mm/huge_memory.c

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