mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/kthread.h>
  20 #include <linux/khugepaged.h>
  21 #include <linux/freezer.h>
  22 #include <linux/mman.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/migrate.h>
  25 #include <linux/hashtable.h>
  26
  27 #include <asm/tlb.h>
  28 #include <asm/pgalloc.h>
  29 #include "internal.h"
  30
  31 /*
  32  * By default transparent hugepage support is disabled in order that avoid
  33  * to risk increase the memory footprint of applications without a guaranteed
  34  * benefit. When transparent hugepage support is enabled, is for all mappings,
  35  * and khugepaged scans all mappings.
  36  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  37  * for all hugepage allocations.
  38  */
  39 unsigned long transparent_hugepage_flags __read_mostly =
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  41         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  42 #endif
  43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  44         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  45 #endif
  46         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  47         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  48         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  49
  50 /* default scan 8*512 pte (or vmas) every 30 second */
  51 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  52 static unsigned int khugepaged_pages_collapsed;
  53 static unsigned int khugepaged_full_scans;
  54 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  55 /* during fragmentation poll the hugepage allocator once every minute */
  56 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  57 static struct task_struct *khugepaged_thread __read_mostly;
  58 static DEFINE_MUTEX(khugepaged_mutex);
  59 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  60 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  61 /*
  62  * default collapse hugepages if there is at least one pte mapped like
  63  * it would have happened if the vma was large enough during page
  64  * fault.
  65  */
  66 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  67
  68 static int khugepaged(void *none);
  69 static int khugepaged_slab_init(void);
  70
  71 #define MM_SLOTS_HASH_BITS 10
  72 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  73
  74 static struct kmem_cache *mm_slot_cache __read_mostly;
  75
  76 /**
  77  * struct mm_slot - hash lookup from mm to mm_slot
  78  * @hash: hash collision list
  79  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  80  * @mm: the mm that this information is valid for
  81  */
  82 struct mm_slot {
  83         struct hlist_node hash;
  84         struct list_head mm_node;
  85         struct mm_struct *mm;
  86 };
  87
  88 /**
  89  * struct khugepaged_scan - cursor for scanning
  90  * @mm_head: the head of the mm list to scan
  91  * @mm_slot: the current mm_slot we are scanning
  92  * @address: the next address inside that to be scanned
  93  *
  94  * There is only the one khugepaged_scan instance of this cursor structure.
  95  */
  96 struct khugepaged_scan {
  97         struct list_head mm_head;
  98         struct mm_slot *mm_slot;
  99         unsigned long address;
 100 };
 101 static struct khugepaged_scan khugepaged_scan = {
 102         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 103 };
 104
 105
 106 static int set_recommended_min_free_kbytes(void)
 107 {
 108         struct zone *zone;
 109         int nr_zones = 0;
 110         unsigned long recommended_min;
 111
 112         if (!khugepaged_enabled())
 113                 return 0;
 114
 115         for_each_populated_zone(zone)
 116                 nr_zones++;
 117
 118         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 119         recommended_min = pageblock_nr_pages * nr_zones * 2;
 120
 121         /*
 122          * Make sure that on average at least two pageblocks are almost free
 123          * of another type, one for a migratetype to fall back to and a
 124          * second to avoid subsequent fallbacks of other types There are 3
 125          * MIGRATE_TYPES we care about.
 126          */
 127         recommended_min += pageblock_nr_pages * nr_zones *
 128                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 129
 130         /* don't ever allow to reserve more than 5% of the lowmem */
 131         recommended_min = min(recommended_min,
 132                               (unsigned long) nr_free_buffer_pages() / 20);
 133         recommended_min <<= (PAGE_SHIFT-10);
 134
 135         if (recommended_min > min_free_kbytes) {
 136                 if (user_min_free_kbytes >= 0)
 137                         pr_info("raising min_free_kbytes from %d to %lu "
 138                                 "to help transparent hugepage allocations\n",
 139                                 min_free_kbytes, recommended_min);
 140
 141                 min_free_kbytes = recommended_min;
 142         }
 143         setup_per_zone_wmarks();
 144         return 0;
 145 }
 146 late_initcall(set_recommended_min_free_kbytes);
 147
 148 static int start_khugepaged(void)
 149 {
 150         int err = 0;
 151         if (khugepaged_enabled()) {
 152                 if (!khugepaged_thread)
 153                         khugepaged_thread = kthread_run(khugepaged, NULL,
 154                                                         "khugepaged");
 155                 if (unlikely(IS_ERR(khugepaged_thread))) {
 156                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 157                         err = PTR_ERR(khugepaged_thread);
 158                         khugepaged_thread = NULL;
 159                 }
 160
 161                 if (!list_empty(&khugepaged_scan.mm_head))
 162                         wake_up_interruptible(&khugepaged_wait);
 163
 164                 set_recommended_min_free_kbytes();
 165         } else if (khugepaged_thread) {
 166                 kthread_stop(khugepaged_thread);
 167                 khugepaged_thread = NULL;
 168         }
 169
 170         return err;
 171 }
 172
 173 static atomic_t huge_zero_refcount;
 174 struct page *huge_zero_page __read_mostly;
 175
 176 static inline bool is_huge_zero_pmd(pmd_t pmd)
 177 {
 178         return is_huge_zero_page(pmd_page(pmd));
 179 }
 180
 181 static struct page *get_huge_zero_page(void)
 182 {
 183         struct page *zero_page;
 184 retry:
 185         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 186                 return ACCESS_ONCE(huge_zero_page);
 187
 188         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 189                         HPAGE_PMD_ORDER);
 190         if (!zero_page) {
 191                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 192                 return NULL;
 193         }
 194         count_vm_event(THP_ZERO_PAGE_ALLOC);
 195         preempt_disable();
 196         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 197                 preempt_enable();
 198                 __free_pages(zero_page, compound_order(zero_page));
 199                 goto retry;
 200         }
 201
 202         /* We take additional reference here. It will be put back by shrinker */
 203         atomic_set(&huge_zero_refcount, 2);
 204         preempt_enable();
 205         return ACCESS_ONCE(huge_zero_page);
 206 }
 207
 208 static void put_huge_zero_page(void)
 209 {
 210         /*
 211          * Counter should never go to zero here. Only shrinker can put
 212          * last reference.
 213          */
 214         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 215 }
 216
 217 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 218                                         struct shrink_control *sc)
 219 {
 220         /* we can free zero page only if last reference remains */
 221         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 222 }
 223
 224 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 225                                        struct shrink_control *sc)
 226 {
 227         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 228                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 229                 BUG_ON(zero_page == NULL);
 230                 __free_pages(zero_page, compound_order(zero_page));
 231                 return HPAGE_PMD_NR;
 232         }
 233
 234         return 0;
 235 }
 236
 237 static struct shrinker huge_zero_page_shrinker = {
 238         .count_objects = shrink_huge_zero_page_count,
 239         .scan_objects = shrink_huge_zero_page_scan,
 240         .seeks = DEFAULT_SEEKS,
 241 };
 242
 243 #ifdef CONFIG_SYSFS
 244
 245 static ssize_t double_flag_show(struct kobject *kobj,
 246                                 struct kobj_attribute *attr, char *buf,
 247                                 enum transparent_hugepage_flag enabled,
 248                                 enum transparent_hugepage_flag req_madv)
 249 {
 250         if (test_bit(enabled, &transparent_hugepage_flags)) {
 251                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 252                 return sprintf(buf, "[always] madvise never\n");
 253         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 254                 return sprintf(buf, "always [madvise] never\n");
 255         else
 256                 return sprintf(buf, "always madvise [never]\n");
 257 }
 258 static ssize_t double_flag_store(struct kobject *kobj,
 259                                  struct kobj_attribute *attr,
 260                                  const char *buf, size_t count,
 261                                  enum transparent_hugepage_flag enabled,
 262                                  enum transparent_hugepage_flag req_madv)
 263 {
 264         if (!memcmp("always", buf,
 265                     min(sizeof("always")-1, count))) {
 266                 set_bit(enabled, &transparent_hugepage_flags);
 267                 clear_bit(req_madv, &transparent_hugepage_flags);
 268         } else if (!memcmp("madvise", buf,
 269                            min(sizeof("madvise")-1, count))) {
 270                 clear_bit(enabled, &transparent_hugepage_flags);
 271                 set_bit(req_madv, &transparent_hugepage_flags);
 272         } else if (!memcmp("never", buf,
 273                            min(sizeof("never")-1, count))) {
 274                 clear_bit(enabled, &transparent_hugepage_flags);
 275                 clear_bit(req_madv, &transparent_hugepage_flags);
 276         } else
 277                 return -EINVAL;
 278
 279         return count;
 280 }
 281
 282 static ssize_t enabled_show(struct kobject *kobj,
 283                             struct kobj_attribute *attr, char *buf)
 284 {
 285         return double_flag_show(kobj, attr, buf,
 286                                 TRANSPARENT_HUGEPAGE_FLAG,
 287                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 288 }
 289 static ssize_t enabled_store(struct kobject *kobj,
 290                              struct kobj_attribute *attr,
 291                              const char *buf, size_t count)
 292 {
 293         ssize_t ret;
 294
 295         ret = double_flag_store(kobj, attr, buf, count,
 296                                 TRANSPARENT_HUGEPAGE_FLAG,
 297                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 298
 299         if (ret > 0) {
 300                 int err;
 301
 302                 mutex_lock(&khugepaged_mutex);
 303                 err = start_khugepaged();
 304                 mutex_unlock(&khugepaged_mutex);
 305
 306                 if (err)
 307                         ret = err;
 308         }
 309
 310         return ret;
 311 }
 312 static struct kobj_attribute enabled_attr =
 313         __ATTR(enabled, 0644, enabled_show, enabled_store);
 314
 315 static ssize_t single_flag_show(struct kobject *kobj,
 316                                 struct kobj_attribute *attr, char *buf,
 317                                 enum transparent_hugepage_flag flag)
 318 {
 319         return sprintf(buf, "%d\n",
 320                        !!test_bit(flag, &transparent_hugepage_flags));
 321 }
 322
 323 static ssize_t single_flag_store(struct kobject *kobj,
 324                                  struct kobj_attribute *attr,
 325                                  const char *buf, size_t count,
 326                                  enum transparent_hugepage_flag flag)
 327 {
 328         unsigned long value;
 329         int ret;
 330
 331         ret = kstrtoul(buf, 10, &value);
 332         if (ret < 0)
 333                 return ret;
 334         if (value > 1)
 335                 return -EINVAL;
 336
 337         if (value)
 338                 set_bit(flag, &transparent_hugepage_flags);
 339         else
 340                 clear_bit(flag, &transparent_hugepage_flags);
 341
 342         return count;
 343 }
 344
 345 /*
 346  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 347  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 348  * memory just to allocate one more hugepage.
