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_pages(zero_page, compound_order(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_pages(zero_page, compound_order(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         struct mem_cgroup *memcg;
 719         pgtable_t pgtable;
 720         spinlock_t *ptl;
 721
 722         VM_BUG_ON_PAGE(!PageCompound(page), page);
 723
 724         if (mem_cgroup_try_charge(page, mm, GFP_TRANSHUGE, &memcg))
 725                 return VM_FAULT_OOM;
 726
 727         pgtable = pte_alloc_one(mm, haddr);
 728         if (unlikely(!pgtable)) {
 729                 mem_cgroup_cancel_charge(page, memcg);
 730                 return VM_FAULT_OOM;
 731         }
 732
 733         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 734         /*
 735          * The memory barrier inside __SetPageUptodate makes sure that
 736          * clear_huge_page writes become visible before the set_pmd_at()
 737          * write.
 738          */
 739         __SetPageUptodate(page);
 740
 741         ptl = pmd_lock(mm, pmd);
 742         if (unlikely(!pmd_none(*pmd))) {
 743                 spin_unlock(ptl);
 744                 mem_cgroup_cancel_charge(page, memcg);
 745                 put_page(page);
 746                 pte_free(mm, pgtable);
 747         } else {
 748                 pmd_t entry;
 749                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 750                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 751                 page_add_new_anon_rmap(page, vma, haddr);
 752                 mem_cgroup_commit_charge(page, memcg, false);
 753                 lru_cache_add_active_or_unevictable(page, vma);
 754                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 755                 set_pmd_at(mm, haddr, pmd, entry);
 756                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 757                 atomic_long_inc(&mm->nr_ptes);
 758                 spin_unlock(ptl);
 759         }
 760
 761         return 0;
 762 }
 763
 764 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 765 {
 766         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 767 }
 768
 769 static inline struct page *alloc_hugepage_vma(int defrag,
 770                                               struct vm_area_struct *vma,
 771                                               unsigned long haddr, int nd,
 772                                               gfp_t extra_gfp)
 773 {
 774         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 775                                HPAGE_PMD_ORDER, vma, haddr, nd);
 776 }
 777
 778 /* Caller must hold page table lock. */
 779 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 780                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 781                 struct page *zero_page)
 782 {
 783         pmd_t entry;
 784         if (!pmd_none(*pmd))
 785                 return false;
 786         entry = mk_pmd(zero_page, vma->vm_page_prot);
 787         entry = pmd_mkhuge(entry);
 788         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 789         set_pmd_at(mm, haddr, pmd, entry);
 790         atomic_long_inc(&mm->nr_ptes);
 791         return true;
 792 }
 793
 794 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 795                                unsigned long address, pmd_t *pmd,
 796                                unsigned int flags)
 797 {
 798         struct page *page;
 799         unsigned long haddr = address & HPAGE_PMD_MASK;
 800
 801         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 802                 return VM_FAULT_FALLBACK;
 803         if (unlikely(anon_vma_prepare(vma)))
 804                 return VM_FAULT_OOM;
 805         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 806                 return VM_FAULT_OOM;
 807         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 808                         transparent_hugepage_use_zero_page()) {
 809                 spinlock_t *ptl;
 810                 pgtable_t pgtable;
 811                 struct page *zero_page;
 812                 bool set;
 813                 pgtable = pte_alloc_one(mm, haddr);
 814                 if (unlikely(!pgtable))
 815                         return VM_FAULT_OOM;
 816                 zero_page = get_huge_zero_page();
 817                 if (unlikely(!zero_page)) {
 818                         pte_free(mm, pgtable);
 819                         count_vm_event(THP_FAULT_FALLBACK);
 820                         return VM_FAULT_FALLBACK;
 821                 }
 822                 ptl = pmd_lock(mm, pmd);
 823                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 824                                 zero_page);
