mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/kthread.h>
  20 #include <linux/khugepaged.h>
  21 #include <linux/freezer.h>
  22 #include <linux/mman.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/migrate.h>
  25 #include <linux/hashtable.h>
  26 #include <linux/userfaultfd_k.h>
  27
  28 #include <asm/tlb.h>
  29 #include <asm/pgalloc.h>
  30 #include "internal.h"
  31
  32 /*
  33  * By default transparent hugepage support is disabled in order that avoid
  34  * to risk increase the memory footprint of applications without a guaranteed
  35  * benefit. When transparent hugepage support is enabled, is for all mappings,
  36  * and khugepaged scans all mappings.
  37  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  38  * for all hugepage allocations.
  39  */
  40 unsigned long transparent_hugepage_flags __read_mostly =
  41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  42         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  43 #endif
  44 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  45         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  46 #endif
  47         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  48         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  49         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  50
  51 /* default scan 8*512 pte (or vmas) every 30 second */
  52 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  53 static unsigned int khugepaged_pages_collapsed;
  54 static unsigned int khugepaged_full_scans;
  55 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  56 /* during fragmentation poll the hugepage allocator once every minute */
  57 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  58 static struct task_struct *khugepaged_thread __read_mostly;
  59 static DEFINE_MUTEX(khugepaged_mutex);
  60 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  61 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  62 /*
  63  * default collapse hugepages if there is at least one pte mapped like
  64  * it would have happened if the vma was large enough during page
  65  * fault.
  66  */
  67 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  68
  69 static int khugepaged(void *none);
  70 static int khugepaged_slab_init(void);
  71 static void khugepaged_slab_exit(void);
  72
  73 #define MM_SLOTS_HASH_BITS 10
  74 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  75
  76 static struct kmem_cache *mm_slot_cache __read_mostly;
  77
  78 /**
  79  * struct mm_slot - hash lookup from mm to mm_slot
  80  * @hash: hash collision list
  81  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  82  * @mm: the mm that this information is valid for
  83  */
  84 struct mm_slot {
  85         struct hlist_node hash;
  86         struct list_head mm_node;
  87         struct mm_struct *mm;
  88 };
  89
  90 /**
  91  * struct khugepaged_scan - cursor for scanning
  92  * @mm_head: the head of the mm list to scan
  93  * @mm_slot: the current mm_slot we are scanning
  94  * @address: the next address inside that to be scanned
  95  *
  96  * There is only the one khugepaged_scan instance of this cursor structure.
  97  */
  98 struct khugepaged_scan {
  99         struct list_head mm_head;
 100         struct mm_slot *mm_slot;
 101         unsigned long address;
 102 };
 103 static struct khugepaged_scan khugepaged_scan = {
 104         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 105 };
 106
 107
 108 static int set_recommended_min_free_kbytes(void)
 109 {
 110         struct zone *zone;
 111         int nr_zones = 0;
 112         unsigned long recommended_min;
 113
 114         for_each_populated_zone(zone)
 115                 nr_zones++;
 116
 117         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 118         recommended_min = pageblock_nr_pages * nr_zones * 2;
 119
 120         /*
 121          * Make sure that on average at least two pageblocks are almost free
 122          * of another type, one for a migratetype to fall back to and a
 123          * second to avoid subsequent fallbacks of other types There are 3
 124          * MIGRATE_TYPES we care about.
 125          */
 126         recommended_min += pageblock_nr_pages * nr_zones *
 127                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 128
 129         /* don't ever allow to reserve more than 5% of the lowmem */
 130         recommended_min = min(recommended_min,
 131                               (unsigned long) nr_free_buffer_pages() / 20);
 132         recommended_min <<= (PAGE_SHIFT-10);
 133
 134         if (recommended_min > min_free_kbytes) {
 135                 if (user_min_free_kbytes >= 0)
 136                         pr_info("raising min_free_kbytes from %d to %lu "
 137                                 "to help transparent hugepage allocations\n",
 138                                 min_free_kbytes, recommended_min);
 139
 140                 min_free_kbytes = recommended_min;
 141         }
 142         setup_per_zone_wmarks();
 143         return 0;
 144 }
 145
 146 static int start_stop_khugepaged(void)
 147 {
 148         int err = 0;
 149         if (khugepaged_enabled()) {
 150                 if (!khugepaged_thread)
 151                         khugepaged_thread = kthread_run(khugepaged, NULL,
 152                                                         "khugepaged");
 153                 if (unlikely(IS_ERR(khugepaged_thread))) {
 154                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 155                         err = PTR_ERR(khugepaged_thread);
 156                         khugepaged_thread = NULL;
 157                         goto fail;
 158                 }
 159
 160                 if (!list_empty(&khugepaged_scan.mm_head))
 161                         wake_up_interruptible(&khugepaged_wait);
 162
 163                 set_recommended_min_free_kbytes();
 164         } else if (khugepaged_thread) {
 165                 kthread_stop(khugepaged_thread);
 166                 khugepaged_thread = NULL;
 167         }
 168 fail:
 169         return err;
 170 }
 171
 172 static atomic_t huge_zero_refcount;
 173 struct page *huge_zero_page __read_mostly;
 174
 175 static inline bool is_huge_zero_pmd(pmd_t pmd)
 176 {
 177         return is_huge_zero_page(pmd_page(pmd));
 178 }
 179
 180 static struct page *get_huge_zero_page(void)
 181 {
 182         struct page *zero_page;
 183 retry:
 184         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 185                 return READ_ONCE(huge_zero_page);
 186
 187         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 188                         HPAGE_PMD_ORDER);
 189         if (!zero_page) {
 190                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 191                 return NULL;
 192         }
 193         count_vm_event(THP_ZERO_PAGE_ALLOC);
 194         preempt_disable();
 195         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 196                 preempt_enable();
 197                 __free_pages(zero_page, compound_order(zero_page));
 198                 goto retry;
 199         }
 200
 201         /* We take additional reference here. It will be put back by shrinker */
 202         atomic_set(&huge_zero_refcount, 2);
 203         preempt_enable();
 204         return READ_ONCE(huge_zero_page);
 205 }
 206
 207 static void put_huge_zero_page(void)
 208 {
 209         /*
 210          * Counter should never go to zero here. Only shrinker can put
 211          * last reference.
 212          */
 213         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 214 }
 215
 216 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 217                                         struct shrink_control *sc)
 218 {
 219         /* we can free zero page only if last reference remains */
 220         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 221 }
 222
 223 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 224                                        struct shrink_control *sc)
 225 {
 226         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 227                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 228                 BUG_ON(zero_page == NULL);
 229                 __free_pages(zero_page, compound_order(zero_page));
 230                 return HPAGE_PMD_NR;
 231         }
 232
 233         return 0;
 234 }
 235
 236 static struct shrinker huge_zero_page_shrinker = {
 237         .count_objects = shrink_huge_zero_page_count,
 238         .scan_objects = shrink_huge_zero_page_scan,
 239         .seeks = DEFAULT_SEEKS,
 240 };
 241
 242 #ifdef CONFIG_SYSFS
 243
 244 static ssize_t double_flag_show(struct kobject *kobj,
 245                                 struct kobj_attribute *attr, char *buf,
 246                                 enum transparent_hugepage_flag enabled,
 247                                 enum transparent_hugepage_flag req_madv)
 248 {
 249         if (test_bit(enabled, &transparent_hugepage_flags)) {
 250                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 251                 return sprintf(buf, "[always] madvise never\n");
 252         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 253                 return sprintf(buf, "always [madvise] never\n");
 254         else
 255                 return sprintf(buf, "always madvise [never]\n");
 256 }
 257 static ssize_t double_flag_store(struct kobject *kobj,
 258                                  struct kobj_attribute *attr,
 259                                  const char *buf, size_t count,
 260                                  enum transparent_hugepage_flag enabled,
 261                                  enum transparent_hugepage_flag req_madv)
 262 {
 263         if (!memcmp("always", buf,
 264                     min(sizeof("always")-1, count))) {
 265                 set_bit(enabled, &transparent_hugepage_flags);
 266                 clear_bit(req_madv, &transparent_hugepage_flags);
 267         } else if (!memcmp("madvise", buf,
 268                            min(sizeof("madvise")-1, count))) {
 269                 clear_bit(enabled, &transparent_hugepage_flags);
 270                 set_bit(req_madv, &transparent_hugepage_flags);
 271         } else if (!memcmp("never", buf,
 272                            min(sizeof("never")-1, count))) {
 273                 clear_bit(enabled, &transparent_hugepage_flags);
 274                 clear_bit(req_madv, &transparent_hugepage_flags);
 275         } else
 276                 return -EINVAL;
 277
 278         return count;
 279 }
 280
 281 static ssize_t enabled_show(struct kobject *kobj,
 282                             struct kobj_attribute *attr, char *buf)
 283 {
 284         return double_flag_show(kobj, attr, buf,
 285                                 TRANSPARENT_HUGEPAGE_FLAG,
 286                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 287 }
 288 static ssize_t enabled_store(struct kobject *kobj,
 289                              struct kobj_attribute *attr,
 290                              const char *buf, size_t count)
 291 {
 292         ssize_t ret;
 293
 294         ret = double_flag_store(kobj, attr, buf, count,
 295                                 TRANSPARENT_HUGEPAGE_FLAG,
 296                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 297
 298         if (ret > 0) {
 299                 int err;
 300
 301                 mutex_lock(&khugepaged_mutex);
 302                 err = start_stop_khugepaged();
 303                 mutex_unlock(&khugepaged_mutex);
 304
 305                 if (err)
 306                         ret = err;
 307         }
 308
 309         return ret;
 310 }
 311 static struct kobj_attribute enabled_attr =
 312         __ATTR(enabled, 0644, enabled_show, enabled_store);
 313
 314 static ssize_t single_flag_show(struct kobject *kobj,
 315                                 struct kobj_attribute *attr, char *buf,
 316                                 enum transparent_hugepage_flag flag)
 317 {
 318         return sprintf(buf, "%d\n",
 319                        !!test_bit(flag, &transparent_hugepage_flags));
 320 }
 321
 322 static ssize_t single_flag_store(struct kobject *kobj,
 323                                  struct kobj_attribute *attr,
 324                                  const char *buf, size_t count,
 325                                  enum transparent_hugepage_flag flag)
 326 {
 327         unsigned long value;
 328         int ret;
 329
 330         ret = kstrtoul(buf, 10, &value);
 331         if (ret < 0)
 332                 return ret;
 333         if (value > 1)
 334                 return -EINVAL;
 335
 336         if (value)
 337                 set_bit(flag, &transparent_hugepage_flags);
 338         else
 339                 clear_bit(flag, &transparent_hugepage_flags);
 340
 341         return count;
 342 }
 343
 344 /*
 345  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 346  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 347  * memory just to allocate one more hugepage.