 349  */
 350 static ssize_t defrag_show(struct kobject *kobj,
 351                            struct kobj_attribute *attr, char *buf)
 352 {
 353         return double_flag_show(kobj, attr, buf,
 354                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 355                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 356 }
 357 static ssize_t defrag_store(struct kobject *kobj,
 358                             struct kobj_attribute *attr,
 359                             const char *buf, size_t count)
 360 {
 361         return double_flag_store(kobj, attr, buf, count,
 362                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 363                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 364 }
 365 static struct kobj_attribute defrag_attr =
 366         __ATTR(defrag, 0644, defrag_show, defrag_store);
 367
 368 static ssize_t use_zero_page_show(struct kobject *kobj,
 369                 struct kobj_attribute *attr, char *buf)
 370 {
 371         return single_flag_show(kobj, attr, buf,
 372                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 373 }
 374 static ssize_t use_zero_page_store(struct kobject *kobj,
 375                 struct kobj_attribute *attr, const char *buf, size_t count)
 376 {
 377         return single_flag_store(kobj, attr, buf, count,
 378                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 379 }
 380 static struct kobj_attribute use_zero_page_attr =
 381         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 382 #ifdef CONFIG_DEBUG_VM
 383 static ssize_t debug_cow_show(struct kobject *kobj,
 384                                 struct kobj_attribute *attr, char *buf)
 385 {
 386         return single_flag_show(kobj, attr, buf,
 387                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 388 }
 389 static ssize_t debug_cow_store(struct kobject *kobj,
 390                                struct kobj_attribute *attr,
 391                                const char *buf, size_t count)
 392 {
 393         return single_flag_store(kobj, attr, buf, count,
 394                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 395 }
 396 static struct kobj_attribute debug_cow_attr =
 397         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 398 #endif /* CONFIG_DEBUG_VM */
 399
 400 static struct attribute *hugepage_attr[] = {
 401         &enabled_attr.attr,
 402         &defrag_attr.attr,
 403         &use_zero_page_attr.attr,
 404 #ifdef CONFIG_DEBUG_VM
 405         &debug_cow_attr.attr,
 406 #endif
 407         NULL,
 408 };
 409
 410 static struct attribute_group hugepage_attr_group = {
 411         .attrs = hugepage_attr,
 412 };
 413
 414 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 415                                          struct kobj_attribute *attr,
 416                                          char *buf)
 417 {
 418         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 419 }
 420
 421 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 422                                           struct kobj_attribute *attr,
 423                                           const char *buf, size_t count)
 424 {
 425         unsigned long msecs;
 426         int err;
 427
 428         err = kstrtoul(buf, 10, &msecs);
 429         if (err || msecs > UINT_MAX)
 430                 return -EINVAL;
 431
 432         khugepaged_scan_sleep_millisecs = msecs;
 433         wake_up_interruptible(&khugepaged_wait);
 434
 435         return count;
 436 }
 437 static struct kobj_attribute scan_sleep_millisecs_attr =
 438         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 439                scan_sleep_millisecs_store);
 440
 441 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 442                                           struct kobj_attribute *attr,
 443                                           char *buf)
 444 {
 445         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 446 }
 447
 448 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 449                                            struct kobj_attribute *attr,
 450                                            const char *buf, size_t count)
 451 {
 452         unsigned long msecs;
 453         int err;
 454
 455         err = kstrtoul(buf, 10, &msecs);
 456         if (err || msecs > UINT_MAX)
 457                 return -EINVAL;
 458
 459         khugepaged_alloc_sleep_millisecs = msecs;
 460         wake_up_interruptible(&khugepaged_wait);
 461
 462         return count;
 463 }
 464 static struct kobj_attribute alloc_sleep_millisecs_attr =
 465         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 466                alloc_sleep_millisecs_store);
 467
 468 static ssize_t pages_to_scan_show(struct kobject *kobj,
 469                                   struct kobj_attribute *attr,
 470                                   char *buf)
 471 {
 472         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 473 }
 474 static ssize_t pages_to_scan_store(struct kobject *kobj,
 475                                    struct kobj_attribute *attr,
 476                                    const char *buf, size_t count)
 477 {
 478         int err;
 479         unsigned long pages;
 480
 481         err = kstrtoul(buf, 10, &pages);
 482         if (err || !pages || pages > UINT_MAX)
 483                 return -EINVAL;
 484
 485         khugepaged_pages_to_scan = pages;
 486
 487         return count;
 488 }
 489 static struct kobj_attribute pages_to_scan_attr =
 490         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 491                pages_to_scan_store);
 492
 493 static ssize_t pages_collapsed_show(struct kobject *kobj,
 494                                     struct kobj_attribute *attr,
 495                                     char *buf)
 496 {
 497         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 498 }
 499 static struct kobj_attribute pages_collapsed_attr =
 500         __ATTR_RO(pages_collapsed);
 501
 502 static ssize_t full_scans_show(struct kobject *kobj,
 503                                struct kobj_attribute *attr,
 504                                char *buf)
 505 {
 506         return sprintf(buf, "%u\n", khugepaged_full_scans);
 507 }
 508 static struct kobj_attribute full_scans_attr =
 509         __ATTR_RO(full_scans);
 510
 511 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 512                                       struct kobj_attribute *attr, char *buf)
 513 {
 514         return single_flag_show(kobj, attr, buf,
 515                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 516 }
 517 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 518                                        struct kobj_attribute *attr,
 519                                        const char *buf, size_t count)
 520 {
 521         return single_flag_store(kobj, attr, buf, count,
 522                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 523 }
 524 static struct kobj_attribute khugepaged_defrag_attr =
 525         __ATTR(defrag, 0644, khugepaged_defrag_show,
 526                khugepaged_defrag_store);
 527
 528 /*
 529  * max_ptes_none controls if khugepaged should collapse hugepages over
 530  * any unmapped ptes in turn potentially increasing the memory
 531  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 532  * reduce the available free memory in the system as it
 533  * runs. Increasing max_ptes_none will instead potentially reduce the
 534  * free memory in the system during the khugepaged scan.
 535  */
 536 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 537                                              struct kobj_attribute *attr,
 538                                              char *buf)
 539 {
 540         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 541 }
 542 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 543                                               struct kobj_attribute *attr,
 544                                               const char *buf, size_t count)
 545 {
 546         int err;
 547         unsigned long max_ptes_none;
 548
 549         err = kstrtoul(buf, 10, &max_ptes_none);
 550         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 551                 return -EINVAL;
 552
 553         khugepaged_max_ptes_none = max_ptes_none;
 554
 555         return count;
 556 }
 557 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 558         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 559                khugepaged_max_ptes_none_store);
 560
 561 static struct attribute *khugepaged_attr[] = {
 562         &khugepaged_defrag_attr.attr,
 563         &khugepaged_max_ptes_none_attr.attr,
 564         &pages_to_scan_attr.attr,
 565         &pages_collapsed_attr.attr,
 566         &full_scans_attr.attr,
 567         &scan_sleep_millisecs_attr.attr,
 568         &alloc_sleep_millisecs_attr.attr,
 569         NULL,
 570 };
 571
 572 static struct attribute_group khugepaged_attr_group = {
 573         .attrs = khugepaged_attr,
 574         .name = "khugepaged",
 575 };
 576
 577 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 578 {
 579         int err;
 580
 581         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 582         if (unlikely(!*hugepage_kobj)) {
 583                 pr_err("failed to create transparent hugepage kobject\n");
 584                 return -ENOMEM;
 585         }
 586
 587         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 588         if (err) {
 589                 pr_err("failed to register transparent hugepage group\n");
 590                 goto delete_obj;
 591         }
 592
 593         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 594         if (err) {
 595                 pr_err("failed to register transparent hugepage group\n");
 596                 goto remove_hp_group;
 597         }
 598
 599         return 0;
 600
 601 remove_hp_group:
 602         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 603 delete_obj:
 604         kobject_put(*hugepage_kobj);
 605         return err;
 606 }
 607
 608 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 609 {
 610         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 611         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 612         kobject_put(hugepage_kobj);
 613 }
 614 #else
 615 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 616 {
 617         return 0;
 618 }
 619
 620 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 621 {
 622 }
 623 #endif /* CONFIG_SYSFS */
 624
 625 static int __init hugepage_init(void)
 626 {
 627         int err;
 628         struct kobject *hugepage_kobj;
 629
 630         if (!has_transparent_hugepage()) {
 631                 transparent_hugepage_flags = 0;
 632                 return -EINVAL;
 633         }
 634
 635         err = hugepage_init_sysfs(&hugepage_kobj);
 636         if (err)
 637                 return err;
 638
 639         err = khugepaged_slab_init();
 640         if (err)
 641                 goto out;
 642
 643         register_shrinker(&huge_zero_page_shrinker);
 644
 645         /*
 646          * By default disable transparent hugepages on smaller systems,
 647          * where the extra memory used could hurt more than TLB overhead
 648          * is likely to save.  The admin can still enable it through /sys.
 649          */
 650         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 651                 transparent_hugepage_flags = 0;
 652
 653         start_khugepaged();
 654
 655         return 0;
 656 out:
 657         hugepage_exit_sysfs(hugepage_kobj);
 658         return err;
 659 }
 660 subsys_initcall(hugepage_init);
 661
 662 static int __init setup_transparent_hugepage(char *str)
 663 {
 664         int ret = 0;
 665         if (!str)
 666                 goto out;
 667         if (!strcmp(str, "always")) {
 668                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 669                         &transparent_hugepage_flags);
 670                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 671                           &transparent_hugepage_flags);
 672                 ret = 1;
 673         } else if (!strcmp(str, "madvise")) {
 674                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 675                           &transparent_hugepage_flags);
 676                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 677                         &transparent_hugepage_flags);
 678                 ret = 1;
 679         } else if (!strcmp(str, "never")) {
 680                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 681                           &transparent_hugepage_flags);
 682                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 683                           &transparent_hugepage_flags);
 684                 ret = 1;
 685         }
 686 out:
 687         if (!ret)
 688                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 689         return ret;
 690 }
 691 __setup("transparent_hugepage=", setup_transparent_hugepage);
 692
 693 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 694 {
 695         if (likely(vma->vm_flags & VM_WRITE))
 696                 pmd = pmd_mkwrite(pmd);
 697         return pmd;
 698 }
 699
 700 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 701 {
 702         pmd_t entry;
 703         entry = mk_pmd(page, prot);
 704         entry = pmd_mkhuge(entry);
 705         return entry;
 706 }
 707
 708 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 709                                         struct vm_area_struct *vma,
 710                                         unsigned long haddr, pmd_t *pmd,
 711                                         struct page *page)
 712 {
 713         struct mem_cgroup *memcg;
 714         pgtable_t pgtable;
 715         spinlock_t *ptl;
 716
 717         VM_BUG_ON_PAGE(!PageCompound(page), page);
 718
 719         if (mem_cgroup_try_charge(page, mm, GFP_TRANSHUGE, &memcg))
 720                 return VM_FAULT_OOM;
 721
 722         pgtable = pte_alloc_one(mm, haddr);
 723         if (unlikely(!pgtable)) {
 724                 mem_cgroup_cancel_charge(page, memcg);
 725                 return VM_FAULT_OOM;
 726         }
 727
 728         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 729         /*
 730          * The memory barrier inside __SetPageUptodate makes sure that
 731          * clear_huge_page writes become visible before the set_pmd_at()
 732          * write.
