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/dax.h>
  20 #include <linux/kthread.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/mman.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/migrate.h>
  26 #include <linux/hashtable.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/page_idle.h>
  29
  30 #include <asm/tlb.h>
  31 #include <asm/pgalloc.h>
  32 #include "internal.h"
  33
  34 /*
  35  * By default transparent hugepage support is disabled in order that avoid
  36  * to risk increase the memory footprint of applications without a guaranteed
  37  * benefit. When transparent hugepage support is enabled, is for all mappings,
  38  * and khugepaged scans all mappings.
  39  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  40  * for all hugepage allocations.
  41  */
  42 unsigned long transparent_hugepage_flags __read_mostly =
  43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  44         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  45 #endif
  46 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  47         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  48 #endif
  49         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  50         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  51         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  52
  53 /* default scan 8*512 pte (or vmas) every 30 second */
  54 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  55 static unsigned int khugepaged_pages_collapsed;
  56 static unsigned int khugepaged_full_scans;
  57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  58 /* during fragmentation poll the hugepage allocator once every minute */
  59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  60 static struct task_struct *khugepaged_thread __read_mostly;
  61 static DEFINE_MUTEX(khugepaged_mutex);
  62 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  63 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  64 /*
  65  * default collapse hugepages if there is at least one pte mapped like
  66  * it would have happened if the vma was large enough during page
  67  * fault.
  68  */
  69 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  70
  71 static int khugepaged(void *none);
  72 static int khugepaged_slab_init(void);
  73 static void khugepaged_slab_exit(void);
  74
  75 #define MM_SLOTS_HASH_BITS 10
  76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  77
  78 static struct kmem_cache *mm_slot_cache __read_mostly;
  79
  80 /**
  81  * struct mm_slot - hash lookup from mm to mm_slot
  82  * @hash: hash collision list
  83  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  84  * @mm: the mm that this information is valid for
  85  */
  86 struct mm_slot {
  87         struct hlist_node hash;
  88         struct list_head mm_node;
  89         struct mm_struct *mm;
  90 };
  91
  92 /**
  93  * struct khugepaged_scan - cursor for scanning
  94  * @mm_head: the head of the mm list to scan
  95  * @mm_slot: the current mm_slot we are scanning
  96  * @address: the next address inside that to be scanned
  97  *
  98  * There is only the one khugepaged_scan instance of this cursor structure.
  99  */
 100 struct khugepaged_scan {
 101         struct list_head mm_head;
 102         struct mm_slot *mm_slot;
 103         unsigned long address;
 104 };
 105 static struct khugepaged_scan khugepaged_scan = {
 106         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 107 };
 108
 109
 110 static void set_recommended_min_free_kbytes(void)
 111 {
 112         struct zone *zone;
 113         int nr_zones = 0;
 114         unsigned long recommended_min;
 115
 116         for_each_populated_zone(zone)
 117                 nr_zones++;
 118
 119         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 120         recommended_min = pageblock_nr_pages * nr_zones * 2;
 121
 122         /*
 123          * Make sure that on average at least two pageblocks are almost free
 124          * of another type, one for a migratetype to fall back to and a
 125          * second to avoid subsequent fallbacks of other types There are 3
 126          * MIGRATE_TYPES we care about.
 127          */
 128         recommended_min += pageblock_nr_pages * nr_zones *
 129                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 130
 131         /* don't ever allow to reserve more than 5% of the lowmem */
 132         recommended_min = min(recommended_min,
 133                               (unsigned long) nr_free_buffer_pages() / 20);
 134         recommended_min <<= (PAGE_SHIFT-10);
 135
 136         if (recommended_min > min_free_kbytes) {
 137                 if (user_min_free_kbytes >= 0)
 138                         pr_info("raising min_free_kbytes from %d to %lu 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 }
 145
 146 static int start_stop_khugepaged(void)
 147 {
 148         int err = 0;
 149         if (khugepaged_enabled()) {
 150                 if (!khugepaged_thread)
 151                         khugepaged_thread = kthread_run(khugepaged, NULL,
 152                                                         "khugepaged");
 153                 if (IS_ERR(khugepaged_thread)) {
 154                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 155                         err = PTR_ERR(khugepaged_thread);
 156                         khugepaged_thread = NULL;
 157                         goto fail;
 158                 }
 159
 160                 if (!list_empty(&khugepaged_scan.mm_head))
 161                         wake_up_interruptible(&khugepaged_wait);
 162
 163                 set_recommended_min_free_kbytes();
 164         } else if (khugepaged_thread) {
 165                 kthread_stop(khugepaged_thread);
 166                 khugepaged_thread = NULL;
 167         }
 168 fail:
 169         return err;
 170 }
 171
 172 static atomic_t huge_zero_refcount;
 173 struct page *huge_zero_page __read_mostly;
 174
 175 struct page *get_huge_zero_page(void)
 176 {
 177         struct page *zero_page;
 178 retry:
 179         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 180                 return READ_ONCE(huge_zero_page);
 181
 182         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 183                         HPAGE_PMD_ORDER);
 184         if (!zero_page) {
 185                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 186                 return NULL;
 187         }
 188         count_vm_event(THP_ZERO_PAGE_ALLOC);
 189         preempt_disable();
 190         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 191                 preempt_enable();
 192                 __free_pages(zero_page, compound_order(zero_page));
 193                 goto retry;
 194         }
 195
 196         /* We take additional reference here. It will be put back by shrinker */
 197         atomic_set(&huge_zero_refcount, 2);
 198         preempt_enable();
 199         return READ_ONCE(huge_zero_page);
 200 }
 201
 202 static void put_huge_zero_page(void)
 203 {
 204         /*
 205          * Counter should never go to zero here. Only shrinker can put
 206          * last reference.
 207          */
 208         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 209 }
 210
 211 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 212                                         struct shrink_control *sc)
 213 {
 214         /* we can free zero page only if last reference remains */
 215         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 216 }
 217
 218 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 219                                        struct shrink_control *sc)
 220 {
 221         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 222                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 223                 BUG_ON(zero_page == NULL);
 224                 __free_pages(zero_page, compound_order(zero_page));
 225                 return HPAGE_PMD_NR;
 226         }
 227
 228         return 0;
 229 }
 230
 231 static struct shrinker huge_zero_page_shrinker = {
 232         .count_objects = shrink_huge_zero_page_count,
 233         .scan_objects = shrink_huge_zero_page_scan,
 234         .seeks = DEFAULT_SEEKS,
 235 };
 236
 237 #ifdef CONFIG_SYSFS
 238
 239 static ssize_t double_flag_show(struct kobject *kobj,
 240                                 struct kobj_attribute *attr, char *buf,
 241                                 enum transparent_hugepage_flag enabled,
 242                                 enum transparent_hugepage_flag req_madv)
 243 {
 244         if (test_bit(enabled, &transparent_hugepage_flags)) {
 245                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 246                 return sprintf(buf, "[always] madvise never\n");
 247         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 248                 return sprintf(buf, "always [madvise] never\n");
 249         else
 250                 return sprintf(buf, "always madvise [never]\n");
 251 }
 252 static ssize_t double_flag_store(struct kobject *kobj,
 253                                  struct kobj_attribute *attr,
 254                                  const char *buf, size_t count,
 255                                  enum transparent_hugepage_flag enabled,
 256                                  enum transparent_hugepage_flag req_madv)
 257 {
 258         if (!memcmp("always", buf,
 259                     min(sizeof("always")-1, count))) {
 260                 set_bit(enabled, &transparent_hugepage_flags);
 261                 clear_bit(req_madv, &transparent_hugepage_flags);
 262         } else if (!memcmp("madvise", buf,
 263                            min(sizeof("madvise")-1, count))) {
 264                 clear_bit(enabled, &transparent_hugepage_flags);
 265                 set_bit(req_madv, &transparent_hugepage_flags);
 266         } else if (!memcmp("never", buf,
 267                            min(sizeof("never")-1, count))) {
 268                 clear_bit(enabled, &transparent_hugepage_flags);
 269                 clear_bit(req_madv, &transparent_hugepage_flags);
 270         } else
 271                 return -EINVAL;
 272
 273         return count;
 274 }
 275
 276 static ssize_t enabled_show(struct kobject *kobj,
 277                             struct kobj_attribute *attr, char *buf)
 278 {
 279         return double_flag_show(kobj, attr, buf,
 280                                 TRANSPARENT_HUGEPAGE_FLAG,
 281                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 282 }
 283 static ssize_t enabled_store(struct kobject *kobj,
 284                              struct kobj_attribute *attr,
 285                              const char *buf, size_t count)
 286 {
 287         ssize_t ret;
 288
 289         ret = double_flag_store(kobj, attr, buf, count,
 290                                 TRANSPARENT_HUGEPAGE_FLAG,
 291                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 292
 293         if (ret > 0) {
 294                 int err;
 295
 296                 mutex_lock(&khugepaged_mutex);
 297                 err = start_stop_khugepaged();
 298                 mutex_unlock(&khugepaged_mutex);
 299
 300                 if (err)
 301                         ret = err;
 302         }
 303
 304         return ret;
 305 }
 306 static struct kobj_attribute enabled_attr =
 307         __ATTR(enabled, 0644, enabled_show, enabled_store);
 308
 309 static ssize_t single_flag_show(struct kobject *kobj,
 310                                 struct kobj_attribute *attr, char *buf,
 311                                 enum transparent_hugepage_flag flag)
 312 {
 313         return sprintf(buf, "%d\n",
 314                        !!test_bit(flag, &transparent_hugepage_flags));
 315 }
 316
 317 static ssize_t single_flag_store(struct kobject *kobj,
 318                                  struct kobj_attribute *attr,
 319                                  const char *buf, size_t count,
 320                                  enum transparent_hugepage_flag flag)
 321 {
 322         unsigned long value;
 323         int ret;
 324
 325         ret = kstrtoul(buf, 10, &value);
 326         if (ret < 0)
 327                 return ret;
 328         if (value > 1)
 329                 return -EINVAL;
 330
 331         if (value)
 332                 set_bit(flag, &transparent_hugepage_flags);
 333         else
 334                 clear_bit(flag, &transparent_hugepage_flags);
 335
 336         return count;
 337 }
 338
 339 /*
 340  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 341  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 342  * memory just to allocate one more hugepage.
