
1372         update_mmu_cache_pmd(vma, addr, pmdp);
1373         unlock_page(page);
1374 out_unlock:
1375         spin_unlock(ptl);
1376
1377 out:
1378         if (anon_vma)
1379                 page_unlock_anon_vma_read(anon_vma);
1380
1381         if (page_nid != -1)
1382                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1383
1384         return 0;
1385 }
1386
1387 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1388                  pmd_t *pmd, unsigned long addr)
1389 {
1390         spinlock_t *ptl;
1391         int ret = 0;
1392
1393         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1394                 struct page *page;
1395                 pgtable_t pgtable;
1396                 pmd_t orig_pmd;
1397                 /*
1398                  * For architectures like ppc64 we look at deposited pgtable
1399                  * when calling pmdp_get_and_clear. So do the
1400                  * pgtable_trans_huge_withdraw after finishing pmdp related
1401                  * operations.
1402                  */
1403                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1404                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1405                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1406                 if (is_huge_zero_pmd(orig_pmd)) {
1407                         atomic_long_dec(&tlb->mm->nr_ptes);
1408                         spin_unlock(ptl);
1409                         put_huge_zero_page();
1410                 } else {
1411                         page = pmd_page(orig_pmd);
1412                         page_remove_rmap(page);
1413                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1414                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1415                         VM_BUG_ON_PAGE(!PageHead(page), page);
1416                         atomic_long_dec(&tlb->mm->nr_ptes);
1417                         spin_unlock(ptl);
1418                         tlb_remove_page(tlb, page);
1419                 }
1420                 pte_free(tlb->mm, pgtable);
1421                 ret = 1;
1422         }
1423         return ret;
1424 }
1425
1426 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1427                 unsigned long addr, unsigned long end,
1428                 unsigned char *vec)
1429 {
1430         spinlock_t *ptl;
1431         int ret = 0;
1432
1433         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1434                 /*
1435                  * All logical pages in the range are present
1436                  * if backed by a huge page.
1437                  */
1438                 spin_unlock(ptl);
1439                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1440                 ret = 1;
1441         }
1442
1443         return ret;
1444 }
1445
1446 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1447                   unsigned long old_addr,
1448                   unsigned long new_addr, unsigned long old_end,
1449                   pmd_t *old_pmd, pmd_t *new_pmd)
1450 {
1451         spinlock_t *old_ptl, *new_ptl;
1452         int ret = 0;
1453         pmd_t pmd;
1454
1455         struct mm_struct *mm = vma->vm_mm;
1456
1457         if ((old_addr & ~HPAGE_PMD_MASK) ||
1458             (new_addr & ~HPAGE_PMD_MASK) ||
1459             old_end - old_addr < HPAGE_PMD_SIZE ||
1460             (new_vma->vm_flags & VM_NOHUGEPAGE))
1461                 goto out;
1462
1463         /*
1464          * The destination pmd shouldn't be established, free_pgtables()
1465          * should have release it.
1466          */
1467         if (WARN_ON(!pmd_none(*new_pmd))) {
1468                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1469                 goto out;
1470         }
1471
1472         /*
1473          * We don't have to worry about the ordering of src and dst
1474          * ptlocks because exclusive mmap_sem prevents deadlock.
1475          */
1476         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1477         if (ret == 1) {
1478                 new_ptl = pmd_lockptr(mm, new_pmd);
1479                 if (new_ptl != old_ptl)
1480                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1481                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1482                 VM_BUG_ON(!pmd_none(*new_pmd));
1483
1484                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1485                         pgtable_t pgtable;
1486                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1487                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1488                 }
1489                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1490                 if (new_ptl != old_ptl)
1491                         spin_unlock(new_ptl);
1492                 spin_unlock(old_ptl);
1493         }
1494 out:
1495         return ret;
1496 }
1497
1498 /*
1499  * Returns
1500  *  - 0 if PMD could not be locked
1501  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1502  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1503  */
1504 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1505                 unsigned long addr, pgprot_t newprot, int prot_numa)
1506 {
1507         struct mm_struct *mm = vma->vm_mm;
1508         spinlock_t *ptl;
1509         int ret = 0;
1510
1511         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1512                 pmd_t entry;
1513                 ret = 1;
1514                 if (!prot_numa) {
1515                         entry = pmdp_get_and_clear(mm, addr, pmd);
1516                         if (pmd_numa(entry))
1517                                 entry = pmd_mknonnuma(entry);
1518                         entry = pmd_modify(entry, newprot);
1519                         ret = HPAGE_PMD_NR;
1520                         set_pmd_at(mm, addr, pmd, entry);
1521                         BUG_ON(pmd_write(entry));
1522                 } else {
1523                         struct page *page = pmd_page(*pmd);
1524
1525                         /*
1526                          * Do not trap faults against the zero page. The
1527                          * read-only data is likely to be read-cached on the
1528                          * local CPU cache and it is less useful to know about
1529                          * local vs remote hits on the zero page.
1530                          */
1531                         if (!is_huge_zero_page(page) &&
1532                             !pmd_numa(*pmd)) {
1533                                 pmdp_set_numa(mm, addr, pmd);
1534                                 ret = HPAGE_PMD_NR;
1535                         }
1536                 }
1537                 spin_unlock(ptl);
1538         }
1539
1540         return ret;
1541 }
1542
1543 /*
1544  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1545  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1546  *
1547  * Note that if it returns 1, this routine returns without unlocking page
1548  * table locks. So callers must unlock them.
1549  */
1550 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1551                 spinlock_t **ptl)
1552 {
1553         *ptl = pmd_lock(vma->vm_mm, pmd);
1554         if (likely(pmd_trans_huge(*pmd))) {
1555                 if (unlikely(pmd_trans_splitting(*pmd))) {
1556                         spin_unlock(*ptl);
1557                         wait_split_huge_page(vma->anon_vma, pmd);
1558                         return -1;
1559                 } else {
1560                         /* Thp mapped by 'pmd' is stable, so we can
1561                          * handle it as it is. */
1562                         return 1;
1563                 }
1564         }
1565         spin_unlock(*ptl);
1566         return 0;
1567 }
1568
1569 /*
1570  * This function returns whether a given @page is mapped onto the @address
1571  * in the virtual space of @mm.
1572  *
1573  * When it's true, this function returns *pmd with holding the page table lock
1574  * and passing it back to the caller via @ptl.
1575  * If it's false, returns NULL without holding the page table lock.