 733          */
 734         __SetPageUptodate(page);
 735
 736         ptl = pmd_lock(mm, pmd);
 737         if (unlikely(!pmd_none(*pmd))) {
 738                 spin_unlock(ptl);
 739                 mem_cgroup_cancel_charge(page, memcg);
 740                 put_page(page);
 741                 pte_free(mm, pgtable);
 742         } else {
 743                 pmd_t entry;
 744                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 745                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 746                 page_add_new_anon_rmap(page, vma, haddr);
 747                 mem_cgroup_commit_charge(page, memcg, false);
 748                 lru_cache_add_active_or_unevictable(page, vma);
 749                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 750                 set_pmd_at(mm, haddr, pmd, entry);
 751                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 752                 atomic_long_inc(&mm->nr_ptes);
 753                 spin_unlock(ptl);
 754         }
 755
 756         return 0;
 757 }
 758
 759 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 760 {
 761         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 762 }
 763
 764 static inline struct page *alloc_hugepage_vma(int defrag,
 765                                               struct vm_area_struct *vma,
 766                                               unsigned long haddr, int nd,
 767                                               gfp_t extra_gfp)
 768 {
 769         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 770                                HPAGE_PMD_ORDER, vma, haddr, nd);
 771 }
 772
 773 /* Caller must hold page table lock. */
 774 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 775                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 776                 struct page *zero_page)
 777 {
 778         pmd_t entry;
 779         if (!pmd_none(*pmd))
 780                 return false;
 781         entry = mk_pmd(zero_page, vma->vm_page_prot);
 782         entry = pmd_mkhuge(entry);
 783         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 784         set_pmd_at(mm, haddr, pmd, entry);
 785         atomic_long_inc(&mm->nr_ptes);
 786         return true;
 787 }
 788
 789 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 790                                unsigned long address, pmd_t *pmd,
 791                                unsigned int flags)
 792 {
 793         struct page *page;
 794         unsigned long haddr = address & HPAGE_PMD_MASK;
 795
 796         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 797                 return VM_FAULT_FALLBACK;
 798         if (unlikely(anon_vma_prepare(vma)))
 799                 return VM_FAULT_OOM;
 800         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 801                 return VM_FAULT_OOM;
 802         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 803                         transparent_hugepage_use_zero_page()) {
 804                 spinlock_t *ptl;
 805                 pgtable_t pgtable;
 806                 struct page *zero_page;
 807                 bool set;
 808                 pgtable = pte_alloc_one(mm, haddr);
 809                 if (unlikely(!pgtable))
 810                         return VM_FAULT_OOM;
 811                 zero_page = get_huge_zero_page();
 812                 if (unlikely(!zero_page)) {
 813                         pte_free(mm, pgtable);
 814                         count_vm_event(THP_FAULT_FALLBACK);
 815                         return VM_FAULT_FALLBACK;
 816                 }
 817                 ptl = pmd_lock(mm, pmd);
 818                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 819                                 zero_page);
 820                 spin_unlock(ptl);
 821                 if (!set) {
 822                         pte_free(mm, pgtable);
 823                         put_huge_zero_page();
 824                 }
 825                 return 0;
 826         }
 827         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 828                         vma, haddr, numa_node_id(), 0);
 829         if (unlikely(!page)) {
 830                 count_vm_event(THP_FAULT_FALLBACK);
 831                 return VM_FAULT_FALLBACK;
 832         }
 833         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 834                 put_page(page);
 835                 count_vm_event(THP_FAULT_FALLBACK);
 836                 return VM_FAULT_FALLBACK;
 837         }
 838
 839         count_vm_event(THP_FAULT_ALLOC);
 840         return 0;
 841 }
 842
 843 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 844                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 845                   struct vm_area_struct *vma)
 846 {
 847         spinlock_t *dst_ptl, *src_ptl;
 848         struct page *src_page;
 849         pmd_t pmd;
 850         pgtable_t pgtable;
 851         int ret;
 852
 853         ret = -ENOMEM;
 854         pgtable = pte_alloc_one(dst_mm, addr);
 855         if (unlikely(!pgtable))
 856                 goto out;
 857
 858         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 859         src_ptl = pmd_lockptr(src_mm, src_pmd);
 860         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 861
 862         ret = -EAGAIN;
 863         pmd = *src_pmd;
 864         if (unlikely(!pmd_trans_huge(pmd))) {
 865                 pte_free(dst_mm, pgtable);
 866                 goto out_unlock;
 867         }
 868         /*
 869          * When page table lock is held, the huge zero pmd should not be
 870          * under splitting since we don't split the page itself, only pmd to
 871          * a page table.
 872          */
 873         if (is_huge_zero_pmd(pmd)) {
 874                 struct page *zero_page;
 875                 bool set;
 876                 /*
 877                  * get_huge_zero_page() will never allocate a new page here,
 878                  * since we already have a zero page to copy. It just takes a
 879                  * reference.
 880                  */
 881                 zero_page = get_huge_zero_page();
 882                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 883                                 zero_page);
 884                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 885                 ret = 0;
 886                 goto out_unlock;
 887         }
 888
 889         if (unlikely(pmd_trans_splitting(pmd))) {
 890                 /* split huge page running from under us */
 891                 spin_unlock(src_ptl);
 892                 spin_unlock(dst_ptl);
 893                 pte_free(dst_mm, pgtable);
 894
 895                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 896                 goto out;
 897         }
 898         src_page = pmd_page(pmd);
 899         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 900         get_page(src_page);
 901         page_dup_rmap(src_page);
 902         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 903
 904         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 905         pmd = pmd_mkold(pmd_wrprotect(pmd));
 906         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 907         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 908         atomic_long_inc(&dst_mm->nr_ptes);
 909
 910         ret = 0;
 911 out_unlock:
 912         spin_unlock(src_ptl);
 913         spin_unlock(dst_ptl);
 914 out:
 915         return ret;
 916 }
 917
 918 void huge_pmd_set_accessed(struct mm_struct *mm,
 919                            struct vm_area_struct *vma,
 920                            unsigned long address,
 921                            pmd_t *pmd, pmd_t orig_pmd,
 922                            int dirty)
 923 {
 924         spinlock_t *ptl;
 925         pmd_t entry;
 926         unsigned long haddr;
 927
 928         ptl = pmd_lock(mm, pmd);
 929         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 930                 goto unlock;
 931
 932         entry = pmd_mkyoung(orig_pmd);
 933         haddr = address & HPAGE_PMD_MASK;
 934         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 935                 update_mmu_cache_pmd(vma, address, pmd);
 936
 937 unlock:
 938         spin_unlock(ptl);
 939 }
 940
 941 /*
 942  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 943  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 944  * the source page gets split and a tail freed before copy completes.
 945  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 946  */
 947 static void get_user_huge_page(struct page *page)
 948 {
 949         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 950                 struct page *endpage = page + HPAGE_PMD_NR;
 951
 952                 atomic_add(HPAGE_PMD_NR, &page->_count);
 953                 while (++page < endpage)
 954                         get_huge_page_tail(page);
 955         } else {
 956                 get_page(page);
 957         }
 958 }
 959
 960 static void put_user_huge_page(struct page *page)
 961 {
 962         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 963                 struct page *endpage = page + HPAGE_PMD_NR;
 964
 965                 while (page < endpage)
 966                         put_page(page++);
 967         } else {
 968                 put_page(page);
 969         }
 970 }
 971
 972 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 973                                         struct vm_area_struct *vma,
 974                                         unsigned long address,
 975                                         pmd_t *pmd, pmd_t orig_pmd,
 976                                         struct page *page,
 977                                         unsigned long haddr)
 978 {
 979         struct mem_cgroup *memcg;
 980         spinlock_t *ptl;
 981         pgtable_t pgtable;
 982         pmd_t _pmd;
 983         int ret = 0, i;
 984         struct page **pages;
 985         unsigned long mmun_start;       /* For mmu_notifiers */
 986         unsigned long mmun_end;         /* For mmu_notifiers */
 987
 988         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 989                         GFP_KERNEL);
 990         if (unlikely(!pages)) {
 991                 ret |= VM_FAULT_OOM;
 992                 goto out;
 993         }
 994
 995         for (i = 0; i < HPAGE_PMD_NR; i++) {
 996                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 997                                                __GFP_OTHER_NODE,
 998                                                vma, address, page_to_nid(page));
 999                 if (unlikely(!pages[i] ||
1000                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1001                                                    &memcg))) {
1002                         if (pages[i])
1003                                 put_page(pages[i]);
1004                         while (--i >= 0) {
1005                                 memcg = (void *)page_private(pages[i]);
1006                                 set_page_private(pages[i], 0);
1007                                 mem_cgroup_cancel_charge(pages[i], memcg);
1008                                 put_page(pages[i]);
1009                         }
1010                         kfree(pages);
1011                         ret |= VM_FAULT_OOM;
1012                         goto out;
1013                 }
1014                 set_page_private(pages[i], (unsigned long)memcg);
1015         }
1016
1017         for (i = 0; i < HPAGE_PMD_NR; i++) {
1018                 copy_user_highpage(pages[i], page + i,
1019                                    haddr + PAGE_SIZE * i, vma);
1020                 __SetPageUptodate(pages[i]);
1021                 cond_resched();
1022         }
1023
1024         mmun_start = haddr;
1025         mmun_end   = haddr + HPAGE_PMD_SIZE;
1026         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1027
1028         ptl = pmd_lock(mm, pmd);
1029         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1030                 goto out_free_pages;
1031         VM_BUG_ON_PAGE(!PageHead(page), page);
1032
1033         pmdp_clear_flush_notify(vma, haddr, pmd);
1034         /* leave pmd empty until pte is filled */
1035
1036         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1037         pmd_populate(mm, &_pmd, pgtable);
1038
1039         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1040                 pte_t *pte, entry;
1041                 entry = mk_pte(pages[i], vma->vm_page_prot);
1042                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1043                 memcg = (void *)page_private(pages[i]);
1044                 set_page_private(pages[i], 0);
1045                 page_add_new_anon_rmap(pages[i], vma, haddr);
1046                 mem_cgroup_commit_charge(pages[i], memcg, false);
1047                 lru_cache_add_active_or_unevictable(pages[i], vma);
1048                 pte = pte_offset_map(&_pmd, haddr);
1049                 VM_BUG_ON(!pte_none(*pte));
1050                 set_pte_at(mm, haddr, pte, entry);
1051                 pte_unmap(pte);
1052         }
1053         kfree(pages);
1054
1055         smp_wmb(); /* make pte visible before pmd */
1056         pmd_populate(mm, pmd, pgtable);
1057         page_remove_rmap(page);
1058         spin_unlock(ptl);
1059
1060         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1061
1062         ret |= VM_FAULT_WRITE;
1063         put_page(page);
1064
1065 out:
1066         return ret;
1067
1068 out_free_pages:
1069         spin_unlock(ptl);
1070         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1071         for (i = 0; i < HPAGE_PMD_NR; i++) {
1072                 memcg = (void *)page_private(pages[i]);
1073                 set_page_private(pages[i], 0);
1074                 mem_cgroup_cancel_charge(pages[i], memcg);
1075                 put_page(pages[i]);
1076         }
1077         kfree(pages);
1078         goto out;
1079 }
1080
1081 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1082                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1083 {
1084         spinlock_t *ptl;
1085         int ret = 0;
1086         struct page *page = NULL, *new_page;
1087         struct mem_cgroup *memcg;
1088         unsigned long haddr;
1089         unsigned long mmun_start;       /* For mmu_notifiers */
1090         unsigned long mmun_end;         /* For mmu_notifiers */
1091
1092         ptl = pmd_lockptr(mm, pmd);
1093         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1094         haddr = address & HPAGE_PMD_MASK;
1095         if (is_huge_zero_pmd(orig_pmd))
1096                 goto alloc;
1097         spin_lock(ptl);
1098         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1099                 goto out_unlock;
1100
1101         page = pmd_page(orig_pmd);
1102         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1103         if (page_mapcount(page) == 1) {
1104                 pmd_t entry;
1105                 entry = pmd_mkyoung(orig_pmd);
1106                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1107                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1108                         update_mmu_cache_pmd(vma, address, pmd);
1109                 ret |= VM_FAULT_WRITE;
1110                 goto out_unlock;
1111         }
1112         get_user_huge_page(page);
1113         spin_unlock(ptl);
1114 alloc:
1115         if (transparent_hugepage_enabled(vma) &&
1116             !transparent_hugepage_debug_cow())
1117                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1118                                               vma, haddr, numa_node_id(), 0);
1119         else
1120                 new_page = NULL;
1121
1122         if (unlikely(!new_page)) {
1123                 if (!page) {
1124                         split_huge_page_pmd(vma, address, pmd);
1125                         ret |= VM_FAULT_FALLBACK;
1126                 } else {
1127                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1128                                         pmd, orig_pmd, page, haddr);
1129                         if (ret & VM_FAULT_OOM) {
1130                                 split_huge_page(page);
1131                                 ret |= VM_FAULT_FALLBACK;
1132                         }
1133                         put_user_huge_page(page);
1134                 }
1135                 count_vm_event(THP_FAULT_FALLBACK);
1136                 goto out;
1137         }
1138
1139         if (unlikely(mem_cgroup_try_charge(new_page, mm,
1140                                            GFP_TRANSHUGE, &memcg))) {
1141                 put_page(new_page);
1142                 if (page) {
1143                         split_huge_page(page);
1144                         put_user_huge_page(page);
1145                 } else
1146                         split_huge_page_pmd(vma, address, pmd);
1147                 ret |= VM_FAULT_FALLBACK;
1148                 count_vm_event(THP_FAULT_FALLBACK);
1149                 goto out;
1150         }
1151
1152         count_vm_event(THP_FAULT_ALLOC);
1153
1154         if (!page)
1155                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1156         else
1157                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1158         __SetPageUptodate(new_page);
1159
1160         mmun_start = haddr;
1161         mmun_end   = haddr + HPAGE_PMD_SIZE;
1162         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1163
1164         spin_lock(ptl);
1165         if (page)
1166                 put_user_huge_page(page);
1167         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1168                 spin_unlock(ptl);
1169                 mem_cgroup_cancel_charge(new_page, memcg);
1170                 put_page(new_page);
1171                 goto out_mn;
1172         } else {
1173                 pmd_t entry;
1174                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1175                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1176                 pmdp_clear_flush_notify(vma, haddr, pmd);
1177                 page_add_new_anon_rmap(new_page, vma, haddr);
1178                 mem_cgroup_commit_charge(new_page, memcg, false);
1179                 lru_cache_add_active_or_unevictable(new_page, vma);
1180                 set_pmd_at(mm, haddr, pmd, entry);
1181                 update_mmu_cache_pmd(vma, address, pmd);
1182                 if (!page) {
1183                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1184                         put_huge_zero_page();
1185                 } else {
1186                         VM_BUG_ON_PAGE(!PageHead(page), page);
1187                         page_remove_rmap(page);
1188                         put_page(page);
1189                 }
1190                 ret |= VM_FAULT_WRITE;
1191         }
1192         spin_unlock(ptl);
1193 out_mn:
1194         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1195 out:
1196         return ret;
1197 out_unlock:
1198         spin_unlock(ptl);
1199         return ret;
1200 }
1201
1202 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1203                                    unsigned long addr,
1204                                    pmd_t *pmd,
1205                                    unsigned int flags)
1206 {
1207         struct mm_struct *mm = vma->vm_mm;
1208         struct page *page = NULL;
1209
1210         assert_spin_locked(pmd_lockptr(mm, pmd));
1211
1212         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1213                 goto out;
1214
1215         /* Avoid dumping huge zero page */
1216         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1217                 return ERR_PTR(-EFAULT);
1218
1219         /* Full NUMA hinting faults to serialise migration in fault paths */
1220         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1221                 goto out;
1222
1223         page = pmd_page(*pmd);
1224         VM_BUG_ON_PAGE(!PageHead(page), page);
1225         if (flags & FOLL_TOUCH) {
1226                 pmd_t _pmd;
1227                 /*
1228                  * We should set the dirty bit only for FOLL_WRITE but
1229                  * for now the dirty bit in the pmd is meaningless.