 825                 spin_unlock(ptl);
 826                 if (!set) {
 827                         pte_free(mm, pgtable);
 828                         put_huge_zero_page();
 829                 }
 830                 return 0;
 831         }
 832         page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 833                         vma, haddr, numa_node_id(), 0);
 834         if (unlikely(!page)) {
 835                 count_vm_event(THP_FAULT_FALLBACK);
 836                 return VM_FAULT_FALLBACK;
 837         }
 838         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
 839                 put_page(page);
 840                 count_vm_event(THP_FAULT_FALLBACK);
 841                 return VM_FAULT_FALLBACK;
 842         }
 843
 844         count_vm_event(THP_FAULT_ALLOC);
 845         return 0;
 846 }
 847
 848 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 849                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 850                   struct vm_area_struct *vma)
 851 {
 852         spinlock_t *dst_ptl, *src_ptl;
 853         struct page *src_page;
 854         pmd_t pmd;
 855         pgtable_t pgtable;
 856         int ret;
 857
 858         ret = -ENOMEM;
 859         pgtable = pte_alloc_one(dst_mm, addr);
 860         if (unlikely(!pgtable))
 861                 goto out;
 862
 863         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 864         src_ptl = pmd_lockptr(src_mm, src_pmd);
 865         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 866
 867         ret = -EAGAIN;
 868         pmd = *src_pmd;
 869         if (unlikely(!pmd_trans_huge(pmd))) {
 870                 pte_free(dst_mm, pgtable);
 871                 goto out_unlock;
 872         }
 873         /*
 874          * When page table lock is held, the huge zero pmd should not be
 875          * under splitting since we don't split the page itself, only pmd to
 876          * a page table.
 877          */
 878         if (is_huge_zero_pmd(pmd)) {
 879                 struct page *zero_page;
 880                 bool set;
 881                 /*
 882                  * get_huge_zero_page() will never allocate a new page here,
 883                  * since we already have a zero page to copy. It just takes a
 884                  * reference.
 885                  */
 886                 zero_page = get_huge_zero_page();
 887                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 888                                 zero_page);
 889                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 890                 ret = 0;
 891                 goto out_unlock;
 892         }
 893
 894         if (unlikely(pmd_trans_splitting(pmd))) {
 895                 /* split huge page running from under us */
 896                 spin_unlock(src_ptl);
 897                 spin_unlock(dst_ptl);
 898                 pte_free(dst_mm, pgtable);
 899
 900                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 901                 goto out;
 902         }
 903         src_page = pmd_page(pmd);
 904         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 905         get_page(src_page);
 906         page_dup_rmap(src_page);
 907         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 908
 909         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 910         pmd = pmd_mkold(pmd_wrprotect(pmd));
 911         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 912         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 913         atomic_long_inc(&dst_mm->nr_ptes);
 914
 915         ret = 0;
 916 out_unlock:
 917         spin_unlock(src_ptl);
 918         spin_unlock(dst_ptl);
 919 out:
 920         return ret;
 921 }
 922
 923 void huge_pmd_set_accessed(struct mm_struct *mm,
 924                            struct vm_area_struct *vma,
 925                            unsigned long address,
 926                            pmd_t *pmd, pmd_t orig_pmd,
 927                            int dirty)
 928 {
 929         spinlock_t *ptl;
 930         pmd_t entry;
 931         unsigned long haddr;
 932
 933         ptl = pmd_lock(mm, pmd);
 934         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 935                 goto unlock;
 936
 937         entry = pmd_mkyoung(orig_pmd);
 938         haddr = address & HPAGE_PMD_MASK;
 939         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 940                 update_mmu_cache_pmd(vma, address, pmd);
 941
 942 unlock:
 943         spin_unlock(ptl);
 944 }
 945
 946 /*
 947  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 948  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 949  * the source page gets split and a tail freed before copy completes.