 348  */
 349 static ssize_t defrag_show(struct kobject *kobj,
 350                            struct kobj_attribute *attr, char *buf)
 351 {
 352         return double_flag_show(kobj, attr, buf,
 353                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 354                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 355 }
 356 static ssize_t defrag_store(struct kobject *kobj,
 357                             struct kobj_attribute *attr,
 358                             const char *buf, size_t count)
 359 {
 360         return double_flag_store(kobj, attr, buf, count,
 361                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 362                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 363 }
 364 static struct kobj_attribute defrag_attr =
 365         __ATTR(defrag, 0644, defrag_show, defrag_store);
 366
 367 static ssize_t use_zero_page_show(struct kobject *kobj,
 368                 struct kobj_attribute *attr, char *buf)
 369 {
 370         return single_flag_show(kobj, attr, buf,
 371                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 372 }
 373 static ssize_t use_zero_page_store(struct kobject *kobj,
 374                 struct kobj_attribute *attr, const char *buf, size_t count)
 375 {
 376         return single_flag_store(kobj, attr, buf, count,
 377                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 378 }
 379 static struct kobj_attribute use_zero_page_attr =
 380         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 381 #ifdef CONFIG_DEBUG_VM
 382 static ssize_t debug_cow_show(struct kobject *kobj,
 383                                 struct kobj_attribute *attr, char *buf)
 384 {
 385         return single_flag_show(kobj, attr, buf,
 386                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 387 }
 388 static ssize_t debug_cow_store(struct kobject *kobj,
 389                                struct kobj_attribute *attr,
 390                                const char *buf, size_t count)
 391 {
 392         return single_flag_store(kobj, attr, buf, count,
 393                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 394 }
 395 static struct kobj_attribute debug_cow_attr =
 396         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 397 #endif /* CONFIG_DEBUG_VM */
 398
 399 static struct attribute *hugepage_attr[] = {
 400         &enabled_attr.attr,
 401         &defrag_attr.attr,
 402         &use_zero_page_attr.attr,
 403 #ifdef CONFIG_DEBUG_VM
 404         &debug_cow_attr.attr,
 405 #endif
 406         NULL,
 407 };
 408
 409 static struct attribute_group hugepage_attr_group = {
 410         .attrs = hugepage_attr,
 411 };
 412
 413 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 414                                          struct kobj_attribute *attr,
 415                                          char *buf)
 416 {
 417         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 418 }
 419
 420 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 421                                           struct kobj_attribute *attr,
 422                                           const char *buf, size_t count)
 423 {
 424         unsigned long msecs;
 425         int err;
 426
 427         err = kstrtoul(buf, 10, &msecs);
 428         if (err || msecs > UINT_MAX)
 429                 return -EINVAL;
 430
 431         khugepaged_scan_sleep_millisecs = msecs;
 432         wake_up_interruptible(&khugepaged_wait);
 433
 434         return count;
 435 }
 436 static struct kobj_attribute scan_sleep_millisecs_attr =
 437         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 438                scan_sleep_millisecs_store);
 439
 440 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 441                                           struct kobj_attribute *attr,
 442                                           char *buf)
 443 {
 444         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 445 }
 446
 447 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 448                                            struct kobj_attribute *attr,
 449                                            const char *buf, size_t count)
 450 {
 451         unsigned long msecs;
 452         int err;
 453
 454         err = kstrtoul(buf, 10, &msecs);
 455         if (err || msecs > UINT_MAX)
 456                 return -EINVAL;
 457
 458         khugepaged_alloc_sleep_millisecs = msecs;
 459         wake_up_interruptible(&khugepaged_wait);
 460
 461         return count;
 462 }
 463 static struct kobj_attribute alloc_sleep_millisecs_attr =
 464         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 465                alloc_sleep_millisecs_store);
 466
 467 static ssize_t pages_to_scan_show(struct kobject *kobj,
 468                                   struct kobj_attribute *attr,
 469                                   char *buf)
 470 {
 471         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 472 }
 473 static ssize_t pages_to_scan_store(struct kobject *kobj,
 474                                    struct kobj_attribute *attr,
 475                                    const char *buf, size_t count)
 476 {
 477         int err;
 478         unsigned long pages;
 479
 480         err = kstrtoul(buf, 10, &pages);
 481         if (err || !pages || pages > UINT_MAX)
 482                 return -EINVAL;
 483
 484         khugepaged_pages_to_scan = pages;
 485
 486         return count;
 487 }
 488 static struct kobj_attribute pages_to_scan_attr =
 489         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 490                pages_to_scan_store);
 491
 492 static ssize_t pages_collapsed_show(struct kobject *kobj,
 493                                     struct kobj_attribute *attr,
 494                                     char *buf)
 495 {
 496         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 497 }
 498 static struct kobj_attribute pages_collapsed_attr =
 499         __ATTR_RO(pages_collapsed);
 500
 501 static ssize_t full_scans_show(struct kobject *kobj,
 502                                struct kobj_attribute *attr,
 503                                char *buf)
 504 {
 505         return sprintf(buf, "%u\n", khugepaged_full_scans);
 506 }
 507 static struct kobj_attribute full_scans_attr =
 508         __ATTR_RO(full_scans);
 509
 510 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 511                                       struct kobj_attribute *attr, char *buf)
 512 {
 513         return single_flag_show(kobj, attr, buf,
 514                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 515 }
 516 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 517                                        struct kobj_attribute *attr,
 518                                        const char *buf, size_t count)
 519 {
 520         return single_flag_store(kobj, attr, buf, count,
 521                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 522 }
 523 static struct kobj_attribute khugepaged_defrag_attr =
 524         __ATTR(defrag, 0644, khugepaged_defrag_show,
 525                khugepaged_defrag_store);
 526
 527 /*
 528  * max_ptes_none controls if khugepaged should collapse hugepages over
 529  * any unmapped ptes in turn potentially increasing the memory
 530  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 531  * reduce the available free memory in the system as it
 532  * runs. Increasing max_ptes_none will instead potentially reduce the
 533  * free memory in the system during the khugepaged scan.
 534  */
 535 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 536                                              struct kobj_attribute *attr,
 537                                              char *buf)
 538 {
 539         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 540 }
 541 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 542                                               struct kobj_attribute *attr,
 543                                               const char *buf, size_t count)
 544 {
 545         int err;
 546         unsigned long max_ptes_none;
 547
 548         err = kstrtoul(buf, 10, &max_ptes_none);
 549         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 550                 return -EINVAL;
 551
 552         khugepaged_max_ptes_none = max_ptes_none;
 553
 554         return count;
 555 }
 556 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 557         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 558                khugepaged_max_ptes_none_store);
 559
 560 static struct attribute *khugepaged_attr[] = {
 561         &khugepaged_defrag_attr.attr,
 562         &khugepaged_max_ptes_none_attr.attr,
 563         &pages_to_scan_attr.attr,
 564         &pages_collapsed_attr.attr,
 565         &full_scans_attr.attr,
 566         &scan_sleep_millisecs_attr.attr,
 567         &alloc_sleep_millisecs_attr.attr,
 568         NULL,
 569 };
 570
 571 static struct attribute_group khugepaged_attr_group = {
 572         .attrs = khugepaged_attr,
 573         .name = "khugepaged",
 574 };
 575
 576 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 577 {
 578         int err;
 579
 580         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 581         if (unlikely(!*hugepage_kobj)) {
 582                 pr_err("failed to create transparent hugepage kobject\n");
 583                 return -ENOMEM;
 584         }
 585
 586         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 587         if (err) {
 588                 pr_err("failed to register transparent hugepage group\n");
 589                 goto delete_obj;
 590         }
 591
 592         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 593         if (err) {
 594                 pr_err("failed to register transparent hugepage group\n");
 595                 goto remove_hp_group;
 596         }
 597
 598         return 0;
 599
 600 remove_hp_group:
 601         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 602 delete_obj:
 603         kobject_put(*hugepage_kobj);
 604         return err;
 605 }
 606
 607 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 608 {
 609         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 610         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 611         kobject_put(hugepage_kobj);
 612 }
 613 #else
 614 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 615 {
 616         return 0;
 617 }
 618
 619 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 620 {
 621 }
 622 #endif /* CONFIG_SYSFS */
 623
 624 static int __init hugepage_init(void)
 625 {
 626         int err;
 627         struct kobject *hugepage_kobj;
 628
 629         if (!has_transparent_hugepage()) {
 630                 transparent_hugepage_flags = 0;
 631                 return -EINVAL;
 632         }
 633
 634         err = hugepage_init_sysfs(&hugepage_kobj);
 635         if (err)
 636                 goto err_sysfs;
 637
 638         err = khugepaged_slab_init();
 639         if (err)
 640                 goto err_slab;
 641
 642         err = register_shrinker(&huge_zero_page_shrinker);
 643         if (err)
 644                 goto err_hzp_shrinker;
 645
 646         /*
 647          * By default disable transparent hugepages on smaller systems,
 648          * where the extra memory used could hurt more than TLB overhead
 649          * is likely to save.  The admin can still enable it through /sys.
 650          */
 651         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 652                 transparent_hugepage_flags = 0;
 653                 return 0;
 654         }
 655
 656         err = start_stop_khugepaged();
 657         if (err)
 658                 goto err_khugepaged;
 659
 660         return 0;
 661 err_khugepaged:
 662         unregister_shrinker(&huge_zero_page_shrinker);
 663 err_hzp_shrinker:
 664         khugepaged_slab_exit();
 665 err_slab:
 666         hugepage_exit_sysfs(hugepage_kobj);
 667 err_sysfs:
 668         return err;
 669 }
 670 subsys_initcall(hugepage_init);
 671
 672 static int __init setup_transparent_hugepage(char *str)
 673 {
 674         int ret = 0;
 675         if (!str)
 676                 goto out;
 677         if (!strcmp(str, "always")) {
 678                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 679                         &transparent_hugepage_flags);
 680                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 681                           &transparent_hugepage_flags);
 682                 ret = 1;
 683         } else if (!strcmp(str, "madvise")) {
 684                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 685                           &transparent_hugepage_flags);
 686                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 687                         &transparent_hugepage_flags);
 688                 ret = 1;
 689         } else if (!strcmp(str, "never")) {
 690                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 691                           &transparent_hugepage_flags);
 692                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 693                           &transparent_hugepage_flags);
 694                 ret = 1;
 695         }
 696 out:
 697         if (!ret)
 698                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 699         return ret;
 700 }
 701 __setup("transparent_hugepage=", setup_transparent_hugepage);
 702
 703 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 704 {
 705         if (likely(vma->vm_flags & VM_WRITE))
 706                 pmd = pmd_mkwrite(pmd);
 707         return pmd;
 708 }
 709
 710 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 711 {
 712         pmd_t entry;
 713         entry = mk_pmd(page, prot);
 714         entry = pmd_mkhuge(entry);
 715         return entry;
 716 }
 717
 718 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 719                                         struct vm_area_struct *vma,
 720                                         unsigned long haddr, pmd_t *pmd,
 721                                         struct page *page, gfp_t gfp,
 722                                         unsigned int flags)
 723 {
 724         struct mem_cgroup *memcg;
 725         pgtable_t pgtable;
 726         spinlock_t *ptl;
 727
 728         VM_BUG_ON_PAGE(!PageCompound(page), page);
 729
 730         if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
 731                 put_page(page);
 732                 count_vm_event(THP_FAULT_FALLBACK);
 733                 return VM_FAULT_FALLBACK;
 734         }
 735
 736         pgtable = pte_alloc_one(mm, haddr);
 737         if (unlikely(!pgtable)) {
 738                 mem_cgroup_cancel_charge(page, memcg);
 739                 put_page(page);
 740                 return VM_FAULT_OOM;
 741         }
 742
 743         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 744         /*
 745          * The memory barrier inside __SetPageUptodate makes sure that
 746          * clear_huge_page writes become visible before the set_pmd_at()
 747          * write.