1230                  * And if the dirty bit will become meaningful and
1231                  * we'll only set it with FOLL_WRITE, an atomic
1232                  * set_bit will be required on the pmd to set the
1233                  * young bit, instead of the current set_pmd_at.
1234                  */
1235                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1236                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1237                                           pmd, _pmd,  1))
1238                         update_mmu_cache_pmd(vma, addr, pmd);
1239         }
1240         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1241                 if (page->mapping && trylock_page(page)) {
1242                         lru_add_drain();
1243                         if (page->mapping)
1244                                 mlock_vma_page(page);
1245                         unlock_page(page);
1246                 }
1247         }
1248         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1249         VM_BUG_ON_PAGE(!PageCompound(page), page);
1250         if (flags & FOLL_GET)
1251                 get_page_foll(page);
1252
1253 out:
1254         return page;
1255 }
1256
1257 /* NUMA hinting page fault entry point for trans huge pmds */
1258 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1259                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1260 {
1261         spinlock_t *ptl;
1262         struct anon_vma *anon_vma = NULL;
1263         struct page *page;
1264         unsigned long haddr = addr & HPAGE_PMD_MASK;
1265         int page_nid = -1, this_nid = numa_node_id();
1266         int target_nid, last_cpupid = -1;
1267         bool page_locked;
1268         bool migrated = false;
1269         int flags = 0;
1270
1271         ptl = pmd_lock(mm, pmdp);
1272         if (unlikely(!pmd_same(pmd, *pmdp)))
1273                 goto out_unlock;
1274
1275         /*
1276          * If there are potential migrations, wait for completion and retry
1277          * without disrupting NUMA hinting information. Do not relock and
1278          * check_same as the page may no longer be mapped.
1279          */
1280         if (unlikely(pmd_trans_migrating(*pmdp))) {
1281                 spin_unlock(ptl);
1282                 wait_migrate_huge_page(vma->anon_vma, pmdp);
1283                 goto out;
1284         }
1285
1286         page = pmd_page(pmd);
1287         BUG_ON(is_huge_zero_page(page));
1288         page_nid = page_to_nid(page);
1289         last_cpupid = page_cpupid_last(page);
1290         count_vm_numa_event(NUMA_HINT_FAULTS);
1291         if (page_nid == this_nid) {
1292                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1293                 flags |= TNF_FAULT_LOCAL;
1294         }
1295
1296         /*
1297          * Avoid grouping on DSO/COW pages in specific and RO pages
1298          * in general, RO pages shouldn't hurt as much anyway since
1299          * they can be in shared cache state.
1300          */
1301         if (!pmd_write(pmd))
1302                 flags |= TNF_NO_GROUP;
1303
1304         /*
1305          * Acquire the page lock to serialise THP migrations but avoid dropping
1306          * page_table_lock if at all possible
1307          */
1308         page_locked = trylock_page(page);
1309         target_nid = mpol_misplaced(page, vma, haddr);
1310         if (target_nid == -1) {
1311                 /* If the page was locked, there are no parallel migrations */
1312                 if (page_locked)
1313                         goto clear_pmdnuma;
1314         }
1315
1316         /* Migration could have started since the pmd_trans_migrating check */
1317         if (!page_locked) {
1318                 spin_unlock(ptl);
1319                 wait_on_page_locked(page);
1320                 page_nid = -1;
1321                 goto out;
1322         }
1323
1324         /*
1325          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1326          * to serialises splits
1327          */
1328         get_page(page);
1329         spin_unlock(ptl);
1330         anon_vma = page_lock_anon_vma_read(page);
1331
1332         /* Confirm the PMD did not change while page_table_lock was released */
1333         spin_lock(ptl);
1334         if (unlikely(!pmd_same(pmd, *pmdp))) {
1335                 unlock_page(page);
1336                 put_page(page);
1337                 page_nid = -1;
1338                 goto out_unlock;
1339         }
1340
1341         /* Bail if we fail to protect against THP splits for any reason */
1342         if (unlikely(!anon_vma)) {
1343                 put_page(page);
1344                 page_nid = -1;
1345                 goto clear_pmdnuma;
1346         }
1347
1348         /*
1349          * Migrate the THP to the requested node, returns with page unlocked
1350          * and pmd_numa cleared.
1351          */
1352         spin_unlock(ptl);
1353         migrated = migrate_misplaced_transhuge_page(mm, vma,
1354                                 pmdp, pmd, addr, page, target_nid);
1355         if (migrated) {
1356                 flags |= TNF_MIGRATED;
1357                 page_nid = target_nid;
1358         }
1359
1360         goto out;
1361 clear_pmdnuma:
1362         BUG_ON(!PageLocked(page));
1363         pmd = pmd_mknonnuma(pmd);
1364         set_pmd_at(mm, haddr, pmdp, pmd);
1365         VM_BUG_ON(pmd_numa(*pmdp));
1366         update_mmu_cache_pmd(vma, addr, pmdp);
1367         unlock_page(page);
1368 out_unlock:
1369         spin_unlock(ptl);
1370
1371 out:
1372         if (anon_vma)
1373                 page_unlock_anon_vma_read(anon_vma);
1374
1375         if (page_nid != -1)
1376                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1377
1378         return 0;
1379 }
1380
1381 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1382                  pmd_t *pmd, unsigned long addr)
1383 {
1384         spinlock_t *ptl;
1385         int ret = 0;
1386
1387         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1388                 struct page *page;
1389                 pgtable_t pgtable;
1390                 pmd_t orig_pmd;
1391                 /*
1392                  * For architectures like ppc64 we look at deposited pgtable
1393                  * when calling pmdp_get_and_clear. So do the
1394                  * pgtable_trans_huge_withdraw after finishing pmdp related
1395                  * operations.
1396                  */
1397                 orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd,
1398                                                    tlb->fullmm);
1399                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1400                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1401                 if (is_huge_zero_pmd(orig_pmd)) {
1402                         atomic_long_dec(&tlb->mm->nr_ptes);
1403                         spin_unlock(ptl);
1404                         put_huge_zero_page();
1405                 } else {
1406                         page = pmd_page(orig_pmd);
1407                         page_remove_rmap(page);
1408                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1409                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1410                         VM_BUG_ON_PAGE(!PageHead(page), page);
1411                         atomic_long_dec(&tlb->mm->nr_ptes);
1412                         spin_unlock(ptl);
1413                         tlb_remove_page(tlb, page);
1414                 }
1415                 pte_free(tlb->mm, pgtable);
1416                 ret = 1;
1417         }
1418         return ret;
1419 }
1420
1421 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1422                 unsigned long addr, unsigned long end,
1423                 unsigned char *vec)
1424 {
1425         spinlock_t *ptl;
1426         int ret = 0;
1427
1428         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1429                 /*
1430                  * All logical pages in the range are present
1431                  * if backed by a huge page.
1432                  */
1433                 spin_unlock(ptl);
1434                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1435                 ret = 1;
1436         }
1437
1438         return ret;
1439 }
1440
1441 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1442                   unsigned long old_addr,
1443                   unsigned long new_addr, unsigned long old_end,
1444                   pmd_t *old_pmd, pmd_t *new_pmd)
1445 {
1446         spinlock_t *old_ptl, *new_ptl;
1447         int ret = 0;
1448         pmd_t pmd;
1449
1450         struct mm_struct *mm = vma->vm_mm;
1451
1452         if ((old_addr & ~HPAGE_PMD_MASK) ||
1453             (new_addr & ~HPAGE_PMD_MASK) ||
1454             old_end - old_addr < HPAGE_PMD_SIZE ||
1455             (new_vma->vm_flags & VM_NOHUGEPAGE))
1456                 goto out;
1457
1458         /*
1459          * The destination pmd shouldn't be established, free_pgtables()
1460          * should have release it.
1461          */
1462         if (WARN_ON(!pmd_none(*new_pmd))) {
1463                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1464                 goto out;
1465         }
1466
1467         /*
1468          * We don't have to worry about the ordering of src and dst
1469          * ptlocks because exclusive mmap_sem prevents deadlock.
1470          */
1471         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1472         if (ret == 1) {
1473                 new_ptl = pmd_lockptr(mm, new_pmd);
1474                 if (new_ptl != old_ptl)
1475                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1476                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1477                 VM_BUG_ON(!pmd_none(*new_pmd));
1478
1479                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1480                         pgtable_t pgtable;
1481                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1482                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1483                 }
1484                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1485                 if (new_ptl != old_ptl)
1486                         spin_unlock(new_ptl);
1487                 spin_unlock(old_ptl);
1488         }
1489 out:
1490         return ret;
1491 }
1492
1493 /*
1494  * Returns
1495  *  - 0 if PMD could not be locked
1496  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1497  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1498  */
1499 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1500                 unsigned long addr, pgprot_t newprot, int prot_numa)
1501 {
1502         struct mm_struct *mm = vma->vm_mm;
1503         spinlock_t *ptl;
1504         int ret = 0;
1505
1506         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1507                 pmd_t entry;
1508                 ret = 1;
1509                 if (!prot_numa) {
1510                         entry = pmdp_get_and_clear_notify(mm, addr, pmd);
1511                         if (pmd_numa(entry))
1512                                 entry = pmd_mknonnuma(entry);
1513                         entry = pmd_modify(entry, newprot);
1514                         ret = HPAGE_PMD_NR;
1515                         set_pmd_at(mm, addr, pmd, entry);
1516                         BUG_ON(pmd_write(entry));
1517                 } else {
1518                         struct page *page = pmd_page(*pmd);
1519
1520                         /*
1521                          * Do not trap faults against the zero page. The
1522                          * read-only data is likely to be read-cached on the
1523                          * local CPU cache and it is less useful to know about
1524                          * local vs remote hits on the zero page.