1576  */
1577 pmd_t *page_check_address_pmd(struct page *page,
1578                               struct mm_struct *mm,
1579                               unsigned long address,
1580                               enum page_check_address_pmd_flag flag,
1581                               spinlock_t **ptl)
1582 {
1583         pgd_t *pgd;
1584         pud_t *pud;
1585         pmd_t *pmd;
1586
1587         if (address & ~HPAGE_PMD_MASK)
1588                 return NULL;
1589
1590         pgd = pgd_offset(mm, address);
1591         if (!pgd_present(*pgd))
1592                 return NULL;
1593         pud = pud_offset(pgd, address);
1594         if (!pud_present(*pud))
1595                 return NULL;
1596         pmd = pmd_offset(pud, address);
1597
1598         *ptl = pmd_lock(mm, pmd);
1599         if (!pmd_present(*pmd))
1600                 goto unlock;
1601         if (pmd_page(*pmd) != page)
1602                 goto unlock;
1603         /*
1604          * split_vma() may create temporary aliased mappings. There is
1605          * no risk as long as all huge pmd are found and have their
1606          * splitting bit set before __split_huge_page_refcount
1607          * runs. Finding the same huge pmd more than once during the
1608          * same rmap walk is not a problem.
1609          */
1610         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1611             pmd_trans_splitting(*pmd))
1612                 goto unlock;
1613         if (pmd_trans_huge(*pmd)) {
1614                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1615                           !pmd_trans_splitting(*pmd));
1616                 return pmd;
1617         }
1618 unlock:
1619         spin_unlock(*ptl);
1620         return NULL;
1621 }
1622
1623 static int __split_huge_page_splitting(struct page *page,
1624                                        struct vm_area_struct *vma,
1625                                        unsigned long address)
1626 {
1627         struct mm_struct *mm = vma->vm_mm;
1628         spinlock_t *ptl;
1629         pmd_t *pmd;
1630         int ret = 0;
1631         /* For mmu_notifiers */
1632         const unsigned long mmun_start = address;
1633         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1634
1635         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1636         pmd = page_check_address_pmd(page, mm, address,
1637                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1638         if (pmd) {
1639                 /*
1640                  * We can't temporarily set the pmd to null in order
1641                  * to split it, the pmd must remain marked huge at all
1642                  * times or the VM won't take the pmd_trans_huge paths
1643                  * and it won't wait on the anon_vma->root->rwsem to
1644                  * serialize against split_huge_page*.
1645                  */
1646                 pmdp_splitting_flush(vma, address, pmd);
1647                 ret = 1;
1648                 spin_unlock(ptl);
1649         }
1650         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1651
1652         return ret;
1653 }
1654
1655 static void __split_huge_page_refcount(struct page *page,
1656                                        struct list_head *list)
1657 {
1658         int i;
1659         struct zone *zone = page_zone(page);
1660         struct lruvec *lruvec;
1661         int tail_count = 0;
1662
1663         /* prevent PageLRU to go away from under us, and freeze lru stats */
1664         spin_lock_irq(&zone->lru_lock);
1665         lruvec = mem_cgroup_page_lruvec(page, zone);
1666
1667         compound_lock(page);
1668         /* complete memcg works before add pages to LRU */
1669         mem_cgroup_split_huge_fixup(page);
1670
1671         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1672                 struct page *page_tail = page + i;
1673
1674                 /* tail_page->_mapcount cannot change */
1675                 BUG_ON(page_mapcount(page_tail) < 0);
1676                 tail_count += page_mapcount(page_tail);
1677                 /* check for overflow */
1678                 BUG_ON(tail_count < 0);
1679                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1680                 /*
1681                  * tail_page->_count is zero and not changing from
1682                  * under us. But get_page_unless_zero() may be running
1683                  * from under us on the tail_page. If we used
1684                  * atomic_set() below instead of atomic_add(), we
1685                  * would then run atomic_set() concurrently with
1686                  * get_page_unless_zero(), and atomic_set() is
1687                  * implemented in C not using locked ops. spin_unlock
1688                  * on x86 sometime uses locked ops because of PPro
1689                  * errata 66, 92, so unless somebody can guarantee
1690                  * atomic_set() here would be safe on all archs (and
1691                  * not only on x86), it's safer to use atomic_add().
1692                  */
1693                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1694                            &page_tail->_count);
1695
1696                 /* after clearing PageTail the gup refcount can be released */
1697                 smp_mb__after_atomic();
1698
1699                 /*
1700                  * retain hwpoison flag of the poisoned tail page:
1701                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1702                  *   by the memory-failure.
1703                  */
1704                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1705                 page_tail->flags |= (page->flags &
1706                                      ((1L << PG_referenced) |
1707                                       (1L << PG_swapbacked) |
1708                                       (1L << PG_mlocked) |
1709                                       (1L << PG_uptodate) |
1710                                       (1L << PG_active) |
1711                                       (1L << PG_unevictable)));
1712                 page_tail->flags |= (1L << PG_dirty);
1713
1714                 /* clear PageTail before overwriting first_page */
1715                 smp_wmb();
1716
1717                 /*
1718                  * __split_huge_page_splitting() already set the
1719                  * splitting bit in all pmd that could map this
1720                  * hugepage, that will ensure no CPU can alter the
1721                  * mapcount on the head page. The mapcount is only
1722                  * accounted in the head page and it has to be
1723                  * transferred to all tail pages in the below code. So
1724                  * for this code to be safe, the split the mapcount
1725                  * can't change. But that doesn't mean userland can't
1726                  * keep changing and reading the page contents while
1727                  * we transfer the mapcount, so the pmd splitting
1728                  * status is achieved setting a reserved bit in the
1729                  * pmd, not by clearing the present bit.
1730                 */
1731                 page_tail->_mapcount = page->_mapcount;
1732
1733                 BUG_ON(page_tail->mapping);
1734                 page_tail->mapping = page->mapping;
1735
1736                 page_tail->index = page->index + i;
1737                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1738
1739                 BUG_ON(!PageAnon(page_tail));
1740                 BUG_ON(!PageUptodate(page_tail));
1741                 BUG_ON(!PageDirty(page_tail));
1742                 BUG_ON(!PageSwapBacked(page_tail));
1743
1744                 lru_add_page_tail(page, page_tail, lruvec, list);
1745         }
1746         atomic_sub(tail_count, &page->_count);
1747         BUG_ON(atomic_read(&page->_count) <= 0);
1748
1749         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1750
1751         ClearPageCompound(page);
1752         compound_unlock(page);
1753         spin_unlock_irq(&zone->lru_lock);
1754
1755         for (i = 1; i < HPAGE_PMD_NR; i++) {
1756                 struct page *page_tail = page + i;
1757                 BUG_ON(page_count(page_tail) <= 0);
1758                 /*
1759                  * Tail pages may be freed if there wasn't any mapping
1760                  * like if add_to_swap() is running on a lru page that
1761                  * had its mapping zapped. And freeing these pages
1762                  * requires taking the lru_lock so we do the put_page
1763                  * of the tail pages after the split is complete.