 748          */
 749         __SetPageUptodate(page);
 750
 751         ptl = pmd_lock(mm, pmd);
 752         if (unlikely(!pmd_none(*pmd))) {
 753                 spin_unlock(ptl);
 754                 mem_cgroup_cancel_charge(page, memcg);
 755                 put_page(page);
 756                 pte_free(mm, pgtable);
 757         } else {
 758                 pmd_t entry;
 759
 760                 /* Deliver the page fault to userland */
 761                 if (userfaultfd_missing(vma)) {
 762                         int ret;
 763
 764                         spin_unlock(ptl);
 765                         mem_cgroup_cancel_charge(page, memcg);
 766                         put_page(page);
 767                         pte_free(mm, pgtable);
 768                         ret = handle_userfault(vma, haddr, flags,
 769                                                VM_UFFD_MISSING);
 770                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 771                         return ret;
 772                 }
 773
 774                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 775                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 776                 page_add_new_anon_rmap(page, vma, haddr);
 777                 mem_cgroup_commit_charge(page, memcg, false);
 778                 lru_cache_add_active_or_unevictable(page, vma);
 779                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 780                 set_pmd_at(mm, haddr, pmd, entry);
 781                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 782                 atomic_long_inc(&mm->nr_ptes);
 783                 spin_unlock(ptl);
 784                 count_vm_event(THP_FAULT_ALLOC);
 785         }
 786
 787         return 0;
 788 }
 789
 790 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 791 {
 792         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 793 }
 794
 795 /* Caller must hold page table lock. */
 796 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 797                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 798                 struct page *zero_page)
 799 {
 800         pmd_t entry;
 801         entry = mk_pmd(zero_page, vma->vm_page_prot);
 802         entry = pmd_mkhuge(entry);
 803         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 804         set_pmd_at(mm, haddr, pmd, entry);
 805         atomic_long_inc(&mm->nr_ptes);
 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                 bool set;
 828                 int ret;
 829                 pgtable = pte_alloc_one(mm, haddr);
 830                 if (unlikely(!pgtable))
 831                         return VM_FAULT_OOM;
 832                 zero_page = get_huge_zero_page();
 833                 if (unlikely(!zero_page)) {
 834                         pte_free(mm, pgtable);
 835                         count_vm_event(THP_FAULT_FALLBACK);
 836                         return VM_FAULT_FALLBACK;
 837                 }
 838                 ptl = pmd_lock(mm, pmd);
 839                 ret = 0;
 840                 set = false;
 841                 if (pmd_none(*pmd)) {
 842                         if (userfaultfd_missing(vma)) {
 843                                 spin_unlock(ptl);
 844                                 ret = handle_userfault(vma, haddr, 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                                 set = true;
 853                         }
 854                 } else
 855                         spin_unlock(ptl);
 856                 if (!set) {
 857                         pte_free(mm, pgtable);
 858                         put_huge_zero_page();
 859                 }
 860                 return ret;
 861         }
 862         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 863         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 864         if (unlikely(!page)) {
 865                 count_vm_event(THP_FAULT_FALLBACK);
 866                 return VM_FAULT_FALLBACK;
 867         }
 868         return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page, gfp, flags);
 869 }
 870
 871 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 872                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 873                   struct vm_area_struct *vma)
 874 {
 875         spinlock_t *dst_ptl, *src_ptl;
 876         struct page *src_page;
 877         pmd_t pmd;
 878         pgtable_t pgtable;
 879         int ret;
 880
 881         ret = -ENOMEM;
 882         pgtable = pte_alloc_one(dst_mm, addr);
 883         if (unlikely(!pgtable))
 884                 goto out;
 885
 886         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 887         src_ptl = pmd_lockptr(src_mm, src_pmd);
 888         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 889
 890         ret = -EAGAIN;
 891         pmd = *src_pmd;
 892         if (unlikely(!pmd_trans_huge(pmd))) {
 893                 pte_free(dst_mm, pgtable);
 894                 goto out_unlock;
 895         }
 896         /*
 897          * When page table lock is held, the huge zero pmd should not be
 898          * under splitting since we don't split the page itself, only pmd to
 899          * a page table.
 900          */
 901         if (is_huge_zero_pmd(pmd)) {
 902                 struct page *zero_page;
 903                 /*
 904                  * get_huge_zero_page() will never allocate a new page here,
 905                  * since we already have a zero page to copy. It just takes a
 906                  * reference.
 907                  */
 908                 zero_page = get_huge_zero_page();
 909                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 910                                 zero_page);
 911                 ret = 0;
 912                 goto out_unlock;
 913         }
 914
 915         if (unlikely(pmd_trans_splitting(pmd))) {
 916                 /* split huge page running from under us */
 917                 spin_unlock(src_ptl);
 918                 spin_unlock(dst_ptl);
 919                 pte_free(dst_mm, pgtable);
 920
 921                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 922                 goto out;
 923         }
 924         src_page = pmd_page(pmd);
 925         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 926         get_page(src_page);
 927         page_dup_rmap(src_page);
 928         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 929
 930         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 931         pmd = pmd_mkold(pmd_wrprotect(pmd));
 932         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 933         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 934         atomic_long_inc(&dst_mm->nr_ptes);
 935
 936         ret = 0;
 937 out_unlock:
 938         spin_unlock(src_ptl);
 939         spin_unlock(dst_ptl);
 940 out:
 941         return ret;
 942 }
 943
 944 void huge_pmd_set_accessed(struct mm_struct *mm,
 945                            struct vm_area_struct *vma,
 946                            unsigned long address,
 947                            pmd_t *pmd, pmd_t orig_pmd,
 948                            int dirty)
 949 {
 950         spinlock_t *ptl;
 951         pmd_t entry;
 952         unsigned long haddr;
 953
 954         ptl = pmd_lock(mm, pmd);
 955         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 956                 goto unlock;
 957
 958         entry = pmd_mkyoung(orig_pmd);
 959         haddr = address & HPAGE_PMD_MASK;
 960         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 961                 update_mmu_cache_pmd(vma, address, pmd);
 962
 963 unlock:
 964         spin_unlock(ptl);
 965 }
 966
 967 /*
 968  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 969  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 970  * the source page gets split and a tail freed before copy completes.
 971  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 972  */
 973 static void get_user_huge_page(struct page *page)
 974 {
 975         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 976                 struct page *endpage = page + HPAGE_PMD_NR;
 977
 978                 atomic_add(HPAGE_PMD_NR, &page->_count);
 979                 while (++page < endpage)
 980                         get_huge_page_tail(page);
 981         } else {
 982                 get_page(page);
 983         }
 984 }
 985
 986 static void put_user_huge_page(struct page *page)
 987 {
 988         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 989                 struct page *endpage = page + HPAGE_PMD_NR;
 990
 991                 while (page < endpage)
 992                         put_page(page++);
 993         } else {
 994                 put_page(page);
 995         }
 996 }
 997
 998 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 999                                         struct vm_area_struct *vma,
1000                                         unsigned long address,
1001                                         pmd_t *pmd, pmd_t orig_pmd,
1002                                         struct page *page,
1003                                         unsigned long haddr)
1004 {
1005         struct mem_cgroup *memcg;
1006         spinlock_t *ptl;
1007         pgtable_t pgtable;
1008         pmd_t _pmd;
1009         int ret = 0, i;
1010         struct page **pages;
1011         unsigned long mmun_start;       /* For mmu_notifiers */
1012         unsigned long mmun_end;         /* For mmu_notifiers */
1013
1014         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1015                         GFP_KERNEL);
1016         if (unlikely(!pages)) {
1017                 ret |= VM_FAULT_OOM;
1018                 goto out;
1019         }
1020
1021         for (i = 0; i < HPAGE_PMD_NR; i++) {
1022                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1023                                                __GFP_OTHER_NODE,
1024                                                vma, address, page_to_nid(page));
1025                 if (unlikely(!pages[i] ||
1026                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1027                                                    &memcg))) {
1028                         if (pages[i])
1029                                 put_page(pages[i]);
1030                         while (--i >= 0) {
1031                                 memcg = (void *)page_private(pages[i]);
1032                                 set_page_private(pages[i], 0);
1033                                 mem_cgroup_cancel_charge(pages[i], memcg);
1034                                 put_page(pages[i]);
1035                         }
1036                         kfree(pages);
1037                         ret |= VM_FAULT_OOM;
1038                         goto out;
1039                 }
1040                 set_page_private(pages[i], (unsigned long)memcg);
1041         }
1042
1043         for (i = 0; i < HPAGE_PMD_NR; i++) {
1044                 copy_user_highpage(pages[i], page + i,
1045                                    haddr + PAGE_SIZE * i, vma);
1046                 __SetPageUptodate(pages[i]);
1047                 cond_resched();
1048         }
1049
1050         mmun_start = haddr;
1051         mmun_end   = haddr + HPAGE_PMD_SIZE;
1052         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1053
1054         ptl = pmd_lock(mm, pmd);
1055         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1056                 goto out_free_pages;
1057         VM_BUG_ON_PAGE(!PageHead(page), page);
1058
1059         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1060         /* leave pmd empty until pte is filled */
1061
1062         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1063         pmd_populate(mm, &_pmd, pgtable);
1064
1065         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1066                 pte_t *pte, entry;
1067                 entry = mk_pte(pages[i], vma->vm_page_prot);
1068                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1069                 memcg = (void *)page_private(pages[i]);
1070                 set_page_private(pages[i], 0);
1071                 page_add_new_anon_rmap(pages[i], vma, haddr);
1072                 mem_cgroup_commit_charge(pages[i], memcg, false);
1073                 lru_cache_add_active_or_unevictable(pages[i], vma);
1074                 pte = pte_offset_map(&_pmd, haddr);
1075                 VM_BUG_ON(!pte_none(*pte));
1076                 set_pte_at(mm, haddr, pte, entry);
1077                 pte_unmap(pte);
1078         }
1079         kfree(pages);
1080
1081         smp_wmb(); /* make pte visible before pmd */
1082         pmd_populate(mm, pmd, pgtable);
1083         page_remove_rmap(page);
1084         spin_unlock(ptl);
1085
1086         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1087
1088         ret |= VM_FAULT_WRITE;
1089         put_page(page);
1090
1091 out:
1092         return ret;
1093
1094 out_free_pages:
1095         spin_unlock(ptl);
1096         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1097         for (i = 0; i < HPAGE_PMD_NR; i++) {
1098                 memcg = (void *)page_private(pages[i]);
1099                 set_page_private(pages[i], 0);
1100                 mem_cgroup_cancel_charge(pages[i], memcg);
1101                 put_page(pages[i]);
1102         }
1103         kfree(pages);
1104         goto out;
1105 }
1106
1107 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1108                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1109 {
1110         spinlock_t *ptl;
1111         int ret = 0;
1112         struct page *page = NULL, *new_page;
1113         struct mem_cgroup *memcg;
1114         unsigned long haddr;
1115         unsigned long mmun_start;       /* For mmu_notifiers */
1116         unsigned long mmun_end;         /* For mmu_notifiers */
1117         gfp_t huge_gfp;                 /* for allocation and charge */
1118
1119         ptl = pmd_lockptr(mm, pmd);
1120         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1121         haddr = address & HPAGE_PMD_MASK;
1122         if (is_huge_zero_pmd(orig_pmd))
1123                 goto alloc;
1124         spin_lock(ptl);
1125         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1126                 goto out_unlock;
1127
1128         page = pmd_page(orig_pmd);
1129         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1130         if (page_mapcount(page) == 1) {
1131                 pmd_t entry;
1132                 entry = pmd_mkyoung(orig_pmd);
1133                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1134                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1135                         update_mmu_cache_pmd(vma, address, pmd);
1136                 ret |= VM_FAULT_WRITE;
1137                 goto out_unlock;
1138         }
1139         get_user_huge_page(page);
1140         spin_unlock(ptl);
1141 alloc:
1142         if (transparent_hugepage_enabled(vma) &&
1143             !transparent_hugepage_debug_cow()) {
1144                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1145                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1146         } else
1147                 new_page = NULL;
1148
1149         if (unlikely(!new_page)) {
1150                 if (!page) {
1151                         split_huge_page_pmd(vma, address, pmd);
1152                         ret |= VM_FAULT_FALLBACK;
1153                 } else {
1154                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1155                                         pmd, orig_pmd, page, haddr);
1156                         if (ret & VM_FAULT_OOM) {
1157                                 split_huge_page(page);
1158                                 ret |= VM_FAULT_FALLBACK;
1159                         }
1160                         put_user_huge_page(page);
1161                 }
1162                 count_vm_event(THP_FAULT_FALLBACK);
1163                 goto out;
1164         }
1165
1166         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1167                 put_page(new_page);
1168                 if (page) {
1169                         split_huge_page(page);
1170                         put_user_huge_page(page);
1171                 } else
1172                         split_huge_page_pmd(vma, address, pmd);
1173                 ret |= VM_FAULT_FALLBACK;
1174                 count_vm_event(THP_FAULT_FALLBACK);
1175                 goto out;
1176         }
1177
1178         count_vm_event(THP_FAULT_ALLOC);
1179
1180         if (!page)
1181                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1182         else
1183                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1184         __SetPageUptodate(new_page);
1185
1186         mmun_start = haddr;
1187         mmun_end   = haddr + HPAGE_PMD_SIZE;
1188         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1189
1190         spin_lock(ptl);
1191         if (page)
1192                 put_user_huge_page(page);
1193         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1194                 spin_unlock(ptl);
1195                 mem_cgroup_cancel_charge(new_page, memcg);
1196                 put_page(new_page);
1197                 goto out_mn;
1198         } else {
1199                 pmd_t entry;
1200                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1201                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1202                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1203                 page_add_new_anon_rmap(new_page, vma, haddr);
1204                 mem_cgroup_commit_charge(new_page, memcg, false);
1205                 lru_cache_add_active_or_unevictable(new_page, vma);
1206                 set_pmd_at(mm, haddr, pmd, entry);
1207                 update_mmu_cache_pmd(vma, address, pmd);
1208                 if (!page) {
1209                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1210                         put_huge_zero_page();
1211                 } else {
1212                         VM_BUG_ON_PAGE(!PageHead(page), page);
1213                         page_remove_rmap(page);
1214                         put_page(page);
1215                 }
1216                 ret |= VM_FAULT_WRITE;
1217         }
1218         spin_unlock(ptl);
1219 out_mn:
1220         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1221 out:
1222         return ret;
1223 out_unlock:
1224         spin_unlock(ptl);
1225         return ret;
1226 }
1227
1228 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1229                                    unsigned long addr,
1230                                    pmd_t *pmd,
1231                                    unsigned int flags)
1232 {
1233         struct mm_struct *mm = vma->vm_mm;
1234         struct page *page = NULL;
1235
1236         assert_spin_locked(pmd_lockptr(mm, pmd));
1237
1238         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1239                 goto out;
1240
1241         /* Avoid dumping huge zero page */
1242         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1243                 return ERR_PTR(-EFAULT);
1244
1245         /* Full NUMA hinting faults to serialise migration in fault paths */
1246         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1247                 goto out;
1248
1249         page = pmd_page(*pmd);
1250         VM_BUG_ON_PAGE(!PageHead(page), page);
1251         if (flags & FOLL_TOUCH) {
1252                 pmd_t _pmd;
1253                 /*
1254                  * We should set the dirty bit only for FOLL_WRITE but
1255                  * for now the dirty bit in the pmd is meaningless.