1525                          */
1526                         if (!is_huge_zero_page(page) &&
1527                             !pmd_numa(*pmd)) {
1528                                 pmdp_set_numa(mm, addr, pmd);
1529                                 ret = HPAGE_PMD_NR;
1530                         }
1531                 }
1532                 spin_unlock(ptl);
1533         }
1534
1535         return ret;
1536 }
1537
1538 /*
1539  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1540  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1541  *
1542  * Note that if it returns 1, this routine returns without unlocking page
1543  * table locks. So callers must unlock them.
1544  */
1545 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1546                 spinlock_t **ptl)
1547 {
1548         *ptl = pmd_lock(vma->vm_mm, pmd);
1549         if (likely(pmd_trans_huge(*pmd))) {
1550                 if (unlikely(pmd_trans_splitting(*pmd))) {
1551                         spin_unlock(*ptl);
1552                         wait_split_huge_page(vma->anon_vma, pmd);
1553                         return -1;
1554                 } else {
1555                         /* Thp mapped by 'pmd' is stable, so we can
1556                          * handle it as it is. */
1557                         return 1;
1558                 }
1559         }
1560         spin_unlock(*ptl);
1561         return 0;
1562 }
1563
1564 /*
1565  * This function returns whether a given @page is mapped onto the @address
1566  * in the virtual space of @mm.
1567  *
1568  * When it's true, this function returns *pmd with holding the page table lock
1569  * and passing it back to the caller via @ptl.
1570  * If it's false, returns NULL without holding the page table lock.
1571  */
1572 pmd_t *page_check_address_pmd(struct page *page,
1573                               struct mm_struct *mm,
1574                               unsigned long address,
1575                               enum page_check_address_pmd_flag flag,
1576                               spinlock_t **ptl)
1577 {
1578         pgd_t *pgd;
1579         pud_t *pud;
1580         pmd_t *pmd;
1581
1582         if (address & ~HPAGE_PMD_MASK)
1583                 return NULL;
1584
1585         pgd = pgd_offset(mm, address);
1586         if (!pgd_present(*pgd))
1587                 return NULL;
1588         pud = pud_offset(pgd, address);
1589         if (!pud_present(*pud))
1590                 return NULL;
1591         pmd = pmd_offset(pud, address);
1592
1593         *ptl = pmd_lock(mm, pmd);
1594         if (!pmd_present(*pmd))
1595                 goto unlock;
1596         if (pmd_page(*pmd) != page)
1597                 goto unlock;
1598         /*
1599          * split_vma() may create temporary aliased mappings. There is
1600          * no risk as long as all huge pmd are found and have their
1601          * splitting bit set before __split_huge_page_refcount
1602          * runs. Finding the same huge pmd more than once during the
1603          * same rmap walk is not a problem.
1604          */
1605         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1606             pmd_trans_splitting(*pmd))
1607                 goto unlock;
1608         if (pmd_trans_huge(*pmd)) {
1609                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1610                           !pmd_trans_splitting(*pmd));
1611                 return pmd;
1612         }
1613 unlock:
1614         spin_unlock(*ptl);
1615         return NULL;
1616 }
1617
1618 static int __split_huge_page_splitting(struct page *page,
1619                                        struct vm_area_struct *vma,
1620                                        unsigned long address)
1621 {
1622         struct mm_struct *mm = vma->vm_mm;
1623         spinlock_t *ptl;
1624         pmd_t *pmd;
1625         int ret = 0;
1626         /* For mmu_notifiers */
1627         const unsigned long mmun_start = address;
1628         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1629
1630         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1631         pmd = page_check_address_pmd(page, mm, address,
1632                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1633         if (pmd) {
1634                 /*
1635                  * We can't temporarily set the pmd to null in order
1636                  * to split it, the pmd must remain marked huge at all
1637                  * times or the VM won't take the pmd_trans_huge paths
1638                  * and it won't wait on the anon_vma->root->rwsem to
1639                  * serialize against split_huge_page*.
1640                  */
1641                 pmdp_splitting_flush(vma, address, pmd);
1642
1643                 ret = 1;
1644                 spin_unlock(ptl);
1645         }
1646         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1647
1648         return ret;
1649 }
1650
1651 static void __split_huge_page_refcount(struct page *page,
1652                                        struct list_head *list)
1653 {
1654         int i;
1655         struct zone *zone = page_zone(page);
1656         struct lruvec *lruvec;
1657         int tail_count = 0;
1658
1659         /* prevent PageLRU to go away from under us, and freeze lru stats */
1660         spin_lock_irq(&zone->lru_lock);
1661         lruvec = mem_cgroup_page_lruvec(page, zone);
1662
1663         compound_lock(page);
1664         /* complete memcg works before add pages to LRU */
1665         mem_cgroup_split_huge_fixup(page);
1666
1667         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1668                 struct page *page_tail = page + i;
1669
1670                 /* tail_page->_mapcount cannot change */
1671                 BUG_ON(page_mapcount(page_tail) < 0);
1672                 tail_count += page_mapcount(page_tail);
1673                 /* check for overflow */
1674                 BUG_ON(tail_count < 0);
1675                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1676                 /*
1677                  * tail_page->_count is zero and not changing from
1678                  * under us. But get_page_unless_zero() may be running
1679                  * from under us on the tail_page. If we used
1680                  * atomic_set() below instead of atomic_add(), we
1681                  * would then run atomic_set() concurrently with
1682                  * get_page_unless_zero(), and atomic_set() is
1683                  * implemented in C not using locked ops. spin_unlock
1684                  * on x86 sometime uses locked ops because of PPro
1685                  * errata 66, 92, so unless somebody can guarantee
1686                  * atomic_set() here would be safe on all archs (and
1687                  * not only on x86), it's safer to use atomic_add().
1688                  */
1689                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1690                            &page_tail->_count);
1691
1692                 /* after clearing PageTail the gup refcount can be released */
1693                 smp_mb__after_atomic();
1694
1695                 /*
1696                  * retain hwpoison flag of the poisoned tail page:
1697                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1698                  *   by the memory-failure.
1699                  */
1700                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1701                 page_tail->flags |= (page->flags &
1702                                      ((1L << PG_referenced) |
1703                                       (1L << PG_swapbacked) |
1704                                       (1L << PG_mlocked) |
1705                                       (1L << PG_uptodate) |
1706                                       (1L << PG_active) |
1707                                       (1L << PG_unevictable)));
1708                 page_tail->flags |= (1L << PG_dirty);
1709
1710                 /* clear PageTail before overwriting first_page */
1711                 smp_wmb();
1712
1713                 /*
1714                  * __split_huge_page_splitting() already set the
1715                  * splitting bit in all pmd that could map this
1716                  * hugepage, that will ensure no CPU can alter the
1717                  * mapcount on the head page. The mapcount is only
1718                  * accounted in the head page and it has to be
1719                  * transferred to all tail pages in the below code. So
1720                  * for this code to be safe, the split the mapcount
1721                  * can't change. But that doesn't mean userland can't
1722                  * keep changing and reading the page contents while
1723                  * we transfer the mapcount, so the pmd splitting
1724                  * status is achieved setting a reserved bit in the
1725                  * pmd, not by clearing the present bit.
1726                 */
1727                 page_tail->_mapcount = page->_mapcount;
1728
1729                 BUG_ON(page_tail->mapping);
1730                 page_tail->mapping = page->mapping;
1731
1732                 page_tail->index = page->index + i;
1733                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1734
1735                 BUG_ON(!PageAnon(page_tail));
1736                 BUG_ON(!PageUptodate(page_tail));
1737                 BUG_ON(!PageDirty(page_tail));
1738                 BUG_ON(!PageSwapBacked(page_tail));
1739
1740                 lru_add_page_tail(page, page_tail, lruvec, list);
1741         }
1742         atomic_sub(tail_count, &page->_count);
1743         BUG_ON(atomic_read(&page->_count) <= 0);
1744
1745         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1746
1747         ClearPageCompound(page);
1748         compound_unlock(page);
1749         spin_unlock_irq(&zone->lru_lock);
1750
1751         for (i = 1; i < HPAGE_PMD_NR; i++) {
1752                 struct page *page_tail = page + i;
1753                 BUG_ON(page_count(page_tail) <= 0);
1754                 /*
1755                  * Tail pages may be freed if there wasn't any mapping
1756                  * like if add_to_swap() is running on a lru page that
1757                  * had its mapping zapped. And freeing these pages
1758                  * requires taking the lru_lock so we do the put_page
1759                  * of the tail pages after the split is complete.
1760                  */
1761                 put_page(page_tail);
1762         }
1763
1764         /*
1765          * Only the head page (now become a regular page) is required
1766          * to be pinned by the caller.
1767          */
1768         BUG_ON(page_count(page) <= 0);
1769 }
1770
1771 static int __split_huge_page_map(struct page *page,
1772                                  struct vm_area_struct *vma,
1773                                  unsigned long address)
1774 {
1775         struct mm_struct *mm = vma->vm_mm;
1776         spinlock_t *ptl;
1777         pmd_t *pmd, _pmd;
1778         int ret = 0, i;
1779         pgtable_t pgtable;
1780         unsigned long haddr;
1781
1782         pmd = page_check_address_pmd(page, mm, address,
1783                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1784         if (pmd) {
1785                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1786                 pmd_populate(mm, &_pmd, pgtable);
1787                 if (pmd_write(*pmd))
1788                         BUG_ON(page_mapcount(page) != 1);
1789
1790                 haddr = address;
1791                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1792                         pte_t *pte, entry;
1793                         BUG_ON(PageCompound(page+i));
1794                         /*
1795                          * Note that pmd_numa is not transferred deliberately
1796                          * to avoid any possibility that pte_numa leaks to
1797                          * a PROT_NONE VMA by accident.
1798                          */
1799                         entry = mk_pte(page + i, vma->vm_page_prot);
1800                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1801                         if (!pmd_write(*pmd))
1802                                 entry = pte_wrprotect(entry);
1803                         if (!pmd_young(*pmd))
1804                                 entry = pte_mkold(entry);
1805                         pte = pte_offset_map(&_pmd, haddr);
1806                         BUG_ON(!pte_none(*pte));
1807                         set_pte_at(mm, haddr, pte, entry);
1808                         pte_unmap(pte);
1809                 }
1810
1811                 smp_wmb(); /* make pte visible before pmd */
1812                 /*
1813                  * Up to this point the pmd is present and huge and
1814                  * userland has the whole access to the hugepage
1815                  * during the split (which happens in place). If we
1816                  * overwrite the pmd with the not-huge version
1817                  * pointing to the pte here (which of course we could
1818                  * if all CPUs were bug free), userland could trigger
1819                  * a small page size TLB miss on the small sized TLB
1820                  * while the hugepage TLB entry is still established
1821                  * in the huge TLB. Some CPU doesn't like that. See
1822                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1823                  * Erratum 383 on page 93. Intel should be safe but is
1824                  * also warns that it's only safe if the permission
1825                  * and cache attributes of the two entries loaded in
1826                  * the two TLB is identical (which should be the case
1827                  * here). But it is generally safer to never allow
1828                  * small and huge TLB entries for the same virtual
1829                  * address to be loaded simultaneously. So instead of
1830                  * doing "pmd_populate(); flush_tlb_range();" we first
1831                  * mark the current pmd notpresent (atomically because
1832                  * here the pmd_trans_huge and pmd_trans_splitting
1833                  * must remain set at all times on the pmd until the
1834                  * split is complete for this pmd), then we flush the
1835                  * SMP TLB and finally we write the non-huge version
1836                  * of the pmd entry with pmd_populate.