1764                  */
1765                 put_page(page_tail);
1766         }
1767
1768         /*
1769          * Only the head page (now become a regular page) is required
1770          * to be pinned by the caller.
1771          */
1772         BUG_ON(page_count(page) <= 0);
1773 }
1774
1775 static int __split_huge_page_map(struct page *page,
1776                                  struct vm_area_struct *vma,
1777                                  unsigned long address)
1778 {
1779         struct mm_struct *mm = vma->vm_mm;
1780         spinlock_t *ptl;
1781         pmd_t *pmd, _pmd;
1782         int ret = 0, i;
1783         pgtable_t pgtable;
1784         unsigned long haddr;
1785
1786         pmd = page_check_address_pmd(page, mm, address,
1787                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1788         if (pmd) {
1789                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1790                 pmd_populate(mm, &_pmd, pgtable);
1791                 if (pmd_write(*pmd))
1792                         BUG_ON(page_mapcount(page) != 1);
1793
1794                 haddr = address;
1795                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1796                         pte_t *pte, entry;
1797                         BUG_ON(PageCompound(page+i));
1798                         /*
1799                          * Note that pmd_numa is not transferred deliberately
1800                          * to avoid any possibility that pte_numa leaks to
1801                          * a PROT_NONE VMA by accident.
1802                          */
1803                         entry = mk_pte(page + i, vma->vm_page_prot);
1804                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1805                         if (!pmd_write(*pmd))
1806                                 entry = pte_wrprotect(entry);
1807                         if (!pmd_young(*pmd))
1808                                 entry = pte_mkold(entry);
1809                         pte = pte_offset_map(&_pmd, haddr);
1810                         BUG_ON(!pte_none(*pte));
1811                         set_pte_at(mm, haddr, pte, entry);
1812                         pte_unmap(pte);
1813                 }
1814
1815                 smp_wmb(); /* make pte visible before pmd */
1816                 /*
1817                  * Up to this point the pmd is present and huge and
1818                  * userland has the whole access to the hugepage
1819                  * during the split (which happens in place). If we
1820                  * overwrite the pmd with the not-huge version
1821                  * pointing to the pte here (which of course we could
1822                  * if all CPUs were bug free), userland could trigger
1823                  * a small page size TLB miss on the small sized TLB
1824                  * while the hugepage TLB entry is still established
1825                  * in the huge TLB. Some CPU doesn't like that. See
1826                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1827                  * Erratum 383 on page 93. Intel should be safe but is
1828                  * also warns that it's only safe if the permission
1829                  * and cache attributes of the two entries loaded in
1830                  * the two TLB is identical (which should be the case
1831                  * here). But it is generally safer to never allow
1832                  * small and huge TLB entries for the same virtual
1833                  * address to be loaded simultaneously. So instead of
1834                  * doing "pmd_populate(); flush_tlb_range();" we first
1835                  * mark the current pmd notpresent (atomically because
1836                  * here the pmd_trans_huge and pmd_trans_splitting
1837                  * must remain set at all times on the pmd until the
1838                  * split is complete for this pmd), then we flush the
1839                  * SMP TLB and finally we write the non-huge version
1840                  * of the pmd entry with pmd_populate.
1841                  */
1842                 pmdp_invalidate(vma, address, pmd);
1843                 pmd_populate(mm, pmd, pgtable);
1844                 ret = 1;
1845                 spin_unlock(ptl);
1846         }
1847
1848         return ret;
1849 }
1850
1851 /* must be called with anon_vma->root->rwsem held */
1852 static void __split_huge_page(struct page *page,
1853                               struct anon_vma *anon_vma,
1854                               struct list_head *list)
1855 {
1856         int mapcount, mapcount2;
1857         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1858         struct anon_vma_chain *avc;
1859
1860         BUG_ON(!PageHead(page));
1861         BUG_ON(PageTail(page));
1862
1863         mapcount = 0;
1864         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1865                 struct vm_area_struct *vma = avc->vma;
1866                 unsigned long addr = vma_address(page, vma);
1867                 BUG_ON(is_vma_temporary_stack(vma));
1868                 mapcount += __split_huge_page_splitting(page, vma, addr);
1869         }
1870         /*
1871          * It is critical that new vmas are added to the tail of the
1872          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1873          * and establishes a child pmd before
1874          * __split_huge_page_splitting() freezes the parent pmd (so if
1875          * we fail to prevent copy_huge_pmd() from running until the
1876          * whole __split_huge_page() is complete), we will still see
1877          * the newly established pmd of the child later during the
1878          * walk, to be able to set it as pmd_trans_splitting too.
1879          */
1880         if (mapcount != page_mapcount(page)) {
1881                 pr_err("mapcount %d page_mapcount %d\n",
1882                         mapcount, page_mapcount(page));
1883                 BUG();
1884         }
1885
1886         __split_huge_page_refcount(page, list);
1887
1888         mapcount2 = 0;
1889         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1890                 struct vm_area_struct *vma = avc->vma;
1891                 unsigned long addr = vma_address(page, vma);
1892                 BUG_ON(is_vma_temporary_stack(vma));
1893                 mapcount2 += __split_huge_page_map(page, vma, addr);
1894         }
1895         if (mapcount != mapcount2) {
1896                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1897                         mapcount, mapcount2, page_mapcount(page));
1898                 BUG();
1899         }
1900 }
1901
1902 /*
1903  * Split a hugepage into normal pages. This doesn't change the position of head
1904  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1905  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1906  * from the hugepage.
1907  * Return 0 if the hugepage is split successfully otherwise return 1.