1256                  * And if the dirty bit will become meaningful and
1257                  * we'll only set it with FOLL_WRITE, an atomic
1258                  * set_bit will be required on the pmd to set the
1259                  * young bit, instead of the current set_pmd_at.
1260                  */
1261                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1262                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1263                                           pmd, _pmd,  1))
1264                         update_mmu_cache_pmd(vma, addr, pmd);
1265         }
1266         if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1267                 if (page->mapping && trylock_page(page)) {
1268                         lru_add_drain();
1269                         if (page->mapping)
1270                                 mlock_vma_page(page);
1271                         unlock_page(page);
1272                 }
1273         }
1274         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1275         VM_BUG_ON_PAGE(!PageCompound(page), page);
1276         if (flags & FOLL_GET)
1277                 get_page_foll(page);
1278
1279 out:
1280         return page;
1281 }
1282
1283 /* NUMA hinting page fault entry point for trans huge pmds */
1284 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1285                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1286 {
1287         spinlock_t *ptl;
1288         struct anon_vma *anon_vma = NULL;
1289         struct page *page;
1290         unsigned long haddr = addr & HPAGE_PMD_MASK;
1291         int page_nid = -1, this_nid = numa_node_id();
1292         int target_nid, last_cpupid = -1;
1293         bool page_locked;
1294         bool migrated = false;
1295         bool was_writable;
1296         int flags = 0;
1297
1298         /* A PROT_NONE fault should not end up here */
1299         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1300
1301         ptl = pmd_lock(mm, pmdp);
1302         if (unlikely(!pmd_same(pmd, *pmdp)))
1303                 goto out_unlock;
1304
1305         /*
1306          * If there are potential migrations, wait for completion and retry
1307          * without disrupting NUMA hinting information. Do not relock and
1308          * check_same as the page may no longer be mapped.
1309          */
1310         if (unlikely(pmd_trans_migrating(*pmdp))) {
1311                 page = pmd_page(*pmdp);
1312                 spin_unlock(ptl);
1313                 wait_on_page_locked(page);
1314                 goto out;
1315         }
1316
1317         page = pmd_page(pmd);
1318         BUG_ON(is_huge_zero_page(page));
1319         page_nid = page_to_nid(page);
1320         last_cpupid = page_cpupid_last(page);
1321         count_vm_numa_event(NUMA_HINT_FAULTS);
1322         if (page_nid == this_nid) {
1323                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1324                 flags |= TNF_FAULT_LOCAL;
1325         }
1326
1327         /* See similar comment in do_numa_page for explanation */
1328         if (!(vma->vm_flags & VM_WRITE))
1329                 flags |= TNF_NO_GROUP;
1330
1331         /*
1332          * Acquire the page lock to serialise THP migrations but avoid dropping
1333          * page_table_lock if at all possible
1334          */
1335         page_locked = trylock_page(page);
1336         target_nid = mpol_misplaced(page, vma, haddr);
1337         if (target_nid == -1) {
1338                 /* If the page was locked, there are no parallel migrations */
1339                 if (page_locked)
1340                         goto clear_pmdnuma;
1341         }
1342
1343         /* Migration could have started since the pmd_trans_migrating check */
1344         if (!page_locked) {
1345                 spin_unlock(ptl);
1346                 wait_on_page_locked(page);
1347                 page_nid = -1;
1348                 goto out;
1349         }
1350
1351         /*
1352          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1353          * to serialises splits
1354          */
1355         get_page(page);
1356         spin_unlock(ptl);
1357         anon_vma = page_lock_anon_vma_read(page);
1358
1359         /* Confirm the PMD did not change while page_table_lock was released */
1360         spin_lock(ptl);
1361         if (unlikely(!pmd_same(pmd, *pmdp))) {
1362                 unlock_page(page);
1363                 put_page(page);
1364                 page_nid = -1;
1365                 goto out_unlock;
1366         }
1367
1368         /* Bail if we fail to protect against THP splits for any reason */
1369         if (unlikely(!anon_vma)) {
1370                 put_page(page);
1371                 page_nid = -1;
1372                 goto clear_pmdnuma;
1373         }
1374
1375         /*
1376          * Migrate the THP to the requested node, returns with page unlocked
1377          * and access rights restored.
1378          */
1379         spin_unlock(ptl);
1380         migrated = migrate_misplaced_transhuge_page(mm, vma,
1381                                 pmdp, pmd, addr, page, target_nid);
1382         if (migrated) {
1383                 flags |= TNF_MIGRATED;
1384                 page_nid = target_nid;
1385         } else
1386                 flags |= TNF_MIGRATE_FAIL;
1387
1388         goto out;
1389 clear_pmdnuma:
1390         BUG_ON(!PageLocked(page));
1391         was_writable = pmd_write(pmd);
1392         pmd = pmd_modify(pmd, vma->vm_page_prot);
1393         pmd = pmd_mkyoung(pmd);
1394         if (was_writable)
1395                 pmd = pmd_mkwrite(pmd);
1396         set_pmd_at(mm, haddr, pmdp, pmd);
1397         update_mmu_cache_pmd(vma, addr, pmdp);
1398         unlock_page(page);
1399 out_unlock:
1400         spin_unlock(ptl);
1401
1402 out:
1403         if (anon_vma)
1404                 page_unlock_anon_vma_read(anon_vma);
1405
1406         if (page_nid != -1)
1407                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1408
1409         return 0;
1410 }
1411
1412 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1413                  pmd_t *pmd, unsigned long addr)
1414 {
1415         spinlock_t *ptl;
1416         int ret = 0;
1417
1418         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1419                 struct page *page;
1420                 pgtable_t pgtable;
1421                 pmd_t orig_pmd;
1422                 /*
1423                  * For architectures like ppc64 we look at deposited pgtable
1424                  * when calling pmdp_huge_get_and_clear. So do the
1425                  * pgtable_trans_huge_withdraw after finishing pmdp related
1426                  * operations.
1427                  */
1428                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1429                                                         tlb->fullmm);
1430                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1431                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1432                 if (is_huge_zero_pmd(orig_pmd)) {
1433                         atomic_long_dec(&tlb->mm->nr_ptes);
1434                         spin_unlock(ptl);
1435                         put_huge_zero_page();
1436                 } else {
1437                         page = pmd_page(orig_pmd);
1438                         page_remove_rmap(page);
1439                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1440                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1441                         VM_BUG_ON_PAGE(!PageHead(page), page);
1442                         atomic_long_dec(&tlb->mm->nr_ptes);
1443                         spin_unlock(ptl);
1444                         tlb_remove_page(tlb, page);
1445                 }
1446                 pte_free(tlb->mm, pgtable);
1447                 ret = 1;
1448         }
1449         return ret;
1450 }
1451
1452 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1453                   unsigned long old_addr,
1454                   unsigned long new_addr, unsigned long old_end,
1455                   pmd_t *old_pmd, pmd_t *new_pmd)
1456 {
1457         spinlock_t *old_ptl, *new_ptl;
1458         int ret = 0;
1459         pmd_t pmd;
1460
1461         struct mm_struct *mm = vma->vm_mm;
1462
1463         if ((old_addr & ~HPAGE_PMD_MASK) ||
1464             (new_addr & ~HPAGE_PMD_MASK) ||
1465             old_end - old_addr < HPAGE_PMD_SIZE ||
1466             (new_vma->vm_flags & VM_NOHUGEPAGE))
1467                 goto out;
1468
1469         /*
1470          * The destination pmd shouldn't be established, free_pgtables()
1471          * should have release it.
1472          */
1473         if (WARN_ON(!pmd_none(*new_pmd))) {
1474                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1475                 goto out;
1476         }
1477
1478         /*
1479          * We don't have to worry about the ordering of src and dst
1480          * ptlocks because exclusive mmap_sem prevents deadlock.
1481          */
1482         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1483         if (ret == 1) {
1484                 new_ptl = pmd_lockptr(mm, new_pmd);
1485                 if (new_ptl != old_ptl)
1486                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1487                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1488                 VM_BUG_ON(!pmd_none(*new_pmd));
1489
1490                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1491                         pgtable_t pgtable;
1492                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1493                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1494                 }
1495                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1496                 if (new_ptl != old_ptl)
1497                         spin_unlock(new_ptl);
1498                 spin_unlock(old_ptl);
1499         }
1500 out:
1501         return ret;
1502 }
1503
1504 /*
1505  * Returns
1506  *  - 0 if PMD could not be locked
1507  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1508  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1509  */
1510 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1511                 unsigned long addr, pgprot_t newprot, int prot_numa)
1512 {
1513         struct mm_struct *mm = vma->vm_mm;
1514         spinlock_t *ptl;
1515         int ret = 0;
1516
1517         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1518                 pmd_t entry;
1519                 bool preserve_write = prot_numa && pmd_write(*pmd);
1520                 ret = 1;
1521
1522                 /*
1523                  * Avoid trapping faults against the zero page. The read-only
1524                  * data is likely to be read-cached on the local CPU and
1525                  * local/remote hits to the zero page are not interesting.
1526                  */
1527                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1528                         spin_unlock(ptl);
1529                         return ret;
1530                 }
1531
1532                 if (!prot_numa || !pmd_protnone(*pmd)) {
1533                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1534                         entry = pmd_modify(entry, newprot);
1535                         if (preserve_write)
1536                                 entry = pmd_mkwrite(entry);
1537                         ret = HPAGE_PMD_NR;
1538                         set_pmd_at(mm, addr, pmd, entry);
1539                         BUG_ON(!preserve_write && pmd_write(entry));
1540                 }
1541                 spin_unlock(ptl);
1542         }
1543
1544         return ret;
1545 }
1546
1547 /*
1548  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1549  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1550  *
1551  * Note that if it returns 1, this routine returns without unlocking page
1552  * table locks. So callers must unlock them.