1837                  */
1838                 pmdp_invalidate(vma, address, pmd);
1839                 pmd_populate(mm, pmd, pgtable);
1840                 ret = 1;
1841                 spin_unlock(ptl);
1842         }
1843
1844         return ret;
1845 }
1846
1847 /* must be called with anon_vma->root->rwsem held */
1848 static void __split_huge_page(struct page *page,
1849                               struct anon_vma *anon_vma,
1850                               struct list_head *list)
1851 {
1852         int mapcount, mapcount2;
1853         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1854         struct anon_vma_chain *avc;
1855
1856         BUG_ON(!PageHead(page));
1857         BUG_ON(PageTail(page));
1858
1859         mapcount = 0;
1860         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1861                 struct vm_area_struct *vma = avc->vma;
1862                 unsigned long addr = vma_address(page, vma);
1863                 BUG_ON(is_vma_temporary_stack(vma));
1864                 mapcount += __split_huge_page_splitting(page, vma, addr);
1865         }
1866         /*
1867          * It is critical that new vmas are added to the tail of the
1868          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1869          * and establishes a child pmd before
1870          * __split_huge_page_splitting() freezes the parent pmd (so if
1871          * we fail to prevent copy_huge_pmd() from running until the
1872          * whole __split_huge_page() is complete), we will still see
1873          * the newly established pmd of the child later during the
1874          * walk, to be able to set it as pmd_trans_splitting too.
1875          */
1876         if (mapcount != page_mapcount(page)) {
1877                 pr_err("mapcount %d page_mapcount %d\n",
1878                         mapcount, page_mapcount(page));
1879                 BUG();
1880         }
1881
1882         __split_huge_page_refcount(page, list);
1883
1884         mapcount2 = 0;
1885         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1886                 struct vm_area_struct *vma = avc->vma;
1887                 unsigned long addr = vma_address(page, vma);
1888                 BUG_ON(is_vma_temporary_stack(vma));
1889                 mapcount2 += __split_huge_page_map(page, vma, addr);
1890         }
1891         if (mapcount != mapcount2) {
1892                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1893                         mapcount, mapcount2, page_mapcount(page));
1894                 BUG();
1895         }
1896 }
1897
1898 /*
1899  * Split a hugepage into normal pages. This doesn't change the position of head
1900  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1901  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1902  * from the hugepage.
1903  * Return 0 if the hugepage is split successfully otherwise return 1.
1904  */
1905 int split_huge_page_to_list(struct page *page, struct list_head *list)
1906 {
1907         struct anon_vma *anon_vma;
1908         int ret = 1;
1909
1910         BUG_ON(is_huge_zero_page(page));
1911         BUG_ON(!PageAnon(page));
1912
1913         /*
1914          * The caller does not necessarily hold an mmap_sem that would prevent
1915          * the anon_vma disappearing so we first we take a reference to it
1916          * and then lock the anon_vma for write. This is similar to
1917          * page_lock_anon_vma_read except the write lock is taken to serialise
1918          * against parallel split or collapse operations.
1919          */
1920         anon_vma = page_get_anon_vma(page);
1921         if (!anon_vma)
1922                 goto out;
1923         anon_vma_lock_write(anon_vma);
1924
1925         ret = 0;
1926         if (!PageCompound(page))
1927                 goto out_unlock;
1928
1929         BUG_ON(!PageSwapBacked(page));
1930         __split_huge_page(page, anon_vma, list);
1931         count_vm_event(THP_SPLIT);
1932
1933         BUG_ON(PageCompound(page));
1934 out_unlock:
1935         anon_vma_unlock_write(anon_vma);
1936         put_anon_vma(anon_vma);
1937 out:
1938         return ret;
1939 }
1940
1941 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1942
1943 int hugepage_madvise(struct vm_area_struct *vma,
1944                      unsigned long *vm_flags, int advice)
1945 {
1946         switch (advice) {
1947         case MADV_HUGEPAGE:
1948 #ifdef CONFIG_S390
1949                 /*
1950                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1951                  * can't handle this properly after s390_enable_sie, so we simply
1952                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1953                  */
1954                 if (mm_has_pgste(vma->vm_mm))
1955                         return 0;
1956 #endif
1957                 /*
1958                  * Be somewhat over-protective like KSM for now!
1959                  */
1960                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1961                         return -EINVAL;
1962                 *vm_flags &= ~VM_NOHUGEPAGE;
1963                 *vm_flags |= VM_HUGEPAGE;
1964                 /*
1965                  * If the vma become good for khugepaged to scan,
1966                  * register it here without waiting a page fault that
1967                  * may not happen any time soon.
1968                  */
1969                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1970                         return -ENOMEM;
1971                 break;
1972         case MADV_NOHUGEPAGE:
1973                 /*
1974                  * Be somewhat over-protective like KSM for now!
1975                  */
1976                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1977                         return -EINVAL;
1978                 *vm_flags &= ~VM_HUGEPAGE;
1979                 *vm_flags |= VM_NOHUGEPAGE;
1980                 /*
1981                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1982                  * this vma even if we leave the mm registered in khugepaged if
1983                  * it got registered before VM_NOHUGEPAGE was set.
1984                  */
1985                 break;
1986         }
1987
1988         return 0;
1989 }
1990
1991 static int __init khugepaged_slab_init(void)
1992 {
1993         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1994                                           sizeof(struct mm_slot),
1995                                           __alignof__(struct mm_slot), 0, NULL);
1996         if (!mm_slot_cache)
1997                 return -ENOMEM;
1998
1999         return 0;
2000 }
2001
2002 static inline struct mm_slot *alloc_mm_slot(void)
2003 {
2004         if (!mm_slot_cache)     /* initialization failed */
2005                 return NULL;
2006         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2007 }
2008
2009 static inline void free_mm_slot(struct mm_slot *mm_slot)
2010 {
2011         kmem_cache_free(mm_slot_cache, mm_slot);
2012 }
2013
2014 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2015 {
2016         struct mm_slot *mm_slot;
2017
2018         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2019                 if (mm == mm_slot->mm)
2020                         return mm_slot;
2021
2022         return NULL;
2023 }
2024
2025 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2026                                     struct mm_slot *mm_slot)
2027 {
2028         mm_slot->mm = mm;
2029         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2030 }
2031
2032 static inline int khugepaged_test_exit(struct mm_struct *mm)
2033 {
2034         return atomic_read(&mm->mm_users) == 0;
2035 }
2036
2037 int __khugepaged_enter(struct mm_struct *mm)
2038 {
2039         struct mm_slot *mm_slot;
2040         int wakeup;
2041
2042         mm_slot = alloc_mm_slot();
2043         if (!mm_slot)
2044                 return -ENOMEM;
2045
2046         /* __khugepaged_exit() must not run from under us */
2047         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2048         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2049                 free_mm_slot(mm_slot);
2050                 return 0;
2051         }
2052
2053         spin_lock(&khugepaged_mm_lock);
2054         insert_to_mm_slots_hash(mm, mm_slot);
2055         /*
2056          * Insert just behind the scanning cursor, to let the area settle
2057          * down a little.
2058          */
2059         wakeup = list_empty(&khugepaged_scan.mm_head);
2060         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2061         spin_unlock(&khugepaged_mm_lock);
2062
2063         atomic_inc(&mm->mm_count);
2064         if (wakeup)
2065                 wake_up_interruptible(&khugepaged_wait);
2066
2067         return 0;
2068 }
2069
2070 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2071                                unsigned long vm_flags)
2072 {
2073         unsigned long hstart, hend;
2074         if (!vma->anon_vma)
2075                 /*
2076                  * Not yet faulted in so we will register later in the
2077                  * page fault if needed.
2078                  */
2079                 return 0;
2080         if (vma->vm_ops)
2081                 /* khugepaged not yet working on file or special mappings */
2082                 return 0;
2083         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2084         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2085         hend = vma->vm_end & HPAGE_PMD_MASK;
2086         if (hstart < hend)
2087                 return khugepaged_enter(vma, vm_flags);
2088         return 0;
2089 }
2090
2091 void __khugepaged_exit(struct mm_struct *mm)
2092 {
2093         struct mm_slot *mm_slot;
2094         int free = 0;
2095
2096         spin_lock(&khugepaged_mm_lock);
2097         mm_slot = get_mm_slot(mm);
2098         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2099                 hash_del(&mm_slot->hash);
2100                 list_del(&mm_slot->mm_node);
2101                 free = 1;
2102         }
2103         spin_unlock(&khugepaged_mm_lock);
2104
2105         if (free) {
2106                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2107                 free_mm_slot(mm_slot);
2108                 mmdrop(mm);
2109         } else if (mm_slot) {
2110                 /*
2111                  * This is required to serialize against
2112                  * khugepaged_test_exit() (which is guaranteed to run
2113                  * under mmap sem read mode). Stop here (after we
2114                  * return all pagetables will be destroyed) until
2115                  * khugepaged has finished working on the pagetables
2116                  * under the mmap_sem.
2117                  */
2118                 down_write(&mm->mmap_sem);
2119                 up_write(&mm->mmap_sem);
2120         }
2121 }
2122
2123 static void release_pte_page(struct page *page)
2124 {
2125         /* 0 stands for page_is_file_cache(page) == false */
2126         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2127         unlock_page(page);
2128         putback_lru_page(page);
2129 }
2130
2131 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2132 {
2133         while (--_pte >= pte) {
2134                 pte_t pteval = *_pte;
2135                 if (!pte_none(pteval))
2136                         release_pte_page(pte_page(pteval));
2137         }
2138 }
2139
2140 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2141                                         unsigned long address,
2142                                         pte_t *pte)
2143 {
2144         struct page *page;
2145         pte_t *_pte;
2146         int referenced = 0, none = 0;
2147         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2148              _pte++, address += PAGE_SIZE) {
2149                 pte_t pteval = *_pte;
2150                 if (pte_none(pteval)) {
2151                         if (++none <= khugepaged_max_ptes_none)
2152                                 continue;
2153                         else
2154                                 goto out;
2155                 }
2156                 if (!pte_present(pteval) || !pte_write(pteval))
2157                         goto out;
2158                 page = vm_normal_page(vma, address, pteval);
2159                 if (unlikely(!page))
2160                         goto out;
2161
2162                 VM_BUG_ON_PAGE(PageCompound(page), page);
2163                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2164                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2165
2166                 /* cannot use mapcount: can't collapse if there's a gup pin */
2167                 if (page_count(page) != 1)
2168                         goto out;
2169                 /*
2170                  * We can do it before isolate_lru_page because the
2171                  * page can't be freed from under us. NOTE: PG_lock
2172                  * is needed to serialize against split_huge_page
2173                  * when invoked from the VM.
2174                  */
2175                 if (!trylock_page(page))
2176                         goto out;
2177                 /*
2178                  * Isolate the page to avoid collapsing an hugepage
2179                  * currently in use by the VM.
2180                  */
2181                 if (isolate_lru_page(page)) {
2182                         unlock_page(page);
2183                         goto out;
2184                 }
2185                 /* 0 stands for page_is_file_cache(page) == false */
2186                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2187                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2188                 VM_BUG_ON_PAGE(PageLRU(page), page);
2189
2190                 /* If there is no mapped pte young don't collapse the page */
2191                 if (pte_young(pteval) || PageReferenced(page) ||
2192                     mmu_notifier_test_young(vma->vm_mm, address))
2193                         referenced = 1;
2194         }
2195         if (likely(referenced))
2196                 return 1;
2197 out:
2198         release_pte_pages(pte, _pte);
2199         return 0;
2200 }
2201
2202 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2203                                       struct vm_area_struct *vma,
2204                                       unsigned long address,
2205                                       spinlock_t *ptl)
2206 {
2207         pte_t *_pte;
2208         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2209                 pte_t pteval = *_pte;
2210                 struct page *src_page;
2211
2212                 if (pte_none(pteval)) {
2213                         clear_user_highpage(page, address);
2214                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2215                 } else {
2216                         src_page = pte_page(pteval);
2217                         copy_user_highpage(page, src_page, address, vma);
2218                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2219                         release_pte_page(src_page);
2220                         /*
2221                          * ptl mostly unnecessary, but preempt has to
2222                          * be disabled to update the per-cpu stats
2223                          * inside page_remove_rmap().