1908  */
1909 int split_huge_page_to_list(struct page *page, struct list_head *list)
1910 {
1911         struct anon_vma *anon_vma;
1912         int ret = 1;
1913
1914         BUG_ON(is_huge_zero_page(page));
1915         BUG_ON(!PageAnon(page));
1916
1917         /*
1918          * The caller does not necessarily hold an mmap_sem that would prevent
1919          * the anon_vma disappearing so we first we take a reference to it
1920          * and then lock the anon_vma for write. This is similar to
1921          * page_lock_anon_vma_read except the write lock is taken to serialise
1922          * against parallel split or collapse operations.
1923          */
1924         anon_vma = page_get_anon_vma(page);
1925         if (!anon_vma)
1926                 goto out;
1927         anon_vma_lock_write(anon_vma);
1928
1929         ret = 0;
1930         if (!PageCompound(page))
1931                 goto out_unlock;
1932
1933         BUG_ON(!PageSwapBacked(page));
1934         __split_huge_page(page, anon_vma, list);
1935         count_vm_event(THP_SPLIT);
1936
1937         BUG_ON(PageCompound(page));
1938 out_unlock:
1939         anon_vma_unlock_write(anon_vma);
1940         put_anon_vma(anon_vma);
1941 out:
1942         return ret;
1943 }
1944
1945 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1946
1947 int hugepage_madvise(struct vm_area_struct *vma,
1948                      unsigned long *vm_flags, int advice)
1949 {
1950         switch (advice) {
1951         case MADV_HUGEPAGE:
1952 #ifdef CONFIG_S390
1953                 /*
1954                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1955                  * can't handle this properly after s390_enable_sie, so we simply
1956                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1957                  */
1958                 if (mm_has_pgste(vma->vm_mm))
1959                         return 0;
1960 #endif
1961                 /*
1962                  * Be somewhat over-protective like KSM for now!
1963                  */
1964                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1965                         return -EINVAL;
1966                 *vm_flags &= ~VM_NOHUGEPAGE;
1967                 *vm_flags |= VM_HUGEPAGE;
1968                 /*
1969                  * If the vma become good for khugepaged to scan,
1970                  * register it here without waiting a page fault that
1971                  * may not happen any time soon.
1972                  */
1973                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1974                         return -ENOMEM;
1975                 break;
1976         case MADV_NOHUGEPAGE:
1977                 /*
1978                  * Be somewhat over-protective like KSM for now!
1979                  */
1980                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1981                         return -EINVAL;
1982                 *vm_flags &= ~VM_HUGEPAGE;
1983                 *vm_flags |= VM_NOHUGEPAGE;
1984                 /*
1985                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1986                  * this vma even if we leave the mm registered in khugepaged if
1987                  * it got registered before VM_NOHUGEPAGE was set.
1988                  */
1989                 break;
1990         }
1991
1992         return 0;
1993 }
1994
1995 static int __init khugepaged_slab_init(void)
1996 {
1997         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1998                                           sizeof(struct mm_slot),
1999                                           __alignof__(struct mm_slot), 0, NULL);
2000         if (!mm_slot_cache)
2001                 return -ENOMEM;
2002
2003         return 0;
2004 }
2005
2006 static inline struct mm_slot *alloc_mm_slot(void)
2007 {
2008         if (!mm_slot_cache)     /* initialization failed */
2009                 return NULL;
2010         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2011 }
2012
2013 static inline void free_mm_slot(struct mm_slot *mm_slot)
2014 {
2015         kmem_cache_free(mm_slot_cache, mm_slot);
2016 }
2017
2018 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2019 {
2020         struct mm_slot *mm_slot;
2021
2022         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2023                 if (mm == mm_slot->mm)
2024                         return mm_slot;
2025
2026         return NULL;
2027 }
2028
2029 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2030                                     struct mm_slot *mm_slot)
2031 {
2032         mm_slot->mm = mm;
2033         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2034 }
2035
2036 static inline int khugepaged_test_exit(struct mm_struct *mm)
2037 {
2038         return atomic_read(&mm->mm_users) == 0;
2039 }
2040
2041 int __khugepaged_enter(struct mm_struct *mm)
2042 {
2043         struct mm_slot *mm_slot;
2044         int wakeup;
2045
2046         mm_slot = alloc_mm_slot();
2047         if (!mm_slot)
2048                 return -ENOMEM;
2049
2050         /* __khugepaged_exit() must not run from under us */
2051         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2052         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2053                 free_mm_slot(mm_slot);
2054                 return 0;
2055         }
2056
2057         spin_lock(&khugepaged_mm_lock);
2058         insert_to_mm_slots_hash(mm, mm_slot);
2059         /*
2060          * Insert just behind the scanning cursor, to let the area settle
2061          * down a little.
2062          */
2063         wakeup = list_empty(&khugepaged_scan.mm_head);
2064         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2065         spin_unlock(&khugepaged_mm_lock);
2066
2067         atomic_inc(&mm->mm_count);
2068         if (wakeup)
2069                 wake_up_interruptible(&khugepaged_wait);
2070
2071         return 0;
2072 }
2073
2074 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2075 {
2076         unsigned long hstart, hend;
2077         if (!vma->anon_vma)
2078                 /*
2079                  * Not yet faulted in so we will register later in the
2080                  * page fault if needed.
2081                  */
2082                 return 0;
2083         if (vma->vm_ops)
2084                 /* khugepaged not yet working on file or special mappings */
2085                 return 0;
2086         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2087         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2088         hend = vma->vm_end & HPAGE_PMD_MASK;
2089         if (hstart < hend)
2090                 return khugepaged_enter(vma);
2091         return 0;
2092 }
2093
2094 void __khugepaged_exit(struct mm_struct *mm)
2095 {
2096         struct mm_slot *mm_slot;
2097         int free = 0;
2098
2099         spin_lock(&khugepaged_mm_lock);
2100         mm_slot = get_mm_slot(mm);
2101         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2102                 hash_del(&mm_slot->hash);
2103                 list_del(&mm_slot->mm_node);
2104                 free = 1;
2105         }
2106         spin_unlock(&khugepaged_mm_lock);
2107
2108         if (free) {
2109                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2110                 free_mm_slot(mm_slot);
2111                 mmdrop(mm);
2112         } else if (mm_slot) {
2113                 /*
2114                  * This is required to serialize against
2115                  * khugepaged_test_exit() (which is guaranteed to run
2116                  * under mmap sem read mode). Stop here (after we
2117                  * return all pagetables will be destroyed) until
2118                  * khugepaged has finished working on the pagetables
2119                  * under the mmap_sem.