1553  */
1554 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1555                 spinlock_t **ptl)
1556 {
1557         *ptl = pmd_lock(vma->vm_mm, pmd);
1558         if (likely(pmd_trans_huge(*pmd))) {
1559                 if (unlikely(pmd_trans_splitting(*pmd))) {
1560                         spin_unlock(*ptl);
1561                         wait_split_huge_page(vma->anon_vma, pmd);
1562                         return -1;
1563                 } else {
1564                         /* Thp mapped by 'pmd' is stable, so we can
1565                          * handle it as it is. */
1566                         return 1;
1567                 }
1568         }
1569         spin_unlock(*ptl);
1570         return 0;
1571 }
1572
1573 /*
1574  * This function returns whether a given @page is mapped onto the @address
1575  * in the virtual space of @mm.
1576  *
1577  * When it's true, this function returns *pmd with holding the page table lock
1578  * and passing it back to the caller via @ptl.
1579  * If it's false, returns NULL without holding the page table lock.
1580  */
1581 pmd_t *page_check_address_pmd(struct page *page,
1582                               struct mm_struct *mm,
1583                               unsigned long address,
1584                               enum page_check_address_pmd_flag flag,
1585                               spinlock_t **ptl)
1586 {
1587         pgd_t *pgd;
1588         pud_t *pud;
1589         pmd_t *pmd;
1590
1591         if (address & ~HPAGE_PMD_MASK)
1592                 return NULL;
1593
1594         pgd = pgd_offset(mm, address);
1595         if (!pgd_present(*pgd))
1596                 return NULL;
1597         pud = pud_offset(pgd, address);
1598         if (!pud_present(*pud))
1599                 return NULL;
1600         pmd = pmd_offset(pud, address);
1601
1602         *ptl = pmd_lock(mm, pmd);
1603         if (!pmd_present(*pmd))
1604                 goto unlock;
1605         if (pmd_page(*pmd) != page)
1606                 goto unlock;
1607         /*
1608          * split_vma() may create temporary aliased mappings. There is
1609          * no risk as long as all huge pmd are found and have their
1610          * splitting bit set before __split_huge_page_refcount
1611          * runs. Finding the same huge pmd more than once during the
1612          * same rmap walk is not a problem.
1613          */
1614         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1615             pmd_trans_splitting(*pmd))
1616                 goto unlock;
1617         if (pmd_trans_huge(*pmd)) {
1618                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1619                           !pmd_trans_splitting(*pmd));
1620                 return pmd;
1621         }
1622 unlock:
1623         spin_unlock(*ptl);
1624         return NULL;
1625 }
1626
1627 static int __split_huge_page_splitting(struct page *page,
1628                                        struct vm_area_struct *vma,
1629                                        unsigned long address)
1630 {
1631         struct mm_struct *mm = vma->vm_mm;
1632         spinlock_t *ptl;
1633         pmd_t *pmd;
1634         int ret = 0;
1635         /* For mmu_notifiers */
1636         const unsigned long mmun_start = address;
1637         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1638
1639         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1640         pmd = page_check_address_pmd(page, mm, address,
1641                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1642         if (pmd) {
1643                 /*
1644                  * We can't temporarily set the pmd to null in order
1645                  * to split it, the pmd must remain marked huge at all
1646                  * times or the VM won't take the pmd_trans_huge paths
1647                  * and it won't wait on the anon_vma->root->rwsem to
1648                  * serialize against split_huge_page*.
1649                  */
1650                 pmdp_splitting_flush(vma, address, pmd);
1651
1652                 ret = 1;
1653                 spin_unlock(ptl);
1654         }
1655         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1656
1657         return ret;
1658 }
1659
1660 static void __split_huge_page_refcount(struct page *page,
1661                                        struct list_head *list)
1662 {
1663         int i;
1664         struct zone *zone = page_zone(page);
1665         struct lruvec *lruvec;
1666         int tail_count = 0;
1667
1668         /* prevent PageLRU to go away from under us, and freeze lru stats */
1669         spin_lock_irq(&zone->lru_lock);
1670         lruvec = mem_cgroup_page_lruvec(page, zone);
1671
1672         compound_lock(page);
1673         /* complete memcg works before add pages to LRU */
1674         mem_cgroup_split_huge_fixup(page);
1675
1676         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1677                 struct page *page_tail = page + i;
1678
1679                 /* tail_page->_mapcount cannot change */
1680                 BUG_ON(page_mapcount(page_tail) < 0);
1681                 tail_count += page_mapcount(page_tail);
1682                 /* check for overflow */
1683                 BUG_ON(tail_count < 0);
1684                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1685                 /*
1686                  * tail_page->_count is zero and not changing from
1687                  * under us. But get_page_unless_zero() may be running
1688                  * from under us on the tail_page. If we used
1689                  * atomic_set() below instead of atomic_add(), we
1690                  * would then run atomic_set() concurrently with
1691                  * get_page_unless_zero(), and atomic_set() is
1692                  * implemented in C not using locked ops. spin_unlock
1693                  * on x86 sometime uses locked ops because of PPro
1694                  * errata 66, 92, so unless somebody can guarantee
1695                  * atomic_set() here would be safe on all archs (and
1696                  * not only on x86), it's safer to use atomic_add().
1697                  */
1698                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1699                            &page_tail->_count);
1700
1701                 /* after clearing PageTail the gup refcount can be released */
1702                 smp_mb__after_atomic();
1703
1704                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1705                 page_tail->flags |= (page->flags &
1706                                      ((1L << PG_referenced) |
1707                                       (1L << PG_swapbacked) |
1708                                       (1L << PG_mlocked) |
1709                                       (1L << PG_uptodate) |
1710                                       (1L << PG_active) |
1711                                       (1L << PG_unevictable)));
1712                 page_tail->flags |= (1L << PG_dirty);
1713
1714                 /* clear PageTail before overwriting first_page */
1715                 smp_wmb();
1716
1717                 /*
1718                  * __split_huge_page_splitting() already set the
1719                  * splitting bit in all pmd that could map this
1720                  * hugepage, that will ensure no CPU can alter the
1721                  * mapcount on the head page. The mapcount is only
1722                  * accounted in the head page and it has to be
1723                  * transferred to all tail pages in the below code. So
1724                  * for this code to be safe, the split the mapcount
1725                  * can't change. But that doesn't mean userland can't
1726                  * keep changing and reading the page contents while
1727                  * we transfer the mapcount, so the pmd splitting
1728                  * status is achieved setting a reserved bit in the
1729                  * pmd, not by clearing the present bit.
1730                 */
1731                 page_tail->_mapcount = page->_mapcount;
1732
1733                 BUG_ON(page_tail->mapping);
1734                 page_tail->mapping = page->mapping;
1735
1736                 page_tail->index = page->index + i;
1737                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1738
1739                 BUG_ON(!PageAnon(page_tail));
1740                 BUG_ON(!PageUptodate(page_tail));
1741                 BUG_ON(!PageDirty(page_tail));
1742                 BUG_ON(!PageSwapBacked(page_tail));
1743
1744                 lru_add_page_tail(page, page_tail, lruvec, list);
1745         }
1746         atomic_sub(tail_count, &page->_count);
1747         BUG_ON(atomic_read(&page->_count) <= 0);
1748
1749         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1750
1751         ClearPageCompound(page);
1752         compound_unlock(page);
1753         spin_unlock_irq(&zone->lru_lock);
1754
1755         for (i = 1; i < HPAGE_PMD_NR; i++) {
1756                 struct page *page_tail = page + i;
1757                 BUG_ON(page_count(page_tail) <= 0);
1758                 /*
1759                  * Tail pages may be freed if there wasn't any mapping
1760                  * like if add_to_swap() is running on a lru page that
1761                  * had its mapping zapped. And freeing these pages
1762                  * requires taking the lru_lock so we do the put_page
1763                  * of the tail pages after the split is complete.
1764                  */
1765                 put_page(page_tail);
1766         }
1767
1768         /*
1769          * Only the head page (now become a regular page) is required
1770          * to be pinned by the caller.
1771          */
1772         BUG_ON(page_count(page) <= 0);
1773 }
1774
1775 static int __split_huge_page_map(struct page *page,
1776                                  struct vm_area_struct *vma,
1777                                  unsigned long address)
1778 {
1779         struct mm_struct *mm = vma->vm_mm;
1780         spinlock_t *ptl;
1781         pmd_t *pmd, _pmd;
1782         int ret = 0, i;
1783         pgtable_t pgtable;
1784         unsigned long haddr;
1785
1786         pmd = page_check_address_pmd(page, mm, address,
1787                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1788         if (pmd) {
1789                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1790                 pmd_populate(mm, &_pmd, pgtable);
1791                 if (pmd_write(*pmd))
1792                         BUG_ON(page_mapcount(page) != 1);
1793
1794                 haddr = address;
1795                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1796                         pte_t *pte, entry;
1797                         BUG_ON(PageCompound(page+i));
1798                         /*
1799                          * Note that NUMA hinting access restrictions are not
1800                          * transferred to avoid any possibility of altering
1801                          * permissions across VMAs.
1802                          */
1803                         entry = mk_pte(page + i, vma->vm_page_prot);
1804                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1805                         if (!pmd_write(*pmd))
1806                                 entry = pte_wrprotect(entry);
1807                         if (!pmd_young(*pmd))
1808                                 entry = pte_mkold(entry);
1809                         pte = pte_offset_map(&_pmd, haddr);
1810                         BUG_ON(!pte_none(*pte));
1811                         set_pte_at(mm, haddr, pte, entry);
1812                         pte_unmap(pte);
1813                 }
1814
1815                 smp_wmb(); /* make pte visible before pmd */
1816                 /*
1817                  * Up to this point the pmd is present and huge and
1818                  * userland has the whole access to the hugepage
1819                  * during the split (which happens in place). If we
1820                  * overwrite the pmd with the not-huge version
1821                  * pointing to the pte here (which of course we could
1822                  * if all CPUs were bug free), userland could trigger
1823                  * a small page size TLB miss on the small sized TLB
1824                  * while the hugepage TLB entry is still established
1825                  * in the huge TLB. Some CPU doesn't like that. See
1826                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1827                  * Erratum 383 on page 93. Intel should be safe but is
1828                  * also warns that it's only safe if the permission
1829                  * and cache attributes of the two entries loaded in
1830                  * the two TLB is identical (which should be the case
1831                  * here). But it is generally safer to never allow
1832                  * small and huge TLB entries for the same virtual
1833                  * address to be loaded simultaneously. So instead of
1834                  * doing "pmd_populate(); flush_tlb_range();" we first
1835                  * mark the current pmd notpresent (atomically because
1836                  * here the pmd_trans_huge and pmd_trans_splitting
1837                  * must remain set at all times on the pmd until the
1838                  * split is complete for this pmd), then we flush the
1839                  * SMP TLB and finally we write the non-huge version
1840                  * of the pmd entry with pmd_populate.
1841                  */
1842                 pmdp_invalidate(vma, address, pmd);
1843                 pmd_populate(mm, pmd, pgtable);
1844                 ret = 1;
1845                 spin_unlock(ptl);
1846         }
1847
1848         return ret;
1849 }
1850
1851 /* must be called with anon_vma->root->rwsem held */
1852 static void __split_huge_page(struct page *page,
1853                               struct anon_vma *anon_vma,
1854                               struct list_head *list)
1855 {
1856         int mapcount, mapcount2;
1857         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1858         struct anon_vma_chain *avc;
1859
1860         BUG_ON(!PageHead(page));
1861         BUG_ON(PageTail(page));
1862
1863         mapcount = 0;
1864         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1865                 struct vm_area_struct *vma = avc->vma;
1866                 unsigned long addr = vma_address(page, vma);
1867                 BUG_ON(is_vma_temporary_stack(vma));
1868                 mapcount += __split_huge_page_splitting(page, vma, addr);
1869         }
1870         /*
1871          * It is critical that new vmas are added to the tail of the
1872          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1873          * and establishes a child pmd before
1874          * __split_huge_page_splitting() freezes the parent pmd (so if
1875          * we fail to prevent copy_huge_pmd() from running until the
1876          * whole __split_huge_page() is complete), we will still see
1877          * the newly established pmd of the child later during the
1878          * walk, to be able to set it as pmd_trans_splitting too.
