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 "
 139                                 "to help transparent hugepage allocations\n",
 140                                 min_free_kbytes, recommended_min);
 141
 142                 min_free_kbytes = recommended_min;
 143         }
 144         setup_per_zone_wmarks();
 145 }
 146
 147 static int start_stop_khugepaged(void)
 148 {
 149         int err = 0;
 150         if (khugepaged_enabled()) {
 151                 if (!khugepaged_thread)
 152                         khugepaged_thread = kthread_run(khugepaged, NULL,
 153                                                         "khugepaged");
 154                 if (IS_ERR(khugepaged_thread)) {
 155                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 156                         err = PTR_ERR(khugepaged_thread);
 157                         khugepaged_thread = NULL;
 158                         goto fail;
 159                 }
 160
 161                 if (!list_empty(&khugepaged_scan.mm_head))
 162                         wake_up_interruptible(&khugepaged_wait);
 163
 164                 set_recommended_min_free_kbytes();
 165         } else if (khugepaged_thread) {
 166                 kthread_stop(khugepaged_thread);
 167                 khugepaged_thread = NULL;
 168         }
 169 fail:
 170         return err;
 171 }
 172
 173 static atomic_t huge_zero_refcount;
 174 struct page *huge_zero_page __read_mostly;
 175
 176 struct page *get_huge_zero_page(void)
 177 {
 178         struct page *zero_page;
 179 retry:
 180         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 181                 return READ_ONCE(huge_zero_page);
 182
 183         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 184                         HPAGE_PMD_ORDER);
 185         if (!zero_page) {
 186                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 187                 return NULL;
 188         }
 189         count_vm_event(THP_ZERO_PAGE_ALLOC);
 190         preempt_disable();
 191         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 192                 preempt_enable();
 193                 __free_pages(zero_page, compound_order(zero_page));
 194                 goto retry;
 195         }
 196
 197         /* We take additional reference here. It will be put back by shrinker */
 198         atomic_set(&huge_zero_refcount, 2);
 199         preempt_enable();
 200         return READ_ONCE(huge_zero_page);
 201 }
 202
 203 static void put_huge_zero_page(void)
 204 {
 205         /*
 206          * Counter should never go to zero here. Only shrinker can put
 207          * last reference.
 208          */
 209         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 210 }
 211
 212 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 213                                         struct shrink_control *sc)
 214 {
 215         /* we can free zero page only if last reference remains */
 216         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 217 }
 218
 219 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 220                                        struct shrink_control *sc)
 221 {
 222         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 223                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 224                 BUG_ON(zero_page == NULL);
 225                 __free_pages(zero_page, compound_order(zero_page));
 226                 return HPAGE_PMD_NR;
 227         }
 228
 229         return 0;
 230 }
 231
 232 static struct shrinker huge_zero_page_shrinker = {
 233         .count_objects = shrink_huge_zero_page_count,
 234         .scan_objects = shrink_huge_zero_page_scan,
 235         .seeks = DEFAULT_SEEKS,
 236 };
 237
 238 #ifdef CONFIG_SYSFS
 239
 240 static ssize_t double_flag_show(struct kobject *kobj,
 241                                 struct kobj_attribute *attr, char *buf,
 242                                 enum transparent_hugepage_flag enabled,
 243                                 enum transparent_hugepage_flag req_madv)
 244 {
 245         if (test_bit(enabled, &transparent_hugepage_flags)) {
 246                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 247                 return sprintf(buf, "[always] madvise never\n");
 248         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 249                 return sprintf(buf, "always [madvise] never\n");
 250         else
 251                 return sprintf(buf, "always madvise [never]\n");
 252 }
 253 static ssize_t double_flag_store(struct kobject *kobj,
 254                                  struct kobj_attribute *attr,
 255                                  const char *buf, size_t count,
 256                                  enum transparent_hugepage_flag enabled,
 257                                  enum transparent_hugepage_flag req_madv)
 258 {
 259         if (!memcmp("always", buf,
 260                     min(sizeof("always")-1, count))) {
 261                 set_bit(enabled, &transparent_hugepage_flags);
 262                 clear_bit(req_madv, &transparent_hugepage_flags);
 263         } else if (!memcmp("madvise", buf,
 264                            min(sizeof("madvise")-1, count))) {
 265                 clear_bit(enabled, &transparent_hugepage_flags);
 266                 set_bit(req_madv, &transparent_hugepage_flags);
 267         } else if (!memcmp("never", buf,
 268                            min(sizeof("never")-1, count))) {
 269                 clear_bit(enabled, &transparent_hugepage_flags);
 270                 clear_bit(req_madv, &transparent_hugepage_flags);
 271         } else
 272                 return -EINVAL;
 273
 274         return count;
 275 }
 276
 277 static ssize_t enabled_show(struct kobject *kobj,
 278                             struct kobj_attribute *attr, char *buf)
 279 {
 280         return double_flag_show(kobj, attr, buf,
 281                                 TRANSPARENT_HUGEPAGE_FLAG,
 282                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 283 }
 284 static ssize_t enabled_store(struct kobject *kobj,
 285                              struct kobj_attribute *attr,
 286                              const char *buf, size_t count)
 287 {
 288         ssize_t ret;
 289
 290         ret = double_flag_store(kobj, attr, buf, count,
 291                                 TRANSPARENT_HUGEPAGE_FLAG,
 292                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 293
 294         if (ret > 0) {
 295                 int err;
 296
 297                 mutex_lock(&khugepaged_mutex);
 298                 err = start_stop_khugepaged();
 299                 mutex_unlock(&khugepaged_mutex);
 300
 301                 if (err)
 302                         ret = err;
 303         }
 304
 305         return ret;
 306 }
 307 static struct kobj_attribute enabled_attr =
 308         __ATTR(enabled, 0644, enabled_show, enabled_store);
 309
 310 static ssize_t single_flag_show(struct kobject *kobj,
 311                                 struct kobj_attribute *attr, char *buf,
 312                                 enum transparent_hugepage_flag flag)
 313 {
 314         return sprintf(buf, "%d\n",
 315                        !!test_bit(flag, &transparent_hugepage_flags));
 316 }
 317
 318 static ssize_t single_flag_store(struct kobject *kobj,
 319                                  struct kobj_attribute *attr,
 320                                  const char *buf, size_t count,
 321                                  enum transparent_hugepage_flag flag)
 322 {
 323         unsigned long value;
 324         int ret;
 325
 326         ret = kstrtoul(buf, 10, &value);
 327         if (ret < 0)
 328                 return ret;
 329         if (value > 1)
 330                 return -EINVAL;
 331
 332         if (value)
 333                 set_bit(flag, &transparent_hugepage_flags);
 334         else
 335                 clear_bit(flag, &transparent_hugepage_flags);
 336
 337         return count;
 338 }
 339
 340 /*
 341  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 342  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 343  * memory just to allocate one more hugepage.
 344  */
 345 static ssize_t defrag_show(struct kobject *kobj,
 346                            struct kobj_attribute *attr, char *buf)
 347 {
 348         return double_flag_show(kobj, attr, buf,
 349                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 350                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 351 }
 352 static ssize_t defrag_store(struct kobject *kobj,
 353                             struct kobj_attribute *attr,
 354                             const char *buf, size_t count)
 355 {
 356         return double_flag_store(kobj, attr, buf, count,
 357                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 358                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 359 }
 360 static struct kobj_attribute defrag_attr =
 361         __ATTR(defrag, 0644, defrag_show, defrag_store);
 362
 363 static ssize_t use_zero_page_show(struct kobject *kobj,
 364                 struct kobj_attribute *attr, char *buf)
 365 {
 366         return single_flag_show(kobj, attr, buf,
 367                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 368 }
 369 static ssize_t use_zero_page_store(struct kobject *kobj,
 370                 struct kobj_attribute *attr, const char *buf, size_t count)
 371 {
 372         return single_flag_store(kobj, attr, buf, count,
 373                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 374 }
 375 static struct kobj_attribute use_zero_page_attr =
 376         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 377 #ifdef CONFIG_DEBUG_VM
 378 static ssize_t debug_cow_show(struct kobject *kobj,
 379                                 struct kobj_attribute *attr, char *buf)
 380 {
 381         return single_flag_show(kobj, attr, buf,
 382                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 383 }
 384 static ssize_t debug_cow_store(struct kobject *kobj,
 385                                struct kobj_attribute *attr,
 386                                const char *buf, size_t count)
 387 {
 388         return single_flag_store(kobj, attr, buf, count,
 389                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 390 }
 391 static struct kobj_attribute debug_cow_attr =
 392         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 393 #endif /* CONFIG_DEBUG_VM */
 394
 395 static struct attribute *hugepage_attr[] = {
 396         &enabled_attr.attr,
 397         &defrag_attr.attr,
 398         &use_zero_page_attr.attr,
 399 #ifdef CONFIG_DEBUG_VM
 400         &debug_cow_attr.attr,
 401 #endif
 402         NULL,
 403 };
 404
 405 static struct attribute_group hugepage_attr_group = {
 406         .attrs = hugepage_attr,
 407 };
 408
 409 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 410                                          struct kobj_attribute *attr,
 411                                          char *buf)
 412 {
 413         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 414 }
 415
 416 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 417                                           struct kobj_attribute *attr,
 418                                           const char *buf, size_t count)
 419 {
 420         unsigned long msecs;
 421         int err;
 422
 423         err = kstrtoul(buf, 10, &msecs);
 424         if (err || msecs > UINT_MAX)
 425                 return -EINVAL;
 426
 427         khugepaged_scan_sleep_millisecs = msecs;
 428         wake_up_interruptible(&khugepaged_wait);
 429
 430         return count;
 431 }
 432 static struct kobj_attribute scan_sleep_millisecs_attr =
 433         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 434                scan_sleep_millisecs_store);
 435
 436 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 437                                           struct kobj_attribute *attr,
 438                                           char *buf)
 439 {
 440         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 441 }
 442
 443 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 444                                            struct kobj_attribute *attr,
 445                                            const char *buf, size_t count)
 446 {
 447         unsigned long msecs;
 448         int err;
 449
 450         err = kstrtoul(buf, 10, &msecs);
 451         if (err || msecs > UINT_MAX)
 452                 return -EINVAL;
 453
 454         khugepaged_alloc_sleep_millisecs = msecs;
 455         wake_up_interruptible(&khugepaged_wait);
 456
 457         return count;
 458 }
 459 static struct kobj_attribute alloc_sleep_millisecs_attr =
 460         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 461                alloc_sleep_millisecs_store);
 462
 463 static ssize_t pages_to_scan_show(struct kobject *kobj,
 464                                   struct kobj_attribute *attr,
 465                                   char *buf)
 466 {
 467         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 468 }
 469 static ssize_t pages_to_scan_store(struct kobject *kobj,
 470                                    struct kobj_attribute *attr,
 471                                    const char *buf, size_t count)
 472 {
 473         int err;
 474         unsigned long pages;
 475
 476         err = kstrtoul(buf, 10, &pages);
 477         if (err || !pages || pages > UINT_MAX)
 478                 return -EINVAL;
 479
 480         khugepaged_pages_to_scan = pages;
 481
 482         return count;
 483 }
 484 static struct kobj_attribute pages_to_scan_attr =
 485         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 486                pages_to_scan_store);
 487
 488 static ssize_t pages_collapsed_show(struct kobject *kobj,
 489                                     struct kobj_attribute *attr,
 490                                     char *buf)
 491 {
 492         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 493 }
 494 static struct kobj_attribute pages_collapsed_attr =
 495         __ATTR_RO(pages_collapsed);
 496
 497 static ssize_t full_scans_show(struct kobject *kobj,
 498                                struct kobj_attribute *attr,
 499                                char *buf)
 500 {
 501         return sprintf(buf, "%u\n", khugepaged_full_scans);
 502 }
 503 static struct kobj_attribute full_scans_attr =
 504         __ATTR_RO(full_scans);
 505
 506 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 507                                       struct kobj_attribute *attr, char *buf)
 508 {
 509         return single_flag_show(kobj, attr, buf,
 510                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 511 }
 512 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 513                                        struct kobj_attribute *attr,
 514                                        const char *buf, size_t count)
 515 {
 516         return single_flag_store(kobj, attr, buf, count,
 517                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 518 }
 519 static struct kobj_attribute khugepaged_defrag_attr =
 520         __ATTR(defrag, 0644, khugepaged_defrag_show,
 521                khugepaged_defrag_store);
 522
 523 /*
 524  * max_ptes_none controls if khugepaged should collapse hugepages over
 525  * any unmapped ptes in turn potentially increasing the memory
 526  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 527  * reduce the available free memory in the system as it
 528  * runs. Increasing max_ptes_none will instead potentially reduce the
 529  * free memory in the system during the khugepaged scan.