2120                  */
2121                 down_write(&mm->mmap_sem);
2122                 up_write(&mm->mmap_sem);
2123         }
2124 }
2125
2126 static void release_pte_page(struct page *page)
2127 {
2128         /* 0 stands for page_is_file_cache(page) == false */
2129         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2130         unlock_page(page);
2131         putback_lru_page(page);
2132 }
2133
2134 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2135 {
2136         while (--_pte >= pte) {
2137                 pte_t pteval = *_pte;
2138                 if (!pte_none(pteval))
2139                         release_pte_page(pte_page(pteval));
2140         }
2141 }
2142
2143 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2144                                         unsigned long address,
2145                                         pte_t *pte)
2146 {
2147         struct page *page;
2148         pte_t *_pte;
2149         int referenced = 0, none = 0;
2150         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2151              _pte++, address += PAGE_SIZE) {
2152                 pte_t pteval = *_pte;
2153                 if (pte_none(pteval)) {
2154                         if (++none <= khugepaged_max_ptes_none)
2155                                 continue;
2156                         else
2157                                 goto out;
2158                 }
2159                 if (!pte_present(pteval) || !pte_write(pteval))
2160                         goto out;
2161                 page = vm_normal_page(vma, address, pteval);
2162                 if (unlikely(!page))
2163                         goto out;
2164
2165                 VM_BUG_ON_PAGE(PageCompound(page), page);
2166                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2167                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2168
2169                 /* cannot use mapcount: can't collapse if there's a gup pin */
2170                 if (page_count(page) != 1)
2171                         goto out;
2172                 /*
2173                  * We can do it before isolate_lru_page because the
2174                  * page can't be freed from under us. NOTE: PG_lock
2175                  * is needed to serialize against split_huge_page
2176                  * when invoked from the VM.
2177                  */
2178                 if (!trylock_page(page))
2179                         goto out;
2180                 /*
2181                  * Isolate the page to avoid collapsing an hugepage
2182                  * currently in use by the VM.
2183                  */
2184                 if (isolate_lru_page(page)) {
2185                         unlock_page(page);
2186                         goto out;
2187                 }
2188                 /* 0 stands for page_is_file_cache(page) == false */
2189                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2190                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2191                 VM_BUG_ON_PAGE(PageLRU(page), page);
2192
2193                 /* If there is no mapped pte young don't collapse the page */
2194                 if (pte_young(pteval) || PageReferenced(page) ||
2195                     mmu_notifier_test_young(vma->vm_mm, address))
2196                         referenced = 1;
2197         }
2198         if (likely(referenced))
2199                 return 1;
2200 out:
2201         release_pte_pages(pte, _pte);
2202         return 0;
2203 }
2204
2205 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2206                                       struct vm_area_struct *vma,
2207                                       unsigned long address,
2208                                       spinlock_t *ptl)
2209 {
2210         pte_t *_pte;
2211         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2212                 pte_t pteval = *_pte;
2213                 struct page *src_page;
2214
2215                 if (pte_none(pteval)) {
2216                         clear_user_highpage(page, address);
2217                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2218                 } else {
2219                         src_page = pte_page(pteval);
2220                         copy_user_highpage(page, src_page, address, vma);
2221                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2222                         release_pte_page(src_page);
2223                         /*
2224                          * ptl mostly unnecessary, but preempt has to
2225                          * be disabled to update the per-cpu stats
2226                          * inside page_remove_rmap().
2227                          */
2228                         spin_lock(ptl);
2229                         /*
2230                          * paravirt calls inside pte_clear here are
2231                          * superfluous.
2232                          */
2233                         pte_clear(vma->vm_mm, address, _pte);
2234                         page_remove_rmap(src_page);
2235                         spin_unlock(ptl);
2236                         free_page_and_swap_cache(src_page);
2237                 }
2238
2239                 address += PAGE_SIZE;
2240                 page++;
2241         }
2242 }
2243
2244 static void khugepaged_alloc_sleep(void)
2245 {
2246         wait_event_freezable_timeout(khugepaged_wait, false,
2247                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2248 }
2249
2250 static int khugepaged_node_load[MAX_NUMNODES];
2251
2252 static bool khugepaged_scan_abort(int nid)
2253 {
2254         int i;
2255
2256         /*
2257          * If zone_reclaim_mode is disabled, then no extra effort is made to
2258          * allocate memory locally.
2259          */
2260         if (!zone_reclaim_mode)
2261                 return false;
2262
2263         /* If there is a count for this node already, it must be acceptable */
2264         if (khugepaged_node_load[nid])
2265                 return false;
2266
2267         for (i = 0; i < MAX_NUMNODES; i++) {
2268                 if (!khugepaged_node_load[i])
2269                         continue;
2270                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2271                         return true;
2272         }
2273         return false;
2274 }
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_PAGE(*hpage, *hpage);
2325
2326         /*
2327          * Before allocating the hugepage, release the mmap_sem read lock.
2328          * The allocation can take potentially a long time if it involves
2329          * sync compaction, and we do not need to hold the mmap_sem during
2330          * that. We will recheck the vma after taking it again in write mode.