 343  */
 344 static ssize_t defrag_show(struct kobject *kobj,
 345                            struct kobj_attribute *attr, char *buf)
 346 {
 347         return double_flag_show(kobj, attr, buf,
 348                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 349                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 350 }
 351 static ssize_t defrag_store(struct kobject *kobj,
 352                             struct kobj_attribute *attr,
 353                             const char *buf, size_t count)
 354 {
 355         return double_flag_store(kobj, attr, buf, count,
 356                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 357                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 358 }
 359 static struct kobj_attribute defrag_attr =
 360         __ATTR(defrag, 0644, defrag_show, defrag_store);
 361
 362 static ssize_t use_zero_page_show(struct kobject *kobj,
 363                 struct kobj_attribute *attr, char *buf)
 364 {
 365         return single_flag_show(kobj, attr, buf,
 366                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 367 }
 368 static ssize_t use_zero_page_store(struct kobject *kobj,
 369                 struct kobj_attribute *attr, const char *buf, size_t count)
 370 {
 371         return single_flag_store(kobj, attr, buf, count,
 372                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 373 }
 374 static struct kobj_attribute use_zero_page_attr =
 375         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 376 #ifdef CONFIG_DEBUG_VM
 377 static ssize_t debug_cow_show(struct kobject *kobj,
 378                                 struct kobj_attribute *attr, char *buf)
 379 {
 380         return single_flag_show(kobj, attr, buf,
 381                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 382 }
 383 static ssize_t debug_cow_store(struct kobject *kobj,
 384                                struct kobj_attribute *attr,
 385                                const char *buf, size_t count)
 386 {
 387         return single_flag_store(kobj, attr, buf, count,
 388                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 389 }
 390 static struct kobj_attribute debug_cow_attr =
 391         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 392 #endif /* CONFIG_DEBUG_VM */
 393
 394 static struct attribute *hugepage_attr[] = {
 395         &enabled_attr.attr,
 396         &defrag_attr.attr,
 397         &use_zero_page_attr.attr,
 398 #ifdef CONFIG_DEBUG_VM
 399         &debug_cow_attr.attr,
 400 #endif
 401         NULL,
 402 };
 403
 404 static struct attribute_group hugepage_attr_group = {
 405         .attrs = hugepage_attr,
 406 };
 407
 408 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 409                                          struct kobj_attribute *attr,
 410                                          char *buf)
 411 {
 412         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 413 }
 414
 415 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 416                                           struct kobj_attribute *attr,
 417                                           const char *buf, size_t count)
 418 {
 419         unsigned long msecs;
 420         int err;
 421
 422         err = kstrtoul(buf, 10, &msecs);
 423         if (err || msecs > UINT_MAX)
 424                 return -EINVAL;
 425
 426         khugepaged_scan_sleep_millisecs = msecs;
 427         wake_up_interruptible(&khugepaged_wait);
 428
 429         return count;
 430 }
 431 static struct kobj_attribute scan_sleep_millisecs_attr =
 432         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 433                scan_sleep_millisecs_store);
 434
 435 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 436                                           struct kobj_attribute *attr,
 437                                           char *buf)
 438 {
 439         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 440 }
 441
 442 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 443                                            struct kobj_attribute *attr,
 444                                            const char *buf, size_t count)
 445 {
 446         unsigned long msecs;
 447         int err;
 448
 449         err = kstrtoul(buf, 10, &msecs);
 450         if (err || msecs > UINT_MAX)
 451                 return -EINVAL;
 452
 453         khugepaged_alloc_sleep_millisecs = msecs;
 454         wake_up_interruptible(&khugepaged_wait);
 455
 456         return count;
 457 }
 458 static struct kobj_attribute alloc_sleep_millisecs_attr =
 459         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 460                alloc_sleep_millisecs_store);
 461
 462 static ssize_t pages_to_scan_show(struct kobject *kobj,
 463                                   struct kobj_attribute *attr,
 464                                   char *buf)
 465 {
 466         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 467 }
 468 static ssize_t pages_to_scan_store(struct kobject *kobj,
 469                                    struct kobj_attribute *attr,
 470                                    const char *buf, size_t count)
 471 {
 472         int err;
 473         unsigned long pages;
 474
 475         err = kstrtoul(buf, 10, &pages);
 476         if (err || !pages || pages > UINT_MAX)
 477                 return -EINVAL;
 478
 479         khugepaged_pages_to_scan = pages;
 480
 481         return count;
 482 }
 483 static struct kobj_attribute pages_to_scan_attr =
 484         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 485                pages_to_scan_store);
 486
 487 static ssize_t pages_collapsed_show(struct kobject *kobj,
 488                                     struct kobj_attribute *attr,
 489                                     char *buf)
 490 {
 491         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 492 }
 493 static struct kobj_attribute pages_collapsed_attr =
 494         __ATTR_RO(pages_collapsed);
 495
 496 static ssize_t full_scans_show(struct kobject *kobj,
 497                                struct kobj_attribute *attr,
 498                                char *buf)
 499 {
 500         return sprintf(buf, "%u\n", khugepaged_full_scans);
 501 }
 502 static struct kobj_attribute full_scans_attr =
 503         __ATTR_RO(full_scans);
 504
 505 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 506                                       struct kobj_attribute *attr, char *buf)
 507 {
 508         return single_flag_show(kobj, attr, buf,
 509                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 510 }
 511 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 512                                        struct kobj_attribute *attr,
 513                                        const char *buf, size_t count)
 514 {
 515         return single_flag_store(kobj, attr, buf, count,
 516                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 517 }
 518 static struct kobj_attribute khugepaged_defrag_attr =
 519         __ATTR(defrag, 0644, khugepaged_defrag_show,
 520                khugepaged_defrag_store);
 521
 522 /*
 523  * max_ptes_none controls if khugepaged should collapse hugepages over
 524  * any unmapped ptes in turn potentially increasing the memory
 525  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 526  * reduce the available free memory in the system as it
 527  * runs. Increasing max_ptes_none will instead potentially reduce the
 528  * free memory in the system during the khugepaged scan.
 529  */
 530 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 531                                              struct kobj_attribute *attr,
 532                                              char *buf)
 533 {
 534         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 535 }
 536 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 537                                               struct kobj_attribute *attr,
 538                                               const char *buf, size_t count)
 539 {
 540         int err;
 541         unsigned long max_ptes_none;
 542
 543         err = kstrtoul(buf, 10, &max_ptes_none);
 544         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 545                 return -EINVAL;
 546
 547         khugepaged_max_ptes_none = max_ptes_none;
 548
 549         return count;
 550 }
 551 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 552         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 553                khugepaged_max_ptes_none_store);
 554
 555 static struct attribute *khugepaged_attr[] = {
 556         &khugepaged_defrag_attr.attr,
 557         &khugepaged_max_ptes_none_attr.attr,
 558         &pages_to_scan_attr.attr,
 559         &pages_collapsed_attr.attr,
 560         &full_scans_attr.attr,
 561         &scan_sleep_millisecs_attr.attr,
 562         &alloc_sleep_millisecs_attr.attr,
 563         NULL,
 564 };
 565
 566 static struct attribute_group khugepaged_attr_group = {
 567         .attrs = khugepaged_attr,
 568         .name = "khugepaged",
 569 };
 570
 571 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 572 {
 573         int err;
 574
 575         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 576         if (unlikely(!*hugepage_kobj)) {
 577                 pr_err("failed to create transparent hugepage kobject\n");
 578                 return -ENOMEM;
 579         }
 580
 581         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 582         if (err) {
 583                 pr_err("failed to register transparent hugepage group\n");
 584                 goto delete_obj;
 585         }
 586
 587         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 588         if (err) {
 589                 pr_err("failed to register transparent hugepage group\n");
 590                 goto remove_hp_group;
 591         }
 592
 593         return 0;
 594
 595 remove_hp_group:
 596         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 597 delete_obj:
 598         kobject_put(*hugepage_kobj);
 599         return err;
 600 }
 601
 602 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 603 {
 604         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 605         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 606         kobject_put(hugepage_kobj);
 607 }
 608 #else
 609 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 610 {
 611         return 0;
 612 }
 613
 614 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 615 {
 616 }
 617 #endif /* CONFIG_SYSFS */
 618
 619 static int __init hugepage_init(void)
 620 {
 621         int err;
 622         struct kobject *hugepage_kobj;
 623
 624         if (!has_transparent_hugepage()) {
 625                 transparent_hugepage_flags = 0;
 626                 return -EINVAL;
 627         }
 628
 629         err = hugepage_init_sysfs(&hugepage_kobj);
 630         if (err)
 631                 goto err_sysfs;
 632
 633         err = khugepaged_slab_init();
 634         if (err)
 635                 goto err_slab;
 636
 637         err = register_shrinker(&huge_zero_page_shrinker);
 638         if (err)
 639                 goto err_hzp_shrinker;
 640
 641         /*
 642          * By default disable transparent hugepages on smaller systems,
 643          * where the extra memory used could hurt more than TLB overhead
 644          * is likely to save.  The admin can still enable it through /sys.
 645          */
 646         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 647                 transparent_hugepage_flags = 0;
 648                 return 0;
 649         }
 650
 651         err = start_stop_khugepaged();
 652         if (err)
 653                 goto err_khugepaged;
 654
 655         return 0;
 656 err_khugepaged:
 657         unregister_shrinker(&huge_zero_page_shrinker);
 658 err_hzp_shrinker:
 659         khugepaged_slab_exit();
 660 err_slab:
 661         hugepage_exit_sysfs(hugepage_kobj);
 662 err_sysfs:
 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 address, pmd_t *pmd,
 716                                         struct page *page, gfp_t gfp,
 717                                         unsigned int flags)
 718 {
 719         struct mem_cgroup *memcg;
 720         pgtable_t pgtable;
 721         spinlock_t *ptl;
 722         unsigned long haddr = address & HPAGE_PMD_MASK;
 723
 724         VM_BUG_ON_PAGE(!PageCompound(page), page);
 725
 726         if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
 727                 put_page(page);
 728                 count_vm_event(THP_FAULT_FALLBACK);
 729                 return VM_FAULT_FALLBACK;
 730         }
 731
 732         pgtable = pte_alloc_one(mm, haddr);
 733         if (unlikely(!pgtable)) {
 734                 mem_cgroup_cancel_charge(page, memcg);
 735                 put_page(page);
 736                 return VM_FAULT_OOM;
 737         }
 738
 739         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 740         /*
 741          * The memory barrier inside __SetPageUptodate makes sure that
 742          * clear_huge_page writes become visible before the set_pmd_at()
 743          * write.
 744          */
 745         __SetPageUptodate(page);
 746
 747         ptl = pmd_lock(mm, pmd);
 748         if (unlikely(!pmd_none(*pmd))) {
 749                 spin_unlock(ptl);
 750                 mem_cgroup_cancel_charge(page, memcg);
 751                 put_page(page);
 752                 pte_free(mm, pgtable);
 753         } else {
 754                 pmd_t entry;
 755
 756                 /* Deliver the page fault to userland */
 757                 if (userfaultfd_missing(vma)) {
 758                         int ret;
 759
 760                         spin_unlock(ptl);
 761                         mem_cgroup_cancel_charge(page, memcg);
 762                         put_page(page);
 763                         pte_free(mm, pgtable);
 764                         ret = handle_userfault(vma, address, flags,
 765                                                VM_UFFD_MISSING);
 766                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 767                         return ret;
 768                 }
 769
 770                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 771                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 772                 page_add_new_anon_rmap(page, vma, haddr);
 773                 mem_cgroup_commit_charge(page, memcg, false);
 774                 lru_cache_add_active_or_unevictable(page, vma);
 775                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 776                 set_pmd_at(mm, haddr, pmd, entry);
 777                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 778                 atomic_long_inc(&mm->nr_ptes);
 779                 spin_unlock(ptl);
 780                 count_vm_event(THP_FAULT_ALLOC);
 781         }
 782
 783         return 0;
 784 }
 785
 786 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 787 {
 788         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 789 }
 790
 791 /* Caller must hold page table lock. */
 792 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 793                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 794                 struct page *zero_page)
 795 {
 796         pmd_t entry;
 797         if (!pmd_none(*pmd))
 798                 return false;
 799         entry = mk_pmd(zero_page, vma->vm_page_prot);
 800         entry = pmd_mkhuge(entry);
 801         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 802         set_pmd_at(mm, haddr, pmd, entry);
 803         atomic_long_inc(&mm->nr_ptes);
 804         return true;
 805 }
 806
 807 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 808                                unsigned long address, pmd_t *pmd,
 809                                unsigned int flags)
 810 {
 811         gfp_t gfp;
 812         struct page *page;
 813         unsigned long haddr = address & HPAGE_PMD_MASK;
 814
 815         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 816                 return VM_FAULT_FALLBACK;
 817         if (unlikely(anon_vma_prepare(vma)))
 818                 return VM_FAULT_OOM;
 819         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 820                 return VM_FAULT_OOM;
 821         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 822                         transparent_hugepage_use_zero_page()) {
 823                 spinlock_t *ptl;
 824                 pgtable_t pgtable;
 825                 struct page *zero_page;
 826                 bool set;
 827                 int ret;
 828                 pgtable = pte_alloc_one(mm, haddr);
 829                 if (unlikely(!pgtable))
 830                         return VM_FAULT_OOM;
 831                 zero_page = get_huge_zero_page();
 832                 if (unlikely(!zero_page)) {
 833                         pte_free(mm, pgtable);
 834                         count_vm_event(THP_FAULT_FALLBACK);
 835                         return VM_FAULT_FALLBACK;
 836                 }
 837                 ptl = pmd_lock(mm, pmd);
 838                 ret = 0;
 839                 set = false;
 840                 if (pmd_none(*pmd)) {
 841                         if (userfaultfd_missing(vma)) {
 842                                 spin_unlock(ptl);
 843                                 ret = handle_userfault(vma, address, flags,
 844                                                        VM_UFFD_MISSING);
 845                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 846                         } else {
 847                                 set_huge_zero_page(pgtable, mm, vma,
 848                                                    haddr, pmd,
 849                                                    zero_page);
 850                                 spin_unlock(ptl);
 851                                 set = true;
 852                         }
 853                 } else
 854                         spin_unlock(ptl);
 855                 if (!set) {
 856                         pte_free(mm, pgtable);
 857                         put_huge_zero_page();
 858                 }
 859                 return ret;
 860         }
 861         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 862         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 863         if (unlikely(!page)) {
 864                 count_vm_event(THP_FAULT_FALLBACK);
 865                 return VM_FAULT_FALLBACK;
 866         }
 867         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 868                                             flags);
 869 }
 870
 871 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 872                 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
 873 {
 874         struct mm_struct *mm = vma->vm_mm;
 875         pmd_t entry;
 876         spinlock_t *ptl;
 877
 878         ptl = pmd_lock(mm, pmd);
 879         if (pmd_none(*pmd)) {
 880                 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
 881                 if (write) {
 882                         entry = pmd_mkyoung(pmd_mkdirty(entry));
 883                         entry = maybe_pmd_mkwrite(entry, vma);
 884                 }
 885                 set_pmd_at(mm, addr, pmd, entry);
 886                 update_mmu_cache_pmd(vma, addr, pmd);
 887         }
 888         spin_unlock(ptl);
 889 }
 890
 891 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 892                         pmd_t *pmd, unsigned long pfn, bool write)
 893 {
 894         pgprot_t pgprot = vma->vm_page_prot;
 895         /*
 896          * If we had pmd_special, we could avoid all these restrictions,
 897          * but we need to be consistent with PTEs and architectures that
 898          * can't support a 'special' bit.