 950  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 951  */
 952 static void get_user_huge_page(struct page *page)
 953 {
 954         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 955                 struct page *endpage = page + HPAGE_PMD_NR;
 956
 957                 atomic_add(HPAGE_PMD_NR, &page->_count);
 958                 while (++page < endpage)
 959                         get_huge_page_tail(page);
 960         } else {
 961                 get_page(page);
 962         }
 963 }
 964
 965 static void put_user_huge_page(struct page *page)
 966 {
 967         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 968                 struct page *endpage = page + HPAGE_PMD_NR;
 969
 970                 while (page < endpage)
 971                         put_page(page++);
 972         } else {
 973                 put_page(page);
 974         }
 975 }
 976
 977 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 978                                         struct vm_area_struct *vma,
 979                                         unsigned long address,
 980                                         pmd_t *pmd, pmd_t orig_pmd,
 981                                         struct page *page,
 982                                         unsigned long haddr)
 983 {
 984         struct mem_cgroup *memcg;
 985         spinlock_t *ptl;
 986         pgtable_t pgtable;
 987         pmd_t _pmd;
 988         int ret = 0, i;
 989         struct page **pages;
 990         unsigned long mmun_start;       /* For mmu_notifiers */
 991         unsigned long mmun_end;         /* For mmu_notifiers */
 992
 993         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 994                         GFP_KERNEL);
 995         if (unlikely(!pages)) {
 996                 ret |= VM_FAULT_OOM;
 997                 goto out;
 998         }
 999
1000         for (i = 0; i < HPAGE_PMD_NR; i++) {
1001                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1002                                                __GFP_OTHER_NODE,
1003                                                vma, address, page_to_nid(page));
1004                 if (unlikely(!pages[i] ||
1005                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1006                                                    &memcg))) {
1007                         if (pages[i])
1008                                 put_page(pages[i]);
1009                         while (--i >= 0) {
1010                                 memcg = (void *)page_private(pages[i]);
1011                                 set_page_private(pages[i], 0);
1012                                 mem_cgroup_cancel_charge(pages[i], memcg);
1013                                 put_page(pages[i]);
1014                         }
1015                         kfree(pages);
1016                         ret |= VM_FAULT_OOM;
1017                         goto out;
1018                 }
1019                 set_page_private(pages[i], (unsigned long)memcg);
1020         }
1021
1022         for (i = 0; i < HPAGE_PMD_NR; i++) {
1023                 copy_user_highpage(pages[i], page + i,
1024                                    haddr + PAGE_SIZE * i, vma);
1025                 __SetPageUptodate(pages[i]);
1026                 cond_resched();
1027         }
1028
1029         mmun_start = haddr;
1030         mmun_end   = haddr + HPAGE_PMD_SIZE;
1031         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1032
1033         ptl = pmd_lock(mm, pmd);
1034         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1035                 goto out_free_pages;
1036         VM_BUG_ON_PAGE(!PageHead(page), page);
1037
1038         pmdp_clear_flush_notify(vma, haddr, pmd);
1039         /* leave pmd empty until pte is filled */
1040
1041         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1042         pmd_populate(mm, &_pmd, pgtable);
1043
1044         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1045                 pte_t *pte, entry;
1046                 entry = mk_pte(pages[i], vma->vm_page_prot);
1047                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1048                 memcg = (void *)page_private(pages[i]);
1049                 set_page_private(pages[i], 0);
1050                 page_add_new_anon_rmap(pages[i], vma, haddr);
1051                 mem_cgroup_commit_charge(pages[i], memcg, false);
1052                 lru_cache_add_active_or_unevictable(pages[i], vma);
1053                 pte = pte_offset_map(&_pmd, haddr);
1054                 VM_BUG_ON(!pte_none(*pte));
1055                 set_pte_at(mm, haddr, pte, entry);
1056                 pte_unmap(pte);
1057         }
1058         kfree(pages);
1059
1060         smp_wmb(); /* make pte visible before pmd */
1061         pmd_populate(mm, pmd, pgtable);
1062         page_remove_rmap(page);
1063         spin_unlock(ptl);
1064
1065         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1066
1067         ret |= VM_FAULT_WRITE;
1068         put_page(page);
1069
1070 out:
1071         return ret;
1072
1073 out_free_pages:
1074         spin_unlock(ptl);
1075         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1076         for (i = 0; i < HPAGE_PMD_NR; i++) {
1077                 memcg = (void *)page_private(pages[i]);
1078                 set_page_private(pages[i], 0);
1079                 mem_cgroup_cancel_charge(pages[i], memcg);
1080                 put_page(pages[i]);
1081         }
1082         kfree(pages);
1083         goto out;
1084 }
1085
1086 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1087                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1088 {
1089         spinlock_t *ptl;
1090         int ret = 0;
1091         struct page *page = NULL, *new_page;
1092         struct mem_cgroup *memcg;
1093         unsigned long haddr;
1094         unsigned long mmun_start;       /* For mmu_notifiers */