2331          */
2332         up_read(&mm->mmap_sem);
2333
2334         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2335                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2336         if (unlikely(!*hpage)) {
2337                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2338                 *hpage = ERR_PTR(-ENOMEM);
2339                 return NULL;
2340         }
2341
2342         count_vm_event(THP_COLLAPSE_ALLOC);
2343         return *hpage;
2344 }
2345 #else
2346 static int khugepaged_find_target_node(void)
2347 {
2348         return 0;
2349 }
2350
2351 static inline struct page *alloc_hugepage(int defrag)
2352 {
2353         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2354                            HPAGE_PMD_ORDER);
2355 }
2356
2357 static struct page *khugepaged_alloc_hugepage(bool *wait)
2358 {
2359         struct page *hpage;
2360
2361         do {
2362                 hpage = alloc_hugepage(khugepaged_defrag());
2363                 if (!hpage) {
2364                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2365                         if (!*wait)
2366                                 return NULL;
2367
2368                         *wait = false;
2369                         khugepaged_alloc_sleep();
2370                 } else
2371                         count_vm_event(THP_COLLAPSE_ALLOC);
2372         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2373
2374         return hpage;
2375 }
2376
2377 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2378 {
2379         if (!*hpage)
2380                 *hpage = khugepaged_alloc_hugepage(wait);
2381
2382         if (unlikely(!*hpage))
2383                 return false;
2384
2385         return true;
2386 }
2387
2388 static struct page
2389 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2390                        struct vm_area_struct *vma, unsigned long address,
2391                        int node)
2392 {
2393         up_read(&mm->mmap_sem);
2394         VM_BUG_ON(!*hpage);
2395         return  *hpage;
2396 }
2397 #endif
2398
2399 static bool hugepage_vma_check(struct vm_area_struct *vma)
2400 {
2401         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2402             (vma->vm_flags & VM_NOHUGEPAGE))
2403                 return false;
2404
2405         if (!vma->anon_vma || vma->vm_ops)
2406                 return false;
2407         if (is_vma_temporary_stack(vma))
2408                 return false;
2409         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2410         return true;
2411 }
2412
2413 static void collapse_huge_page(struct mm_struct *mm,
2414                                    unsigned long address,
2415                                    struct page **hpage,
2416                                    struct vm_area_struct *vma,
2417                                    int node)
2418 {
2419         pmd_t *pmd, _pmd;
2420         pte_t *pte;
2421         pgtable_t pgtable;
2422         struct page *new_page;
2423         spinlock_t *pmd_ptl, *pte_ptl;
2424         int isolated;
2425         unsigned long hstart, hend;
2426         struct mem_cgroup *memcg;
2427         unsigned long mmun_start;       /* For mmu_notifiers */
2428         unsigned long mmun_end;         /* For mmu_notifiers */
2429
2430         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2431
2432         /* release the mmap_sem read lock. */
2433         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2434         if (!new_page)
2435                 return;
2436
2437         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2438                                            GFP_TRANSHUGE, &memcg)))
2439                 return;
2440
2441         /*
2442          * Prevent all access to pagetables with the exception of
2443          * gup_fast later hanlded by the ptep_clear_flush and the VM
2444          * handled by the anon_vma lock + PG_lock.
2445          */
2446         down_write(&mm->mmap_sem);
2447         if (unlikely(khugepaged_test_exit(mm)))
2448                 goto out;
2449
2450         vma = find_vma(mm, address);
2451         if (!vma)
2452                 goto out;
2453         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2454         hend = vma->vm_end & HPAGE_PMD_MASK;
2455         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2456                 goto out;
2457         if (!hugepage_vma_check(vma))
2458                 goto out;
2459         pmd = mm_find_pmd(mm, address);
2460         if (!pmd)
2461                 goto out;
2462
2463         anon_vma_lock_write(vma->anon_vma);
2464
2465         pte = pte_offset_map(pmd, address);
2466         pte_ptl = pte_lockptr(mm, pmd);
2467
2468         mmun_start = address;
2469         mmun_end   = address + HPAGE_PMD_SIZE;
2470         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2471         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2472         /*
2473          * After this gup_fast can't run anymore. This also removes
2474          * any huge TLB entry from the CPU so we won't allow
2475          * huge and small TLB entries for the same virtual address
2476          * to avoid the risk of CPU bugs in that area.
2477          */
2478         _pmd = pmdp_clear_flush(vma, address, pmd);
2479         spin_unlock(pmd_ptl);
2480         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2481
2482         spin_lock(pte_ptl);
2483         isolated = __collapse_huge_page_isolate(vma, address, pte);
2484         spin_unlock(pte_ptl);
2485
2486         if (unlikely(!isolated)) {
2487                 pte_unmap(pte);
2488                 spin_lock(pmd_ptl);
2489                 BUG_ON(!pmd_none(*pmd));
2490                 /*
2491                  * We can only use set_pmd_at when establishing
2492                  * hugepmds and never for establishing regular pmds that
2493                  * points to regular pagetables. Use pmd_populate for that
2494                  */
2495                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2496                 spin_unlock(pmd_ptl);
2497                 anon_vma_unlock_write(vma->anon_vma);
2498                 goto out;
2499         }
2500
2501         /*
2502          * All pages are isolated and locked so anon_vma rmap
2503          * can't run anymore.
2504          */
2505         anon_vma_unlock_write(vma->anon_vma);
2506
2507         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2508         pte_unmap(pte);
2509         __SetPageUptodate(new_page);
2510         pgtable = pmd_pgtable(_pmd);
2511
2512         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2513         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2514
2515         /*
2516          * spin_lock() below is not the equivalent of smp_wmb(), so
2517          * this is needed to avoid the copy_huge_page writes to become
2518          * visible after the set_pmd_at() write.
2519          */
2520         smp_wmb();
2521
2522         spin_lock(pmd_ptl);
2523         BUG_ON(!pmd_none(*pmd));
2524         page_add_new_anon_rmap(new_page, vma, address);
2525         mem_cgroup_commit_charge(new_page, memcg, false);
2526         lru_cache_add_active_or_unevictable(new_page, vma);
2527         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2528         set_pmd_at(mm, address, pmd, _pmd);
2529         update_mmu_cache_pmd(vma, address, pmd);
2530         spin_unlock(pmd_ptl);
2531
2532         *hpage = NULL;
2533
2534         khugepaged_pages_collapsed++;
2535 out_up_write:
2536         up_write(&mm->mmap_sem);
2537         return;
2538
2539 out:
2540         mem_cgroup_cancel_charge(new_page, memcg);
2541         goto out_up_write;
2542 }
2543
2544 static int khugepaged_scan_pmd(struct mm_struct *mm,
2545                                struct vm_area_struct *vma,
2546                                unsigned long address,
2547                                struct page **hpage)
2548 {
2549         pmd_t *pmd;
2550         pte_t *pte, *_pte;
2551         int ret = 0, referenced = 0, none = 0;
2552         struct page *page;
2553         unsigned long _address;
2554         spinlock_t *ptl;
2555         int node = NUMA_NO_NODE;
2556
2557         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2558
2559         pmd = mm_find_pmd(mm, address);
2560         if (!pmd)
2561                 goto out;
2562
2563         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2564         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2565         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2566              _pte++, _address += PAGE_SIZE) {
2567                 pte_t pteval = *_pte;
2568                 if (pte_none(pteval)) {
2569                         if (++none <= khugepaged_max_ptes_none)
2570                                 continue;
2571                         else
2572                                 goto out_unmap;
2573                 }
2574                 if (!pte_present(pteval) || !pte_write(pteval))
2575                         goto out_unmap;
2576                 page = vm_normal_page(vma, _address, pteval);
2577                 if (unlikely(!page))
2578                         goto out_unmap;
2579                 /*
2580                  * Record which node the original page is from and save this
2581                  * information to khugepaged_node_load[].