 899          */
 900         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 901         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 902                                                 (VM_PFNMAP|VM_MIXEDMAP));
 903         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 904         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
 905
 906         if (addr < vma->vm_start || addr >= vma->vm_end)
 907                 return VM_FAULT_SIGBUS;
 908         if (track_pfn_insert(vma, &pgprot, pfn))
 909                 return VM_FAULT_SIGBUS;
 910         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 911         return VM_FAULT_NOPAGE;
 912 }
 913
 914 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 915                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 916                   struct vm_area_struct *vma)
 917 {
 918         spinlock_t *dst_ptl, *src_ptl;
 919         struct page *src_page;
 920         pmd_t pmd;
 921         pgtable_t pgtable;
 922         int ret;
 923
 924         ret = -ENOMEM;
 925         pgtable = pte_alloc_one(dst_mm, addr);
 926         if (unlikely(!pgtable))
 927                 goto out;
 928
 929         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 930         src_ptl = pmd_lockptr(src_mm, src_pmd);
 931         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 932
 933         ret = -EAGAIN;
 934         pmd = *src_pmd;
 935         if (unlikely(!pmd_trans_huge(pmd))) {
 936                 pte_free(dst_mm, pgtable);
 937                 goto out_unlock;
 938         }
 939         /*
 940          * When page table lock is held, the huge zero pmd should not be
 941          * under splitting since we don't split the page itself, only pmd to
 942          * a page table.
 943          */
 944         if (is_huge_zero_pmd(pmd)) {
 945                 struct page *zero_page;
 946                 /*
 947                  * get_huge_zero_page() will never allocate a new page here,
 948                  * since we already have a zero page to copy. It just takes a
 949                  * reference.
 950                  */
 951                 zero_page = get_huge_zero_page();
 952                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 953                                 zero_page);
 954                 ret = 0;
 955                 goto out_unlock;
 956         }
 957
 958         if (unlikely(pmd_trans_splitting(pmd))) {
 959                 /* split huge page running from under us */
 960                 spin_unlock(src_ptl);
 961                 spin_unlock(dst_ptl);
 962                 pte_free(dst_mm, pgtable);
 963
 964                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 965                 goto out;
 966         }
 967         src_page = pmd_page(pmd);
 968         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 969         get_page(src_page);
 970         page_dup_rmap(src_page);
 971         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 972
 973         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 974         pmd = pmd_mkold(pmd_wrprotect(pmd));
 975         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 976         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 977         atomic_long_inc(&dst_mm->nr_ptes);
 978
 979         ret = 0;
 980 out_unlock:
 981         spin_unlock(src_ptl);
 982         spin_unlock(dst_ptl);
 983 out:
 984         return ret;
 985 }
 986
 987 void huge_pmd_set_accessed(struct mm_struct *mm,
 988                            struct vm_area_struct *vma,
 989                            unsigned long address,
 990                            pmd_t *pmd, pmd_t orig_pmd,
 991                            int dirty)
 992 {
 993         spinlock_t *ptl;
 994         pmd_t entry;
 995         unsigned long haddr;
 996
 997         ptl = pmd_lock(mm, pmd);
 998         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 999                 goto unlock;
1000
1001         entry = pmd_mkyoung(orig_pmd);
1002         haddr = address & HPAGE_PMD_MASK;
1003         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1004                 update_mmu_cache_pmd(vma, address, pmd);
1005
1006 unlock:
1007         spin_unlock(ptl);
1008 }
1009
1010 /*
1011  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1012  * during copy_user_huge_page()'s copy_page_rep(): in the case when
1013  * the source page gets split and a tail freed before copy completes.
1014  * Called under pmd_lock of checked pmd, so safe from splitting itself.
1015  */
1016 static void get_user_huge_page(struct page *page)
1017 {
1018         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1019                 struct page *endpage = page + HPAGE_PMD_NR;
1020
1021                 atomic_add(HPAGE_PMD_NR, &page->_count);
1022                 while (++page < endpage)
1023                         get_huge_page_tail(page);
1024         } else {
1025                 get_page(page);
1026         }
1027 }
1028
1029 static void put_user_huge_page(struct page *page)
1030 {
1031         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1032                 struct page *endpage = page + HPAGE_PMD_NR;
1033
1034                 while (page < endpage)
1035                         put_page(page++);
1036         } else {
1037                 put_page(page);
1038         }
1039 }
1040
1041 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1042                                         struct vm_area_struct *vma,
1043                                         unsigned long address,
1044                                         pmd_t *pmd, pmd_t orig_pmd,
1045                                         struct page *page,
1046                                         unsigned long haddr)
1047 {
1048         struct mem_cgroup *memcg;
1049         spinlock_t *ptl;
1050         pgtable_t pgtable;
1051         pmd_t _pmd;
1052         int ret = 0, i;
1053         struct page **pages;
1054         unsigned long mmun_start;       /* For mmu_notifiers */
1055         unsigned long mmun_end;         /* For mmu_notifiers */
1056
1057         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1058                         GFP_KERNEL);
1059         if (unlikely(!pages)) {
1060                 ret |= VM_FAULT_OOM;
1061                 goto out;
1062         }
1063
1064         for (i = 0; i < HPAGE_PMD_NR; i++) {
1065                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1066                                                __GFP_OTHER_NODE,
1067                                                vma, address, page_to_nid(page));
1068                 if (unlikely(!pages[i] ||
1069                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1070                                                    &memcg))) {
1071                         if (pages[i])
1072                                 put_page(pages[i]);
1073                         while (--i >= 0) {
1074                                 memcg = (void *)page_private(pages[i]);
1075                                 set_page_private(pages[i], 0);
1076                                 mem_cgroup_cancel_charge(pages[i], memcg);
1077                                 put_page(pages[i]);
1078                         }
1079                         kfree(pages);
1080                         ret |= VM_FAULT_OOM;
1081                         goto out;
1082                 }
1083                 set_page_private(pages[i], (unsigned long)memcg);
1084         }
1085
1086         for (i = 0; i < HPAGE_PMD_NR; i++) {
1087                 copy_user_highpage(pages[i], page + i,
1088                                    haddr + PAGE_SIZE * i, vma);
1089                 __SetPageUptodate(pages[i]);
1090                 cond_resched();
1091         }
1092
1093         mmun_start = haddr;
1094         mmun_end   = haddr + HPAGE_PMD_SIZE;
1095         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1096
1097         ptl = pmd_lock(mm, pmd);
1098         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1099                 goto out_free_pages;
1100         VM_BUG_ON_PAGE(!PageHead(page), page);
1101
1102         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1103         /* leave pmd empty until pte is filled */
1104
1105         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1106         pmd_populate(mm, &_pmd, pgtable);
1107
1108         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1109                 pte_t *pte, entry;
1110                 entry = mk_pte(pages[i], vma->vm_page_prot);
1111                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1112                 memcg = (void *)page_private(pages[i]);
1113                 set_page_private(pages[i], 0);
1114                 page_add_new_anon_rmap(pages[i], vma, haddr);
1115                 mem_cgroup_commit_charge(pages[i], memcg, false);
1116                 lru_cache_add_active_or_unevictable(pages[i], vma);
1117                 pte = pte_offset_map(&_pmd, haddr);
1118                 VM_BUG_ON(!pte_none(*pte));
1119                 set_pte_at(mm, haddr, pte, entry);
1120                 pte_unmap(pte);
1121         }
1122         kfree(pages);
1123
1124         smp_wmb(); /* make pte visible before pmd */
1125         pmd_populate(mm, pmd, pgtable);
1126         page_remove_rmap(page);
1127         spin_unlock(ptl);
1128
1129         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1130
1131         ret |= VM_FAULT_WRITE;
1132         put_page(page);
1133
1134 out:
1135         return ret;
1136
1137 out_free_pages:
1138         spin_unlock(ptl);
1139         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1140         for (i = 0; i < HPAGE_PMD_NR; i++) {
1141                 memcg = (void *)page_private(pages[i]);
1142                 set_page_private(pages[i], 0);
1143                 mem_cgroup_cancel_charge(pages[i], memcg);
1144                 put_page(pages[i]);
1145         }
1146         kfree(pages);
1147         goto out;
1148 }
1149
1150 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1151                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1152 {
1153         spinlock_t *ptl;
1154         int ret = 0;
1155         struct page *page = NULL, *new_page;
1156         struct mem_cgroup *memcg;
1157         unsigned long haddr;
1158         unsigned long mmun_start;       /* For mmu_notifiers */
1159         unsigned long mmun_end;         /* For mmu_notifiers */
1160         gfp_t huge_gfp;                 /* for allocation and charge */
1161
1162         ptl = pmd_lockptr(mm, pmd);
1163         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1164         haddr = address & HPAGE_PMD_MASK;
1165         if (is_huge_zero_pmd(orig_pmd))
1166                 goto alloc;
1167         spin_lock(ptl);
1168         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1169                 goto out_unlock;
1170
1171         page = pmd_page(orig_pmd);
1172         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1173         if (page_mapcount(page) == 1) {
1174                 pmd_t entry;
1175                 entry = pmd_mkyoung(orig_pmd);
1176                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1177                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1178                         update_mmu_cache_pmd(vma, address, pmd);
1179                 ret |= VM_FAULT_WRITE;
1180                 goto out_unlock;
1181         }
1182         get_user_huge_page(page);
1183         spin_unlock(ptl);
1184 alloc:
1185         if (transparent_hugepage_enabled(vma) &&
1186             !transparent_hugepage_debug_cow()) {
1187                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1188                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1189         } else
1190                 new_page = NULL;
1191
1192         if (unlikely(!new_page)) {
1193                 if (!page) {
1194                         split_huge_page_pmd(vma, address, pmd);
1195                         ret |= VM_FAULT_FALLBACK;
1196                 } else {
1197                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1198                                         pmd, orig_pmd, page, haddr);
1199                         if (ret & VM_FAULT_OOM) {
1200                                 split_huge_page(page);
1201                                 ret |= VM_FAULT_FALLBACK;
1202                         }
1203                         put_user_huge_page(page);
1204                 }
1205                 count_vm_event(THP_FAULT_FALLBACK);
1206                 goto out;
1207         }
1208
1209         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1210                 put_page(new_page);
1211                 if (page) {
1212                         split_huge_page(page);
1213                         put_user_huge_page(page);
1214                 } else
1215                         split_huge_page_pmd(vma, address, pmd);
1216                 ret |= VM_FAULT_FALLBACK;
1217                 count_vm_event(THP_FAULT_FALLBACK);
1218                 goto out;
1219         }
1220
1221         count_vm_event(THP_FAULT_ALLOC);
1222
1223         if (!page)
1224                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1225         else
1226                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1227         __SetPageUptodate(new_page);
1228
1229         mmun_start = haddr;
1230         mmun_end   = haddr + HPAGE_PMD_SIZE;
1231         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1232
1233         spin_lock(ptl);
1234         if (page)
1235                 put_user_huge_page(page);
1236         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1237                 spin_unlock(ptl);
1238                 mem_cgroup_cancel_charge(new_page, memcg);
1239                 put_page(new_page);
1240                 goto out_mn;
1241         } else {
1242                 pmd_t entry;
1243                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1244                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1245                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1246                 page_add_new_anon_rmap(new_page, vma, haddr);
1247                 mem_cgroup_commit_charge(new_page, memcg, false);
1248                 lru_cache_add_active_or_unevictable(new_page, vma);
1249                 set_pmd_at(mm, haddr, pmd, entry);
1250                 update_mmu_cache_pmd(vma, address, pmd);
1251                 if (!page) {
1252                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1253                         put_huge_zero_page();
1254                 } else {
1255                         VM_BUG_ON_PAGE(!PageHead(page), page);
1256                         page_remove_rmap(page);
1257                         put_page(page);
1258                 }
1259                 ret |= VM_FAULT_WRITE;
1260         }
1261         spin_unlock(ptl);
1262 out_mn:
1263         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1264 out:
1265         return ret;
1266 out_unlock:
1267         spin_unlock(ptl);
1268         return ret;
1269 }
1270
1271 /*
1272  * FOLL_FORCE can write to even unwritable pmd's, but only
1273  * after we've gone through a COW cycle and they are dirty.