1095         unsigned long mmun_end;         /* For mmu_notifiers */
1096
1097         ptl = pmd_lockptr(mm, pmd);
1098         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1099         haddr = address & HPAGE_PMD_MASK;
1100         if (is_huge_zero_pmd(orig_pmd))
1101                 goto alloc;
1102         spin_lock(ptl);
1103         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1104                 goto out_unlock;
1105
1106         page = pmd_page(orig_pmd);
1107         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1108         if (page_mapcount(page) == 1) {
1109                 pmd_t entry;
1110                 entry = pmd_mkyoung(orig_pmd);
1111                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1112                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1113                         update_mmu_cache_pmd(vma, address, pmd);
1114                 ret |= VM_FAULT_WRITE;
1115                 goto out_unlock;
1116         }
1117         get_user_huge_page(page);
1118         spin_unlock(ptl);
1119 alloc:
1120         if (transparent_hugepage_enabled(vma) &&
1121             !transparent_hugepage_debug_cow())
1122                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1123                                               vma, haddr, numa_node_id(), 0);
1124         else
1125                 new_page = NULL;
1126
1127         if (unlikely(!new_page)) {
1128                 if (!page) {
1129                         split_huge_page_pmd(vma, address, pmd);
1130                         ret |= VM_FAULT_FALLBACK;
1131                 } else {
1132                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1133                                         pmd, orig_pmd, page, haddr);
1134                         if (ret & VM_FAULT_OOM) {
1135                                 split_huge_page(page);
1136                                 ret |= VM_FAULT_FALLBACK;
1137                         }
1138                         put_user_huge_page(page);
1139                 }
1140                 count_vm_event(THP_FAULT_FALLBACK);
1141                 goto out;
1142         }
1143
1144         if (unlikely(mem_cgroup_try_charge(new_page, mm,
1145                                            GFP_TRANSHUGE, &memcg))) {
1146                 put_page(new_page);
1147                 if (page) {
1148                         split_huge_page(page);
1149                         put_user_huge_page(page);
1150                 } else
1151                         split_huge_page_pmd(vma, address, pmd);
1152                 ret |= VM_FAULT_FALLBACK;
1153                 count_vm_event(THP_FAULT_FALLBACK);
1154                 goto out;
1155         }
1156
1157         count_vm_event(THP_FAULT_ALLOC);
1158
1159         if (!page)
1160                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1161         else
1162                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1163         __SetPageUptodate(new_page);
1164
1165         mmun_start = haddr;
1166         mmun_end   = haddr + HPAGE_PMD_SIZE;
1167         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1168
1169         spin_lock(ptl);
1170         if (page)
1171                 put_user_huge_page(page);
1172         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1173                 spin_unlock(ptl);
1174                 mem_cgroup_cancel_charge(new_page, memcg);
1175                 put_page(new_page);
1176                 goto out_mn;
1177         } else {
1178                 pmd_t entry;
1179                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1180                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1181                 pmdp_clear_flush_notify(vma, haddr, pmd);
1182                 page_add_new_anon_rmap(new_page, vma, haddr);
1183                 mem_cgroup_commit_charge(new_page, memcg, false);
1184                 lru_cache_add_active_or_unevictable(new_page, vma);
1185                 set_pmd_at(mm, haddr, pmd, entry);
1186                 update_mmu_cache_pmd(vma, address, pmd);
1187                 if (!page) {
1188                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1189                         put_huge_zero_page();
1190                 } else {
1191                         VM_BUG_ON_PAGE(!PageHead(page), page);
1192                         page_remove_rmap(page);
1193                         put_page(page);
1194                 }
1195                 ret |= VM_FAULT_WRITE;
1196         }
1197         spin_unlock(ptl);
1198 out_mn:
1199         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1200 out:
1201         return ret;
1202 out_unlock:
1203         spin_unlock(ptl);
1204         return ret;
1205 }
1206
1207 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1208                                    unsigned long addr,
1209                                    pmd_t *pmd,
1210                                    unsigned int flags)
1211 {
1212         struct mm_struct *mm = vma->vm_mm;
1213         struct page *page = NULL;
1214
1215         assert_spin_locked(pmd_lockptr(mm, pmd));
1216
1217         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1218                 goto out;
1219
1220         /* Avoid dumping huge zero page */
1221         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1222                 return ERR_PTR(-EFAULT);
1223
1224         /* Full NUMA hinting faults to serialise migration in fault paths */
1225         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1226                 goto out;
1227
1228         page = pmd_page(*pmd);
1229         VM_BUG_ON_PAGE(!PageHead(page), page);
1230         if (flags & FOLL_TOUCH) {
1231                 pmd_t _pmd;
1232                 /*
1233                  * We should set the dirty bit only for FOLL_WRITE but
1234                  * for now the dirty bit in the pmd is meaningless.