 530  */
 531 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 532                                              struct kobj_attribute *attr,
 533                                              char *buf)
 534 {
 535         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 536 }
 537 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 538                                               struct kobj_attribute *attr,
 539                                               const char *buf, size_t count)
 540 {
 541         int err;
 542         unsigned long max_ptes_none;
 543
 544         err = kstrtoul(buf, 10, &max_ptes_none);
 545         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 546                 return -EINVAL;
 547
 548         khugepaged_max_ptes_none = max_ptes_none;
 549
 550         return count;
 551 }
 552 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 553         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 554                khugepaged_max_ptes_none_store);
 555
 556 static struct attribute *khugepaged_attr[] = {
 557         &khugepaged_defrag_attr.attr,
 558         &khugepaged_max_ptes_none_attr.attr,
 559         &pages_to_scan_attr.attr,
 560         &pages_collapsed_attr.attr,
 561         &full_scans_attr.attr,
 562         &scan_sleep_millisecs_attr.attr,
 563         &alloc_sleep_millisecs_attr.attr,
 564         NULL,
 565 };
 566
 567 static struct attribute_group khugepaged_attr_group = {
 568         .attrs = khugepaged_attr,
 569         .name = "khugepaged",
 570 };
 571
 572 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 573 {
 574         int err;
 575
 576         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 577         if (unlikely(!*hugepage_kobj)) {
 578                 pr_err("failed to create transparent hugepage kobject\n");
 579                 return -ENOMEM;
 580         }
 581
 582         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 583         if (err) {
 584                 pr_err("failed to register transparent hugepage group\n");
 585                 goto delete_obj;
 586         }
 587
 588         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 589         if (err) {
 590                 pr_err("failed to register transparent hugepage group\n");
 591                 goto remove_hp_group;
 592         }
 593
 594         return 0;
 595
 596 remove_hp_group:
 597         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 598 delete_obj:
 599         kobject_put(*hugepage_kobj);
 600         return err;
 601 }
 602
 603 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 604 {
 605         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 606         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 607         kobject_put(hugepage_kobj);
 608 }
 609 #else
 610 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 611 {
 612         return 0;
 613 }
 614
 615 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 616 {
 617 }
 618 #endif /* CONFIG_SYSFS */
 619
 620 static int __init hugepage_init(void)
 621 {
 622         int err;
 623         struct kobject *hugepage_kobj;
 624
 625         if (!has_transparent_hugepage()) {
 626                 transparent_hugepage_flags = 0;
 627                 return -EINVAL;
 628         }
 629
 630         err = hugepage_init_sysfs(&hugepage_kobj);
 631         if (err)
 632                 goto err_sysfs;
 633
 634         err = khugepaged_slab_init();
 635         if (err)
 636                 goto err_slab;
 637
 638         err = register_shrinker(&huge_zero_page_shrinker);
 639         if (err)
 640                 goto err_hzp_shrinker;
 641
 642         /*
 643          * By default disable transparent hugepages on smaller systems,
 644          * where the extra memory used could hurt more than TLB overhead
 645          * is likely to save.  The admin can still enable it through /sys.
 646          */
 647         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 648                 transparent_hugepage_flags = 0;
 649                 return 0;
 650         }
 651
 652         err = start_stop_khugepaged();
 653         if (err)
 654                 goto err_khugepaged;
 655
 656         return 0;
 657 err_khugepaged:
 658         unregister_shrinker(&huge_zero_page_shrinker);
 659 err_hzp_shrinker:
 660         khugepaged_slab_exit();
 661 err_slab:
 662         hugepage_exit_sysfs(hugepage_kobj);
 663 err_sysfs:
 664         return err;
 665 }
 666 subsys_initcall(hugepage_init);
 667
 668 static int __init setup_transparent_hugepage(char *str)
 669 {
 670         int ret = 0;
 671         if (!str)
 672                 goto out;
 673         if (!strcmp(str, "always")) {
 674                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 675                         &transparent_hugepage_flags);
 676                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 677                           &transparent_hugepage_flags);
 678                 ret = 1;
 679         } else if (!strcmp(str, "madvise")) {
 680                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 681                           &transparent_hugepage_flags);
 682                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 683                         &transparent_hugepage_flags);
 684                 ret = 1;
 685         } else if (!strcmp(str, "never")) {
 686                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 687                           &transparent_hugepage_flags);
 688                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 689                           &transparent_hugepage_flags);
 690                 ret = 1;
 691         }
 692 out:
 693         if (!ret)
 694                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 695         return ret;
 696 }
 697 __setup("transparent_hugepage=", setup_transparent_hugepage);
 698
 699 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 700 {
 701         if (likely(vma->vm_flags & VM_WRITE))
 702                 pmd = pmd_mkwrite(pmd);
 703         return pmd;
 704 }
 705
 706 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 707 {
 708         pmd_t entry;
 709         entry = mk_pmd(page, prot);
 710         entry = pmd_mkhuge(entry);
 711         return entry;
 712 }
 713
 714 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 715                                         struct vm_area_struct *vma,
 716                                         unsigned long address, pmd_t *pmd,
 717                                         struct page *page, gfp_t gfp,
 718                                         unsigned int flags)
 719 {
 720         struct mem_cgroup *memcg;
 721         pgtable_t pgtable;
 722         spinlock_t *ptl;
 723         unsigned long haddr = address & HPAGE_PMD_MASK;
 724
 725         VM_BUG_ON_PAGE(!PageCompound(page), page);
 726
 727         if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
 728                 put_page(page);
 729                 count_vm_event(THP_FAULT_FALLBACK);
 730                 return VM_FAULT_FALLBACK;
 731         }
 732
 733         pgtable = pte_alloc_one(mm, haddr);
 734         if (unlikely(!pgtable)) {
 735                 mem_cgroup_cancel_charge(page, memcg);
 736                 put_page(page);
 737                 return VM_FAULT_OOM;
 738         }
 739
 740         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 741         /*
 742          * The memory barrier inside __SetPageUptodate makes sure that
 743          * clear_huge_page writes become visible before the set_pmd_at()
 744          * write.
 745          */
 746         __SetPageUptodate(page);
 747
 748         ptl = pmd_lock(mm, pmd);
 749         if (unlikely(!pmd_none(*pmd))) {
 750                 spin_unlock(ptl);
 751                 mem_cgroup_cancel_charge(page, memcg);
 752                 put_page(page);
 753                 pte_free(mm, pgtable);
 754         } else {
 755                 pmd_t entry;
 756
 757                 /* Deliver the page fault to userland */
 758                 if (userfaultfd_missing(vma)) {
 759                         int ret;
 760
 761                         spin_unlock(ptl);
 762                         mem_cgroup_cancel_charge(page, memcg);
 763                         put_page(page);
 764                         pte_free(mm, pgtable);
 765                         ret = handle_userfault(vma, address, flags,
 766                                                VM_UFFD_MISSING);
 767                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 768                         return ret;
 769                 }
 770
 771                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 772                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 773                 page_add_new_anon_rmap(page, vma, haddr);
 774                 mem_cgroup_commit_charge(page, memcg, false);
 775                 lru_cache_add_active_or_unevictable(page, vma);
 776                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 777                 set_pmd_at(mm, haddr, pmd, entry);
 778                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 779                 atomic_long_inc(&mm->nr_ptes);
 780                 spin_unlock(ptl);
 781                 count_vm_event(THP_FAULT_ALLOC);
 782         }
 783
 784         return 0;
 785 }
 786
 787 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 788 {
 789         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 790 }
 791
 792 /* Caller must hold page table lock. */
 793 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 794                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 795                 struct page *zero_page)
 796 {
 797         pmd_t entry;
 798         if (!pmd_none(*pmd))
 799                 return false;
 800         entry = mk_pmd(zero_page, vma->vm_page_prot);
 801         entry = pmd_mkhuge(entry);
 802         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 803         set_pmd_at(mm, haddr, pmd, entry);
 804         atomic_long_inc(&mm->nr_ptes);
 805         return true;
 806 }
 807
 808 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 809                                unsigned long address, pmd_t *pmd,
 810                                unsigned int flags)
 811 {
 812         gfp_t gfp;
 813         struct page *page;
 814         unsigned long haddr = address & HPAGE_PMD_MASK;
 815
 816         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 817                 return VM_FAULT_FALLBACK;
 818         if (unlikely(anon_vma_prepare(vma)))
 819                 return VM_FAULT_OOM;
 820         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 821                 return VM_FAULT_OOM;
 822         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 823                         transparent_hugepage_use_zero_page()) {
 824                 spinlock_t *ptl;
 825                 pgtable_t pgtable;
 826                 struct page *zero_page;
 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                 if (pmd_none(*pmd)) {
 840                         if (userfaultfd_missing(vma)) {
 841                                 spin_unlock(ptl);
 842                                 pte_free(mm, pgtable);
 843                                 put_huge_zero_page();
 844                                 ret = handle_userfault(vma, address, flags,
 845                                                        VM_UFFD_MISSING);
 846                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 847                         } else {
 848                                 set_huge_zero_page(pgtable, mm, vma,
 849                                                    haddr, pmd,
 850                                                    zero_page);
 851                                 spin_unlock(ptl);
 852                         }
 853                 } else {
 854                         spin_unlock(ptl);
 855                         pte_free(mm, pgtable);
 856                         put_huge_zero_page();
 857                 }
 858                 return ret;
 859         }
 860         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 861         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 862         if (unlikely(!page)) {
 863                 count_vm_event(THP_FAULT_FALLBACK);
 864                 return VM_FAULT_FALLBACK;
 865         }
 866         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 867                                             flags);
 868 }
 869
 870 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 871                 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
 872 {
 873         struct mm_struct *mm = vma->vm_mm;
 874         pmd_t entry;
 875         spinlock_t *ptl;
 876
 877         ptl = pmd_lock(mm, pmd);
 878         if (pmd_none(*pmd)) {
 879                 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
 880                 if (write) {
 881                         entry = pmd_mkyoung(pmd_mkdirty(entry));
 882                         entry = maybe_pmd_mkwrite(entry, vma);
 883                 }
 884                 set_pmd_at(mm, addr, pmd, entry);
 885                 update_mmu_cache_pmd(vma, addr, pmd);
 886         }
 887         spin_unlock(ptl);
 888 }
 889
 890 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 891                         pmd_t *pmd, unsigned long pfn, bool write)
 892 {
 893         pgprot_t pgprot = vma->vm_page_prot;
 894         /*
 895          * If we had pmd_special, we could avoid all these restrictions,
 896          * but we need to be consistent with PTEs and architectures that
 897          * can't support a 'special' bit.
 898          */
 899         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 900         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 901                                                 (VM_PFNMAP|VM_MIXEDMAP));
 902         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 903         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
 904
 905         if (addr < vma->vm_start || addr >= vma->vm_end)
 906                 return VM_FAULT_SIGBUS;
 907         if (track_pfn_insert(vma, &pgprot, pfn))
 908                 return VM_FAULT_SIGBUS;
 909         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 910         return VM_FAULT_NOPAGE;
 911 }
 912
 913 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 914                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 915                   struct vm_area_struct *vma)
 916 {
 917         spinlock_t *dst_ptl, *src_ptl;
 918         struct page *src_page;
 919         pmd_t pmd;
 920         pgtable_t pgtable;
 921         int ret;
 922
 923         ret = -ENOMEM;
 924         pgtable = pte_alloc_one(dst_mm, addr);
 925         if (unlikely(!pgtable))
 926                 goto out;
 927
 928         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 929         src_ptl = pmd_lockptr(src_mm, src_pmd);
 930         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 931
 932         ret = -EAGAIN;
 933         pmd = *src_pmd;
 934         if (unlikely(!pmd_trans_huge(pmd))) {
 935                 pte_free(dst_mm, pgtable);
 936                 goto out_unlock;
 937         }
 938         /*
 939          * When page table lock is held, the huge zero pmd should not be
 940          * under splitting since we don't split the page itself, only pmd to
 941          * a page table.