1274  */
1275 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1276 {
1277         return pmd_write(pmd) ||
1278                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1279 }
1280
1281 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1282                                    unsigned long addr,
1283                                    pmd_t *pmd,
1284                                    unsigned int flags)
1285 {
1286         struct mm_struct *mm = vma->vm_mm;
1287         struct page *page = NULL;
1288
1289         assert_spin_locked(pmd_lockptr(mm, pmd));
1290
1291         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1292                 goto out;
1293
1294         /* Avoid dumping huge zero page */
1295         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1296                 return ERR_PTR(-EFAULT);
1297
1298         /* Full NUMA hinting faults to serialise migration in fault paths */
1299         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1300                 goto out;
1301
1302         page = pmd_page(*pmd);
1303         VM_BUG_ON_PAGE(!PageHead(page), page);
1304         if (flags & FOLL_TOUCH) {
1305                 pmd_t _pmd;
1306                 _pmd = pmd_mkyoung(*pmd);
1307                 if (flags & FOLL_WRITE)
1308                         _pmd = pmd_mkdirty(_pmd);
1309                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1310                                           pmd, _pmd, flags & FOLL_WRITE))
1311                         update_mmu_cache_pmd(vma, addr, pmd);
1312         }
1313         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1314                 if (page->mapping && trylock_page(page)) {
1315                         lru_add_drain();
1316                         if (page->mapping)
1317                                 mlock_vma_page(page);
1318                         unlock_page(page);
1319                 }
1320         }
1321         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1322         VM_BUG_ON_PAGE(!PageCompound(page), page);
1323         if (flags & FOLL_GET)
1324                 get_page_foll(page);
1325
1326 out:
1327         return page;
1328 }
1329
1330 /* NUMA hinting page fault entry point for trans huge pmds */
1331 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1332                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1333 {
1334         spinlock_t *ptl;
1335         struct anon_vma *anon_vma = NULL;
1336         struct page *page;
1337         unsigned long haddr = addr & HPAGE_PMD_MASK;
1338         int page_nid = -1, this_nid = numa_node_id();
1339         int target_nid, last_cpupid = -1;
1340         bool page_locked;
1341         bool migrated = false;
1342         bool was_writable;
1343         int flags = 0;
1344
1345         /* A PROT_NONE fault should not end up here */
1346         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1347
1348         ptl = pmd_lock(mm, pmdp);
1349         if (unlikely(!pmd_same(pmd, *pmdp)))
1350                 goto out_unlock;
1351
1352         /*
1353          * If there are potential migrations, wait for completion and retry
1354          * without disrupting NUMA hinting information. Do not relock and
1355          * check_same as the page may no longer be mapped.
1356          */
1357         if (unlikely(pmd_trans_migrating(*pmdp))) {
1358                 page = pmd_page(*pmdp);
1359                 if (!get_page_unless_zero(page))
1360                         goto out_unlock;
1361                 spin_unlock(ptl);
1362                 wait_on_page_locked(page);
1363                 put_page(page);
1364                 goto out;
1365         }
1366
1367         page = pmd_page(pmd);
1368         BUG_ON(is_huge_zero_page(page));
1369         page_nid = page_to_nid(page);
1370         last_cpupid = page_cpupid_last(page);
1371         count_vm_numa_event(NUMA_HINT_FAULTS);
1372         if (page_nid == this_nid) {
1373                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1374                 flags |= TNF_FAULT_LOCAL;
1375         }
1376
1377         /* See similar comment in do_numa_page for explanation */
1378         if (!(vma->vm_flags & VM_WRITE))
1379                 flags |= TNF_NO_GROUP;
1380
1381         /*
1382          * Acquire the page lock to serialise THP migrations but avoid dropping
1383          * page_table_lock if at all possible
1384          */
1385         page_locked = trylock_page(page);
1386         target_nid = mpol_misplaced(page, vma, haddr);
1387         if (target_nid == -1) {
1388                 /* If the page was locked, there are no parallel migrations */
1389                 if (page_locked)
1390                         goto clear_pmdnuma;
1391         }
1392
1393         /* Migration could have started since the pmd_trans_migrating check */
1394         if (!page_locked) {
1395                 page_nid = -1;
1396                 if (!get_page_unless_zero(page))
1397                         goto out_unlock;
1398                 spin_unlock(ptl);
1399                 wait_on_page_locked(page);
1400                 put_page(page);
1401                 goto out;
1402         }
1403
1404         /*
1405          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1406          * to serialises splits
1407          */
1408         get_page(page);
1409         spin_unlock(ptl);
1410         anon_vma = page_lock_anon_vma_read(page);
1411
1412         /* Confirm the PMD did not change while page_table_lock was released */
1413         spin_lock(ptl);
1414         if (unlikely(!pmd_same(pmd, *pmdp))) {
1415                 unlock_page(page);
1416                 put_page(page);
1417                 page_nid = -1;
1418                 goto out_unlock;
1419         }
1420
1421         /* Bail if we fail to protect against THP splits for any reason */
1422         if (unlikely(!anon_vma)) {
1423                 put_page(page);
1424                 page_nid = -1;
1425                 goto clear_pmdnuma;
1426         }
1427
1428         /*
1429          * Migrate the THP to the requested node, returns with page unlocked
1430          * and access rights restored.
1431          */
1432         spin_unlock(ptl);
1433         migrated = migrate_misplaced_transhuge_page(mm, vma,
1434                                 pmdp, pmd, addr, page, target_nid);
1435         if (migrated) {
1436                 flags |= TNF_MIGRATED;
1437                 page_nid = target_nid;
1438         } else
1439                 flags |= TNF_MIGRATE_FAIL;
1440
1441         goto out;
1442 clear_pmdnuma:
1443         BUG_ON(!PageLocked(page));
1444         was_writable = pmd_write(pmd);
1445         pmd = pmd_modify(pmd, vma->vm_page_prot);
1446         pmd = pmd_mkyoung(pmd);
1447         if (was_writable)
1448                 pmd = pmd_mkwrite(pmd);
1449         set_pmd_at(mm, haddr, pmdp, pmd);
1450         update_mmu_cache_pmd(vma, addr, pmdp);
1451         unlock_page(page);
1452 out_unlock:
1453         spin_unlock(ptl);
1454
1455 out:
1456         if (anon_vma)
1457                 page_unlock_anon_vma_read(anon_vma);
1458
1459         if (page_nid != -1)
1460                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1461
1462         return 0;
1463 }
1464
1465 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1466                  pmd_t *pmd, unsigned long addr)
1467 {
1468         pmd_t orig_pmd;
1469         spinlock_t *ptl;
1470
1471         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1472                 return 0;
1473         /*
1474          * For architectures like ppc64 we look at deposited pgtable
1475          * when calling pmdp_huge_get_and_clear. So do the
1476          * pgtable_trans_huge_withdraw after finishing pmdp related
1477          * operations.
1478          */
1479         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1480                         tlb->fullmm);
1481         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1482         if (vma_is_dax(vma)) {
1483                 spin_unlock(ptl);
1484                 if (is_huge_zero_pmd(orig_pmd))
1485                         put_huge_zero_page();
1486         } else if (is_huge_zero_pmd(orig_pmd)) {
1487                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1488                 atomic_long_dec(&tlb->mm->nr_ptes);
1489                 spin_unlock(ptl);
1490                 put_huge_zero_page();
1491         } else {
1492                 struct page *page = pmd_page(orig_pmd);
1493                 page_remove_rmap(page);
1494                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1495                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1496                 VM_BUG_ON_PAGE(!PageHead(page), page);
1497                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1498                 atomic_long_dec(&tlb->mm->nr_ptes);
1499                 spin_unlock(ptl);
1500                 tlb_remove_page(tlb, page);
1501         }
1502         return 1;
1503 }
1504
1505 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1506                   unsigned long old_addr,
1507                   unsigned long new_addr, unsigned long old_end,
1508                   pmd_t *old_pmd, pmd_t *new_pmd)
1509 {
1510         spinlock_t *old_ptl, *new_ptl;
1511         int ret = 0;
1512         pmd_t pmd;
1513         bool force_flush = false;
1514         struct mm_struct *mm = vma->vm_mm;
1515
1516         if ((old_addr & ~HPAGE_PMD_MASK) ||
1517             (new_addr & ~HPAGE_PMD_MASK) ||
1518             old_end - old_addr < HPAGE_PMD_SIZE ||
1519             (new_vma->vm_flags & VM_NOHUGEPAGE))
1520                 goto out;
1521
1522         /*
1523          * The destination pmd shouldn't be established, free_pgtables()
1524          * should have release it.
1525          */
1526         if (WARN_ON(!pmd_none(*new_pmd))) {
1527                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1528                 goto out;
1529         }
1530
1531         /*
1532          * We don't have to worry about the ordering of src and dst
1533          * ptlocks because exclusive mmap_sem prevents deadlock.
1534          */
1535         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1536         if (ret == 1) {
1537                 new_ptl = pmd_lockptr(mm, new_pmd);
1538                 if (new_ptl != old_ptl)
1539                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1540                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1541                 if (pmd_present(pmd))
1542                         force_flush = true;
1543                 VM_BUG_ON(!pmd_none(*new_pmd));
1544
1545                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1546                         pgtable_t pgtable;
1547                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1548                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1549                 }
1550                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1551                 if (force_flush)
1552                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1553                 if (new_ptl != old_ptl)
1554                         spin_unlock(new_ptl);
1555                 spin_unlock(old_ptl);
1556         }
1557 out:
1558         return ret;
1559 }
1560
1561 /*
1562  * Returns
1563  *  - 0 if PMD could not be locked
1564  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1565  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1566  */
1567 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1568                 unsigned long addr, pgprot_t newprot, int prot_numa)
1569 {
1570         struct mm_struct *mm = vma->vm_mm;
1571         spinlock_t *ptl;
1572         pmd_t entry;
1573         bool preserve_write;
1574
1575         int ret = 0;
1576
1577         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1578                 return 0;
1579
1580         preserve_write = prot_numa && pmd_write(*pmd);
1581         ret = 1;
1582
1583         /*
1584          * Avoid trapping faults against the zero page. The read-only
1585          * data is likely to be read-cached on the local CPU and
1586          * local/remote hits to the zero page are not interesting.
1587          */
1588         if (prot_numa && is_huge_zero_pmd(*pmd))
1589                 goto unlock;
1590
1591         if (prot_numa && pmd_protnone(*pmd))
1592                 goto unlock;
1593
1594         /*
1595          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1596          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1597          * which is also under down_read(mmap_sem):
1598          *
1599          *      CPU0:                           CPU1:
1600          *                              change_huge_pmd(prot_numa=1)
1601          *                               pmdp_huge_get_and_clear_notify()
1602          * madvise_dontneed()
1603          *  zap_pmd_range()
1604          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1605          *   // skip the pmd
1606          *                               set_pmd_at();
1607          *                               // pmd is re-established
1608          *
1609          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1610          * which may break userspace.
1611          *
1612          * pmdp_invalidate() is required to make sure we don't miss
1613          * dirty/young flags set by hardware.
1614          */
1615         entry = *pmd;
1616         pmdp_invalidate(vma, addr, pmd);
1617
1618         /*
1619          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1620          * corrupt them.