2224                          */
2225                         spin_lock(ptl);
2226                         /*
2227                          * paravirt calls inside pte_clear here are
2228                          * superfluous.
2229                          */
2230                         pte_clear(vma->vm_mm, address, _pte);
2231                         page_remove_rmap(src_page);
2232                         spin_unlock(ptl);
2233                         free_page_and_swap_cache(src_page);
2234                 }
2235
2236                 address += PAGE_SIZE;
2237                 page++;
2238         }
2239 }
2240
2241 static void khugepaged_alloc_sleep(void)
2242 {
2243         wait_event_freezable_timeout(khugepaged_wait, false,
2244                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2245 }
2246
2247 static int khugepaged_node_load[MAX_NUMNODES];
2248
2249 static bool khugepaged_scan_abort(int nid)
2250 {
2251         int i;
2252
2253         /*
2254          * If zone_reclaim_mode is disabled, then no extra effort is made to
2255          * allocate memory locally.
2256          */
2257         if (!zone_reclaim_mode)
2258                 return false;
2259
2260         /* If there is a count for this node already, it must be acceptable */
2261         if (khugepaged_node_load[nid])
2262                 return false;
2263
2264         for (i = 0; i < MAX_NUMNODES; i++) {
2265                 if (!khugepaged_node_load[i])
2266                         continue;
2267                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2268                         return true;
2269         }
2270         return false;
2271 }
2272
2273 #ifdef CONFIG_NUMA
2274 static int khugepaged_find_target_node(void)
2275 {
2276         static int last_khugepaged_target_node = NUMA_NO_NODE;
2277         int nid, target_node = 0, max_value = 0;
2278
2279         /* find first node with max normal pages hit */
2280         for (nid = 0; nid < MAX_NUMNODES; nid++)
2281                 if (khugepaged_node_load[nid] > max_value) {
2282                         max_value = khugepaged_node_load[nid];
2283                         target_node = nid;
2284                 }
2285
2286         /* do some balance if several nodes have the same hit record */
2287         if (target_node <= last_khugepaged_target_node)
2288                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2289                                 nid++)
2290                         if (max_value == khugepaged_node_load[nid]) {
2291                                 target_node = nid;
2292                                 break;
2293                         }
2294
2295         last_khugepaged_target_node = target_node;
2296         return target_node;
2297 }
2298
2299 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2300 {
2301         if (IS_ERR(*hpage)) {
2302                 if (!*wait)
2303                         return false;
2304
2305                 *wait = false;
2306                 *hpage = NULL;
2307                 khugepaged_alloc_sleep();
2308         } else if (*hpage) {
2309                 put_page(*hpage);
2310                 *hpage = NULL;
2311         }
2312
2313         return true;
2314 }
2315
2316 static struct page
2317 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2318                        struct vm_area_struct *vma, unsigned long address,
2319                        int node)
2320 {
2321         VM_BUG_ON_PAGE(*hpage, *hpage);
2322
2323         /*
2324          * Before allocating the hugepage, release the mmap_sem read lock.
2325          * The allocation can take potentially a long time if it involves
2326          * sync compaction, and we do not need to hold the mmap_sem during
2327          * that. We will recheck the vma after taking it again in write mode.
2328          */
2329         up_read(&mm->mmap_sem);
2330
2331         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2332                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2333         if (unlikely(!*hpage)) {
2334                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2335                 *hpage = ERR_PTR(-ENOMEM);
2336                 return NULL;
2337         }
2338
2339         count_vm_event(THP_COLLAPSE_ALLOC);
2340         return *hpage;
2341 }
2342 #else
2343 static int khugepaged_find_target_node(void)
2344 {
2345         return 0;
2346 }
2347
2348 static inline struct page *alloc_hugepage(int defrag)
2349 {
2350         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2351                            HPAGE_PMD_ORDER);
2352 }
2353
2354 static struct page *khugepaged_alloc_hugepage(bool *wait)
2355 {
2356         struct page *hpage;
2357
2358         do {
2359                 hpage = alloc_hugepage(khugepaged_defrag());
2360                 if (!hpage) {
2361                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2362                         if (!*wait)
2363                                 return NULL;
2364
2365                         *wait = false;
2366                         khugepaged_alloc_sleep();
2367                 } else
2368                         count_vm_event(THP_COLLAPSE_ALLOC);
2369         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2370
2371         return hpage;
2372 }
2373
2374 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2375 {
2376         if (!*hpage)
2377                 *hpage = khugepaged_alloc_hugepage(wait);
2378
2379         if (unlikely(!*hpage))
2380                 return false;
2381
2382         return true;
2383 }
2384
2385 static struct page
2386 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2387                        struct vm_area_struct *vma, unsigned long address,
2388                        int node)
2389 {
2390         up_read(&mm->mmap_sem);
2391         VM_BUG_ON(!*hpage);
2392         return  *hpage;
2393 }
2394 #endif
2395
2396 static bool hugepage_vma_check(struct vm_area_struct *vma)
2397 {
2398         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2399             (vma->vm_flags & VM_NOHUGEPAGE))
2400                 return false;
2401
2402         if (!vma->anon_vma || vma->vm_ops)
2403                 return false;
2404         if (is_vma_temporary_stack(vma))
2405                 return false;
2406         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2407         return true;
2408 }
2409
2410 static void collapse_huge_page(struct mm_struct *mm,
2411                                    unsigned long address,
2412                                    struct page **hpage,
2413                                    struct vm_area_struct *vma,
2414                                    int node)
2415 {
2416         pmd_t *pmd, _pmd;
2417         pte_t *pte;
2418         pgtable_t pgtable;
2419         struct page *new_page;
2420         spinlock_t *pmd_ptl, *pte_ptl;
2421         int isolated;
2422         unsigned long hstart, hend;
2423         struct mem_cgroup *memcg;
2424         unsigned long mmun_start;       /* For mmu_notifiers */
2425         unsigned long mmun_end;         /* For mmu_notifiers */
2426
2427         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2428
2429         /* release the mmap_sem read lock. */
2430         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2431         if (!new_page)
2432                 return;
2433
2434         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2435                                            GFP_TRANSHUGE, &memcg)))
2436                 return;
2437
2438         /*
2439          * Prevent all access to pagetables with the exception of
2440          * gup_fast later hanlded by the ptep_clear_flush and the VM
2441          * handled by the anon_vma lock + PG_lock.
2442          */
2443         down_write(&mm->mmap_sem);
2444         if (unlikely(khugepaged_test_exit(mm)))
2445                 goto out;
2446
2447         vma = find_vma(mm, address);
2448         if (!vma)
2449                 goto out;
2450         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2451         hend = vma->vm_end & HPAGE_PMD_MASK;
2452         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2453                 goto out;
2454         if (!hugepage_vma_check(vma))
2455                 goto out;
2456         pmd = mm_find_pmd(mm, address);
2457         if (!pmd)
2458                 goto out;
2459
2460         anon_vma_lock_write(vma->anon_vma);
2461
2462         pte = pte_offset_map(pmd, address);
2463         pte_ptl = pte_lockptr(mm, pmd);
2464
2465         mmun_start = address;
2466         mmun_end   = address + HPAGE_PMD_SIZE;
2467         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2468         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2469         /*
2470          * After this gup_fast can't run anymore. This also removes
2471          * any huge TLB entry from the CPU so we won't allow
2472          * huge and small TLB entries for the same virtual address
2473          * to avoid the risk of CPU bugs in that area.
2474          */
2475         _pmd = pmdp_clear_flush(vma, address, pmd);
2476         spin_unlock(pmd_ptl);
2477         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2478
2479         spin_lock(pte_ptl);
2480         isolated = __collapse_huge_page_isolate(vma, address, pte);
2481         spin_unlock(pte_ptl);
2482
2483         if (unlikely(!isolated)) {
2484                 pte_unmap(pte);
2485                 spin_lock(pmd_ptl);
2486                 BUG_ON(!pmd_none(*pmd));
2487                 /*
2488                  * We can only use set_pmd_at when establishing
2489                  * hugepmds and never for establishing regular pmds that
2490                  * points to regular pagetables. Use pmd_populate for that
2491                  */
2492                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2493                 spin_unlock(pmd_ptl);
2494                 anon_vma_unlock_write(vma->anon_vma);
2495                 goto out;
2496         }
2497
2498         /*
2499          * All pages are isolated and locked so anon_vma rmap
2500          * can't run anymore.
2501          */
2502         anon_vma_unlock_write(vma->anon_vma);
2503
2504         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2505         pte_unmap(pte);
2506         __SetPageUptodate(new_page);
2507         pgtable = pmd_pgtable(_pmd);
2508
2509         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2510         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2511
2512         /*
2513          * spin_lock() below is not the equivalent of smp_wmb(), so
2514          * this is needed to avoid the copy_huge_page writes to become
2515          * visible after the set_pmd_at() write.
2516          */
2517         smp_wmb();
2518
2519         spin_lock(pmd_ptl);
2520         BUG_ON(!pmd_none(*pmd));
2521         page_add_new_anon_rmap(new_page, vma, address);
2522         mem_cgroup_commit_charge(new_page, memcg, false);
2523         lru_cache_add_active_or_unevictable(new_page, vma);
2524         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2525         set_pmd_at(mm, address, pmd, _pmd);
2526         update_mmu_cache_pmd(vma, address, pmd);
2527         spin_unlock(pmd_ptl);
2528
2529         *hpage = NULL;
2530
2531         khugepaged_pages_collapsed++;
2532 out_up_write:
2533         up_write(&mm->mmap_sem);
2534         return;
2535
2536 out:
2537         mem_cgroup_cancel_charge(new_page, memcg);
2538         goto out_up_write;
2539 }
2540
2541 static int khugepaged_scan_pmd(struct mm_struct *mm,
2542                                struct vm_area_struct *vma,
2543                                unsigned long address,
2544                                struct page **hpage)
2545 {
2546         pmd_t *pmd;
2547         pte_t *pte, *_pte;
2548         int ret = 0, referenced = 0, none = 0;
2549         struct page *page;
2550         unsigned long _address;
2551         spinlock_t *ptl;
2552         int node = NUMA_NO_NODE;
2553
2554         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2555
2556         pmd = mm_find_pmd(mm, address);
2557         if (!pmd)
2558                 goto out;
2559
2560         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2561         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2562         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2563              _pte++, _address += PAGE_SIZE) {
2564                 pte_t pteval = *_pte;
2565                 if (pte_none(pteval)) {
2566                         if (++none <= khugepaged_max_ptes_none)
2567                                 continue;
2568                         else
2569                                 goto out_unmap;
2570                 }
2571                 if (!pte_present(pteval) || !pte_write(pteval))
2572                         goto out_unmap;
2573                 page = vm_normal_page(vma, _address, pteval);
2574                 if (unlikely(!page))
2575                         goto out_unmap;
2576                 /*
2577                  * Record which node the original page is from and save this
2578                  * information to khugepaged_node_load[].
2579                  * Khupaged will allocate hugepage from the node has the max
2580                  * hit record.