 942          */
 943         if (is_huge_zero_pmd(pmd)) {
 944                 struct page *zero_page;
 945                 /*
 946                  * get_huge_zero_page() will never allocate a new page here,
 947                  * since we already have a zero page to copy. It just takes a
 948                  * reference.
 949                  */
 950                 zero_page = get_huge_zero_page();
 951                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 952                                 zero_page);
 953                 ret = 0;
 954                 goto out_unlock;
 955         }
 956
 957         if (unlikely(pmd_trans_splitting(pmd))) {
 958                 /* split huge page running from under us */
 959                 spin_unlock(src_ptl);
 960                 spin_unlock(dst_ptl);
 961                 pte_free(dst_mm, pgtable);
 962
 963                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 964                 goto out;
 965         }
 966         src_page = pmd_page(pmd);
 967         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 968         get_page(src_page);
 969         page_dup_rmap(src_page);
 970         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 971
 972         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 973         pmd = pmd_mkold(pmd_wrprotect(pmd));
 974         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 975         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 976         atomic_long_inc(&dst_mm->nr_ptes);
 977
 978         ret = 0;
 979 out_unlock:
 980         spin_unlock(src_ptl);
 981         spin_unlock(dst_ptl);
 982 out:
 983         return ret;
 984 }
 985
 986 void huge_pmd_set_accessed(struct mm_struct *mm,
 987                            struct vm_area_struct *vma,
 988                            unsigned long address,
 989                            pmd_t *pmd, pmd_t orig_pmd,
 990                            int dirty)
 991 {
 992         spinlock_t *ptl;
 993         pmd_t entry;
 994         unsigned long haddr;
 995
 996         ptl = pmd_lock(mm, pmd);
 997         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 998                 goto unlock;
 999
1000         entry = pmd_mkyoung(orig_pmd);
1001         haddr = address & HPAGE_PMD_MASK;
1002         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1003                 update_mmu_cache_pmd(vma, address, pmd);
1004
1005 unlock:
1006         spin_unlock(ptl);
1007 }
1008
1009 /*
1010  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1011  * during copy_user_huge_page()'s copy_page_rep(): in the case when
1012  * the source page gets split and a tail freed before copy completes.
1013  * Called under pmd_lock of checked pmd, so safe from splitting itself.
1014  */
1015 static void get_user_huge_page(struct page *page)
1016 {
1017         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1018                 struct page *endpage = page + HPAGE_PMD_NR;
1019
1020                 atomic_add(HPAGE_PMD_NR, &page->_count);
1021                 while (++page < endpage)
1022                         get_huge_page_tail(page);
1023         } else {
1024                 get_page(page);
1025         }
1026 }
1027
1028 static void put_user_huge_page(struct page *page)
1029 {
1030         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1031                 struct page *endpage = page + HPAGE_PMD_NR;
1032
1033                 while (page < endpage)
1034                         put_page(page++);
1035         } else {
1036                 put_page(page);
1037         }
1038 }
1039
1040 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1041                                         struct vm_area_struct *vma,
1042                                         unsigned long address,
1043                                         pmd_t *pmd, pmd_t orig_pmd,
1044                                         struct page *page,
1045                                         unsigned long haddr)
1046 {
1047         struct mem_cgroup *memcg;
1048         spinlock_t *ptl;
1049         pgtable_t pgtable;
1050         pmd_t _pmd;
1051         int ret = 0, i;
1052         struct page **pages;
1053         unsigned long mmun_start;       /* For mmu_notifiers */
1054         unsigned long mmun_end;         /* For mmu_notifiers */
1055
1056         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1057                         GFP_KERNEL);
1058         if (unlikely(!pages)) {
1059                 ret |= VM_FAULT_OOM;
1060                 goto out;
1061         }
1062
1063         for (i = 0; i < HPAGE_PMD_NR; i++) {
1064                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1065                                                __GFP_OTHER_NODE,
1066                                                vma, address, page_to_nid(page));
1067                 if (unlikely(!pages[i] ||
1068                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1069                                                    &memcg))) {
1070                         if (pages[i])
1071                                 put_page(pages[i]);
1072                         while (--i >= 0) {
1073                                 memcg = (void *)page_private(pages[i]);
1074                                 set_page_private(pages[i], 0);
1075                                 mem_cgroup_cancel_charge(pages[i], memcg);
1076                                 put_page(pages[i]);
1077                         }
1078                         kfree(pages);
1079                         ret |= VM_FAULT_OOM;
1080                         goto out;
1081                 }
1082                 set_page_private(pages[i], (unsigned long)memcg);
1083         }
1084
1085         for (i = 0; i < HPAGE_PMD_NR; i++) {
1086                 copy_user_highpage(pages[i], page + i,
1087                                    haddr + PAGE_SIZE * i, vma);
1088                 __SetPageUptodate(pages[i]);
1089                 cond_resched();
1090         }
1091
1092         mmun_start = haddr;
1093         mmun_end   = haddr + HPAGE_PMD_SIZE;
1094         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1095
1096         ptl = pmd_lock(mm, pmd);
1097         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1098                 goto out_free_pages;
1099         VM_BUG_ON_PAGE(!PageHead(page), page);
1100
1101         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1102         /* leave pmd empty until pte is filled */
1103
1104         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1105         pmd_populate(mm, &_pmd, pgtable);
1106
1107         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1108                 pte_t *pte, entry;
1109                 entry = mk_pte(pages[i], vma->vm_page_prot);
1110                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1111                 memcg = (void *)page_private(pages[i]);
1112                 set_page_private(pages[i], 0);
1113                 page_add_new_anon_rmap(pages[i], vma, haddr);
1114                 mem_cgroup_commit_charge(pages[i], memcg, false);
1115                 lru_cache_add_active_or_unevictable(pages[i], vma);
1116                 pte = pte_offset_map(&_pmd, haddr);
1117                 VM_BUG_ON(!pte_none(*pte));
1118                 set_pte_at(mm, haddr, pte, entry);
1119                 pte_unmap(pte);
1120         }
1121         kfree(pages);
1122
1123         smp_wmb(); /* make pte visible before pmd */
1124         pmd_populate(mm, pmd, pgtable);
1125         page_remove_rmap(page);
1126         spin_unlock(ptl);
1127
1128         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1129
1130         ret |= VM_FAULT_WRITE;
1131         put_page(page);
1132
1133 out:
1134         return ret;
1135
1136 out_free_pages:
1137         spin_unlock(ptl);
1138         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1139         for (i = 0; i < HPAGE_PMD_NR; i++) {
1140                 memcg = (void *)page_private(pages[i]);
1141                 set_page_private(pages[i], 0);
1142                 mem_cgroup_cancel_charge(pages[i], memcg);
1143                 put_page(pages[i]);
1144         }
1145         kfree(pages);
1146         goto out;
1147 }
1148
1149 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1150                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1151 {
1152         spinlock_t *ptl;
1153         int ret = 0;
1154         struct page *page = NULL, *new_page;
1155         struct mem_cgroup *memcg;
1156         unsigned long haddr;
1157         unsigned long mmun_start;       /* For mmu_notifiers */
1158         unsigned long mmun_end;         /* For mmu_notifiers */
1159         gfp_t huge_gfp;                 /* for allocation and charge */
1160
1161         ptl = pmd_lockptr(mm, pmd);
1162         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1163         haddr = address & HPAGE_PMD_MASK;
1164         if (is_huge_zero_pmd(orig_pmd))
1165                 goto alloc;
1166         spin_lock(ptl);
1167         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1168                 goto out_unlock;
1169
1170         page = pmd_page(orig_pmd);
1171         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1172         if (page_mapcount(page) == 1) {
1173                 pmd_t entry;
1174                 entry = pmd_mkyoung(orig_pmd);
1175                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1176                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1177                         update_mmu_cache_pmd(vma, address, pmd);
1178                 ret |= VM_FAULT_WRITE;
1179                 goto out_unlock;
1180         }
1181         get_user_huge_page(page);
1182         spin_unlock(ptl);
1183 alloc:
1184         if (transparent_hugepage_enabled(vma) &&
1185             !transparent_hugepage_debug_cow()) {
1186                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1187                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1188         } else
1189                 new_page = NULL;
1190
1191         if (unlikely(!new_page)) {
1192                 if (!page) {
1193                         split_huge_page_pmd(vma, address, pmd);
1194                         ret |= VM_FAULT_FALLBACK;
1195                 } else {
1196                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1197                                         pmd, orig_pmd, page, haddr);
1198                         if (ret & VM_FAULT_OOM) {
1199                                 split_huge_page(page);
1200                                 ret |= VM_FAULT_FALLBACK;
1201                         }
1202                         put_user_huge_page(page);
1203                 }
1204                 count_vm_event(THP_FAULT_FALLBACK);
1205                 goto out;
1206         }
1207
1208         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1209                 put_page(new_page);
1210                 if (page) {
1211                         split_huge_page(page);
1212                         put_user_huge_page(page);
1213                 } else
1214                         split_huge_page_pmd(vma, address, pmd);
1215                 ret |= VM_FAULT_FALLBACK;
1216                 count_vm_event(THP_FAULT_FALLBACK);
1217                 goto out;
1218         }
1219
1220         count_vm_event(THP_FAULT_ALLOC);
1221
1222         if (!page)
1223                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1224         else
1225                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1226         __SetPageUptodate(new_page);
1227
1228         mmun_start = haddr;
1229         mmun_end   = haddr + HPAGE_PMD_SIZE;
1230         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1231
1232         spin_lock(ptl);
1233         if (page)
1234                 put_user_huge_page(page);
1235         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1236                 spin_unlock(ptl);
1237                 mem_cgroup_cancel_charge(new_page, memcg);
1238                 put_page(new_page);
1239                 goto out_mn;
1240         } else {
1241                 pmd_t entry;
1242                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1243                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1244                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1245                 page_add_new_anon_rmap(new_page, vma, haddr);
1246                 mem_cgroup_commit_charge(new_page, memcg, false);
1247                 lru_cache_add_active_or_unevictable(new_page, vma);
1248                 set_pmd_at(mm, haddr, pmd, entry);
1249                 update_mmu_cache_pmd(vma, address, pmd);
1250                 if (!page) {
1251                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1252                         put_huge_zero_page();
1253                 } else {
1254                         VM_BUG_ON_PAGE(!PageHead(page), page);
1255                         page_remove_rmap(page);
1256                         put_page(page);
1257                 }
1258                 ret |= VM_FAULT_WRITE;
1259         }
1260         spin_unlock(ptl);
1261 out_mn:
1262         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1263 out:
1264         return ret;
1265 out_unlock:
1266         spin_unlock(ptl);
1267         return ret;
1268 }
1269
1270 /*
1271  * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1272  * but only after we've gone through a COW cycle and they are dirty.