2582                  * Khupaged will allocate hugepage from the node has the max
2583                  * hit record.
2584                  */
2585                 node = page_to_nid(page);
2586                 if (khugepaged_scan_abort(node))
2587                         goto out_unmap;
2588                 khugepaged_node_load[node]++;
2589                 VM_BUG_ON_PAGE(PageCompound(page), page);
2590                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2591                         goto out_unmap;
2592                 /* cannot use mapcount: can't collapse if there's a gup pin */
2593                 if (page_count(page) != 1)
2594                         goto out_unmap;
2595                 if (pte_young(pteval) || PageReferenced(page) ||
2596                     mmu_notifier_test_young(vma->vm_mm, address))
2597                         referenced = 1;
2598         }
2599         if (referenced)
2600                 ret = 1;
2601 out_unmap:
2602         pte_unmap_unlock(pte, ptl);
2603         if (ret) {
2604                 node = khugepaged_find_target_node();
2605                 /* collapse_huge_page will return with the mmap_sem released */
2606                 collapse_huge_page(mm, address, hpage, vma, node);
2607         }
2608 out:
2609         return ret;
2610 }
2611
2612 static void collect_mm_slot(struct mm_slot *mm_slot)
2613 {
2614         struct mm_struct *mm = mm_slot->mm;
2615
2616         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2617
2618         if (khugepaged_test_exit(mm)) {
2619                 /* free mm_slot */
2620                 hash_del(&mm_slot->hash);
2621                 list_del(&mm_slot->mm_node);
2622
2623                 /*
2624                  * Not strictly needed because the mm exited already.
2625                  *
2626                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2627                  */
2628
2629                 /* khugepaged_mm_lock actually not necessary for the below */
2630                 free_mm_slot(mm_slot);
2631                 mmdrop(mm);
2632         }
2633 }
2634
2635 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2636                                             struct page **hpage)
2637         __releases(&khugepaged_mm_lock)
2638         __acquires(&khugepaged_mm_lock)
2639 {
2640         struct mm_slot *mm_slot;
2641         struct mm_struct *mm;
2642         struct vm_area_struct *vma;
2643         int progress = 0;
2644
2645         VM_BUG_ON(!pages);
2646         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2647
2648         if (khugepaged_scan.mm_slot)
2649                 mm_slot = khugepaged_scan.mm_slot;
2650         else {
2651                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2652                                      struct mm_slot, mm_node);
2653                 khugepaged_scan.address = 0;
2654                 khugepaged_scan.mm_slot = mm_slot;
2655         }
2656         spin_unlock(&khugepaged_mm_lock);
2657
2658         mm = mm_slot->mm;
2659         down_read(&mm->mmap_sem);
2660         if (unlikely(khugepaged_test_exit(mm)))
2661                 vma = NULL;
2662         else
2663                 vma = find_vma(mm, khugepaged_scan.address);
2664
2665         progress++;
2666         for (; vma; vma = vma->vm_next) {
2667                 unsigned long hstart, hend;
2668
2669                 cond_resched();
2670                 if (unlikely(khugepaged_test_exit(mm))) {
2671                         progress++;
2672                         break;
2673                 }
2674                 if (!hugepage_vma_check(vma)) {
2675 skip:
2676                         progress++;
2677                         continue;
2678                 }
2679                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2680                 hend = vma->vm_end & HPAGE_PMD_MASK;
2681                 if (hstart >= hend)
2682                         goto skip;
2683                 if (khugepaged_scan.address > hend)
2684                         goto skip;
2685                 if (khugepaged_scan.address < hstart)
2686                         khugepaged_scan.address = hstart;
2687                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2688
2689                 while (khugepaged_scan.address < hend) {
2690                         int ret;
2691                         cond_resched();
2692                         if (unlikely(khugepaged_test_exit(mm)))
2693                                 goto breakouterloop;
2694
2695                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2696                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2697                                   hend);
2698                         ret = khugepaged_scan_pmd(mm, vma,
2699                                                   khugepaged_scan.address,
2700                                                   hpage);
2701                         /* move to next address */
2702                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2703                         progress += HPAGE_PMD_NR;
2704                         if (ret)
2705                                 /* we released mmap_sem so break loop */
2706                                 goto breakouterloop_mmap_sem;
2707                         if (progress >= pages)
2708                                 goto breakouterloop;
2709                 }
2710         }
2711 breakouterloop:
2712         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2713 breakouterloop_mmap_sem:
2714
2715         spin_lock(&khugepaged_mm_lock);
2716         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2717         /*
2718          * Release the current mm_slot if this mm is about to die, or
2719          * if we scanned all vmas of this mm.
2720          */
2721         if (khugepaged_test_exit(mm) || !vma) {
2722                 /*
2723                  * Make sure that if mm_users is reaching zero while
2724                  * khugepaged runs here, khugepaged_exit will find
2725                  * mm_slot not pointing to the exiting mm.