1621          */
1622         if (pmd_dirty(*pmd))
1623                 entry = pmd_mkdirty(entry);
1624         if (pmd_young(*pmd))
1625                 entry = pmd_mkyoung(entry);
1626
1627         entry = pmd_modify(entry, newprot);
1628         if (preserve_write)
1629                 entry = pmd_mkwrite(entry);
1630         ret = HPAGE_PMD_NR;
1631         set_pmd_at(mm, addr, pmd, entry);
1632         BUG_ON(!preserve_write && pmd_write(entry));
1633 unlock:
1634         spin_unlock(ptl);
1635         return ret;
1636 }
1637
1638 /*
1639  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1640  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1641  *
1642  * Note that if it returns 1, this routine returns without unlocking page
1643  * table locks. So callers must unlock them.
1644  */
1645 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1646                 spinlock_t **ptl)
1647 {
1648         *ptl = pmd_lock(vma->vm_mm, pmd);
1649         if (likely(pmd_trans_huge(*pmd))) {
1650                 if (unlikely(pmd_trans_splitting(*pmd))) {
1651                         spin_unlock(*ptl);
1652                         wait_split_huge_page(vma->anon_vma, pmd);
1653                         return -1;
1654                 } else {
1655                         /* Thp mapped by 'pmd' is stable, so we can
1656                          * handle it as it is. */
1657                         return 1;
1658                 }
1659         }
1660         spin_unlock(*ptl);
1661         return 0;
1662 }
1663
1664 /*
1665  * This function returns whether a given @page is mapped onto the @address
1666  * in the virtual space of @mm.
1667  *
1668  * When it's true, this function returns *pmd with holding the page table lock
1669  * and passing it back to the caller via @ptl.
1670  * If it's false, returns NULL without holding the page table lock.
1671  */
1672 pmd_t *page_check_address_pmd(struct page *page,
1673                               struct mm_struct *mm,
1674                               unsigned long address,
1675                               enum page_check_address_pmd_flag flag,
1676                               spinlock_t **ptl)
1677 {
1678         pgd_t *pgd;
1679         pud_t *pud;
1680         pmd_t *pmd;
1681
1682         if (address & ~HPAGE_PMD_MASK)
1683                 return NULL;
1684
1685         pgd = pgd_offset(mm, address);
1686         if (!pgd_present(*pgd))
1687                 return NULL;
1688         pud = pud_offset(pgd, address);
1689         if (!pud_present(*pud))
1690                 return NULL;
1691         pmd = pmd_offset(pud, address);
1692
1693         *ptl = pmd_lock(mm, pmd);
1694         if (!pmd_present(*pmd))
1695                 goto unlock;
1696         if (pmd_page(*pmd) != page)
1697                 goto unlock;
1698         /*
1699          * split_vma() may create temporary aliased mappings. There is
1700          * no risk as long as all huge pmd are found and have their
1701          * splitting bit set before __split_huge_page_refcount
1702          * runs. Finding the same huge pmd more than once during the
1703          * same rmap walk is not a problem.
1704          */
1705         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1706             pmd_trans_splitting(*pmd))
1707                 goto unlock;
1708         if (pmd_trans_huge(*pmd)) {
1709                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1710                           !pmd_trans_splitting(*pmd));
1711                 return pmd;
1712         }
1713 unlock:
1714         spin_unlock(*ptl);
1715         return NULL;
1716 }
1717
1718 static int __split_huge_page_splitting(struct page *page,
1719                                        struct vm_area_struct *vma,
1720                                        unsigned long address)
1721 {
1722         struct mm_struct *mm = vma->vm_mm;
1723         spinlock_t *ptl;
1724         pmd_t *pmd;
1725         int ret = 0;
1726         /* For mmu_notifiers */
1727         const unsigned long mmun_start = address;
1728         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1729
1730         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1731         pmd = page_check_address_pmd(page, mm, address,
1732                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1733         if (pmd) {
1734                 /*
1735                  * We can't temporarily set the pmd to null in order
1736                  * to split it, the pmd must remain marked huge at all
1737                  * times or the VM won't take the pmd_trans_huge paths
1738                  * and it won't wait on the anon_vma->root->rwsem to
1739                  * serialize against split_huge_page*.
1740                  */
1741                 pmdp_splitting_flush(vma, address, pmd);
1742
1743                 ret = 1;
1744                 spin_unlock(ptl);
1745         }
1746         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1747
1748         return ret;
1749 }
1750
1751 static void __split_huge_page_refcount(struct page *page,
1752                                        struct list_head *list)
1753 {
1754         int i;
1755         struct zone *zone = page_zone(page);
1756         struct lruvec *lruvec;
1757         int tail_count = 0;
1758
1759         /* prevent PageLRU to go away from under us, and freeze lru stats */
1760         spin_lock_irq(&zone->lru_lock);
1761         lruvec = mem_cgroup_page_lruvec(page, zone);
1762
1763         compound_lock(page);
1764         /* complete memcg works before add pages to LRU */
1765         mem_cgroup_split_huge_fixup(page);
1766
1767         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1768                 struct page *page_tail = page + i;
1769
1770                 /* tail_page->_mapcount cannot change */
1771                 BUG_ON(page_mapcount(page_tail) < 0);
1772                 tail_count += page_mapcount(page_tail);
1773                 /* check for overflow */
1774                 BUG_ON(tail_count < 0);
1775                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1776                 /*
1777                  * tail_page->_count is zero and not changing from
1778                  * under us. But get_page_unless_zero() may be running
1779                  * from under us on the tail_page. If we used
1780                  * atomic_set() below instead of atomic_add(), we
1781                  * would then run atomic_set() concurrently with
1782                  * get_page_unless_zero(), and atomic_set() is
1783                  * implemented in C not using locked ops. spin_unlock
1784                  * on x86 sometime uses locked ops because of PPro
1785                  * errata 66, 92, so unless somebody can guarantee
1786                  * atomic_set() here would be safe on all archs (and
1787                  * not only on x86), it's safer to use atomic_add().
1788                  */
1789                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1790                            &page_tail->_count);
1791
1792                 /* after clearing PageTail the gup refcount can be released */
1793                 smp_mb__after_atomic();
1794
1795                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1796                 page_tail->flags |= (page->flags &
1797                                      ((1L << PG_referenced) |
1798                                       (1L << PG_swapbacked) |
1799                                       (1L << PG_mlocked) |
1800                                       (1L << PG_uptodate) |
1801                                       (1L << PG_active) |
1802                                       (1L << PG_unevictable)));
1803                 page_tail->flags |= (1L << PG_dirty);
1804
1805                 clear_compound_head(page_tail);
1806
1807                 if (page_is_young(page))
1808                         set_page_young(page_tail);
1809                 if (page_is_idle(page))
1810                         set_page_idle(page_tail);
1811
1812                 /*
1813                  * __split_huge_page_splitting() already set the
1814                  * splitting bit in all pmd that could map this
1815                  * hugepage, that will ensure no CPU can alter the
1816                  * mapcount on the head page. The mapcount is only
1817                  * accounted in the head page and it has to be
1818                  * transferred to all tail pages in the below code. So
1819                  * for this code to be safe, the split the mapcount
1820                  * can't change. But that doesn't mean userland can't
1821                  * keep changing and reading the page contents while
1822                  * we transfer the mapcount, so the pmd splitting
1823                  * status is achieved setting a reserved bit in the
1824                  * pmd, not by clearing the present bit.
1825                 */
1826                 page_tail->_mapcount = page->_mapcount;
1827
1828                 BUG_ON(page_tail->mapping);
1829                 page_tail->mapping = page->mapping;
1830
1831                 page_tail->index = page->index + i;
1832                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1833
1834                 BUG_ON(!PageAnon(page_tail));
1835                 BUG_ON(!PageUptodate(page_tail));
1836                 BUG_ON(!PageDirty(page_tail));
1837                 BUG_ON(!PageSwapBacked(page_tail));
1838
1839                 lru_add_page_tail(page, page_tail, lruvec, list);
1840         }
1841         atomic_sub(tail_count, &page->_count);
1842         BUG_ON(atomic_read(&page->_count) <= 0);
1843
1844         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1845
1846         ClearPageCompound(page);
1847         compound_unlock(page);
1848         spin_unlock_irq(&zone->lru_lock);
1849
1850         for (i = 1; i < HPAGE_PMD_NR; i++) {
1851                 struct page *page_tail = page + i;
1852                 BUG_ON(page_count(page_tail) <= 0);
1853                 /*
1854                  * Tail pages may be freed if there wasn't any mapping
1855                  * like if add_to_swap() is running on a lru page that
1856                  * had its mapping zapped. And freeing these pages
1857                  * requires taking the lru_lock so we do the put_page
1858                  * of the tail pages after the split is complete.
1859                  */
1860                 put_page(page_tail);
1861         }
1862
1863         /*
1864          * Only the head page (now become a regular page) is required
1865          * to be pinned by the caller.
1866          */
1867         BUG_ON(page_count(page) <= 0);
1868 }
1869
1870 static int __split_huge_page_map(struct page *page,
1871                                  struct vm_area_struct *vma,
1872                                  unsigned long address)
1873 {
1874         struct mm_struct *mm = vma->vm_mm;
1875         spinlock_t *ptl;
1876         pmd_t *pmd, _pmd;
1877         int ret = 0, i;
1878         pgtable_t pgtable;
1879         unsigned long haddr;
1880
1881         pmd = page_check_address_pmd(page, mm, address,
1882                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1883         if (pmd) {
1884                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1885                 pmd_populate(mm, &_pmd, pgtable);
1886                 if (pmd_write(*pmd))
1887                         BUG_ON(page_mapcount(page) != 1);
1888
1889                 haddr = address;
1890                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1891                         pte_t *pte, entry;
1892                         BUG_ON(PageCompound(page+i));
1893                         /*
1894                          * Note that NUMA hinting access restrictions are not
1895                          * transferred to avoid any possibility of altering
1896                          * permissions across VMAs.
1897                          */
1898                         entry = mk_pte(page + i, vma->vm_page_prot);
1899                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1900                         if (!pmd_write(*pmd))
1901                                 entry = pte_wrprotect(entry);
1902                         if (!pmd_young(*pmd))
1903                                 entry = pte_mkold(entry);
1904                         pte = pte_offset_map(&_pmd, haddr);
1905                         BUG_ON(!pte_none(*pte));
1906                         set_pte_at(mm, haddr, pte, entry);
1907                         pte_unmap(pte);
1908                 }
1909
1910                 smp_wmb(); /* make pte visible before pmd */
1911                 /*
1912                  * Up to this point the pmd is present and huge and
1913                  * userland has the whole access to the hugepage
1914                  * during the split (which happens in place). If we
1915                  * overwrite the pmd with the not-huge version
1916                  * pointing to the pte here (which of course we could
1917                  * if all CPUs were bug free), userland could trigger
1918                  * a small page size TLB miss on the small sized TLB
1919                  * while the hugepage TLB entry is still established
1920                  * in the huge TLB. Some CPU doesn't like that. See
1921                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1922                  * Erratum 383 on page 93. Intel should be safe but is
1923                  * also warns that it's only safe if the permission
1924                  * and cache attributes of the two entries loaded in
1925                  * the two TLB is identical (which should be the case
1926                  * here). But it is generally safer to never allow
1927                  * small and huge TLB entries for the same virtual
1928                  * address to be loaded simultaneously. So instead of
1929                  * doing "pmd_populate(); flush_pmd_tlb_range();" we first
1930                  * mark the current pmd notpresent (atomically because
1931                  * here the pmd_trans_huge and pmd_trans_splitting
1932                  * must remain set at all times on the pmd until the
1933                  * split is complete for this pmd), then we flush the
1934                  * SMP TLB and finally we write the non-huge version
1935                  * of the pmd entry with pmd_populate.
1936                  */
1937                 pmdp_invalidate(vma, address, pmd);
1938                 pmd_populate(mm, pmd, pgtable);
1939                 ret = 1;
1940                 spin_unlock(ptl);
1941         }
1942
1943         return ret;
1944 }
1945
1946 /* must be called with anon_vma->root->rwsem held */
1947 static void __split_huge_page(struct page *page,
1948                               struct anon_vma *anon_vma,
1949                               struct list_head *list)
1950 {
1951         int mapcount, mapcount2;
1952         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1953         struct anon_vma_chain *avc;
1954
1955         BUG_ON(!PageHead(page));
1956         BUG_ON(PageTail(page));
1957
1958         mapcount = 0;
1959         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1960                 struct vm_area_struct *vma = avc->vma;
1961                 unsigned long addr = vma_address(page, vma);
1962                 BUG_ON(is_vma_temporary_stack(vma));
1963                 mapcount += __split_huge_page_splitting(page, vma, addr);
1964         }
1965         /*
1966          * It is critical that new vmas are added to the tail of the
1967          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1968          * and establishes a child pmd before
1969          * __split_huge_page_splitting() freezes the parent pmd (so if
1970          * we fail to prevent copy_huge_pmd() from running until the
1971          * whole __split_huge_page() is complete), we will still see
1972          * the newly established pmd of the child later during the
1973          * walk, to be able to set it as pmd_trans_splitting too.