1273  */
1274 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1275 {
1276         return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1277 }
1278
1279 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1280                                    unsigned long addr,
1281                                    pmd_t *pmd,
1282                                    unsigned int flags)
1283 {
1284         struct mm_struct *mm = vma->vm_mm;
1285         struct page *page = NULL;
1286
1287         assert_spin_locked(pmd_lockptr(mm, pmd));
1288
1289         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1290                 goto out;
1291
1292         /* Avoid dumping huge zero page */
1293         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1294                 return ERR_PTR(-EFAULT);
1295
1296         /* Full NUMA hinting faults to serialise migration in fault paths */
1297         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1298                 goto out;
1299
1300         page = pmd_page(*pmd);
1301         VM_BUG_ON_PAGE(!PageHead(page), page);
1302         if (flags & FOLL_TOUCH) {
1303                 pmd_t _pmd;
1304                 _pmd = pmd_mkyoung(*pmd);
1305                 if (flags & FOLL_WRITE)
1306                         _pmd = pmd_mkdirty(_pmd);
1307                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1308                                           pmd, _pmd, flags & FOLL_WRITE))
1309                         update_mmu_cache_pmd(vma, addr, pmd);
1310         }
1311         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1312                 if (page->mapping && trylock_page(page)) {
1313                         lru_add_drain();
1314                         if (page->mapping)
1315                                 mlock_vma_page(page);
1316                         unlock_page(page);
1317                 }
1318         }
1319         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1320         VM_BUG_ON_PAGE(!PageCompound(page), page);
1321         if (flags & FOLL_GET)
1322                 get_page_foll(page);
1323
1324 out:
1325         return page;
1326 }
1327
1328 /* NUMA hinting page fault entry point for trans huge pmds */
1329 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1330                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1331 {
1332         spinlock_t *ptl;
1333         struct anon_vma *anon_vma = NULL;
1334         struct page *page;
1335         unsigned long haddr = addr & HPAGE_PMD_MASK;
1336         int page_nid = -1, this_nid = numa_node_id();
1337         int target_nid, last_cpupid = -1;
1338         bool page_locked;
1339         bool migrated = false;
1340         bool was_writable;
1341         int flags = 0;
1342
1343         ptl = pmd_lock(mm, pmdp);
1344         if (unlikely(!pmd_same(pmd, *pmdp)))
1345                 goto out_unlock;
1346
1347         /*
1348          * If there are potential migrations, wait for completion and retry
1349          * without disrupting NUMA hinting information. Do not relock and
1350          * check_same as the page may no longer be mapped.
1351          */
1352         if (unlikely(pmd_trans_migrating(*pmdp))) {
1353                 page = pmd_page(*pmdp);
1354                 if (!get_page_unless_zero(page))
1355                         goto out_unlock;
1356                 spin_unlock(ptl);
1357                 wait_on_page_locked(page);
1358                 put_page(page);
1359                 goto out;
1360         }
1361
1362         page = pmd_page(pmd);
1363         BUG_ON(is_huge_zero_page(page));
1364         page_nid = page_to_nid(page);
1365         last_cpupid = page_cpupid_last(page);
1366         count_vm_numa_event(NUMA_HINT_FAULTS);
1367         if (page_nid == this_nid) {
1368                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1369                 flags |= TNF_FAULT_LOCAL;
1370         }
1371
1372         /* See similar comment in do_numa_page for explanation */
1373         if (!(vma->vm_flags & VM_WRITE))
1374                 flags |= TNF_NO_GROUP;
1375
1376         /*
1377          * Acquire the page lock to serialise THP migrations but avoid dropping
1378          * page_table_lock if at all possible
1379          */
1380         page_locked = trylock_page(page);
1381         target_nid = mpol_misplaced(page, vma, haddr);
1382         if (target_nid == -1) {
1383                 /* If the page was locked, there are no parallel migrations */
1384                 if (page_locked)
1385                         goto clear_pmdnuma;
1386         }
1387
1388         /* Migration could have started since the pmd_trans_migrating check */
1389         if (!page_locked) {
1390                 page_nid = -1;
1391                 if (!get_page_unless_zero(page))
1392                         goto out_unlock;
1393                 spin_unlock(ptl);
1394                 wait_on_page_locked(page);
1395                 put_page(page);
1396                 goto out;
1397         }
1398
1399         /*
1400          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1401          * to serialises splits
1402          */
1403         get_page(page);
1404         spin_unlock(ptl);
1405         anon_vma = page_lock_anon_vma_read(page);
1406
1407         /* Confirm the PMD did not change while page_table_lock was released */
1408         spin_lock(ptl);
1409         if (unlikely(!pmd_same(pmd, *pmdp))) {
1410                 unlock_page(page);
1411                 put_page(page);
1412                 page_nid = -1;
1413                 goto out_unlock;
1414         }
1415
1416         /* Bail if we fail to protect against THP splits for any reason */
1417         if (unlikely(!anon_vma)) {
1418                 put_page(page);
1419                 page_nid = -1;
1420                 goto clear_pmdnuma;
1421         }
1422
1423         /*
1424          * Migrate the THP to the requested node, returns with page unlocked
1425          * and access rights restored.
1426          */
1427         spin_unlock(ptl);
1428         migrated = migrate_misplaced_transhuge_page(mm, vma,
1429                                 pmdp, pmd, addr, page, target_nid);
1430         if (migrated) {
1431                 flags |= TNF_MIGRATED;
1432                 page_nid = target_nid;
1433         } else
1434                 flags |= TNF_MIGRATE_FAIL;
1435
1436         goto out;
1437 clear_pmdnuma:
1438         BUG_ON(!PageLocked(page));
1439         was_writable = pmd_write(pmd);
1440         pmd = pmd_modify(pmd, vma->vm_page_prot);
1441         pmd = pmd_mkyoung(pmd);
1442         if (was_writable)
1443                 pmd = pmd_mkwrite(pmd);
1444         set_pmd_at(mm, haddr, pmdp, pmd);
1445         update_mmu_cache_pmd(vma, addr, pmdp);
1446         unlock_page(page);
1447 out_unlock:
1448         spin_unlock(ptl);
1449
1450 out:
1451         if (anon_vma)
1452                 page_unlock_anon_vma_read(anon_vma);
1453
1454         if (page_nid != -1)
1455                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1456
1457         return 0;
1458 }
1459
1460 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1461                  pmd_t *pmd, unsigned long addr)
1462 {
1463         pmd_t orig_pmd;
1464         spinlock_t *ptl;
1465
1466         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1467                 return 0;
1468         /*
1469          * For architectures like ppc64 we look at deposited pgtable
1470          * when calling pmdp_huge_get_and_clear. So do the
1471          * pgtable_trans_huge_withdraw after finishing pmdp related
1472          * operations.
1473          */
1474         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1475                         tlb->fullmm);
1476         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1477         if (vma_is_dax(vma)) {
1478                 spin_unlock(ptl);
1479                 if (is_huge_zero_pmd(orig_pmd))
1480                         put_huge_zero_page();
1481         } else if (is_huge_zero_pmd(orig_pmd)) {
1482                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1483                 atomic_long_dec(&tlb->mm->nr_ptes);
1484                 spin_unlock(ptl);
1485                 put_huge_zero_page();
1486         } else {
1487                 struct page *page = pmd_page(orig_pmd);
1488                 page_remove_rmap(page);
1489                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1490                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1491                 VM_BUG_ON_PAGE(!PageHead(page), page);
1492                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1493                 atomic_long_dec(&tlb->mm->nr_ptes);
1494                 spin_unlock(ptl);
1495                 tlb_remove_page(tlb, page);
1496         }
1497         return 1;
1498 }
1499
1500 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1501                   unsigned long old_addr,
1502                   unsigned long new_addr, unsigned long old_end,
1503                   pmd_t *old_pmd, pmd_t *new_pmd)
1504 {
1505         spinlock_t *old_ptl, *new_ptl;
1506         int ret = 0;
1507         pmd_t pmd;
1508         bool force_flush = false;
1509         struct mm_struct *mm = vma->vm_mm;
1510
1511         if ((old_addr & ~HPAGE_PMD_MASK) ||
1512             (new_addr & ~HPAGE_PMD_MASK) ||
1513             old_end - old_addr < HPAGE_PMD_SIZE ||
1514             (new_vma->vm_flags & VM_NOHUGEPAGE))
1515                 goto out;
1516
1517         /*
1518          * The destination pmd shouldn't be established, free_pgtables()
1519          * should have release it.
1520          */
1521         if (WARN_ON(!pmd_none(*new_pmd))) {
1522                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1523                 goto out;
1524         }
1525
1526         /*
1527          * We don't have to worry about the ordering of src and dst
1528          * ptlocks because exclusive mmap_sem prevents deadlock.
1529          */
1530         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1531         if (ret == 1) {
1532                 new_ptl = pmd_lockptr(mm, new_pmd);
1533                 if (new_ptl != old_ptl)
1534                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1535                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1536                 if (pmd_present(pmd))
1537                         force_flush = true;
1538                 VM_BUG_ON(!pmd_none(*new_pmd));
1539
1540                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1541                         pgtable_t pgtable;
1542                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1543                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1544                 }
1545                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1546                 if (force_flush)
1547                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1548                 if (new_ptl != old_ptl)
1549                         spin_unlock(new_ptl);
1550                 spin_unlock(old_ptl);
1551         }
1552 out:
1553         return ret;
1554 }
1555
1556 /*
1557  * Returns
1558  *  - 0 if PMD could not be locked
1559  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1560  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1561  */
1562 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1563                 unsigned long addr, pgprot_t newprot, int prot_numa)
1564 {
1565         struct mm_struct *mm = vma->vm_mm;
1566         spinlock_t *ptl;
1567         pmd_t entry;
1568         bool preserve_write;
1569
1570         int ret = 0;
1571
1572         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1573                 return 0;
1574
1575         preserve_write = prot_numa && pmd_write(*pmd);
1576         ret = 1;
1577
1578         /*
1579          * Avoid trapping faults against the zero page. The read-only
1580          * data is likely to be read-cached on the local CPU and
1581          * local/remote hits to the zero page are not interesting.
1582          */
1583         if (prot_numa && is_huge_zero_pmd(*pmd))
1584                 goto unlock;
1585
1586         if (prot_numa && pmd_protnone(*pmd))
1587                 goto unlock;
1588
1589         /*
1590          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1591          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1592          * which is also under down_read(mmap_sem):
1593          *
1594          *      CPU0:                           CPU1:
1595          *                              change_huge_pmd(prot_numa=1)
1596          *                               pmdp_huge_get_and_clear_notify()
1597          * madvise_dontneed()
1598          *  zap_pmd_range()
1599          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1600          *   // skip the pmd
1601          *                               set_pmd_at();
1602          *                               // pmd is re-established
1603          *
1604          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1605          * which may break userspace.
1606          *
1607          * pmdp_invalidate() is required to make sure we don't miss
1608          * dirty/young flags set by hardware.
1609          */
1610         entry = *pmd;
1611         pmdp_invalidate(vma, addr, pmd);
1612
1613         /*
1614          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1615          * corrupt them.
1616          */
1617         if (pmd_dirty(*pmd))
1618                 entry = pmd_mkdirty(entry);
1619         if (pmd_young(*pmd))
1620                 entry = pmd_mkyoung(entry);
1621
1622         entry = pmd_modify(entry, newprot);
1623         if (preserve_write)
1624                 entry = pmd_mkwrite(entry);
1625         ret = HPAGE_PMD_NR;
1626         set_pmd_at(mm, addr, pmd, entry);
1627         BUG_ON(!preserve_write && pmd_write(entry));
1628 unlock:
1629         spin_unlock(ptl);
1630         return ret;
1631 }
1632
1633 /*
1634  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1635  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1636  *
1637  * Note that if it returns 1, this routine returns without unlocking page
1638  * table locks. So callers must unlock them.
1639  */
1640 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1641                 spinlock_t **ptl)
1642 {
1643         *ptl = pmd_lock(vma->vm_mm, pmd);
1644         if (likely(pmd_trans_huge(*pmd))) {
1645                 if (unlikely(pmd_trans_splitting(*pmd))) {
1646                         spin_unlock(*ptl);
1647                         wait_split_huge_page(vma->anon_vma, pmd);
1648                         return -1;
1649                 } else {
1650                         /* Thp mapped by 'pmd' is stable, so we can
1651                          * handle it as it is. */
1652                         return 1;
1653                 }
1654         }
1655         spin_unlock(*ptl);
1656         return 0;
1657 }
1658
1659 /*
1660  * This function returns whether a given @page is mapped onto the @address
1661  * in the virtual space of @mm.