1235                  * And if the dirty bit will become meaningful and
1236                  * we'll only set it with FOLL_WRITE, an atomic
1237                  * set_bit will be required on the pmd to set the
1238                  * young bit, instead of the current set_pmd_at.
1239                  */
1240                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1241                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1242                                           pmd, _pmd,  1))
1243                         update_mmu_cache_pmd(vma, addr, pmd);
1244         }
1245         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1246                 if (page->mapping && trylock_page(page)) {
1247                         lru_add_drain();
1248                         if (page->mapping)
1249                                 mlock_vma_page(page);
1250                         unlock_page(page);
1251                 }
1252         }
1253         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1254         VM_BUG_ON_PAGE(!PageCompound(page), page);
1255         if (flags & FOLL_GET)
1256                 get_page_foll(page);
1257
1258 out:
1259         return page;
1260 }
1261
1262 /* NUMA hinting page fault entry point for trans huge pmds */
1263 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1264                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1265 {
1266         spinlock_t *ptl;
1267         struct anon_vma *anon_vma = NULL;
1268         struct page *page;
1269         unsigned long haddr = addr & HPAGE_PMD_MASK;
1270         int page_nid = -1, this_nid = numa_node_id();
1271         int target_nid, last_cpupid = -1;
1272         bool page_locked;
1273         bool migrated = false;
1274         int flags = 0;
1275
1276         ptl = pmd_lock(mm, pmdp);
1277         if (unlikely(!pmd_same(pmd, *pmdp)))
1278                 goto out_unlock;
1279
1280         /*
1281          * If there are potential migrations, wait for completion and retry
1282          * without disrupting NUMA hinting information. Do not relock and
1283          * check_same as the page may no longer be mapped.
1284          */
1285         if (unlikely(pmd_trans_migrating(*pmdp))) {
1286                 spin_unlock(ptl);
1287                 wait_migrate_huge_page(vma->anon_vma, pmdp);
1288                 goto out;
1289         }
1290
1291         page = pmd_page(pmd);
1292         BUG_ON(is_huge_zero_page(page));
1293         page_nid = page_to_nid(page);
1294         last_cpupid = page_cpupid_last(page);
1295         count_vm_numa_event(NUMA_HINT_FAULTS);
1296         if (page_nid == this_nid) {
1297                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1298                 flags |= TNF_FAULT_LOCAL;
1299         }
1300
1301         /*
1302          * Avoid grouping on DSO/COW pages in specific and RO pages
1303          * in general, RO pages shouldn't hurt as much anyway since
1304          * they can be in shared cache state.
1305          */
1306         if (!pmd_write(pmd))
1307                 flags |= TNF_NO_GROUP;
1308
1309         /*
1310          * Acquire the page lock to serialise THP migrations but avoid dropping
1311          * page_table_lock if at all possible
1312          */
1313         page_locked = trylock_page(page);
1314         target_nid = mpol_misplaced(page, vma, haddr);
1315         if (target_nid == -1) {
1316                 /* If the page was locked, there are no parallel migrations */
1317                 if (page_locked)
1318                         goto clear_pmdnuma;
1319         }
1320
1321         /* Migration could have started since the pmd_trans_migrating check */
1322         if (!page_locked) {
1323                 spin_unlock(ptl);
1324                 wait_on_page_locked(page);
1325                 page_nid = -1;
1326                 goto out;
1327         }
1328
1329         /*
1330          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1331          * to serialises splits
1332          */
1333         get_page(page);
1334         spin_unlock(ptl);
1335         anon_vma = page_lock_anon_vma_read(page);
1336
1337         /* Confirm the PMD did not change while page_table_lock was released */
1338         spin_lock(ptl);
1339         if (unlikely(!pmd_same(pmd, *pmdp))) {
1340                 unlock_page(page);
1341                 put_page(page);
1342                 page_nid = -1;
1343                 goto out_unlock;
1344         }
1345
1346         /* Bail if we fail to protect against THP splits for any reason */
1347         if (unlikely(!anon_vma)) {
1348                 put_page(page);
1349                 page_nid = -1;
1350                 goto clear_pmdnuma;
1351         }
1352
1353         /*
1354          * Migrate the THP to the requested node, returns with page unlocked
1355          * and pmd_numa cleared.