1974          */
1975         if (mapcount != page_mapcount(page)) {
1976                 pr_err("mapcount %d page_mapcount %d\n",
1977                         mapcount, page_mapcount(page));
1978                 BUG();
1979         }
1980
1981         __split_huge_page_refcount(page, list);
1982
1983         mapcount2 = 0;
1984         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1985                 struct vm_area_struct *vma = avc->vma;
1986                 unsigned long addr = vma_address(page, vma);
1987                 BUG_ON(is_vma_temporary_stack(vma));
1988                 mapcount2 += __split_huge_page_map(page, vma, addr);
1989         }
1990         if (mapcount != mapcount2) {
1991                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1992                         mapcount, mapcount2, page_mapcount(page));
1993                 BUG();
1994         }
1995 }
1996
1997 /*
1998  * Split a hugepage into normal pages. This doesn't change the position of head
1999  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
2000  * @list. Both head page and tail pages will inherit mapping, flags, and so on
2001  * from the hugepage.
2002  * Return 0 if the hugepage is split successfully otherwise return 1.
2003  */
2004 int split_huge_page_to_list(struct page *page, struct list_head *list)
2005 {
2006         struct anon_vma *anon_vma;
2007         int ret = 1;
2008
2009         BUG_ON(is_huge_zero_page(page));
2010         BUG_ON(!PageAnon(page));
2011
2012         /*
2013          * The caller does not necessarily hold an mmap_sem that would prevent
2014          * the anon_vma disappearing so we first we take a reference to it
2015          * and then lock the anon_vma for write. This is similar to
2016          * page_lock_anon_vma_read except the write lock is taken to serialise
2017          * against parallel split or collapse operations.
2018          */
2019         anon_vma = page_get_anon_vma(page);
2020         if (!anon_vma)
2021                 goto out;
2022         anon_vma_lock_write(anon_vma);
2023
2024         ret = 0;
2025         if (!PageCompound(page))
2026                 goto out_unlock;
2027
2028         BUG_ON(!PageSwapBacked(page));
2029         __split_huge_page(page, anon_vma, list);
2030         count_vm_event(THP_SPLIT);
2031
2032         BUG_ON(PageCompound(page));
2033 out_unlock:
2034         anon_vma_unlock_write(anon_vma);
2035         put_anon_vma(anon_vma);
2036 out:
2037         return ret;
2038 }
2039
2040 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
2041
2042 int hugepage_madvise(struct vm_area_struct *vma,
2043                      unsigned long *vm_flags, int advice)
2044 {
2045         switch (advice) {
2046         case MADV_HUGEPAGE:
2047 #ifdef CONFIG_S390
2048                 /*
2049                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2050                  * can't handle this properly after s390_enable_sie, so we simply
2051                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
2052                  */
2053                 if (mm_has_pgste(vma->vm_mm))
2054                         return 0;
2055 #endif
2056                 /*
2057                  * Be somewhat over-protective like KSM for now!
2058                  */
2059                 if (*vm_flags & VM_NO_THP)
2060                         return -EINVAL;
2061                 *vm_flags &= ~VM_NOHUGEPAGE;
2062                 *vm_flags |= VM_HUGEPAGE;
2063                 /*
2064                  * If the vma become good for khugepaged to scan,
2065                  * register it here without waiting a page fault that
2066                  * may not happen any time soon.
2067                  */
2068                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2069                         return -ENOMEM;
2070                 break;
2071         case MADV_NOHUGEPAGE:
2072                 /*
2073                  * Be somewhat over-protective like KSM for now!
2074                  */
2075                 if (*vm_flags & VM_NO_THP)
2076                         return -EINVAL;
2077                 *vm_flags &= ~VM_HUGEPAGE;
2078                 *vm_flags |= VM_NOHUGEPAGE;
2079                 /*
2080                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2081                  * this vma even if we leave the mm registered in khugepaged if
2082                  * it got registered before VM_NOHUGEPAGE was set.
2083                  */
2084                 break;
2085         }
2086
2087         return 0;
2088 }
2089
2090 static int __init khugepaged_slab_init(void)
2091 {
2092         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2093                                           sizeof(struct mm_slot),
2094                                           __alignof__(struct mm_slot), 0, NULL);
2095         if (!mm_slot_cache)
2096                 return -ENOMEM;
2097
2098         return 0;
2099 }
2100
2101 static void __init khugepaged_slab_exit(void)
2102 {
2103         kmem_cache_destroy(mm_slot_cache);
2104 }
2105
2106 static inline struct mm_slot *alloc_mm_slot(void)
2107 {
2108         if (!mm_slot_cache)     /* initialization failed */
2109                 return NULL;
2110         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2111 }
2112
2113 static inline void free_mm_slot(struct mm_slot *mm_slot)
2114 {
2115         kmem_cache_free(mm_slot_cache, mm_slot);
2116 }
2117
2118 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2119 {
2120         struct mm_slot *mm_slot;
2121
2122         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2123                 if (mm == mm_slot->mm)
2124                         return mm_slot;
2125
2126         return NULL;
2127 }
2128
2129 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2130                                     struct mm_slot *mm_slot)
2131 {
2132         mm_slot->mm = mm;
2133         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2134 }
2135
2136 static inline int khugepaged_test_exit(struct mm_struct *mm)
2137 {
2138         return atomic_read(&mm->mm_users) == 0;
2139 }
2140
2141 int __khugepaged_enter(struct mm_struct *mm)
2142 {
2143         struct mm_slot *mm_slot;
2144         int wakeup;
2145
2146         mm_slot = alloc_mm_slot();
2147         if (!mm_slot)
2148                 return -ENOMEM;
2149
2150         /* __khugepaged_exit() must not run from under us */
2151         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2152         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2153                 free_mm_slot(mm_slot);
2154                 return 0;
2155         }
2156
2157         spin_lock(&khugepaged_mm_lock);
2158         insert_to_mm_slots_hash(mm, mm_slot);
2159         /*
2160          * Insert just behind the scanning cursor, to let the area settle
2161          * down a little.
2162          */
2163         wakeup = list_empty(&khugepaged_scan.mm_head);
2164         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2165         spin_unlock(&khugepaged_mm_lock);
2166
2167         atomic_inc(&mm->mm_count);
2168         if (wakeup)
2169                 wake_up_interruptible(&khugepaged_wait);
2170
2171         return 0;
2172 }
2173
2174 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2175                                unsigned long vm_flags)
2176 {
2177         unsigned long hstart, hend;
2178         if (!vma->anon_vma)
2179                 /*
2180                  * Not yet faulted in so we will register later in the
2181                  * page fault if needed.
2182                  */
2183                 return 0;
2184         if (vma->vm_ops || (vm_flags & VM_NO_THP))
2185                 /* khugepaged not yet working on file or special mappings */
2186                 return 0;
2187         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2188         hend = vma->vm_end & HPAGE_PMD_MASK;
2189         if (hstart < hend)
2190                 return khugepaged_enter(vma, vm_flags);
2191         return 0;
2192 }
2193
2194 void __khugepaged_exit(struct mm_struct *mm)
2195 {
2196         struct mm_slot *mm_slot;
2197         int free = 0;
2198
2199         spin_lock(&khugepaged_mm_lock);
2200         mm_slot = get_mm_slot(mm);
2201         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2202                 hash_del(&mm_slot->hash);
2203                 list_del(&mm_slot->mm_node);
2204                 free = 1;
2205         }
2206         spin_unlock(&khugepaged_mm_lock);
2207
2208         if (free) {
2209                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2210                 free_mm_slot(mm_slot);
2211                 mmdrop(mm);
2212         } else if (mm_slot) {
2213                 /*
2214                  * This is required to serialize against
2215                  * khugepaged_test_exit() (which is guaranteed to run
2216                  * under mmap sem read mode). Stop here (after we
2217                  * return all pagetables will be destroyed) until
2218                  * khugepaged has finished working on the pagetables
2219                  * under the mmap_sem.
2220                  */
2221                 down_write(&mm->mmap_sem);
2222                 up_write(&mm->mmap_sem);
2223         }
2224 }
2225
2226 static void release_pte_page(struct page *page)
2227 {
2228         /* 0 stands for page_is_file_cache(page) == false */
2229         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2230         unlock_page(page);
2231         putback_lru_page(page);
2232 }
2233
2234 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2235 {
2236         while (--_pte >= pte) {
2237                 pte_t pteval = *_pte;
2238                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2239                         release_pte_page(pte_page(pteval));
2240         }
2241 }
2242
2243 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2244                                         unsigned long address,
2245                                         pte_t *pte)
2246 {
2247         struct page *page;
2248         pte_t *_pte;
2249         int none_or_zero = 0;
2250         bool referenced = false, writable = false;
2251         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2252              _pte++, address += PAGE_SIZE) {
2253                 pte_t pteval = *_pte;
2254                 if (pte_none(pteval) || (pte_present(pteval) &&
2255                                 is_zero_pfn(pte_pfn(pteval)))) {
2256                         if (!userfaultfd_armed(vma) &&
2257                             ++none_or_zero <= khugepaged_max_ptes_none)
2258                                 continue;
2259                         else
2260                                 goto out;
2261                 }
2262                 if (!pte_present(pteval))
2263                         goto out;
2264                 page = vm_normal_page(vma, address, pteval);
2265                 if (unlikely(!page))
2266                         goto out;
2267
2268                 VM_BUG_ON_PAGE(PageCompound(page), page);
2269                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2270                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2271
2272                 /*
2273                  * We can do it before isolate_lru_page because the
2274                  * page can't be freed from under us. NOTE: PG_lock
2275                  * is needed to serialize against split_huge_page
2276                  * when invoked from the VM.
2277                  */
2278                 if (!trylock_page(page))
2279                         goto out;
2280
2281                 /*
2282                  * cannot use mapcount: can't collapse if there's a gup pin.
2283                  * The page must only be referenced by the scanned process
2284                  * and page swap cache.
2285                  */
2286                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2287                         unlock_page(page);
2288                         goto out;
2289                 }
2290                 if (pte_write(pteval)) {
2291                         writable = true;
2292                 } else {
2293                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2294                                 unlock_page(page);
2295                                 goto out;
2296                         }
2297                         /*
2298                          * Page is not in the swap cache. It can be collapsed
2299                          * into a THP.
2300                          */
2301                 }
2302
2303                 /*
2304                  * Isolate the page to avoid collapsing an hugepage
2305                  * currently in use by the VM.
2306                  */
2307                 if (isolate_lru_page(page)) {
2308                         unlock_page(page);
2309                         goto out;
2310                 }
2311                 /* 0 stands for page_is_file_cache(page) == false */
2312                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2313                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2314                 VM_BUG_ON_PAGE(PageLRU(page), page);
2315
2316                 /* If there is no mapped pte young don't collapse the page */
2317                 if (pte_young(pteval) ||
2318                     page_is_young(page) || PageReferenced(page) ||
2319                     mmu_notifier_test_young(vma->vm_mm, address))
2320                         referenced = true;
2321         }
2322         if (likely(referenced && writable))
2323                 return 1;
2324 out:
2325         release_pte_pages(pte, _pte);
2326         return 0;
2327 }
2328
2329 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2330                                       struct vm_area_struct *vma,
2331                                       unsigned long address,
2332                                       spinlock_t *ptl)
2333 {
2334         pte_t *_pte;
2335         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2336                 pte_t pteval = *_pte;
2337                 struct page *src_page;
2338
2339                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2340                         clear_user_highpage(page, address);
2341                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2342                         if (is_zero_pfn(pte_pfn(pteval))) {
2343                                 /*
2344                                  * ptl mostly unnecessary.
2345                                  */
2346                                 spin_lock(ptl);
2347                                 /*
2348                                  * paravirt calls inside pte_clear here are
2349                                  * superfluous.