1879          */
1880         if (mapcount != page_mapcount(page)) {
1881                 pr_err("mapcount %d page_mapcount %d\n",
1882                         mapcount, page_mapcount(page));
1883                 BUG();
1884         }
1885
1886         __split_huge_page_refcount(page, list);
1887
1888         mapcount2 = 0;
1889         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1890                 struct vm_area_struct *vma = avc->vma;
1891                 unsigned long addr = vma_address(page, vma);
1892                 BUG_ON(is_vma_temporary_stack(vma));
1893                 mapcount2 += __split_huge_page_map(page, vma, addr);
1894         }
1895         if (mapcount != mapcount2) {
1896                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1897                         mapcount, mapcount2, page_mapcount(page));
1898                 BUG();
1899         }
1900 }
1901
1902 /*
1903  * Split a hugepage into normal pages. This doesn't change the position of head
1904  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1905  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1906  * from the hugepage.
1907  * Return 0 if the hugepage is split successfully otherwise return 1.
1908  */
1909 int split_huge_page_to_list(struct page *page, struct list_head *list)
1910 {
1911         struct anon_vma *anon_vma;
1912         int ret = 1;
1913
1914         BUG_ON(is_huge_zero_page(page));
1915         BUG_ON(!PageAnon(page));
1916
1917         /*
1918          * The caller does not necessarily hold an mmap_sem that would prevent
1919          * the anon_vma disappearing so we first we take a reference to it
1920          * and then lock the anon_vma for write. This is similar to
1921          * page_lock_anon_vma_read except the write lock is taken to serialise
1922          * against parallel split or collapse operations.
1923          */
1924         anon_vma = page_get_anon_vma(page);
1925         if (!anon_vma)
1926                 goto out;
1927         anon_vma_lock_write(anon_vma);
1928
1929         ret = 0;
1930         if (!PageCompound(page))
1931                 goto out_unlock;
1932
1933         BUG_ON(!PageSwapBacked(page));
1934         __split_huge_page(page, anon_vma, list);
1935         count_vm_event(THP_SPLIT);
1936
1937         BUG_ON(PageCompound(page));
1938 out_unlock:
1939         anon_vma_unlock_write(anon_vma);
1940         put_anon_vma(anon_vma);
1941 out:
1942         return ret;
1943 }
1944
1945 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1946
1947 int hugepage_madvise(struct vm_area_struct *vma,
1948                      unsigned long *vm_flags, int advice)
1949 {
1950         switch (advice) {
1951         case MADV_HUGEPAGE:
1952 #ifdef CONFIG_S390
1953                 /*
1954                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1955                  * can't handle this properly after s390_enable_sie, so we simply
1956                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1957                  */
1958                 if (mm_has_pgste(vma->vm_mm))
1959                         return 0;
1960 #endif
1961                 /*
1962                  * Be somewhat over-protective like KSM for now!
1963                  */
1964                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1965                         return -EINVAL;
1966                 *vm_flags &= ~VM_NOHUGEPAGE;
1967                 *vm_flags |= VM_HUGEPAGE;
1968                 /*
1969                  * If the vma become good for khugepaged to scan,
1970                  * register it here without waiting a page fault that
1971                  * may not happen any time soon.
1972                  */
1973                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1974                         return -ENOMEM;
1975                 break;
1976         case MADV_NOHUGEPAGE:
1977                 /*
1978                  * Be somewhat over-protective like KSM for now!
1979                  */
1980                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1981                         return -EINVAL;
1982                 *vm_flags &= ~VM_HUGEPAGE;
1983                 *vm_flags |= VM_NOHUGEPAGE;
1984                 /*
1985                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1986                  * this vma even if we leave the mm registered in khugepaged if
1987                  * it got registered before VM_NOHUGEPAGE was set.
1988                  */
1989                 break;
1990         }
1991
1992         return 0;
1993 }
1994
1995 static int __init khugepaged_slab_init(void)
1996 {
1997         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1998                                           sizeof(struct mm_slot),
1999                                           __alignof__(struct mm_slot), 0, NULL);
2000         if (!mm_slot_cache)
2001                 return -ENOMEM;
2002
2003         return 0;
2004 }
2005
2006 static void __init khugepaged_slab_exit(void)
2007 {
2008         kmem_cache_destroy(mm_slot_cache);
2009 }
2010
2011 static inline struct mm_slot *alloc_mm_slot(void)
2012 {
2013         if (!mm_slot_cache)     /* initialization failed */
2014                 return NULL;
2015         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2016 }
2017
2018 static inline void free_mm_slot(struct mm_slot *mm_slot)
2019 {
2020         kmem_cache_free(mm_slot_cache, mm_slot);
2021 }
2022
2023 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2024 {
2025         struct mm_slot *mm_slot;
2026
2027         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2028                 if (mm == mm_slot->mm)
2029                         return mm_slot;
2030
2031         return NULL;
2032 }
2033
2034 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2035                                     struct mm_slot *mm_slot)
2036 {
2037         mm_slot->mm = mm;
2038         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2039 }
2040
2041 static inline int khugepaged_test_exit(struct mm_struct *mm)
2042 {
2043         return atomic_read(&mm->mm_users) == 0;
2044 }
2045
2046 int __khugepaged_enter(struct mm_struct *mm)
2047 {
2048         struct mm_slot *mm_slot;
2049         int wakeup;
2050
2051         mm_slot = alloc_mm_slot();
2052         if (!mm_slot)
2053                 return -ENOMEM;
2054
2055         /* __khugepaged_exit() must not run from under us */
2056         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2057         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2058                 free_mm_slot(mm_slot);
2059                 return 0;
2060         }
2061
2062         spin_lock(&khugepaged_mm_lock);
2063         insert_to_mm_slots_hash(mm, mm_slot);
2064         /*
2065          * Insert just behind the scanning cursor, to let the area settle
2066          * down a little.
2067          */
2068         wakeup = list_empty(&khugepaged_scan.mm_head);
2069         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2070         spin_unlock(&khugepaged_mm_lock);
2071
2072         atomic_inc(&mm->mm_count);
2073         if (wakeup)
2074                 wake_up_interruptible(&khugepaged_wait);
2075
2076         return 0;
2077 }
2078
2079 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2080                                unsigned long vm_flags)
2081 {
2082         unsigned long hstart, hend;
2083         if (!vma->anon_vma)
2084                 /*
2085                  * Not yet faulted in so we will register later in the
2086                  * page fault if needed.
2087                  */
2088                 return 0;
2089         if (vma->vm_ops)
2090                 /* khugepaged not yet working on file or special mappings */
2091                 return 0;
2092         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2093         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2094         hend = vma->vm_end & HPAGE_PMD_MASK;
2095         if (hstart < hend)
2096                 return khugepaged_enter(vma, vm_flags);
2097         return 0;
2098 }
2099
2100 void __khugepaged_exit(struct mm_struct *mm)
2101 {
2102         struct mm_slot *mm_slot;
2103         int free = 0;
2104
2105         spin_lock(&khugepaged_mm_lock);
2106         mm_slot = get_mm_slot(mm);
2107         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2108                 hash_del(&mm_slot->hash);
2109                 list_del(&mm_slot->mm_node);
2110                 free = 1;
2111         }
2112         spin_unlock(&khugepaged_mm_lock);
2113
2114         if (free) {
2115                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2116                 free_mm_slot(mm_slot);
2117                 mmdrop(mm);
2118         } else if (mm_slot) {
2119                 /*
2120                  * This is required to serialize against
2121                  * khugepaged_test_exit() (which is guaranteed to run
2122                  * under mmap sem read mode). Stop here (after we
2123                  * return all pagetables will be destroyed) until
2124                  * khugepaged has finished working on the pagetables
2125                  * under the mmap_sem.
2126                  */
2127                 down_write(&mm->mmap_sem);
2128                 up_write(&mm->mmap_sem);
2129         }
2130 }
2131
2132 static void release_pte_page(struct page *page)
2133 {
2134         /* 0 stands for page_is_file_cache(page) == false */
2135         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2136         unlock_page(page);
2137         putback_lru_page(page);
2138 }
2139
2140 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2141 {
2142         while (--_pte >= pte) {
2143                 pte_t pteval = *_pte;
2144                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2145                         release_pte_page(pte_page(pteval));
2146         }
2147 }
2148
2149 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2150                                         unsigned long address,
2151                                         pte_t *pte)
2152 {
2153         struct page *page;
2154         pte_t *_pte;
2155         int none_or_zero = 0;
2156         bool referenced = false, writable = false;
2157         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2158              _pte++, address += PAGE_SIZE) {
2159                 pte_t pteval = *_pte;
2160                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2161                         if (++none_or_zero <= khugepaged_max_ptes_none)
2162                                 continue;
2163                         else
2164                                 goto out;
2165                 }
2166                 if (!pte_present(pteval))
2167                         goto out;
2168                 page = vm_normal_page(vma, address, pteval);
2169                 if (unlikely(!page))
2170                         goto out;
2171
2172                 VM_BUG_ON_PAGE(PageCompound(page), page);
2173                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2174                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2175
2176                 /*
2177                  * We can do it before isolate_lru_page because the
2178                  * page can't be freed from under us. NOTE: PG_lock
2179                  * is needed to serialize against split_huge_page
2180                  * when invoked from the VM.
2181                  */
2182                 if (!trylock_page(page))
2183                         goto out;
2184
2185                 /*
2186                  * cannot use mapcount: can't collapse if there's a gup pin.
2187                  * The page must only be referenced by the scanned process
2188                  * and page swap cache.
2189                  */
2190                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2191                         unlock_page(page);
2192                         goto out;
2193                 }
2194                 if (pte_write(pteval)) {
2195                         writable = true;
2196                 } else {
2197                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2198                                 unlock_page(page);
2199                                 goto out;
2200                         }
2201                         /*
2202                          * Page is not in the swap cache. It can be collapsed
2203                          * into a THP.
2204                          */
2205                 }
2206
2207                 /*
2208                  * Isolate the page to avoid collapsing an hugepage
2209                  * currently in use by the VM.
2210                  */
2211                 if (isolate_lru_page(page)) {
2212                         unlock_page(page);
2213                         goto out;
2214                 }
2215                 /* 0 stands for page_is_file_cache(page) == false */
2216                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2217                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2218                 VM_BUG_ON_PAGE(PageLRU(page), page);
2219
2220                 /* If there is no mapped pte young don't collapse the page */
2221                 if (pte_young(pteval) || PageReferenced(page) ||
2222                     mmu_notifier_test_young(vma->vm_mm, address))
2223                         referenced = true;
2224         }
2225         if (likely(referenced && writable))
2226                 return 1;
2227 out:
2228         release_pte_pages(pte, _pte);
2229         return 0;
2230 }
2231
2232 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2233                                       struct vm_area_struct *vma,
2234                                       unsigned long address,
2235                                       spinlock_t *ptl)
2236 {
2237         pte_t *_pte;
2238         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2239                 pte_t pteval = *_pte;
2240                 struct page *src_page;
2241
2242                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2243                         clear_user_highpage(page, address);
2244                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2245                         if (is_zero_pfn(pte_pfn(pteval))) {
2246                                 /*
2247                                  * ptl mostly unnecessary.
2248                                  */
2249                                 spin_lock(ptl);
2250                                 /*
2251                                  * paravirt calls inside pte_clear here are
2252                                  * superfluous.
2253                                  */
2254                                 pte_clear(vma->vm_mm, address, _pte);
2255                                 spin_unlock(ptl);
2256                         }
2257                 } else {
2258                         src_page = pte_page(pteval);
2259                         copy_user_highpage(page, src_page, address, vma);
2260                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2261                         release_pte_page(src_page);
2262                         /*
2263                          * ptl mostly unnecessary, but preempt has to
2264                          * be disabled to update the per-cpu stats
2265                          * inside page_remove_rmap().