1662  *
1663  * When it's true, this function returns *pmd with holding the page table lock
1664  * and passing it back to the caller via @ptl.
1665  * If it's false, returns NULL without holding the page table lock.
1666  */
1667 pmd_t *page_check_address_pmd(struct page *page,
1668                               struct mm_struct *mm,
1669                               unsigned long address,
1670                               enum page_check_address_pmd_flag flag,
1671                               spinlock_t **ptl)
1672 {
1673         pgd_t *pgd;
1674         pud_t *pud;
1675         pmd_t *pmd;
1676
1677         if (address & ~HPAGE_PMD_MASK)
1678                 return NULL;
1679
1680         pgd = pgd_offset(mm, address);
1681         if (!pgd_present(*pgd))
1682                 return NULL;
1683         pud = pud_offset(pgd, address);
1684         if (!pud_present(*pud))
1685                 return NULL;
1686         pmd = pmd_offset(pud, address);
1687
1688         *ptl = pmd_lock(mm, pmd);
1689         if (!pmd_present(*pmd))
1690                 goto unlock;
1691         if (pmd_page(*pmd) != page)
1692                 goto unlock;
1693         /*
1694          * split_vma() may create temporary aliased mappings. There is
1695          * no risk as long as all huge pmd are found and have their
1696          * splitting bit set before __split_huge_page_refcount
1697          * runs. Finding the same huge pmd more than once during the
1698          * same rmap walk is not a problem.
1699          */
1700         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1701             pmd_trans_splitting(*pmd))
1702                 goto unlock;
1703         if (pmd_trans_huge(*pmd)) {
1704                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1705                           !pmd_trans_splitting(*pmd));
1706                 return pmd;
1707         }
1708 unlock:
1709         spin_unlock(*ptl);
1710         return NULL;
1711 }
1712
1713 static int __split_huge_page_splitting(struct page *page,
1714                                        struct vm_area_struct *vma,
1715                                        unsigned long address)
1716 {
1717         struct mm_struct *mm = vma->vm_mm;
1718         spinlock_t *ptl;
1719         pmd_t *pmd;
1720         int ret = 0;
1721         /* For mmu_notifiers */
1722         const unsigned long mmun_start = address;
1723         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1724
1725         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1726         pmd = page_check_address_pmd(page, mm, address,
1727                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1728         if (pmd) {
1729                 /*
1730                  * We can't temporarily set the pmd to null in order
1731                  * to split it, the pmd must remain marked huge at all
1732                  * times or the VM won't take the pmd_trans_huge paths
1733                  * and it won't wait on the anon_vma->root->rwsem to
1734                  * serialize against split_huge_page*.
1735                  */
1736                 pmdp_splitting_flush(vma, address, pmd);
1737
1738                 ret = 1;
1739                 spin_unlock(ptl);
1740         }
1741         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1742
1743         return ret;
1744 }
1745
1746 static void __split_huge_page_refcount(struct page *page,
1747                                        struct list_head *list)
1748 {
1749         int i;
1750         struct zone *zone = page_zone(page);
1751         struct lruvec *lruvec;
1752         int tail_count = 0;
1753
1754         /* prevent PageLRU to go away from under us, and freeze lru stats */
1755         spin_lock_irq(&zone->lru_lock);
1756         lruvec = mem_cgroup_page_lruvec(page, zone);
1757
1758         compound_lock(page);
1759         /* complete memcg works before add pages to LRU */
1760         mem_cgroup_split_huge_fixup(page);
1761
1762         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1763                 struct page *page_tail = page + i;
1764
1765                 /* tail_page->_mapcount cannot change */
1766                 BUG_ON(page_mapcount(page_tail) < 0);
1767                 tail_count += page_mapcount(page_tail);
1768                 /* check for overflow */
1769                 BUG_ON(tail_count < 0);
1770                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1771                 /*
1772                  * tail_page->_count is zero and not changing from
1773                  * under us. But get_page_unless_zero() may be running
1774                  * from under us on the tail_page. If we used
1775                  * atomic_set() below instead of atomic_add(), we
1776                  * would then run atomic_set() concurrently with
1777                  * get_page_unless_zero(), and atomic_set() is
1778                  * implemented in C not using locked ops. spin_unlock
1779                  * on x86 sometime uses locked ops because of PPro
1780                  * errata 66, 92, so unless somebody can guarantee
1781                  * atomic_set() here would be safe on all archs (and
1782                  * not only on x86), it's safer to use atomic_add().
1783                  */
1784                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1785                            &page_tail->_count);
1786
1787                 /* after clearing PageTail the gup refcount can be released */
1788                 smp_mb__after_atomic();
1789
1790                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1791                 page_tail->flags |= (page->flags &
1792                                      ((1L << PG_referenced) |
1793                                       (1L << PG_swapbacked) |
1794                                       (1L << PG_mlocked) |
1795                                       (1L << PG_uptodate) |
1796                                       (1L << PG_active) |
1797                                       (1L << PG_unevictable)));
1798                 page_tail->flags |= (1L << PG_dirty);
1799
1800                 clear_compound_head(page_tail);
1801
1802                 if (page_is_young(page))
1803                         set_page_young(page_tail);
1804                 if (page_is_idle(page))
1805                         set_page_idle(page_tail);
1806
1807                 /*
1808                  * __split_huge_page_splitting() already set the
1809                  * splitting bit in all pmd that could map this
1810                  * hugepage, that will ensure no CPU can alter the
1811                  * mapcount on the head page. The mapcount is only
1812                  * accounted in the head page and it has to be
1813                  * transferred to all tail pages in the below code. So
1814                  * for this code to be safe, the split the mapcount
1815                  * can't change. But that doesn't mean userland can't
1816                  * keep changing and reading the page contents while
1817                  * we transfer the mapcount, so the pmd splitting
1818                  * status is achieved setting a reserved bit in the
1819                  * pmd, not by clearing the present bit.
1820                 */
1821                 page_tail->_mapcount = page->_mapcount;
1822
1823                 BUG_ON(page_tail->mapping);
1824                 page_tail->mapping = page->mapping;
1825
1826                 page_tail->index = page->index + i;
1827                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1828
1829                 BUG_ON(!PageAnon(page_tail));
1830                 BUG_ON(!PageUptodate(page_tail));
1831                 BUG_ON(!PageDirty(page_tail));
1832                 BUG_ON(!PageSwapBacked(page_tail));
1833
1834                 lru_add_page_tail(page, page_tail, lruvec, list);
1835         }
1836         atomic_sub(tail_count, &page->_count);
1837         BUG_ON(atomic_read(&page->_count) <= 0);
1838
1839         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1840
1841         ClearPageCompound(page);
1842         compound_unlock(page);
1843         spin_unlock_irq(&zone->lru_lock);
1844
1845         for (i = 1; i < HPAGE_PMD_NR; i++) {
1846                 struct page *page_tail = page + i;
1847                 BUG_ON(page_count(page_tail) <= 0);
1848                 /*
1849                  * Tail pages may be freed if there wasn't any mapping
1850                  * like if add_to_swap() is running on a lru page that
1851                  * had its mapping zapped. And freeing these pages
1852                  * requires taking the lru_lock so we do the put_page
1853                  * of the tail pages after the split is complete.
1854                  */
1855                 put_page(page_tail);
1856         }
1857
1858         /*
1859          * Only the head page (now become a regular page) is required
1860          * to be pinned by the caller.
1861          */
1862         BUG_ON(page_count(page) <= 0);
1863 }
1864
1865 static int __split_huge_page_map(struct page *page,
1866                                  struct vm_area_struct *vma,
1867                                  unsigned long address)
1868 {
1869         struct mm_struct *mm = vma->vm_mm;
1870         spinlock_t *ptl;
1871         pmd_t *pmd, _pmd;
1872         int ret = 0, i;
1873         pgtable_t pgtable;
1874         unsigned long haddr;
1875
1876         pmd = page_check_address_pmd(page, mm, address,
1877                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1878         if (pmd) {
1879                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1880                 pmd_populate(mm, &_pmd, pgtable);
1881                 if (pmd_write(*pmd))
1882                         BUG_ON(page_mapcount(page) != 1);
1883
1884                 haddr = address;
1885                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1886                         pte_t *pte, entry;
1887                         BUG_ON(PageCompound(page+i));
1888                         /*
1889                          * Note that NUMA hinting access restrictions are not
1890                          * transferred to avoid any possibility of altering
1891                          * permissions across VMAs.
1892                          */
1893                         entry = mk_pte(page + i, vma->vm_page_prot);
1894                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1895                         if (!pmd_write(*pmd))
1896                                 entry = pte_wrprotect(entry);
1897                         if (!pmd_young(*pmd))
1898                                 entry = pte_mkold(entry);
1899                         pte = pte_offset_map(&_pmd, haddr);
1900                         BUG_ON(!pte_none(*pte));
1901                         set_pte_at(mm, haddr, pte, entry);
1902                         pte_unmap(pte);
1903                 }
1904
1905                 smp_wmb(); /* make pte visible before pmd */
1906                 /*
1907                  * Up to this point the pmd is present and huge and
1908                  * userland has the whole access to the hugepage
1909                  * during the split (which happens in place). If we
1910                  * overwrite the pmd with the not-huge version
1911                  * pointing to the pte here (which of course we could
1912                  * if all CPUs were bug free), userland could trigger
1913                  * a small page size TLB miss on the small sized TLB
1914                  * while the hugepage TLB entry is still established
1915                  * in the huge TLB. Some CPU doesn't like that. See
1916                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1917                  * Erratum 383 on page 93. Intel should be safe but is
1918                  * also warns that it's only safe if the permission
1919                  * and cache attributes of the two entries loaded in
1920                  * the two TLB is identical (which should be the case
1921                  * here). But it is generally safer to never allow
1922                  * small and huge TLB entries for the same virtual
1923                  * address to be loaded simultaneously. So instead of
1924                  * doing "pmd_populate(); flush_pmd_tlb_range();" we first
1925                  * mark the current pmd notpresent (atomically because
1926                  * here the pmd_trans_huge and pmd_trans_splitting
1927                  * must remain set at all times on the pmd until the
1928                  * split is complete for this pmd), then we flush the
1929                  * SMP TLB and finally we write the non-huge version
1930                  * of the pmd entry with pmd_populate.
1931                  */
1932                 pmdp_invalidate(vma, address, pmd);
1933                 pmd_populate(mm, pmd, pgtable);
1934                 ret = 1;
1935                 spin_unlock(ptl);
1936         }
1937
1938         return ret;
1939 }
1940
1941 /* must be called with anon_vma->root->rwsem held */
1942 static void __split_huge_page(struct page *page,
1943                               struct anon_vma *anon_vma,
1944                               struct list_head *list)
1945 {
1946         int mapcount, mapcount2;
1947         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1948         struct anon_vma_chain *avc;
1949
1950         BUG_ON(!PageHead(page));
1951         BUG_ON(PageTail(page));
1952
1953         mapcount = 0;
1954         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1955                 struct vm_area_struct *vma = avc->vma;
1956                 unsigned long addr = vma_address(page, vma);
1957                 BUG_ON(is_vma_temporary_stack(vma));
1958                 mapcount += __split_huge_page_splitting(page, vma, addr);
1959         }
1960         /*
1961          * It is critical that new vmas are added to the tail of the
1962          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1963          * and establishes a child pmd before
1964          * __split_huge_page_splitting() freezes the parent pmd (so if
1965          * we fail to prevent copy_huge_pmd() from running until the
1966          * whole __split_huge_page() is complete), we will still see
1967          * the newly established pmd of the child later during the
1968          * walk, to be able to set it as pmd_trans_splitting too.