2350                                  */
2351                                 pte_clear(vma->vm_mm, address, _pte);
2352                                 spin_unlock(ptl);
2353                         }
2354                 } else {
2355                         src_page = pte_page(pteval);
2356                         copy_user_highpage(page, src_page, address, vma);
2357                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2358                         release_pte_page(src_page);
2359                         /*
2360                          * ptl mostly unnecessary, but preempt has to
2361                          * be disabled to update the per-cpu stats
2362                          * inside page_remove_rmap().
2363                          */
2364                         spin_lock(ptl);
2365                         /*
2366                          * paravirt calls inside pte_clear here are
2367                          * superfluous.
2368                          */
2369                         pte_clear(vma->vm_mm, address, _pte);
2370                         page_remove_rmap(src_page);
2371                         spin_unlock(ptl);
2372                         free_page_and_swap_cache(src_page);
2373                 }
2374
2375                 address += PAGE_SIZE;
2376                 page++;
2377         }
2378 }
2379
2380 static void khugepaged_alloc_sleep(void)
2381 {
2382         DEFINE_WAIT(wait);
2383
2384         add_wait_queue(&khugepaged_wait, &wait);
2385         freezable_schedule_timeout_interruptible(
2386                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2387         remove_wait_queue(&khugepaged_wait, &wait);
2388 }
2389
2390 static int khugepaged_node_load[MAX_NUMNODES];
2391
2392 static bool khugepaged_scan_abort(int nid)
2393 {
2394         int i;
2395
2396         /*
2397          * If zone_reclaim_mode is disabled, then no extra effort is made to
2398          * allocate memory locally.
2399          */
2400         if (!zone_reclaim_mode)
2401                 return false;
2402
2403         /* If there is a count for this node already, it must be acceptable */
2404         if (khugepaged_node_load[nid])
2405                 return false;
2406
2407         for (i = 0; i < MAX_NUMNODES; i++) {
2408                 if (!khugepaged_node_load[i])
2409                         continue;
2410                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2411                         return true;
2412         }
2413         return false;
2414 }
2415
2416 #ifdef CONFIG_NUMA
2417 static int khugepaged_find_target_node(void)
2418 {
2419         static int last_khugepaged_target_node = NUMA_NO_NODE;
2420         int nid, target_node = 0, max_value = 0;
2421
2422         /* find first node with max normal pages hit */
2423         for (nid = 0; nid < MAX_NUMNODES; nid++)
2424                 if (khugepaged_node_load[nid] > max_value) {
2425                         max_value = khugepaged_node_load[nid];
2426                         target_node = nid;
2427                 }
2428
2429         /* do some balance if several nodes have the same hit record */
2430         if (target_node <= last_khugepaged_target_node)
2431                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2432                                 nid++)
2433                         if (max_value == khugepaged_node_load[nid]) {
2434                                 target_node = nid;
2435                                 break;
2436                         }
2437
2438         last_khugepaged_target_node = target_node;
2439         return target_node;
2440 }
2441
2442 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2443 {
2444         if (IS_ERR(*hpage)) {
2445                 if (!*wait)
2446                         return false;
2447
2448                 *wait = false;
2449                 *hpage = NULL;
2450                 khugepaged_alloc_sleep();
2451         } else if (*hpage) {
2452                 put_page(*hpage);
2453                 *hpage = NULL;
2454         }
2455
2456         return true;
2457 }
2458
2459 static struct page *
2460 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2461                        unsigned long address, int node)
2462 {
2463         VM_BUG_ON_PAGE(*hpage, *hpage);
2464
2465         /*
2466          * Before allocating the hugepage, release the mmap_sem read lock.
2467          * The allocation can take potentially a long time if it involves
2468          * sync compaction, and we do not need to hold the mmap_sem during
2469          * that. We will recheck the vma after taking it again in write mode.
2470          */
2471         up_read(&mm->mmap_sem);
2472
2473         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2474         if (unlikely(!*hpage)) {
2475                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2476                 *hpage = ERR_PTR(-ENOMEM);
2477                 return NULL;
2478         }
2479
2480         count_vm_event(THP_COLLAPSE_ALLOC);
2481         return *hpage;
2482 }
2483 #else
2484 static int khugepaged_find_target_node(void)
2485 {
2486         return 0;
2487 }
2488
2489 static inline struct page *alloc_hugepage(int defrag)
2490 {
2491         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2492                            HPAGE_PMD_ORDER);
2493 }
2494
2495 static struct page *khugepaged_alloc_hugepage(bool *wait)
2496 {
2497         struct page *hpage;
2498
2499         do {
2500                 hpage = alloc_hugepage(khugepaged_defrag());
2501                 if (!hpage) {
2502                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2503                         if (!*wait)
2504                                 return NULL;
2505
2506                         *wait = false;
2507                         khugepaged_alloc_sleep();
2508                 } else
2509                         count_vm_event(THP_COLLAPSE_ALLOC);
2510         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2511
2512         return hpage;
2513 }
2514
2515 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2516 {
2517         if (!*hpage)
2518                 *hpage = khugepaged_alloc_hugepage(wait);
2519
2520         if (unlikely(!*hpage))
2521                 return false;
2522
2523         return true;
2524 }
2525
2526 static struct page *
2527 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2528                        unsigned long address, int node)
2529 {
2530         up_read(&mm->mmap_sem);
2531         VM_BUG_ON(!*hpage);
2532
2533         return  *hpage;
2534 }
2535 #endif
2536
2537 static bool hugepage_vma_check(struct vm_area_struct *vma)
2538 {
2539         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2540             (vma->vm_flags & VM_NOHUGEPAGE))
2541                 return false;
2542
2543         if (!vma->anon_vma || vma->vm_ops)
2544                 return false;
2545         if (is_vma_temporary_stack(vma))
2546                 return false;
2547         return !(vma->vm_flags & VM_NO_THP);
2548 }
2549
2550 static void collapse_huge_page(struct mm_struct *mm,
2551                                    unsigned long address,
2552                                    struct page **hpage,
2553                                    struct vm_area_struct *vma,
2554                                    int node)
2555 {
2556         pmd_t *pmd, _pmd;
2557         pte_t *pte;
2558         pgtable_t pgtable;
2559         struct page *new_page;
2560         spinlock_t *pmd_ptl, *pte_ptl;
2561         int isolated;
2562         unsigned long hstart, hend;
2563         struct mem_cgroup *memcg;
2564         unsigned long mmun_start;       /* For mmu_notifiers */
2565         unsigned long mmun_end;         /* For mmu_notifiers */
2566         gfp_t gfp;
2567
2568         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2569
2570         /* Only allocate from the target node */
2571         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2572                 __GFP_THISNODE;
2573
2574         /* release the mmap_sem read lock. */
2575         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2576         if (!new_page)
2577                 return;
2578
2579         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2580                                            gfp, &memcg)))
2581                 return;
2582
2583         /*
2584          * Prevent all access to pagetables with the exception of
2585          * gup_fast later hanlded by the ptep_clear_flush and the VM
2586          * handled by the anon_vma lock + PG_lock.
2587          */
2588         down_write(&mm->mmap_sem);
2589         if (unlikely(khugepaged_test_exit(mm)))
2590                 goto out;
2591
2592         vma = find_vma(mm, address);
2593         if (!vma)
2594                 goto out;
2595         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2596         hend = vma->vm_end & HPAGE_PMD_MASK;
2597         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2598                 goto out;
2599         if (!hugepage_vma_check(vma))
2600                 goto out;
2601         pmd = mm_find_pmd(mm, address);
2602         if (!pmd)
2603                 goto out;
2604
2605         anon_vma_lock_write(vma->anon_vma);
2606
2607         pte = pte_offset_map(pmd, address);
2608         pte_ptl = pte_lockptr(mm, pmd);
2609
2610         mmun_start = address;
2611         mmun_end   = address + HPAGE_PMD_SIZE;
2612         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2613         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2614         /*
2615          * After this gup_fast can't run anymore. This also removes
2616          * any huge TLB entry from the CPU so we won't allow
2617          * huge and small TLB entries for the same virtual address
2618          * to avoid the risk of CPU bugs in that area.
2619          */
2620         _pmd = pmdp_collapse_flush(vma, address, pmd);
2621         spin_unlock(pmd_ptl);
2622         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2623
2624         spin_lock(pte_ptl);
2625         isolated = __collapse_huge_page_isolate(vma, address, pte);
2626         spin_unlock(pte_ptl);
2627
2628         if (unlikely(!isolated)) {
2629                 pte_unmap(pte);
2630                 spin_lock(pmd_ptl);
2631                 BUG_ON(!pmd_none(*pmd));
2632                 /*
2633                  * We can only use set_pmd_at when establishing
2634                  * hugepmds and never for establishing regular pmds that
2635                  * points to regular pagetables. Use pmd_populate for that
2636                  */
2637                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2638                 spin_unlock(pmd_ptl);
2639                 anon_vma_unlock_write(vma->anon_vma);
2640                 goto out;
2641         }
2642
2643         /*
2644          * All pages are isolated and locked so anon_vma rmap
2645          * can't run anymore.
2646          */
2647         anon_vma_unlock_write(vma->anon_vma);
2648
2649         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2650         pte_unmap(pte);
2651         __SetPageUptodate(new_page);
2652         pgtable = pmd_pgtable(_pmd);
2653
2654         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2655         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2656
2657         /*
2658          * spin_lock() below is not the equivalent of smp_wmb(), so
2659          * this is needed to avoid the copy_huge_page writes to become
2660          * visible after the set_pmd_at() write.
2661          */
2662         smp_wmb();
2663
2664         spin_lock(pmd_ptl);
2665         BUG_ON(!pmd_none(*pmd));
2666         page_add_new_anon_rmap(new_page, vma, address);
2667         mem_cgroup_commit_charge(new_page, memcg, false);
2668         lru_cache_add_active_or_unevictable(new_page, vma);
2669         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2670         set_pmd_at(mm, address, pmd, _pmd);
2671         update_mmu_cache_pmd(vma, address, pmd);
2672         spin_unlock(pmd_ptl);
2673
2674         *hpage = NULL;
2675
2676         khugepaged_pages_collapsed++;
2677 out_up_write:
2678         up_write(&mm->mmap_sem);
2679         return;
2680
2681 out:
2682         mem_cgroup_cancel_charge(new_page, memcg);
2683         goto out_up_write;
2684 }
2685
2686 static int khugepaged_scan_pmd(struct mm_struct *mm,
2687                                struct vm_area_struct *vma,
2688                                unsigned long address,
2689                                struct page **hpage)
2690 {
2691         pmd_t *pmd;
2692         pte_t *pte, *_pte;
2693         int ret = 0, none_or_zero = 0;
2694         struct page *page;
2695         unsigned long _address;
2696         spinlock_t *ptl;
2697         int node = NUMA_NO_NODE;
2698         bool writable = false, referenced = false;
2699
2700         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2701
2702         pmd = mm_find_pmd(mm, address);
2703         if (!pmd)
2704                 goto out;
2705
2706         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2707         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2708         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2709              _pte++, _address += PAGE_SIZE) {
2710                 pte_t pteval = *_pte;
2711                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2712                         if (!userfaultfd_armed(vma) &&
2713                             ++none_or_zero <= khugepaged_max_ptes_none)
2714                                 continue;
2715                         else
2716                                 goto out_unmap;
2717                 }
2718                 if (!pte_present(pteval))
2719                         goto out_unmap;
2720                 if (pte_write(pteval))
2721                         writable = true;
2722
2723                 page = vm_normal_page(vma, _address, pteval);
2724                 if (unlikely(!page))
2725                         goto out_unmap;
2726                 /*
2727                  * Record which node the original page is from and save this
2728                  * information to khugepaged_node_load[].
2729                  * Khupaged will allocate hugepage from the node has the max
2730                  * hit record.
2731                  */
2732                 node = page_to_nid(page);
2733                 if (khugepaged_scan_abort(node))
2734                         goto out_unmap;
2735                 khugepaged_node_load[node]++;
2736                 VM_BUG_ON_PAGE(PageCompound(page), page);
2737                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2738                         goto out_unmap;
2739                 /*
2740                  * cannot use mapcount: can't collapse if there's a gup pin.
2741                  * The page must only be referenced by the scanned process
2742                  * and page swap cache.