2266                          */
2267                         spin_lock(ptl);
2268                         /*
2269                          * paravirt calls inside pte_clear here are
2270                          * superfluous.
2271                          */
2272                         pte_clear(vma->vm_mm, address, _pte);
2273                         page_remove_rmap(src_page);
2274                         spin_unlock(ptl);
2275                         free_page_and_swap_cache(src_page);
2276                 }
2277
2278                 address += PAGE_SIZE;
2279                 page++;
2280         }
2281 }
2282
2283 static void khugepaged_alloc_sleep(void)
2284 {
2285         wait_event_freezable_timeout(khugepaged_wait, false,
2286                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2287 }
2288
2289 static int khugepaged_node_load[MAX_NUMNODES];
2290
2291 static bool khugepaged_scan_abort(int nid)
2292 {
2293         int i;
2294
2295         /*
2296          * If zone_reclaim_mode is disabled, then no extra effort is made to
2297          * allocate memory locally.
2298          */
2299         if (!zone_reclaim_mode)
2300                 return false;
2301
2302         /* If there is a count for this node already, it must be acceptable */
2303         if (khugepaged_node_load[nid])
2304                 return false;
2305
2306         for (i = 0; i < MAX_NUMNODES; i++) {
2307                 if (!khugepaged_node_load[i])
2308                         continue;
2309                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2310                         return true;
2311         }
2312         return false;
2313 }
2314
2315 #ifdef CONFIG_NUMA
2316 static int khugepaged_find_target_node(void)
2317 {
2318         static int last_khugepaged_target_node = NUMA_NO_NODE;
2319         int nid, target_node = 0, max_value = 0;
2320
2321         /* find first node with max normal pages hit */
2322         for (nid = 0; nid < MAX_NUMNODES; nid++)
2323                 if (khugepaged_node_load[nid] > max_value) {
2324                         max_value = khugepaged_node_load[nid];
2325                         target_node = nid;
2326                 }
2327
2328         /* do some balance if several nodes have the same hit record */
2329         if (target_node <= last_khugepaged_target_node)
2330                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2331                                 nid++)
2332                         if (max_value == khugepaged_node_load[nid]) {
2333                                 target_node = nid;
2334                                 break;
2335                         }
2336
2337         last_khugepaged_target_node = target_node;
2338         return target_node;
2339 }
2340
2341 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2342 {
2343         if (IS_ERR(*hpage)) {
2344                 if (!*wait)
2345                         return false;
2346
2347                 *wait = false;
2348                 *hpage = NULL;
2349                 khugepaged_alloc_sleep();
2350         } else if (*hpage) {
2351                 put_page(*hpage);
2352                 *hpage = NULL;
2353         }
2354
2355         return true;
2356 }
2357
2358 static struct page *
2359 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2360                        struct vm_area_struct *vma, unsigned long address,
2361                        int node)
2362 {
2363         VM_BUG_ON_PAGE(*hpage, *hpage);
2364
2365         /*
2366          * Before allocating the hugepage, release the mmap_sem read lock.
2367          * The allocation can take potentially a long time if it involves
2368          * sync compaction, and we do not need to hold the mmap_sem during
2369          * that. We will recheck the vma after taking it again in write mode.
2370          */
2371         up_read(&mm->mmap_sem);
2372
2373         *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2374         if (unlikely(!*hpage)) {
2375                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2376                 *hpage = ERR_PTR(-ENOMEM);
2377                 return NULL;
2378         }
2379
2380         count_vm_event(THP_COLLAPSE_ALLOC);
2381         return *hpage;
2382 }
2383 #else
2384 static int khugepaged_find_target_node(void)
2385 {
2386         return 0;
2387 }
2388
2389 static inline struct page *alloc_hugepage(int defrag)
2390 {
2391         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2392                            HPAGE_PMD_ORDER);
2393 }
2394
2395 static struct page *khugepaged_alloc_hugepage(bool *wait)
2396 {
2397         struct page *hpage;
2398
2399         do {
2400                 hpage = alloc_hugepage(khugepaged_defrag());
2401                 if (!hpage) {
2402                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2403                         if (!*wait)
2404                                 return NULL;
2405
2406                         *wait = false;
2407                         khugepaged_alloc_sleep();
2408                 } else
2409                         count_vm_event(THP_COLLAPSE_ALLOC);
2410         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2411
2412         return hpage;
2413 }
2414
2415 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2416 {
2417         if (!*hpage)
2418                 *hpage = khugepaged_alloc_hugepage(wait);
2419
2420         if (unlikely(!*hpage))
2421                 return false;
2422
2423         return true;
2424 }
2425
2426 static struct page *
2427 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2428                        struct vm_area_struct *vma, unsigned long address,
2429                        int node)
2430 {
2431         up_read(&mm->mmap_sem);
2432         VM_BUG_ON(!*hpage);
2433
2434         return  *hpage;
2435 }
2436 #endif
2437
2438 static bool hugepage_vma_check(struct vm_area_struct *vma)
2439 {
2440         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2441             (vma->vm_flags & VM_NOHUGEPAGE))
2442                 return false;
2443
2444         if (!vma->anon_vma || vma->vm_ops)
2445                 return false;
2446         if (is_vma_temporary_stack(vma))
2447                 return false;
2448         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2449         return true;
2450 }
2451
2452 static void collapse_huge_page(struct mm_struct *mm,
2453                                    unsigned long address,
2454                                    struct page **hpage,
2455                                    struct vm_area_struct *vma,
2456                                    int node)
2457 {
2458         pmd_t *pmd, _pmd;
2459         pte_t *pte;
2460         pgtable_t pgtable;
2461         struct page *new_page;
2462         spinlock_t *pmd_ptl, *pte_ptl;
2463         int isolated;
2464         unsigned long hstart, hend;
2465         struct mem_cgroup *memcg;
2466         unsigned long mmun_start;       /* For mmu_notifiers */
2467         unsigned long mmun_end;         /* For mmu_notifiers */
2468         gfp_t gfp;
2469
2470         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2471
2472         /* Only allocate from the target node */
2473         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2474                 __GFP_THISNODE;
2475
2476         /* release the mmap_sem read lock. */
2477         new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2478         if (!new_page)
2479                 return;
2480
2481         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2482                                            gfp, &memcg)))
2483                 return;
2484
2485         /*
2486          * Prevent all access to pagetables with the exception of
2487          * gup_fast later hanlded by the ptep_clear_flush and the VM
2488          * handled by the anon_vma lock + PG_lock.
2489          */
2490         down_write(&mm->mmap_sem);
2491         if (unlikely(khugepaged_test_exit(mm)))
2492                 goto out;
2493
2494         vma = find_vma(mm, address);
2495         if (!vma)
2496                 goto out;
2497         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2498         hend = vma->vm_end & HPAGE_PMD_MASK;
2499         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2500                 goto out;
2501         if (!hugepage_vma_check(vma))
2502                 goto out;
2503         pmd = mm_find_pmd(mm, address);
2504         if (!pmd)
2505                 goto out;
2506
2507         anon_vma_lock_write(vma->anon_vma);
2508
2509         pte = pte_offset_map(pmd, address);
2510         pte_ptl = pte_lockptr(mm, pmd);
2511
2512         mmun_start = address;
2513         mmun_end   = address + HPAGE_PMD_SIZE;
2514         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2515         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2516         /*
2517          * After this gup_fast can't run anymore. This also removes
2518          * any huge TLB entry from the CPU so we won't allow
2519          * huge and small TLB entries for the same virtual address
2520          * to avoid the risk of CPU bugs in that area.
2521          */
2522         _pmd = pmdp_collapse_flush(vma, address, pmd);
2523         spin_unlock(pmd_ptl);
2524         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2525
2526         spin_lock(pte_ptl);
2527         isolated = __collapse_huge_page_isolate(vma, address, pte);
2528         spin_unlock(pte_ptl);
2529
2530         if (unlikely(!isolated)) {
2531                 pte_unmap(pte);
2532                 spin_lock(pmd_ptl);
2533                 BUG_ON(!pmd_none(*pmd));
2534                 /*
2535                  * We can only use set_pmd_at when establishing
2536                  * hugepmds and never for establishing regular pmds that
2537                  * points to regular pagetables. Use pmd_populate for that
2538                  */
2539                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2540                 spin_unlock(pmd_ptl);
2541                 anon_vma_unlock_write(vma->anon_vma);
2542                 goto out;
2543         }
2544
2545         /*
2546          * All pages are isolated and locked so anon_vma rmap
2547          * can't run anymore.
2548          */
2549         anon_vma_unlock_write(vma->anon_vma);
2550
2551         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2552         pte_unmap(pte);
2553         __SetPageUptodate(new_page);
2554         pgtable = pmd_pgtable(_pmd);
2555
2556         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2557         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2558
2559         /*
2560          * spin_lock() below is not the equivalent of smp_wmb(), so
2561          * this is needed to avoid the copy_huge_page writes to become
2562          * visible after the set_pmd_at() write.
2563          */
2564         smp_wmb();
2565
2566         spin_lock(pmd_ptl);
2567         BUG_ON(!pmd_none(*pmd));
2568         page_add_new_anon_rmap(new_page, vma, address);
2569         mem_cgroup_commit_charge(new_page, memcg, false);
2570         lru_cache_add_active_or_unevictable(new_page, vma);
2571         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2572         set_pmd_at(mm, address, pmd, _pmd);
2573         update_mmu_cache_pmd(vma, address, pmd);
2574         spin_unlock(pmd_ptl);
2575
2576         *hpage = NULL;
2577
2578         khugepaged_pages_collapsed++;
2579 out_up_write:
2580         up_write(&mm->mmap_sem);
2581         return;
2582
2583 out:
2584         mem_cgroup_cancel_charge(new_page, memcg);
2585         goto out_up_write;
2586 }
2587
2588 static int khugepaged_scan_pmd(struct mm_struct *mm,
2589                                struct vm_area_struct *vma,
2590                                unsigned long address,
2591                                struct page **hpage)
2592 {
2593         pmd_t *pmd;
2594         pte_t *pte, *_pte;
2595         int ret = 0, none_or_zero = 0;
2596         struct page *page;
2597         unsigned long _address;
2598         spinlock_t *ptl;
2599         int node = NUMA_NO_NODE;
2600         bool writable = false, referenced = false;
2601
2602         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2603
2604         pmd = mm_find_pmd(mm, address);
2605         if (!pmd)
2606                 goto out;
2607
2608         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2609         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2610         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2611              _pte++, _address += PAGE_SIZE) {
2612                 pte_t pteval = *_pte;
2613                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2614                         if (++none_or_zero <= khugepaged_max_ptes_none)
2615                                 continue;
2616                         else
2617                                 goto out_unmap;
2618                 }
2619                 if (!pte_present(pteval))
2620                         goto out_unmap;
2621                 if (pte_write(pteval))
2622                         writable = true;
2623
2624                 page = vm_normal_page(vma, _address, pteval);
2625                 if (unlikely(!page))
2626                         goto out_unmap;
2627                 /*
2628                  * Record which node the original page is from and save this
2629                  * information to khugepaged_node_load[].
2630                  * Khupaged will allocate hugepage from the node has the max
2631                  * hit record.
2632                  */
2633                 node = page_to_nid(page);
2634                 if (khugepaged_scan_abort(node))
2635                         goto out_unmap;
2636                 khugepaged_node_load[node]++;
2637                 VM_BUG_ON_PAGE(PageCompound(page), page);
2638                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2639                         goto out_unmap;
2640                 /*
2641                  * cannot use mapcount: can't collapse if there's a gup pin.
2642                  * The page must only be referenced by the scanned process
2643                  * and page swap cache.