1969          */
1970         if (mapcount != page_mapcount(page)) {
1971                 pr_err("mapcount %d page_mapcount %d\n",
1972                         mapcount, page_mapcount(page));
1973                 BUG();
1974         }
1975
1976         __split_huge_page_refcount(page, list);
1977
1978         mapcount2 = 0;
1979         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1980                 struct vm_area_struct *vma = avc->vma;
1981                 unsigned long addr = vma_address(page, vma);
1982                 BUG_ON(is_vma_temporary_stack(vma));
1983                 mapcount2 += __split_huge_page_map(page, vma, addr);
1984         }
1985         if (mapcount != mapcount2) {
1986                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1987                         mapcount, mapcount2, page_mapcount(page));
1988                 BUG();
1989         }
1990 }
1991
1992 /*
1993  * Split a hugepage into normal pages. This doesn't change the position of head
1994  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1995  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1996  * from the hugepage.
1997  * Return 0 if the hugepage is split successfully otherwise return 1.
1998  */
1999 int split_huge_page_to_list(struct page *page, struct list_head *list)
2000 {
2001         struct anon_vma *anon_vma;
2002         int ret = 1;
2003
2004         BUG_ON(is_huge_zero_page(page));
2005         BUG_ON(!PageAnon(page));
2006
2007         /*
2008          * The caller does not necessarily hold an mmap_sem that would prevent
2009          * the anon_vma disappearing so we first we take a reference to it
2010          * and then lock the anon_vma for write. This is similar to
2011          * page_lock_anon_vma_read except the write lock is taken to serialise
2012          * against parallel split or collapse operations.
2013          */
2014         anon_vma = page_get_anon_vma(page);
2015         if (!anon_vma)
2016                 goto out;
2017         anon_vma_lock_write(anon_vma);
2018
2019         ret = 0;
2020         if (!PageCompound(page))
2021                 goto out_unlock;
2022
2023         BUG_ON(!PageSwapBacked(page));
2024         __split_huge_page(page, anon_vma, list);
2025         count_vm_event(THP_SPLIT);
2026
2027         BUG_ON(PageCompound(page));
2028 out_unlock:
2029         anon_vma_unlock_write(anon_vma);
2030         put_anon_vma(anon_vma);
2031 out:
2032         return ret;
2033 }
2034
2035 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
2036
2037 int hugepage_madvise(struct vm_area_struct *vma,
2038                      unsigned long *vm_flags, int advice)
2039 {
2040         switch (advice) {
2041         case MADV_HUGEPAGE:
2042 #ifdef CONFIG_S390
2043                 /*
2044                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2045                  * can't handle this properly after s390_enable_sie, so we simply
2046                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
2047                  */
2048                 if (mm_has_pgste(vma->vm_mm))
2049                         return 0;
2050 #endif
2051                 /*
2052                  * Be somewhat over-protective like KSM for now!
2053                  */
2054                 if (*vm_flags & VM_NO_THP)
2055                         return -EINVAL;
2056                 *vm_flags &= ~VM_NOHUGEPAGE;
2057                 *vm_flags |= VM_HUGEPAGE;
2058                 /*
2059                  * If the vma become good for khugepaged to scan,
2060                  * register it here without waiting a page fault that
2061                  * may not happen any time soon.
2062                  */
2063                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2064                         return -ENOMEM;
2065                 break;
2066         case MADV_NOHUGEPAGE:
2067                 /*
2068                  * Be somewhat over-protective like KSM for now!
2069                  */
2070                 if (*vm_flags & VM_NO_THP)
2071                         return -EINVAL;
2072                 *vm_flags &= ~VM_HUGEPAGE;
2073                 *vm_flags |= VM_NOHUGEPAGE;
2074                 /*
2075                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2076                  * this vma even if we leave the mm registered in khugepaged if
2077                  * it got registered before VM_NOHUGEPAGE was set.
2078                  */
2079                 break;
2080         }
2081
2082         return 0;
2083 }
2084
2085 static int __init khugepaged_slab_init(void)
2086 {
2087         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2088                                           sizeof(struct mm_slot),
2089                                           __alignof__(struct mm_slot), 0, NULL);
2090         if (!mm_slot_cache)
2091                 return -ENOMEM;
2092
2093         return 0;
2094 }
2095
2096 static void __init khugepaged_slab_exit(void)
2097 {
2098         kmem_cache_destroy(mm_slot_cache);
2099 }
2100
2101 static inline struct mm_slot *alloc_mm_slot(void)
2102 {
2103         if (!mm_slot_cache)     /* initialization failed */
2104                 return NULL;
2105         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2106 }
2107
2108 static inline void free_mm_slot(struct mm_slot *mm_slot)
2109 {
2110         kmem_cache_free(mm_slot_cache, mm_slot);
2111 }
2112
2113 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2114 {
2115         struct mm_slot *mm_slot;
2116
2117         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2118                 if (mm == mm_slot->mm)
2119                         return mm_slot;
2120
2121         return NULL;
2122 }
2123
2124 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2125                                     struct mm_slot *mm_slot)
2126 {
2127         mm_slot->mm = mm;
2128         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2129 }
2130
2131 static inline int khugepaged_test_exit(struct mm_struct *mm)
2132 {
2133         return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
2134 }
2135
2136 int __khugepaged_enter(struct mm_struct *mm)
2137 {
2138         struct mm_slot *mm_slot;
2139         int wakeup;
2140
2141         mm_slot = alloc_mm_slot();
2142         if (!mm_slot)
2143                 return -ENOMEM;
2144
2145         /* __khugepaged_exit() must not run from under us */
2146         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
2147         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2148                 free_mm_slot(mm_slot);
2149                 return 0;
2150         }
2151
2152         spin_lock(&khugepaged_mm_lock);
2153         insert_to_mm_slots_hash(mm, mm_slot);
2154         /*
2155          * Insert just behind the scanning cursor, to let the area settle
2156          * down a little.
2157          */
2158         wakeup = list_empty(&khugepaged_scan.mm_head);
2159         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2160         spin_unlock(&khugepaged_mm_lock);
2161
2162         atomic_inc(&mm->mm_count);
2163         if (wakeup)
2164                 wake_up_interruptible(&khugepaged_wait);
2165
2166         return 0;
2167 }
2168
2169 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2170                                unsigned long vm_flags)
2171 {
2172         unsigned long hstart, hend;
2173         if (!vma->anon_vma)
2174                 /*
2175                  * Not yet faulted in so we will register later in the
2176                  * page fault if needed.
2177                  */
2178                 return 0;
2179         if (vma->vm_ops || (vm_flags & VM_NO_THP))
2180                 /* khugepaged not yet working on file or special mappings */
2181                 return 0;
2182         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2183         hend = vma->vm_end & HPAGE_PMD_MASK;
2184         if (hstart < hend)
2185                 return khugepaged_enter(vma, vm_flags);
2186         return 0;
2187 }
2188
2189 void __khugepaged_exit(struct mm_struct *mm)
2190 {
2191         struct mm_slot *mm_slot;
2192         int free = 0;
2193
2194         spin_lock(&khugepaged_mm_lock);
2195         mm_slot = get_mm_slot(mm);
2196         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2197                 hash_del(&mm_slot->hash);
2198                 list_del(&mm_slot->mm_node);
2199                 free = 1;
2200         }
2201         spin_unlock(&khugepaged_mm_lock);
2202
2203         if (free) {
2204                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2205                 free_mm_slot(mm_slot);
2206                 mmdrop(mm);
2207         } else if (mm_slot) {
2208                 /*
2209                  * This is required to serialize against
2210                  * khugepaged_test_exit() (which is guaranteed to run
2211                  * under mmap sem read mode). Stop here (after we
2212                  * return all pagetables will be destroyed) until
2213                  * khugepaged has finished working on the pagetables
2214                  * under the mmap_sem.
2215                  */
2216                 down_write(&mm->mmap_sem);
2217                 up_write(&mm->mmap_sem);
2218         }
2219 }
2220
2221 static void release_pte_page(struct page *page)
2222 {
2223         /* 0 stands for page_is_file_cache(page) == false */
2224         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2225         unlock_page(page);
2226         putback_lru_page(page);
2227 }
2228
2229 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2230 {
2231         while (--_pte >= pte) {
2232                 pte_t pteval = *_pte;
2233                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2234                         release_pte_page(pte_page(pteval));
2235         }
2236 }
2237
2238 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2239                                         unsigned long address,
2240                                         pte_t *pte)
2241 {
2242         struct page *page;
2243         pte_t *_pte;
2244         int none_or_zero = 0;
2245         bool referenced = false, writable = false;
2246         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2247              _pte++, address += PAGE_SIZE) {
2248                 pte_t pteval = *_pte;
2249                 if (pte_none(pteval) || (pte_present(pteval) &&
2250                                 is_zero_pfn(pte_pfn(pteval)))) {
2251                         if (!userfaultfd_armed(vma) &&
2252                             ++none_or_zero <= khugepaged_max_ptes_none)
2253                                 continue;
2254                         else
2255                                 goto out;
2256                 }
2257                 if (!pte_present(pteval))
2258                         goto out;
2259                 page = vm_normal_page(vma, address, pteval);
2260                 if (unlikely(!page))
2261                         goto out;
2262
2263                 VM_BUG_ON_PAGE(PageCompound(page), page);
2264                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2265                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2266
2267                 /*
2268                  * We can do it before isolate_lru_page because the
2269                  * page can't be freed from under us. NOTE: PG_lock
2270                  * is needed to serialize against split_huge_page
2271                  * when invoked from the VM.
2272                  */
2273                 if (!trylock_page(page))
2274                         goto out;
2275
2276                 /*
2277                  * cannot use mapcount: can't collapse if there's a gup pin.
2278                  * The page must only be referenced by the scanned process
2279                  * and page swap cache.
2280                  */
2281                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2282                         unlock_page(page);
2283                         goto out;
2284                 }
2285                 if (pte_write(pteval)) {
2286                         writable = true;
2287                 } else {
2288                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2289                                 unlock_page(page);
2290                                 goto out;
2291                         }
2292                         /*
2293                          * Page is not in the swap cache. It can be collapsed
2294                          * into a THP.
2295                          */
2296                 }
2297
2298                 /*
2299                  * Isolate the page to avoid collapsing an hugepage
2300                  * currently in use by the VM.
2301                  */
2302                 if (isolate_lru_page(page)) {
2303                         unlock_page(page);
2304                         goto out;
2305                 }
2306                 /* 0 stands for page_is_file_cache(page) == false */
2307                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2308                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2309                 VM_BUG_ON_PAGE(PageLRU(page), page);
2310
2311                 /* If there is no mapped pte young don't collapse the page */
2312                 if (pte_young(pteval) ||
2313                     page_is_young(page) || PageReferenced(page) ||
2314                     mmu_notifier_test_young(vma->vm_mm, address))
2315                         referenced = true;
2316         }
2317         if (likely(referenced && writable))
2318                 return 1;
2319 out:
2320         release_pte_pages(pte, _pte);
2321         return 0;
2322 }
2323
2324 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2325                                       struct vm_area_struct *vma,
2326                                       unsigned long address,
2327                                       spinlock_t *ptl)
2328 {
2329         pte_t *_pte;
2330         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2331                 pte_t pteval = *_pte;
2332                 struct page *src_page;
2333
2334                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2335                         clear_user_highpage(page, address);
2336                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2337                         if (is_zero_pfn(pte_pfn(pteval))) {
2338                                 /*
2339                                  * ptl mostly unnecessary.
2340                                  */
2341                                 spin_lock(ptl);
2342                                 /*
2343                                  * paravirt calls inside pte_clear here are
2344                                  * superfluous.
2345                                  */
2346                                 pte_clear(vma->vm_mm, address, _pte);
2347                                 spin_unlock(ptl);
2348                         }
2349                 } else {
2350                         src_page = pte_page(pteval);
2351                         copy_user_highpage(page, src_page, address, vma);
2352                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2353                         release_pte_page(src_page);
2354                         /*
2355                          * ptl mostly unnecessary, but preempt has to
2356                          * be disabled to update the per-cpu stats
2357                          * inside page_remove_rmap().
2358                          */
2359                         spin_lock(ptl);
2360                         /*
2361                          * paravirt calls inside pte_clear here are
2362                          * superfluous.