1356          */
1357         spin_unlock(ptl);
1358         migrated = migrate_misplaced_transhuge_page(mm, vma,
1359                                 pmdp, pmd, addr, page, target_nid);
1360         if (migrated) {
1361                 flags |= TNF_MIGRATED;
1362                 page_nid = target_nid;
1363         }
1364
1365         goto out;
1366 clear_pmdnuma:
1367         BUG_ON(!PageLocked(page));
1368         pmd = pmd_mknonnuma(pmd);
1369         set_pmd_at(mm, haddr, pmdp, pmd);
1370         VM_BUG_ON(pmd_numa(*pmdp));
1371         update_mmu_cache_pmd(vma, addr, pmdp);
1372         unlock_page(page);
1373 out_unlock:
1374         spin_unlock(ptl);
1375
1376 out:
1377         if (anon_vma)
1378                 page_unlock_anon_vma_read(anon_vma);
1379
1380         if (page_nid != -1)
1381                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1382
1383         return 0;
1384 }
1385
1386 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1387                  pmd_t *pmd, unsigned long addr)
1388 {
1389         spinlock_t *ptl;
1390         int ret = 0;
1391
1392         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1393                 struct page *page;
1394                 pgtable_t pgtable;
1395                 pmd_t orig_pmd;
1396                 /*
1397                  * For architectures like ppc64 we look at deposited pgtable
1398                  * when calling pmdp_get_and_clear. So do the
1399                  * pgtable_trans_huge_withdraw after finishing pmdp related
1400                  * operations.
1401                  */
1402                 orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd,
1403                                                    tlb->fullmm);
1404                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1405                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1406                 if (is_huge_zero_pmd(orig_pmd)) {
1407                         atomic_long_dec(&tlb->mm->nr_ptes);
1408                         spin_unlock(ptl);
1409                         put_huge_zero_page();
1410                 } else {
1411                         page = pmd_page(orig_pmd);
1412                         page_remove_rmap(page);
1413                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1414                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1415                         VM_BUG_ON_PAGE(!PageHead(page), page);
1416                         atomic_long_dec(&tlb->mm->nr_ptes);
1417                         spin_unlock(ptl);
1418                         tlb_remove_page(tlb, page);
1419                 }
1420                 pte_free(tlb->mm, pgtable);
1421                 ret = 1;
1422         }
1423         return ret;
1424 }
1425
1426 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1427                 unsigned long addr, unsigned long end,
1428                 unsigned char *vec)
1429 {
1430         spinlock_t *ptl;
1431         int ret = 0;
1432
1433         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1434                 /*
1435                  * All logical pages in the range are present
1436                  * if backed by a huge page.
1437                  */
1438                 spin_unlock(ptl);
1439                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1440                 ret = 1;
1441         }
1442
1443         return ret;
1444 }
1445
1446 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1447                   unsigned long old_addr,
1448                   unsigned long new_addr, unsigned long old_end,
1449                   pmd_t *old_pmd, pmd_t *new_pmd)
1450 {
1451         spinlock_t *old_ptl, *new_ptl;
1452         int ret = 0;
1453         pmd_t pmd;
1454
1455         struct mm_struct *mm = vma->vm_mm;
1456
1457         if ((old_addr & ~HPAGE_PMD_MASK) ||
1458             (new_addr & ~HPAGE_PMD_MASK) ||
1459             old_end - old_addr < HPAGE_PMD_SIZE ||
1460             (new_vma->vm_flags & VM_NOHUGEPAGE))
1461                 goto out;
1462
1463         /*
1464          * The destination pmd shouldn't be established, free_pgtables()
1465          * should have release it.