2743                  */
2744                 if (page_count(page) != 1 + !!PageSwapCache(page))
2745                         goto out_unmap;
2746                 if (pte_young(pteval) ||
2747                     page_is_young(page) || PageReferenced(page) ||
2748                     mmu_notifier_test_young(vma->vm_mm, address))
2749                         referenced = true;
2750         }
2751         if (referenced && writable)
2752                 ret = 1;
2753 out_unmap:
2754         pte_unmap_unlock(pte, ptl);
2755         if (ret) {
2756                 node = khugepaged_find_target_node();
2757                 /* collapse_huge_page will return with the mmap_sem released */
2758                 collapse_huge_page(mm, address, hpage, vma, node);
2759         }
2760 out:
2761         return ret;
2762 }
2763
2764 static void collect_mm_slot(struct mm_slot *mm_slot)
2765 {
2766         struct mm_struct *mm = mm_slot->mm;
2767
2768         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2769
2770         if (khugepaged_test_exit(mm)) {
2771                 /* free mm_slot */
2772                 hash_del(&mm_slot->hash);
2773                 list_del(&mm_slot->mm_node);
2774
2775                 /*
2776                  * Not strictly needed because the mm exited already.
2777                  *
2778                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2779                  */
2780
2781                 /* khugepaged_mm_lock actually not necessary for the below */
2782                 free_mm_slot(mm_slot);
2783                 mmdrop(mm);
2784         }
2785 }
2786
2787 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2788                                             struct page **hpage)
2789         __releases(&khugepaged_mm_lock)
2790         __acquires(&khugepaged_mm_lock)
2791 {
2792         struct mm_slot *mm_slot;
2793         struct mm_struct *mm;
2794         struct vm_area_struct *vma;
2795         int progress = 0;
2796
2797         VM_BUG_ON(!pages);
2798         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2799
2800         if (khugepaged_scan.mm_slot)
2801                 mm_slot = khugepaged_scan.mm_slot;
2802         else {
2803                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2804                                      struct mm_slot, mm_node);
2805                 khugepaged_scan.address = 0;
2806                 khugepaged_scan.mm_slot = mm_slot;
2807         }
2808         spin_unlock(&khugepaged_mm_lock);
2809
2810         mm = mm_slot->mm;
2811         down_read(&mm->mmap_sem);
2812         if (unlikely(khugepaged_test_exit(mm)))
2813                 vma = NULL;
2814         else
2815                 vma = find_vma(mm, khugepaged_scan.address);
2816
2817         progress++;
2818         for (; vma; vma = vma->vm_next) {
2819                 unsigned long hstart, hend;
2820
2821                 cond_resched();
2822                 if (unlikely(khugepaged_test_exit(mm))) {
2823                         progress++;
2824                         break;
2825                 }
2826                 if (!hugepage_vma_check(vma)) {
2827 skip:
2828                         progress++;
2829                         continue;
2830                 }
2831                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2832                 hend = vma->vm_end & HPAGE_PMD_MASK;
2833                 if (hstart >= hend)
2834                         goto skip;
2835                 if (khugepaged_scan.address > hend)
2836                         goto skip;
2837                 if (khugepaged_scan.address < hstart)
2838                         khugepaged_scan.address = hstart;
2839                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2840
2841                 while (khugepaged_scan.address < hend) {
2842                         int ret;
2843                         cond_resched();
2844                         if (unlikely(khugepaged_test_exit(mm)))
2845                                 goto breakouterloop;
2846
2847                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2848                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2849                                   hend);
2850                         ret = khugepaged_scan_pmd(mm, vma,
2851                                                   khugepaged_scan.address,
2852                                                   hpage);
2853                         /* move to next address */
2854                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2855                         progress += HPAGE_PMD_NR;
2856                         if (ret)
2857                                 /* we released mmap_sem so break loop */
2858                                 goto breakouterloop_mmap_sem;
2859                         if (progress >= pages)
2860                                 goto breakouterloop;
2861                 }
2862         }
2863 breakouterloop:
2864         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2865 breakouterloop_mmap_sem:
2866
2867         spin_lock(&khugepaged_mm_lock);
2868         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2869         /*
2870          * Release the current mm_slot if this mm is about to die, or
2871          * if we scanned all vmas of this mm.
2872          */
2873         if (khugepaged_test_exit(mm) || !vma) {
2874                 /*
2875                  * Make sure that if mm_users is reaching zero while
2876                  * khugepaged runs here, khugepaged_exit will find
2877                  * mm_slot not pointing to the exiting mm.
2878                  */
2879                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2880                         khugepaged_scan.mm_slot = list_entry(
2881                                 mm_slot->mm_node.next,
2882                                 struct mm_slot, mm_node);
2883                         khugepaged_scan.address = 0;
2884                 } else {
2885                         khugepaged_scan.mm_slot = NULL;
2886                         khugepaged_full_scans++;
2887                 }
2888
2889                 collect_mm_slot(mm_slot);
2890         }
2891
2892         return progress;
2893 }
2894
2895 static int khugepaged_has_work(void)
2896 {
2897         return !list_empty(&khugepaged_scan.mm_head) &&
2898                 khugepaged_enabled();
2899 }
2900
2901 static int khugepaged_wait_event(void)
2902 {
2903         return !list_empty(&khugepaged_scan.mm_head) ||
2904                 kthread_should_stop();
2905 }
2906
2907 static void khugepaged_do_scan(void)
2908 {
2909         struct page *hpage = NULL;
2910         unsigned int progress = 0, pass_through_head = 0;
2911         unsigned int pages = khugepaged_pages_to_scan;
2912         bool wait = true;
2913
2914         barrier(); /* write khugepaged_pages_to_scan to local stack */
2915
2916         while (progress < pages) {
2917                 if (!khugepaged_prealloc_page(&hpage, &wait))
2918                         break;
2919
2920                 cond_resched();
2921
2922                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2923                         break;
2924
2925                 spin_lock(&khugepaged_mm_lock);
2926                 if (!khugepaged_scan.mm_slot)
2927                         pass_through_head++;
2928                 if (khugepaged_has_work() &&
2929                     pass_through_head < 2)
2930                         progress += khugepaged_scan_mm_slot(pages - progress,
2931                                                             &hpage);
2932                 else
2933                         progress = pages;
2934                 spin_unlock(&khugepaged_mm_lock);
2935         }
2936
2937         if (!IS_ERR_OR_NULL(hpage))
2938                 put_page(hpage);
2939 }
2940
2941 static void khugepaged_wait_work(void)
2942 {
2943         if (khugepaged_has_work()) {
2944                 if (!khugepaged_scan_sleep_millisecs)
2945                         return;
2946
2947                 wait_event_freezable_timeout(khugepaged_wait,
2948                                              kthread_should_stop(),
2949                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2950                 return;
2951         }
2952
2953         if (khugepaged_enabled())
2954                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2955 }
2956
2957 static int khugepaged(void *none)
2958 {
2959         struct mm_slot *mm_slot;
2960
2961         set_freezable();
2962         set_user_nice(current, MAX_NICE);
2963
2964         while (!kthread_should_stop()) {
2965                 khugepaged_do_scan();
2966                 khugepaged_wait_work();
2967         }
2968
2969         spin_lock(&khugepaged_mm_lock);
2970         mm_slot = khugepaged_scan.mm_slot;
2971         khugepaged_scan.mm_slot = NULL;
2972         if (mm_slot)
2973                 collect_mm_slot(mm_slot);
2974         spin_unlock(&khugepaged_mm_lock);
2975         return 0;
2976 }
2977
2978 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2979                 unsigned long haddr, pmd_t *pmd)
2980 {
2981         struct mm_struct *mm = vma->vm_mm;
2982         pgtable_t pgtable;
2983         pmd_t _pmd;
2984         int i;
2985
2986         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2987         /* leave pmd empty until pte is filled */
2988
2989         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2990         pmd_populate(mm, &_pmd, pgtable);
2991
2992         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2993                 pte_t *pte, entry;
2994                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2995                 entry = pte_mkspecial(entry);
2996                 pte = pte_offset_map(&_pmd, haddr);
2997                 VM_BUG_ON(!pte_none(*pte));
2998                 set_pte_at(mm, haddr, pte, entry);
2999                 pte_unmap(pte);
3000         }
3001         smp_wmb(); /* make pte visible before pmd */
3002         pmd_populate(mm, pmd, pgtable);
3003         put_huge_zero_page();
3004 }
3005
3006 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
3007                 pmd_t *pmd)
3008 {
3009         spinlock_t *ptl;
3010         struct page *page = NULL;
3011         struct mm_struct *mm = vma->vm_mm;
3012         unsigned long haddr = address & HPAGE_PMD_MASK;
3013         unsigned long mmun_start;       /* For mmu_notifiers */
3014         unsigned long mmun_end;         /* For mmu_notifiers */
3015
3016         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
3017
3018         mmun_start = haddr;
3019         mmun_end   = haddr + HPAGE_PMD_SIZE;
3020 again:
3021         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3022         ptl = pmd_lock(mm, pmd);
3023         if (unlikely(!pmd_trans_huge(*pmd)))
3024                 goto unlock;
3025         if (vma_is_dax(vma)) {
3026                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
3027                 if (is_huge_zero_pmd(_pmd))
3028                         put_huge_zero_page();
3029         } else if (is_huge_zero_pmd(*pmd)) {
3030                 __split_huge_zero_page_pmd(vma, haddr, pmd);
3031         } else {
3032                 page = pmd_page(*pmd);
3033                 VM_BUG_ON_PAGE(!page_count(page), page);
3034                 get_page(page);
3035         }
3036  unlock:
3037         spin_unlock(ptl);
3038         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3039
3040         if (!page)
3041                 return;
3042
3043         split_huge_page(page);
3044         put_page(page);
3045
3046         /*
3047          * We don't always have down_write of mmap_sem here: a racing
3048          * do_huge_pmd_wp_page() might have copied-on-write to another
3049          * huge page before our split_huge_page() got the anon_vma lock.
3050          */
3051         if (unlikely(pmd_trans_huge(*pmd)))
3052                 goto again;
3053 }
3054
3055 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3056                 pmd_t *pmd)
3057 {
3058         struct vm_area_struct *vma;
3059
3060         vma = find_vma(mm, address);
3061         BUG_ON(vma == NULL);
3062         split_huge_page_pmd(vma, address, pmd);
3063 }
3064
3065 static void split_huge_page_address(struct mm_struct *mm,
3066                                     unsigned long address)
3067 {
3068         pgd_t *pgd;
3069         pud_t *pud;
3070         pmd_t *pmd;
3071
3072         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3073
3074         pgd = pgd_offset(mm, address);
3075         if (!pgd_present(*pgd))
3076                 return;
3077
3078         pud = pud_offset(pgd, address);
3079         if (!pud_present(*pud))
3080                 return;
3081
3082         pmd = pmd_offset(pud, address);
3083         if (!pmd_present(*pmd))
3084                 return;
3085         /*
3086          * Caller holds the mmap_sem write mode, so a huge pmd cannot
3087          * materialize from under us.
3088          */
3089         split_huge_page_pmd_mm(mm, address, pmd);
3090 }
3091
3092 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3093                              unsigned long start,
3094                              unsigned long end,
3095                              long adjust_next)
3096 {
3097         /*
3098          * If the new start address isn't hpage aligned and it could
3099          * previously contain an hugepage: check if we need to split
3100          * an huge pmd.
3101          */
3102         if (start & ~HPAGE_PMD_MASK &&
3103             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3104             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3105                 split_huge_page_address(vma->vm_mm, start);
3106
3107         /*
3108          * If the new end address isn't hpage aligned and it could
3109          * previously contain an hugepage: check if we need to split
3110          * an huge pmd.
3111          */
3112         if (end & ~HPAGE_PMD_MASK &&
3113             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3114             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3115                 split_huge_page_address(vma->vm_mm, end);
3116
3117         /*
3118          * If we're also updating the vma->vm_next->vm_start, if the new
3119          * vm_next->vm_start isn't page aligned and it could previously
3120          * contain an hugepage: check if we need to split an huge pmd.
3121          */
3122         if (adjust_next > 0) {
3123                 struct vm_area_struct *next = vma->vm_next;
3124                 unsigned long nstart = next->vm_start;
3125                 nstart += adjust_next << PAGE_SHIFT;
3126                 if (nstart & ~HPAGE_PMD_MASK &&
3127                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3128                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3129                         split_huge_page_address(next->vm_mm, nstart);
3130         }
3131 }