2363                          */
2364                         pte_clear(vma->vm_mm, address, _pte);
2365                         page_remove_rmap(src_page);
2366                         spin_unlock(ptl);
2367                         free_page_and_swap_cache(src_page);
2368                 }
2369
2370                 address += PAGE_SIZE;
2371                 page++;
2372         }
2373 }
2374
2375 static void khugepaged_alloc_sleep(void)
2376 {
2377         DEFINE_WAIT(wait);
2378
2379         add_wait_queue(&khugepaged_wait, &wait);
2380         freezable_schedule_timeout_interruptible(
2381                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2382         remove_wait_queue(&khugepaged_wait, &wait);
2383 }
2384
2385 static int khugepaged_node_load[MAX_NUMNODES];
2386
2387 static bool khugepaged_scan_abort(int nid)
2388 {
2389         int i;
2390
2391         /*
2392          * If zone_reclaim_mode is disabled, then no extra effort is made to
2393          * allocate memory locally.
2394          */
2395         if (!zone_reclaim_mode)
2396                 return false;
2397
2398         /* If there is a count for this node already, it must be acceptable */
2399         if (khugepaged_node_load[nid])
2400                 return false;
2401
2402         for (i = 0; i < MAX_NUMNODES; i++) {
2403                 if (!khugepaged_node_load[i])
2404                         continue;
2405                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2406                         return true;
2407         }
2408         return false;
2409 }
2410
2411 #ifdef CONFIG_NUMA
2412 static int khugepaged_find_target_node(void)
2413 {
2414         static int last_khugepaged_target_node = NUMA_NO_NODE;
2415         int nid, target_node = 0, max_value = 0;
2416
2417         /* find first node with max normal pages hit */
2418         for (nid = 0; nid < MAX_NUMNODES; nid++)
2419                 if (khugepaged_node_load[nid] > max_value) {
2420                         max_value = khugepaged_node_load[nid];
2421                         target_node = nid;
2422                 }
2423
2424         /* do some balance if several nodes have the same hit record */
2425         if (target_node <= last_khugepaged_target_node)
2426                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2427                                 nid++)
2428                         if (max_value == khugepaged_node_load[nid]) {
2429                                 target_node = nid;
2430                                 break;
2431                         }
2432
2433         last_khugepaged_target_node = target_node;
2434         return target_node;
2435 }
2436
2437 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2438 {
2439         if (IS_ERR(*hpage)) {
2440                 if (!*wait)
2441                         return false;
2442
2443                 *wait = false;
2444                 *hpage = NULL;
2445                 khugepaged_alloc_sleep();
2446         } else if (*hpage) {
2447                 put_page(*hpage);
2448                 *hpage = NULL;
2449         }
2450
2451         return true;
2452 }
2453
2454 static struct page *
2455 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2456                        unsigned long address, int node)
2457 {
2458         VM_BUG_ON_PAGE(*hpage, *hpage);
2459
2460         /*
2461          * Before allocating the hugepage, release the mmap_sem read lock.
2462          * The allocation can take potentially a long time if it involves
2463          * sync compaction, and we do not need to hold the mmap_sem during
2464          * that. We will recheck the vma after taking it again in write mode.
2465          */
2466         up_read(&mm->mmap_sem);
2467
2468         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2469         if (unlikely(!*hpage)) {
2470                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2471                 *hpage = ERR_PTR(-ENOMEM);
2472                 return NULL;
2473         }
2474
2475         count_vm_event(THP_COLLAPSE_ALLOC);
2476         return *hpage;
2477 }
2478 #else
2479 static int khugepaged_find_target_node(void)
2480 {
2481         return 0;
2482 }
2483
2484 static inline struct page *alloc_hugepage(int defrag)
2485 {
2486         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2487                            HPAGE_PMD_ORDER);
2488 }
2489
2490 static struct page *khugepaged_alloc_hugepage(bool *wait)
2491 {
2492         struct page *hpage;
2493
2494         do {
2495                 hpage = alloc_hugepage(khugepaged_defrag());
2496                 if (!hpage) {
2497                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2498                         if (!*wait)
2499                                 return NULL;
2500
2501                         *wait = false;
2502                         khugepaged_alloc_sleep();
2503                 } else
2504                         count_vm_event(THP_COLLAPSE_ALLOC);
2505         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2506
2507         return hpage;
2508 }
2509
2510 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2511 {
2512         if (!*hpage)
2513                 *hpage = khugepaged_alloc_hugepage(wait);
2514
2515         if (unlikely(!*hpage))
2516                 return false;
2517
2518         return true;
2519 }
2520
2521 static struct page *
2522 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2523                        unsigned long address, int node)
2524 {
2525         up_read(&mm->mmap_sem);
2526         VM_BUG_ON(!*hpage);
2527
2528         return  *hpage;
2529 }
2530 #endif
2531
2532 static bool hugepage_vma_check(struct vm_area_struct *vma)
2533 {
2534         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2535             (vma->vm_flags & VM_NOHUGEPAGE))
2536                 return false;
2537
2538         if (!vma->anon_vma || vma->vm_ops)
2539                 return false;
2540         if (is_vma_temporary_stack(vma))
2541                 return false;
2542         return !(vma->vm_flags & VM_NO_THP);
2543 }
2544
2545 static void collapse_huge_page(struct mm_struct *mm,
2546                                    unsigned long address,
2547                                    struct page **hpage,
2548                                    struct vm_area_struct *vma,
2549                                    int node)
2550 {
2551         pmd_t *pmd, _pmd;
2552         pte_t *pte;
2553         pgtable_t pgtable;
2554         struct page *new_page;
2555         spinlock_t *pmd_ptl, *pte_ptl;
2556         int isolated;
2557         unsigned long hstart, hend;
2558         struct mem_cgroup *memcg;
2559         unsigned long mmun_start;       /* For mmu_notifiers */
2560         unsigned long mmun_end;         /* For mmu_notifiers */
2561         gfp_t gfp;
2562
2563         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2564
2565         /* Only allocate from the target node */
2566         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2567                 __GFP_THISNODE;
2568
2569         /* release the mmap_sem read lock. */
2570         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2571         if (!new_page)
2572                 return;
2573
2574         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2575                                            gfp, &memcg)))
2576                 return;
2577
2578         /*
2579          * Prevent all access to pagetables with the exception of
2580          * gup_fast later hanlded by the ptep_clear_flush and the VM
2581          * handled by the anon_vma lock + PG_lock.
2582          */
2583         down_write(&mm->mmap_sem);
2584         if (unlikely(khugepaged_test_exit(mm)))
2585                 goto out;
2586
2587         vma = find_vma(mm, address);
2588         if (!vma)
2589                 goto out;
2590         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2591         hend = vma->vm_end & HPAGE_PMD_MASK;
2592         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2593                 goto out;
2594         if (!hugepage_vma_check(vma))
2595                 goto out;
2596         pmd = mm_find_pmd(mm, address);
2597         if (!pmd)
2598                 goto out;
2599
2600         anon_vma_lock_write(vma->anon_vma);
2601
2602         pte = pte_offset_map(pmd, address);
2603         pte_ptl = pte_lockptr(mm, pmd);
2604
2605         mmun_start = address;
2606         mmun_end   = address + HPAGE_PMD_SIZE;
2607         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2608         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2609         /*
2610          * After this gup_fast can't run anymore. This also removes
2611          * any huge TLB entry from the CPU so we won't allow
2612          * huge and small TLB entries for the same virtual address
2613          * to avoid the risk of CPU bugs in that area.
2614          */
2615         _pmd = pmdp_collapse_flush(vma, address, pmd);
2616         spin_unlock(pmd_ptl);
2617         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2618
2619         spin_lock(pte_ptl);
2620         isolated = __collapse_huge_page_isolate(vma, address, pte);
2621         spin_unlock(pte_ptl);
2622
2623         if (unlikely(!isolated)) {
2624                 pte_unmap(pte);
2625                 spin_lock(pmd_ptl);
2626                 BUG_ON(!pmd_none(*pmd));
2627                 /*
2628                  * We can only use set_pmd_at when establishing
2629                  * hugepmds and never for establishing regular pmds that
2630                  * points to regular pagetables. Use pmd_populate for that
2631                  */
2632                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2633                 spin_unlock(pmd_ptl);
2634                 anon_vma_unlock_write(vma->anon_vma);
2635                 goto out;
2636         }
2637
2638         /*
2639          * All pages are isolated and locked so anon_vma rmap
2640          * can't run anymore.
2641          */
2642         anon_vma_unlock_write(vma->anon_vma);
2643
2644         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2645         pte_unmap(pte);
2646         __SetPageUptodate(new_page);
2647         pgtable = pmd_pgtable(_pmd);
2648
2649         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2650         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2651
2652         /*
2653          * spin_lock() below is not the equivalent of smp_wmb(), so
2654          * this is needed to avoid the copy_huge_page writes to become
2655          * visible after the set_pmd_at() write.
2656          */
2657         smp_wmb();
2658
2659         spin_lock(pmd_ptl);
2660         BUG_ON(!pmd_none(*pmd));
2661         page_add_new_anon_rmap(new_page, vma, address);
2662         mem_cgroup_commit_charge(new_page, memcg, false);
2663         lru_cache_add_active_or_unevictable(new_page, vma);
2664         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2665         set_pmd_at(mm, address, pmd, _pmd);
2666         update_mmu_cache_pmd(vma, address, pmd);
2667         spin_unlock(pmd_ptl);
2668
2669         *hpage = NULL;
2670
2671         khugepaged_pages_collapsed++;
2672 out_up_write:
2673         up_write(&mm->mmap_sem);
2674         return;
2675
2676 out:
2677         mem_cgroup_cancel_charge(new_page, memcg);
2678         goto out_up_write;
2679 }
2680
2681 static int khugepaged_scan_pmd(struct mm_struct *mm,
2682                                struct vm_area_struct *vma,
2683                                unsigned long address,
2684                                struct page **hpage)
2685 {
2686         pmd_t *pmd;
2687         pte_t *pte, *_pte;
2688         int ret = 0, none_or_zero = 0;
2689         struct page *page;
2690         unsigned long _address;
2691         spinlock_t *ptl;
2692         int node = NUMA_NO_NODE;
2693         bool writable = false, referenced = false;
2694
2695         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2696
2697         pmd = mm_find_pmd(mm, address);
2698         if (!pmd)
2699                 goto out;
2700
2701         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2702         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2703         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2704              _pte++, _address += PAGE_SIZE) {
2705                 pte_t pteval = *_pte;
2706                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2707                         if (!userfaultfd_armed(vma) &&
2708                             ++none_or_zero <= khugepaged_max_ptes_none)
2709                                 continue;
2710                         else
2711                                 goto out_unmap;
2712                 }
2713                 if (!pte_present(pteval))
2714                         goto out_unmap;
2715                 if (pte_write(pteval))
2716                         writable = true;
2717
2718                 page = vm_normal_page(vma, _address, pteval);
2719                 if (unlikely(!page))
2720                         goto out_unmap;
2721                 /*
2722                  * Record which node the original page is from and save this
2723                  * information to khugepaged_node_load[].
2724                  * Khupaged will allocate hugepage from the node has the max
2725                  * hit record.
2726                  */
2727                 node = page_to_nid(page);
2728                 if (khugepaged_scan_abort(node))
2729                         goto out_unmap;
2730                 khugepaged_node_load[node]++;
2731                 VM_BUG_ON_PAGE(PageCompound(page), page);
2732                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2733                         goto out_unmap;
2734                 /*
2735                  * cannot use mapcount: can't collapse if there's a gup pin.
2736                  * The page must only be referenced by the scanned process
2737                  * and page swap cache.