2726                  */
2727                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2728                         khugepaged_scan.mm_slot = list_entry(
2729                                 mm_slot->mm_node.next,
2730                                 struct mm_slot, mm_node);
2731                         khugepaged_scan.address = 0;
2732                 } else {
2733                         khugepaged_scan.mm_slot = NULL;
2734                         khugepaged_full_scans++;
2735                 }
2736
2737                 collect_mm_slot(mm_slot);
2738         }
2739
2740         return progress;
2741 }
2742
2743 static int khugepaged_has_work(void)
2744 {
2745         return !list_empty(&khugepaged_scan.mm_head) &&
2746                 khugepaged_enabled();
2747 }
2748
2749 static int khugepaged_wait_event(void)
2750 {
2751         return !list_empty(&khugepaged_scan.mm_head) ||
2752                 kthread_should_stop();
2753 }
2754
2755 static void khugepaged_do_scan(void)
2756 {
2757         struct page *hpage = NULL;
2758         unsigned int progress = 0, pass_through_head = 0;
2759         unsigned int pages = khugepaged_pages_to_scan;
2760         bool wait = true;
2761
2762         barrier(); /* write khugepaged_pages_to_scan to local stack */
2763
2764         while (progress < pages) {
2765                 if (!khugepaged_prealloc_page(&hpage, &wait))
2766                         break;
2767
2768                 cond_resched();
2769
2770                 if (unlikely(kthread_should_stop() || freezing(current)))
2771                         break;
2772
2773                 spin_lock(&khugepaged_mm_lock);
2774                 if (!khugepaged_scan.mm_slot)
2775                         pass_through_head++;
2776                 if (khugepaged_has_work() &&
2777                     pass_through_head < 2)
2778                         progress += khugepaged_scan_mm_slot(pages - progress,
2779                                                             &hpage);
2780                 else
2781                         progress = pages;
2782                 spin_unlock(&khugepaged_mm_lock);
2783         }
2784
2785         if (!IS_ERR_OR_NULL(hpage))
2786                 put_page(hpage);
2787 }
2788
2789 static void khugepaged_wait_work(void)
2790 {
2791         try_to_freeze();
2792
2793         if (khugepaged_has_work()) {
2794                 if (!khugepaged_scan_sleep_millisecs)
2795                         return;
2796
2797                 wait_event_freezable_timeout(khugepaged_wait,
2798                                              kthread_should_stop(),
2799                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2800                 return;
2801         }
2802
2803         if (khugepaged_enabled())
2804                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2805 }
2806
2807 static int khugepaged(void *none)
2808 {
2809         struct mm_slot *mm_slot;
2810
2811         set_freezable();
2812         set_user_nice(current, MAX_NICE);
2813
2814         while (!kthread_should_stop()) {
2815                 khugepaged_do_scan();
2816                 khugepaged_wait_work();
2817         }
2818
2819         spin_lock(&khugepaged_mm_lock);
2820         mm_slot = khugepaged_scan.mm_slot;
2821         khugepaged_scan.mm_slot = NULL;
2822         if (mm_slot)
2823                 collect_mm_slot(mm_slot);
2824         spin_unlock(&khugepaged_mm_lock);
2825         return 0;
2826 }
2827
2828 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2829                 unsigned long haddr, pmd_t *pmd)
2830 {
2831         struct mm_struct *mm = vma->vm_mm;
2832         pgtable_t pgtable;
2833         pmd_t _pmd;
2834         int i;
2835
2836         pmdp_clear_flush(vma, haddr, pmd);
2837         /* leave pmd empty until pte is filled */
2838
2839         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2840         pmd_populate(mm, &_pmd, pgtable);
2841
2842         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2843                 pte_t *pte, entry;
2844                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2845                 entry = pte_mkspecial(entry);
2846                 pte = pte_offset_map(&_pmd, haddr);
2847                 VM_BUG_ON(!pte_none(*pte));
2848                 set_pte_at(mm, haddr, pte, entry);
2849                 pte_unmap(pte);
2850         }
2851         smp_wmb(); /* make pte visible before pmd */
2852         pmd_populate(mm, pmd, pgtable);
2853         put_huge_zero_page();
2854 }
2855
2856 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2857                 pmd_t *pmd)
2858 {
2859         spinlock_t *ptl;
2860         struct page *page;
2861         struct mm_struct *mm = vma->vm_mm;
2862         unsigned long haddr = address & HPAGE_PMD_MASK;
2863         unsigned long mmun_start;       /* For mmu_notifiers */
2864         unsigned long mmun_end;         /* For mmu_notifiers */
2865
2866         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2867
2868         mmun_start = haddr;
2869         mmun_end   = haddr + HPAGE_PMD_SIZE;
2870 again:
2871         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2872         ptl = pmd_lock(mm, pmd);
2873         if (unlikely(!pmd_trans_huge(*pmd))) {
2874                 spin_unlock(ptl);
2875                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2876                 return;
2877         }
2878         if (is_huge_zero_pmd(*pmd)) {
2879                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2880                 spin_unlock(ptl);
2881                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2882                 return;
2883         }
2884         page = pmd_page(*pmd);
2885         VM_BUG_ON_PAGE(!page_count(page), page);
2886         get_page(page);
2887         spin_unlock(ptl);
2888         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2889
2890         split_huge_page(page);
2891
2892         put_page(page);
2893
2894         /*
2895          * We don't always have down_write of mmap_sem here: a racing
2896          * do_huge_pmd_wp_page() might have copied-on-write to another
2897          * huge page before our split_huge_page() got the anon_vma lock.
2898          */
2899         if (unlikely(pmd_trans_huge(*pmd)))
2900                 goto again;
2901 }
2902
2903 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2904                 pmd_t *pmd)
2905 {
2906         struct vm_area_struct *vma;
2907
2908         vma = find_vma(mm, address);
2909         BUG_ON(vma == NULL);
2910         split_huge_page_pmd(vma, address, pmd);
2911 }
2912
2913 static void split_huge_page_address(struct mm_struct *mm,
2914                                     unsigned long address)
2915 {
2916         pgd_t *pgd;
2917         pud_t *pud;
2918         pmd_t *pmd;
2919
2920         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2921
2922         pgd = pgd_offset(mm, address);
2923         if (!pgd_present(*pgd))
2924                 return;
2925
2926         pud = pud_offset(pgd, address);
2927         if (!pud_present(*pud))
2928                 return;
2929
2930         pmd = pmd_offset(pud, address);
2931         if (!pmd_present(*pmd))
2932                 return;
2933         /*
2934          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2935          * materialize from under us.
2936          */
2937         split_huge_page_pmd_mm(mm, address, pmd);
2938 }
2939
2940 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2941                              unsigned long start,
2942                              unsigned long end,
2943                              long adjust_next)
2944 {
2945         /*
2946          * If the new start address isn't hpage aligned and it could
2947          * previously contain an hugepage: check if we need to split
2948          * an huge pmd.
2949          */
2950         if (start & ~HPAGE_PMD_MASK &&
2951             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2952             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2953                 split_huge_page_address(vma->vm_mm, start);
2954
2955         /*
2956          * If the new end address isn't hpage aligned and it could
2957          * previously contain an hugepage: check if we need to split
2958          * an huge pmd.
2959          */
2960         if (end & ~HPAGE_PMD_MASK &&
2961             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2962             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2963                 split_huge_page_address(vma->vm_mm, end);
2964
2965         /*
2966          * If we're also updating the vma->vm_next->vm_start, if the new
2967          * vm_next->vm_start isn't page aligned and it could previously
2968          * contain an hugepage: check if we need to split an huge pmd.
2969          */
2970         if (adjust_next > 0) {
2971                 struct vm_area_struct *next = vma->vm_next;
2972                 unsigned long nstart = next->vm_start;
2973                 nstart += adjust_next << PAGE_SHIFT;
2974                 if (nstart & ~HPAGE_PMD_MASK &&
2975                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2976                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2977                         split_huge_page_address(next->vm_mm, nstart);
2978         }
2979 }