2644                  */
2645                 if (page_count(page) != 1 + !!PageSwapCache(page))
2646                         goto out_unmap;
2647                 if (pte_young(pteval) || PageReferenced(page) ||
2648                     mmu_notifier_test_young(vma->vm_mm, address))
2649                         referenced = true;
2650         }
2651         if (referenced && writable)
2652                 ret = 1;
2653 out_unmap:
2654         pte_unmap_unlock(pte, ptl);
2655         if (ret) {
2656                 node = khugepaged_find_target_node();
2657                 /* collapse_huge_page will return with the mmap_sem released */
2658                 collapse_huge_page(mm, address, hpage, vma, node);
2659         }
2660 out:
2661         return ret;
2662 }
2663
2664 static void collect_mm_slot(struct mm_slot *mm_slot)
2665 {
2666         struct mm_struct *mm = mm_slot->mm;
2667
2668         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2669
2670         if (khugepaged_test_exit(mm)) {
2671                 /* free mm_slot */
2672                 hash_del(&mm_slot->hash);
2673                 list_del(&mm_slot->mm_node);
2674
2675                 /*
2676                  * Not strictly needed because the mm exited already.
2677                  *
2678                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2679                  */
2680
2681                 /* khugepaged_mm_lock actually not necessary for the below */
2682                 free_mm_slot(mm_slot);
2683                 mmdrop(mm);
2684         }
2685 }
2686
2687 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2688                                             struct page **hpage)
2689         __releases(&khugepaged_mm_lock)
2690         __acquires(&khugepaged_mm_lock)
2691 {
2692         struct mm_slot *mm_slot;
2693         struct mm_struct *mm;
2694         struct vm_area_struct *vma;
2695         int progress = 0;
2696
2697         VM_BUG_ON(!pages);
2698         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2699
2700         if (khugepaged_scan.mm_slot)
2701                 mm_slot = khugepaged_scan.mm_slot;
2702         else {
2703                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2704                                      struct mm_slot, mm_node);
2705                 khugepaged_scan.address = 0;
2706                 khugepaged_scan.mm_slot = mm_slot;
2707         }
2708         spin_unlock(&khugepaged_mm_lock);
2709
2710         mm = mm_slot->mm;
2711         down_read(&mm->mmap_sem);
2712         if (unlikely(khugepaged_test_exit(mm)))
2713                 vma = NULL;
2714         else
2715                 vma = find_vma(mm, khugepaged_scan.address);
2716
2717         progress++;
2718         for (; vma; vma = vma->vm_next) {
2719                 unsigned long hstart, hend;
2720
2721                 cond_resched();
2722                 if (unlikely(khugepaged_test_exit(mm))) {
2723                         progress++;
2724                         break;
2725                 }
2726                 if (!hugepage_vma_check(vma)) {
2727 skip:
2728                         progress++;
2729                         continue;
2730                 }
2731                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2732                 hend = vma->vm_end & HPAGE_PMD_MASK;
2733                 if (hstart >= hend)
2734                         goto skip;
2735                 if (khugepaged_scan.address > hend)
2736                         goto skip;
2737                 if (khugepaged_scan.address < hstart)
2738                         khugepaged_scan.address = hstart;
2739                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2740
2741                 while (khugepaged_scan.address < hend) {
2742                         int ret;
2743                         cond_resched();
2744                         if (unlikely(khugepaged_test_exit(mm)))
2745                                 goto breakouterloop;
2746
2747                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2748                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2749                                   hend);
2750                         ret = khugepaged_scan_pmd(mm, vma,
2751                                                   khugepaged_scan.address,
2752                                                   hpage);
2753                         /* move to next address */
2754                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2755                         progress += HPAGE_PMD_NR;
2756                         if (ret)
2757                                 /* we released mmap_sem so break loop */
2758                                 goto breakouterloop_mmap_sem;
2759                         if (progress >= pages)
2760                                 goto breakouterloop;
2761                 }
2762         }
2763 breakouterloop:
2764         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2765 breakouterloop_mmap_sem:
2766
2767         spin_lock(&khugepaged_mm_lock);
2768         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2769         /*
2770          * Release the current mm_slot if this mm is about to die, or
2771          * if we scanned all vmas of this mm.
2772          */
2773         if (khugepaged_test_exit(mm) || !vma) {
2774                 /*
2775                  * Make sure that if mm_users is reaching zero while
2776                  * khugepaged runs here, khugepaged_exit will find
2777                  * mm_slot not pointing to the exiting mm.
2778                  */
2779                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2780                         khugepaged_scan.mm_slot = list_entry(
2781                                 mm_slot->mm_node.next,
2782                                 struct mm_slot, mm_node);
2783                         khugepaged_scan.address = 0;
2784                 } else {
2785                         khugepaged_scan.mm_slot = NULL;
2786                         khugepaged_full_scans++;
2787                 }
2788
2789                 collect_mm_slot(mm_slot);
2790         }
2791
2792         return progress;
2793 }
2794
2795 static int khugepaged_has_work(void)
2796 {
2797         return !list_empty(&khugepaged_scan.mm_head) &&
2798                 khugepaged_enabled();
2799 }
2800
2801 static int khugepaged_wait_event(void)
2802 {
2803         return !list_empty(&khugepaged_scan.mm_head) ||
2804                 kthread_should_stop();
2805 }
2806
2807 static void khugepaged_do_scan(void)
2808 {
2809         struct page *hpage = NULL;
2810         unsigned int progress = 0, pass_through_head = 0;
2811         unsigned int pages = khugepaged_pages_to_scan;
2812         bool wait = true;
2813
2814         barrier(); /* write khugepaged_pages_to_scan to local stack */
2815
2816         while (progress < pages) {
2817                 if (!khugepaged_prealloc_page(&hpage, &wait))
2818                         break;
2819
2820                 cond_resched();
2821
2822                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2823                         break;
2824
2825                 spin_lock(&khugepaged_mm_lock);
2826                 if (!khugepaged_scan.mm_slot)
2827                         pass_through_head++;
2828                 if (khugepaged_has_work() &&
2829                     pass_through_head < 2)
2830                         progress += khugepaged_scan_mm_slot(pages - progress,
2831                                                             &hpage);
2832                 else
2833                         progress = pages;
2834                 spin_unlock(&khugepaged_mm_lock);
2835         }
2836
2837         if (!IS_ERR_OR_NULL(hpage))
2838                 put_page(hpage);
2839 }
2840
2841 static void khugepaged_wait_work(void)
2842 {
2843         if (khugepaged_has_work()) {
2844                 if (!khugepaged_scan_sleep_millisecs)
2845                         return;
2846
2847                 wait_event_freezable_timeout(khugepaged_wait,
2848                                              kthread_should_stop(),
2849                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2850                 return;
2851         }
2852
2853         if (khugepaged_enabled())
2854                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2855 }
2856
2857 static int khugepaged(void *none)
2858 {
2859         struct mm_slot *mm_slot;
2860
2861         set_freezable();
2862         set_user_nice(current, MAX_NICE);
2863
2864         while (!kthread_should_stop()) {
2865                 khugepaged_do_scan();
2866                 khugepaged_wait_work();
2867         }
2868
2869         spin_lock(&khugepaged_mm_lock);
2870         mm_slot = khugepaged_scan.mm_slot;
2871         khugepaged_scan.mm_slot = NULL;
2872         if (mm_slot)
2873                 collect_mm_slot(mm_slot);
2874         spin_unlock(&khugepaged_mm_lock);
2875         return 0;
2876 }
2877
2878 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2879                 unsigned long haddr, pmd_t *pmd)
2880 {
2881         struct mm_struct *mm = vma->vm_mm;
2882         pgtable_t pgtable;
2883         pmd_t _pmd;
2884         int i;
2885
2886         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2887         /* leave pmd empty until pte is filled */
2888
2889         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2890         pmd_populate(mm, &_pmd, pgtable);
2891
2892         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2893                 pte_t *pte, entry;
2894                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2895                 entry = pte_mkspecial(entry);
2896                 pte = pte_offset_map(&_pmd, haddr);
2897                 VM_BUG_ON(!pte_none(*pte));
2898                 set_pte_at(mm, haddr, pte, entry);
2899                 pte_unmap(pte);
2900         }
2901         smp_wmb(); /* make pte visible before pmd */
2902         pmd_populate(mm, pmd, pgtable);
2903         put_huge_zero_page();
2904 }
2905
2906 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2907                 pmd_t *pmd)
2908 {
2909         spinlock_t *ptl;
2910         struct page *page;
2911         struct mm_struct *mm = vma->vm_mm;
2912         unsigned long haddr = address & HPAGE_PMD_MASK;
2913         unsigned long mmun_start;       /* For mmu_notifiers */
2914         unsigned long mmun_end;         /* For mmu_notifiers */
2915
2916         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2917
2918         mmun_start = haddr;
2919         mmun_end   = haddr + HPAGE_PMD_SIZE;
2920 again:
2921         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2922         ptl = pmd_lock(mm, pmd);
2923         if (unlikely(!pmd_trans_huge(*pmd))) {
2924                 spin_unlock(ptl);
2925                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2926                 return;
2927         }
2928         if (is_huge_zero_pmd(*pmd)) {
2929                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2930                 spin_unlock(ptl);
2931                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2932                 return;
2933         }
2934         page = pmd_page(*pmd);
2935         VM_BUG_ON_PAGE(!page_count(page), page);
2936         get_page(page);
2937         spin_unlock(ptl);
2938         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2939
2940         split_huge_page(page);
2941
2942         put_page(page);
2943
2944         /*
2945          * We don't always have down_write of mmap_sem here: a racing
2946          * do_huge_pmd_wp_page() might have copied-on-write to another
2947          * huge page before our split_huge_page() got the anon_vma lock.
2948          */
2949         if (unlikely(pmd_trans_huge(*pmd)))
2950                 goto again;
2951 }
2952
2953 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2954                 pmd_t *pmd)
2955 {
2956         struct vm_area_struct *vma;
2957
2958         vma = find_vma(mm, address);
2959         BUG_ON(vma == NULL);
2960         split_huge_page_pmd(vma, address, pmd);
2961 }
2962
2963 static void split_huge_page_address(struct mm_struct *mm,
2964                                     unsigned long address)
2965 {
2966         pgd_t *pgd;
2967         pud_t *pud;
2968         pmd_t *pmd;
2969
2970         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2971
2972         pgd = pgd_offset(mm, address);
2973         if (!pgd_present(*pgd))
2974                 return;
2975
2976         pud = pud_offset(pgd, address);
2977         if (!pud_present(*pud))
2978                 return;
2979
2980         pmd = pmd_offset(pud, address);
2981         if (!pmd_present(*pmd))
2982                 return;
2983         /*
2984          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2985          * materialize from under us.
2986          */
2987         split_huge_page_pmd_mm(mm, address, pmd);
2988 }
2989
2990 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2991                              unsigned long start,
2992                              unsigned long end,
2993                              long adjust_next)
2994 {
2995         /*
2996          * If the new start address isn't hpage aligned and it could
2997          * previously contain an hugepage: check if we need to split
2998          * an huge pmd.
2999          */
3000         if (start & ~HPAGE_PMD_MASK &&
3001             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3002             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3003                 split_huge_page_address(vma->vm_mm, start);
3004
3005         /*
3006          * If the new end address isn't hpage aligned and it could
3007          * previously contain an hugepage: check if we need to split
3008          * an huge pmd.
3009          */
3010         if (end & ~HPAGE_PMD_MASK &&
3011             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3012             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3013                 split_huge_page_address(vma->vm_mm, end);
3014
3015         /*
3016          * If we're also updating the vma->vm_next->vm_start, if the new
3017          * vm_next->vm_start isn't page aligned and it could previously
3018          * contain an hugepage: check if we need to split an huge pmd.
3019          */
3020         if (adjust_next > 0) {
3021                 struct vm_area_struct *next = vma->vm_next;
3022                 unsigned long nstart = next->vm_start;
3023                 nstart += adjust_next << PAGE_SHIFT;
3024                 if (nstart & ~HPAGE_PMD_MASK &&
3025                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3026                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3027                         split_huge_page_address(next->vm_mm, nstart);
3028         }
3029 }