2738                  */
2739                 if (page_count(page) != 1 + !!PageSwapCache(page))
2740                         goto out_unmap;
2741                 if (pte_young(pteval) ||
2742                     page_is_young(page) || PageReferenced(page) ||
2743                     mmu_notifier_test_young(vma->vm_mm, address))
2744                         referenced = true;
2745         }
2746         if (referenced && writable)
2747                 ret = 1;
2748 out_unmap:
2749         pte_unmap_unlock(pte, ptl);
2750         if (ret) {
2751                 node = khugepaged_find_target_node();
2752                 /* collapse_huge_page will return with the mmap_sem released */
2753                 collapse_huge_page(mm, address, hpage, vma, node);
2754         }
2755 out:
2756         return ret;
2757 }
2758
2759 static void collect_mm_slot(struct mm_slot *mm_slot)
2760 {
2761         struct mm_struct *mm = mm_slot->mm;
2762
2763         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2764
2765         if (khugepaged_test_exit(mm)) {
2766                 /* free mm_slot */
2767                 hash_del(&mm_slot->hash);
2768                 list_del(&mm_slot->mm_node);
2769
2770                 /*
2771                  * Not strictly needed because the mm exited already.
2772                  *
2773                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2774                  */
2775
2776                 /* khugepaged_mm_lock actually not necessary for the below */
2777                 free_mm_slot(mm_slot);
2778                 mmdrop(mm);
2779         }
2780 }
2781
2782 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2783                                             struct page **hpage)
2784         __releases(&khugepaged_mm_lock)
2785         __acquires(&khugepaged_mm_lock)
2786 {
2787         struct mm_slot *mm_slot;
2788         struct mm_struct *mm;
2789         struct vm_area_struct *vma;
2790         int progress = 0;
2791
2792         VM_BUG_ON(!pages);
2793         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2794
2795         if (khugepaged_scan.mm_slot)
2796                 mm_slot = khugepaged_scan.mm_slot;
2797         else {
2798                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2799                                      struct mm_slot, mm_node);
2800                 khugepaged_scan.address = 0;
2801                 khugepaged_scan.mm_slot = mm_slot;
2802         }
2803         spin_unlock(&khugepaged_mm_lock);
2804
2805         mm = mm_slot->mm;
2806         down_read(&mm->mmap_sem);
2807         if (unlikely(khugepaged_test_exit(mm)))
2808                 vma = NULL;
2809         else
2810                 vma = find_vma(mm, khugepaged_scan.address);
2811
2812         progress++;
2813         for (; vma; vma = vma->vm_next) {
2814                 unsigned long hstart, hend;
2815
2816                 cond_resched();
2817                 if (unlikely(khugepaged_test_exit(mm))) {
2818                         progress++;
2819                         break;
2820                 }
2821                 if (!hugepage_vma_check(vma)) {
2822 skip:
2823                         progress++;
2824                         continue;
2825                 }
2826                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2827                 hend = vma->vm_end & HPAGE_PMD_MASK;
2828                 if (hstart >= hend)
2829                         goto skip;
2830                 if (khugepaged_scan.address > hend)
2831                         goto skip;
2832                 if (khugepaged_scan.address < hstart)
2833                         khugepaged_scan.address = hstart;
2834                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2835
2836                 while (khugepaged_scan.address < hend) {
2837                         int ret;
2838                         cond_resched();
2839                         if (unlikely(khugepaged_test_exit(mm)))
2840                                 goto breakouterloop;
2841
2842                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2843                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2844                                   hend);
2845                         ret = khugepaged_scan_pmd(mm, vma,
2846                                                   khugepaged_scan.address,
2847                                                   hpage);
2848                         /* move to next address */
2849                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2850                         progress += HPAGE_PMD_NR;
2851                         if (ret)
2852                                 /* we released mmap_sem so break loop */
2853                                 goto breakouterloop_mmap_sem;
2854                         if (progress >= pages)
2855                                 goto breakouterloop;
2856                 }
2857         }
2858 breakouterloop:
2859         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2860 breakouterloop_mmap_sem:
2861
2862         spin_lock(&khugepaged_mm_lock);
2863         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2864         /*
2865          * Release the current mm_slot if this mm is about to die, or
2866          * if we scanned all vmas of this mm.
2867          */
2868         if (khugepaged_test_exit(mm) || !vma) {
2869                 /*
2870                  * Make sure that if mm_users is reaching zero while
2871                  * khugepaged runs here, khugepaged_exit will find
2872                  * mm_slot not pointing to the exiting mm.
2873                  */
2874                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2875                         khugepaged_scan.mm_slot = list_entry(
2876                                 mm_slot->mm_node.next,
2877                                 struct mm_slot, mm_node);
2878                         khugepaged_scan.address = 0;
2879                 } else {
2880                         khugepaged_scan.mm_slot = NULL;
2881                         khugepaged_full_scans++;
2882                 }
2883
2884                 collect_mm_slot(mm_slot);
2885         }
2886
2887         return progress;
2888 }
2889
2890 static int khugepaged_has_work(void)
2891 {
2892         return !list_empty(&khugepaged_scan.mm_head) &&
2893                 khugepaged_enabled();
2894 }
2895
2896 static int khugepaged_wait_event(void)
2897 {
2898         return !list_empty(&khugepaged_scan.mm_head) ||
2899                 kthread_should_stop();
2900 }
2901
2902 static void khugepaged_do_scan(void)
2903 {
2904         struct page *hpage = NULL;
2905         unsigned int progress = 0, pass_through_head = 0;
2906         unsigned int pages = khugepaged_pages_to_scan;
2907         bool wait = true;
2908
2909         barrier(); /* write khugepaged_pages_to_scan to local stack */
2910
2911         while (progress < pages) {
2912                 if (!khugepaged_prealloc_page(&hpage, &wait))
2913                         break;
2914
2915                 cond_resched();
2916
2917                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2918                         break;
2919
2920                 spin_lock(&khugepaged_mm_lock);
2921                 if (!khugepaged_scan.mm_slot)
2922                         pass_through_head++;
2923                 if (khugepaged_has_work() &&
2924                     pass_through_head < 2)
2925                         progress += khugepaged_scan_mm_slot(pages - progress,
2926                                                             &hpage);
2927                 else
2928                         progress = pages;
2929                 spin_unlock(&khugepaged_mm_lock);
2930         }
2931
2932         if (!IS_ERR_OR_NULL(hpage))
2933                 put_page(hpage);
2934 }
2935
2936 static void khugepaged_wait_work(void)
2937 {
2938         if (khugepaged_has_work()) {
2939                 if (!khugepaged_scan_sleep_millisecs)
2940                         return;
2941
2942                 wait_event_freezable_timeout(khugepaged_wait,
2943                                              kthread_should_stop(),
2944                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2945                 return;
2946         }
2947
2948         if (khugepaged_enabled())
2949                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2950 }
2951
2952 static int khugepaged(void *none)
2953 {
2954         struct mm_slot *mm_slot;
2955
2956         set_freezable();
2957         set_user_nice(current, MAX_NICE);
2958
2959         while (!kthread_should_stop()) {
2960                 khugepaged_do_scan();
2961                 khugepaged_wait_work();
2962         }
2963
2964         spin_lock(&khugepaged_mm_lock);
2965         mm_slot = khugepaged_scan.mm_slot;
2966         khugepaged_scan.mm_slot = NULL;
2967         if (mm_slot)
2968                 collect_mm_slot(mm_slot);
2969         spin_unlock(&khugepaged_mm_lock);
2970         return 0;
2971 }
2972
2973 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2974                 unsigned long haddr, pmd_t *pmd)
2975 {
2976         struct mm_struct *mm = vma->vm_mm;
2977         pgtable_t pgtable;
2978         pmd_t _pmd;
2979         int i;
2980
2981         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2982         /* leave pmd empty until pte is filled */
2983
2984         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2985         pmd_populate(mm, &_pmd, pgtable);
2986
2987         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2988                 pte_t *pte, entry;
2989                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2990                 entry = pte_mkspecial(entry);
2991                 pte = pte_offset_map(&_pmd, haddr);
2992                 VM_BUG_ON(!pte_none(*pte));
2993                 set_pte_at(mm, haddr, pte, entry);
2994                 pte_unmap(pte);
2995         }
2996         smp_wmb(); /* make pte visible before pmd */
2997         pmd_populate(mm, pmd, pgtable);
2998         put_huge_zero_page();
2999 }
3000
3001 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
3002                 pmd_t *pmd)
3003 {
3004         spinlock_t *ptl;
3005         struct page *page = NULL;
3006         struct mm_struct *mm = vma->vm_mm;
3007         unsigned long haddr = address & HPAGE_PMD_MASK;
3008         unsigned long mmun_start;       /* For mmu_notifiers */
3009         unsigned long mmun_end;         /* For mmu_notifiers */
3010
3011         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
3012
3013         mmun_start = haddr;
3014         mmun_end   = haddr + HPAGE_PMD_SIZE;
3015 again:
3016         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3017         ptl = pmd_lock(mm, pmd);
3018         if (unlikely(!pmd_trans_huge(*pmd)))
3019                 goto unlock;
3020         if (vma_is_dax(vma)) {
3021                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
3022                 if (is_huge_zero_pmd(_pmd))
3023                         put_huge_zero_page();
3024         } else if (is_huge_zero_pmd(*pmd)) {
3025                 __split_huge_zero_page_pmd(vma, haddr, pmd);
3026         } else {
3027                 page = pmd_page(*pmd);
3028                 VM_BUG_ON_PAGE(!page_count(page), page);
3029                 get_page(page);
3030         }
3031  unlock:
3032         spin_unlock(ptl);
3033         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3034
3035         if (!page)
3036                 return;
3037
3038         split_huge_page(page);
3039         put_page(page);
3040
3041         /*
3042          * We don't always have down_write of mmap_sem here: a racing
3043          * do_huge_pmd_wp_page() might have copied-on-write to another
3044          * huge page before our split_huge_page() got the anon_vma lock.
3045          */
3046         if (unlikely(pmd_trans_huge(*pmd)))
3047                 goto again;
3048 }
3049
3050 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3051                 pmd_t *pmd)
3052 {
3053         struct vm_area_struct *vma;
3054
3055         vma = find_vma(mm, address);
3056         BUG_ON(vma == NULL);
3057         split_huge_page_pmd(vma, address, pmd);
3058 }
3059
3060 static void split_huge_page_address(struct mm_struct *mm,
3061                                     unsigned long address)
3062 {
3063         pgd_t *pgd;
3064         pud_t *pud;
3065         pmd_t *pmd;
3066
3067         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3068
3069         pgd = pgd_offset(mm, address);
3070         if (!pgd_present(*pgd))
3071                 return;
3072
3073         pud = pud_offset(pgd, address);
3074         if (!pud_present(*pud))
3075                 return;
3076
3077         pmd = pmd_offset(pud, address);
3078         if (!pmd_present(*pmd))
3079                 return;
3080         /*
3081          * Caller holds the mmap_sem write mode, so a huge pmd cannot
3082          * materialize from under us.
3083          */
3084         split_huge_page_pmd_mm(mm, address, pmd);
3085 }
3086
3087 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3088                              unsigned long start,
3089                              unsigned long end,
3090                              long adjust_next)
3091 {
3092         /*
3093          * If the new start address isn't hpage aligned and it could
3094          * previously contain an hugepage: check if we need to split
3095          * an huge pmd.
3096          */
3097         if (start & ~HPAGE_PMD_MASK &&
3098             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3099             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3100                 split_huge_page_address(vma->vm_mm, start);
3101
3102         /*
3103          * If the new end address isn't hpage aligned and it could
3104          * previously contain an hugepage: check if we need to split
3105          * an huge pmd.
3106          */
3107         if (end & ~HPAGE_PMD_MASK &&
3108             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3109             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3110                 split_huge_page_address(vma->vm_mm, end);
3111
3112         /*
3113          * If we're also updating the vma->vm_next->vm_start, if the new
3114          * vm_next->vm_start isn't page aligned and it could previously
3115          * contain an hugepage: check if we need to split an huge pmd.
3116          */
3117         if (adjust_next > 0) {
3118                 struct vm_area_struct *next = vma->vm_next;
3119                 unsigned long nstart = next->vm_start;
3120                 nstart += adjust_next << PAGE_SHIFT;
3121                 if (nstart & ~HPAGE_PMD_MASK &&
3122                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3123                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3124                         split_huge_page_address(next->vm_mm, nstart);
3125         }
3126 }