2581                  */
2582                 node = page_to_nid(page);
2583                 if (khugepaged_scan_abort(node))
2584                         goto out_unmap;
2585                 khugepaged_node_load[node]++;
2586                 VM_BUG_ON_PAGE(PageCompound(page), page);
2587                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2588                         goto out_unmap;
2589                 /* cannot use mapcount: can't collapse if there's a gup pin */
2590                 if (page_count(page) != 1)
2591                         goto out_unmap;
2592                 if (pte_young(pteval) || PageReferenced(page) ||
2593                     mmu_notifier_test_young(vma->vm_mm, address))
2594                         referenced = 1;
2595         }
2596         if (referenced)
2597                 ret = 1;
2598 out_unmap:
2599         pte_unmap_unlock(pte, ptl);
2600         if (ret) {
2601                 node = khugepaged_find_target_node();
2602                 /* collapse_huge_page will return with the mmap_sem released */
2603                 collapse_huge_page(mm, address, hpage, vma, node);
2604         }
2605 out:
2606         return ret;
2607 }
2608
2609 static void collect_mm_slot(struct mm_slot *mm_slot)
2610 {
2611         struct mm_struct *mm = mm_slot->mm;
2612
2613         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2614
2615         if (khugepaged_test_exit(mm)) {
2616                 /* free mm_slot */
2617                 hash_del(&mm_slot->hash);
2618                 list_del(&mm_slot->mm_node);
2619
2620                 /*
2621                  * Not strictly needed because the mm exited already.
2622                  *
2623                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2624                  */
2625
2626                 /* khugepaged_mm_lock actually not necessary for the below */
2627                 free_mm_slot(mm_slot);
2628                 mmdrop(mm);
2629         }
2630 }
2631
2632 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2633                                             struct page **hpage)
2634         __releases(&khugepaged_mm_lock)
2635         __acquires(&khugepaged_mm_lock)
2636 {
2637         struct mm_slot *mm_slot;
2638         struct mm_struct *mm;
2639         struct vm_area_struct *vma;
2640         int progress = 0;
2641
2642         VM_BUG_ON(!pages);
2643         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2644
2645         if (khugepaged_scan.mm_slot)
2646                 mm_slot = khugepaged_scan.mm_slot;
2647         else {
2648                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2649                                      struct mm_slot, mm_node);
2650                 khugepaged_scan.address = 0;
2651                 khugepaged_scan.mm_slot = mm_slot;
2652         }
2653         spin_unlock(&khugepaged_mm_lock);
2654
2655         mm = mm_slot->mm;
2656         down_read(&mm->mmap_sem);
2657         if (unlikely(khugepaged_test_exit(mm)))
2658                 vma = NULL;
2659         else
2660                 vma = find_vma(mm, khugepaged_scan.address);
2661
2662         progress++;
2663         for (; vma; vma = vma->vm_next) {
2664                 unsigned long hstart, hend;
2665
2666                 cond_resched();
2667                 if (unlikely(khugepaged_test_exit(mm))) {
2668                         progress++;
2669                         break;
2670                 }
2671                 if (!hugepage_vma_check(vma)) {
2672 skip:
2673                         progress++;
2674                         continue;
2675                 }
2676                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2677                 hend = vma->vm_end & HPAGE_PMD_MASK;
2678                 if (hstart >= hend)
2679                         goto skip;
2680                 if (khugepaged_scan.address > hend)
2681                         goto skip;
2682                 if (khugepaged_scan.address < hstart)
2683                         khugepaged_scan.address = hstart;
2684                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2685
2686                 while (khugepaged_scan.address < hend) {
2687                         int ret;
2688                         cond_resched();
2689                         if (unlikely(khugepaged_test_exit(mm)))
2690                                 goto breakouterloop;
2691
2692                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2693                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2694                                   hend);
2695                         ret = khugepaged_scan_pmd(mm, vma,
2696                                                   khugepaged_scan.address,
2697                                                   hpage);
2698                         /* move to next address */
2699                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2700                         progress += HPAGE_PMD_NR;
2701                         if (ret)
2702                                 /* we released mmap_sem so break loop */
2703                                 goto breakouterloop_mmap_sem;
2704                         if (progress >= pages)
2705                                 goto breakouterloop;
2706                 }
2707         }
2708 breakouterloop:
2709         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2710 breakouterloop_mmap_sem:
2711
2712         spin_lock(&khugepaged_mm_lock);
2713         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2714         /*
2715          * Release the current mm_slot if this mm is about to die, or
2716          * if we scanned all vmas of this mm.
2717          */
2718         if (khugepaged_test_exit(mm) || !vma) {
2719                 /*
2720                  * Make sure that if mm_users is reaching zero while
2721                  * khugepaged runs here, khugepaged_exit will find
2722                  * mm_slot not pointing to the exiting mm.
2723                  */
2724                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2725                         khugepaged_scan.mm_slot = list_entry(
2726                                 mm_slot->mm_node.next,
2727                                 struct mm_slot, mm_node);
2728                         khugepaged_scan.address = 0;
2729                 } else {
2730                         khugepaged_scan.mm_slot = NULL;
2731                         khugepaged_full_scans++;
2732                 }
2733
2734                 collect_mm_slot(mm_slot);
2735         }
2736
2737         return progress;
2738 }
2739
2740 static int khugepaged_has_work(void)
2741 {
2742         return !list_empty(&khugepaged_scan.mm_head) &&
2743                 khugepaged_enabled();
2744 }
2745
2746 static int khugepaged_wait_event(void)
2747 {
2748         return !list_empty(&khugepaged_scan.mm_head) ||
2749                 kthread_should_stop();
2750 }
2751
2752 static void khugepaged_do_scan(void)
2753 {
2754         struct page *hpage = NULL;
2755         unsigned int progress = 0, pass_through_head = 0;
2756         unsigned int pages = khugepaged_pages_to_scan;
2757         bool wait = true;
2758
2759         barrier(); /* write khugepaged_pages_to_scan to local stack */
2760
2761         while (progress < pages) {
2762                 if (!khugepaged_prealloc_page(&hpage, &wait))
2763                         break;
2764
2765                 cond_resched();
2766
2767                 if (unlikely(kthread_should_stop() || freezing(current)))
2768                         break;
2769
2770                 spin_lock(&khugepaged_mm_lock);
2771                 if (!khugepaged_scan.mm_slot)
2772                         pass_through_head++;
2773                 if (khugepaged_has_work() &&
2774                     pass_through_head < 2)
2775                         progress += khugepaged_scan_mm_slot(pages - progress,
2776                                                             &hpage);
2777                 else
2778                         progress = pages;
2779                 spin_unlock(&khugepaged_mm_lock);
2780         }
2781
2782         if (!IS_ERR_OR_NULL(hpage))
2783                 put_page(hpage);
2784 }
2785
2786 static void khugepaged_wait_work(void)
2787 {
2788         try_to_freeze();
2789
2790         if (khugepaged_has_work()) {
2791                 if (!khugepaged_scan_sleep_millisecs)
2792                         return;
2793
2794                 wait_event_freezable_timeout(khugepaged_wait,
2795                                              kthread_should_stop(),
2796                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2797                 return;
2798         }
2799
2800         if (khugepaged_enabled())
2801                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2802 }
2803
2804 static int khugepaged(void *none)
2805 {
2806         struct mm_slot *mm_slot;
2807
2808         set_freezable();
2809         set_user_nice(current, MAX_NICE);
2810
2811         while (!kthread_should_stop()) {
2812                 khugepaged_do_scan();
2813                 khugepaged_wait_work();
2814         }
2815
2816         spin_lock(&khugepaged_mm_lock);
2817         mm_slot = khugepaged_scan.mm_slot;
2818         khugepaged_scan.mm_slot = NULL;
2819         if (mm_slot)
2820                 collect_mm_slot(mm_slot);
2821         spin_unlock(&khugepaged_mm_lock);
2822         return 0;
2823 }
2824
2825 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2826                 unsigned long haddr, pmd_t *pmd)
2827 {
2828         struct mm_struct *mm = vma->vm_mm;
2829         pgtable_t pgtable;
2830         pmd_t _pmd;
2831         int i;
2832
2833         pmdp_clear_flush_notify(vma, haddr, pmd);
2834         /* leave pmd empty until pte is filled */
2835
2836         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2837         pmd_populate(mm, &_pmd, pgtable);
2838
2839         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2840                 pte_t *pte, entry;
2841                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2842                 entry = pte_mkspecial(entry);
2843                 pte = pte_offset_map(&_pmd, haddr);
2844                 VM_BUG_ON(!pte_none(*pte));
2845                 set_pte_at(mm, haddr, pte, entry);
2846                 pte_unmap(pte);
2847         }
2848         smp_wmb(); /* make pte visible before pmd */
2849         pmd_populate(mm, pmd, pgtable);
2850         put_huge_zero_page();
2851 }
2852
2853 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2854                 pmd_t *pmd)
2855 {
2856         spinlock_t *ptl;
2857         struct page *page;
2858         struct mm_struct *mm = vma->vm_mm;
2859         unsigned long haddr = address & HPAGE_PMD_MASK;
2860         unsigned long mmun_start;       /* For mmu_notifiers */
2861         unsigned long mmun_end;         /* For mmu_notifiers */
2862
2863         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2864
2865         mmun_start = haddr;
2866         mmun_end   = haddr + HPAGE_PMD_SIZE;
2867 again:
2868         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2869         ptl = pmd_lock(mm, pmd);
2870         if (unlikely(!pmd_trans_huge(*pmd))) {
2871                 spin_unlock(ptl);
2872                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2873                 return;
2874         }
2875         if (is_huge_zero_pmd(*pmd)) {
2876                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2877                 spin_unlock(ptl);
2878                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2879                 return;
2880         }
2881         page = pmd_page(*pmd);
2882         VM_BUG_ON_PAGE(!page_count(page), page);
2883         get_page(page);
2884         spin_unlock(ptl);
2885         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2886
2887         split_huge_page(page);
2888
2889         put_page(page);
2890
2891         /*
2892          * We don't always have down_write of mmap_sem here: a racing
2893          * do_huge_pmd_wp_page() might have copied-on-write to another
2894          * huge page before our split_huge_page() got the anon_vma lock.
2895          */
2896         if (unlikely(pmd_trans_huge(*pmd)))
2897                 goto again;
2898 }
2899
2900 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2901                 pmd_t *pmd)
2902 {
2903         struct vm_area_struct *vma;
2904
2905         vma = find_vma(mm, address);
2906         BUG_ON(vma == NULL);
2907         split_huge_page_pmd(vma, address, pmd);
2908 }
2909
2910 static void split_huge_page_address(struct mm_struct *mm,
2911                                     unsigned long address)
2912 {
2913         pgd_t *pgd;
2914         pud_t *pud;
2915         pmd_t *pmd;
2916
2917         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2918
2919         pgd = pgd_offset(mm, address);
2920         if (!pgd_present(*pgd))
2921                 return;
2922
2923         pud = pud_offset(pgd, address);
2924         if (!pud_present(*pud))
2925                 return;
2926
2927         pmd = pmd_offset(pud, address);
2928         if (!pmd_present(*pmd))
2929                 return;
2930         /*
2931          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2932          * materialize from under us.
2933          */
2934         split_huge_page_pmd_mm(mm, address, pmd);
2935 }
2936
2937 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2938                              unsigned long start,
2939                              unsigned long end,
2940                              long adjust_next)
2941 {
2942         /*
2943          * If the new start address isn't hpage aligned and it could
2944          * previously contain an hugepage: check if we need to split
2945          * an huge pmd.
2946          */
2947         if (start & ~HPAGE_PMD_MASK &&
2948             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2949             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2950                 split_huge_page_address(vma->vm_mm, start);
2951
2952         /*
2953          * If the new end address isn't hpage aligned and it could
2954          * previously contain an hugepage: check if we need to split
2955          * an huge pmd.
2956          */
2957         if (end & ~HPAGE_PMD_MASK &&
2958             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2959             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2960                 split_huge_page_address(vma->vm_mm, end);
2961
2962         /*
2963          * If we're also updating the vma->vm_next->vm_start, if the new
2964          * vm_next->vm_start isn't page aligned and it could previously
2965          * contain an hugepage: check if we need to split an huge pmd.
2966          */
2967         if (adjust_next > 0) {
2968                 struct vm_area_struct *next = vma->vm_next;
2969                 unsigned long nstart = next->vm_start;
2970                 nstart += adjust_next << PAGE_SHIFT;
2971                 if (nstart & ~HPAGE_PMD_MASK &&
2972                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2973                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2974                         split_huge_page_address(next->vm_mm, nstart);
2975         }
2976 }