1466          */
1467         if (WARN_ON(!pmd_none(*new_pmd))) {
1468                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1469                 goto out;
1470         }
1471
1472         /*
1473          * We don't have to worry about the ordering of src and dst
1474          * ptlocks because exclusive mmap_sem prevents deadlock.
1475          */
1476         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1477         if (ret == 1) {
1478                 new_ptl = pmd_lockptr(mm, new_pmd);
1479                 if (new_ptl != old_ptl)
1480                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1481                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1482                 VM_BUG_ON(!pmd_none(*new_pmd));
1483
1484                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1485                         pgtable_t pgtable;
1486                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1487                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1488                 }
1489                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1490                 if (new_ptl != old_ptl)
1491                         spin_unlock(new_ptl);
1492                 spin_unlock(old_ptl);
1493         }
1494 out:
1495         return ret;
1496 }
1497
1498 /*
1499  * Returns
1500  *  - 0 if PMD could not be locked
1501  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1502  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1503  */
1504 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1505                 unsigned long addr, pgprot_t newprot, int prot_numa)
1506 {
1507         struct mm_struct *mm = vma->vm_mm;
1508         spinlock_t *ptl;
1509         int ret = 0;
1510
1511         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1512                 pmd_t entry;
1513                 ret = 1;
1514                 if (!prot_numa) {
1515                         entry = pmdp_get_and_clear_notify(mm, addr, pmd);
1516                         if (pmd_numa(entry))
1517                                 entry = pmd_mknonnuma(entry);
1518                         entry = pmd_modify(entry, newprot);
1519                         ret = HPAGE_PMD_NR;
1520                         set_pmd_at(mm, addr, pmd, entry);
1521                         BUG_ON(pmd_write(entry));
1522                 } else {
1523                         struct page *page = pmd_page(*pmd);
1524
1525                         /*
1526                          * Do not trap faults against the zero page. The
1527                          * read-only data is likely to be read-cached on the
1528                          * local CPU cache and it is less useful to know about
1529                          * local vs remote hits on the zero page.
1530                          */
1531                         if (!is_huge_zero_page(page) &&
1532                             !pmd_numa(*pmd)) {
1533                                 pmdp_set_numa(mm, addr, pmd);
1534                                 ret = HPAGE_PMD_NR;
1535                         }
1536                 }
1537                 spin_unlock(ptl);
1538         }
1539
1540         return ret;
1541 }
1542
1543 /*
1544  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1545  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1546  *
1547  * Note that if it returns 1, this routine returns without unlocking page
1548  * table locks. So callers must unlock them.
1549  */
1550 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1551                 spinlock_t **ptl)
1552 {
1553         *ptl = pmd_lock(vma->vm_mm, pmd);
1554         if (likely(pmd_trans_huge(*pmd))) {
1555                 if (unlikely(pmd_trans_splitting(*pmd))) {
1556                         spin_unlock(*ptl);
1557                         wait_split_huge_page(vma->anon_vma, pmd);
1558                         return -1;
1559                 } else {
1560                         /* Thp mapped by 'pmd' is stable, so we can
1561                          * handle it as it is. */
1562                         return 1;
1563                 }
1564         }
1565         spin_unlock(*ptl);
1566         return 0;
1567 }
1568
1569 /*
1570  * This function returns whether a given @page is mapped onto the @address
1571  * in the virtual space of @mm.
1572  *
1573  * When it's true, this function returns *pmd with holding the page table lock
1574  * and passing it back to the caller via @ptl.
1575  * If it's false, returns NULL without holding the page table lock.
1576  */
1577 pmd_t *page_check_address_pmd(struct page *page,
1578                               struct mm_struct *mm,
1579                               unsigned long address,
1580                               enum page_check_address_pmd_flag flag,
1581                               spinlock_t **ptl)
1582 {
1583         pgd_t *pgd;
1584         pud_t *pud;
1585         pmd_t *pmd;
1586
1587         if (address & ~HPAGE_PMD_MASK)
1588                 return NULL;
1589
1590         pgd = pgd_offset(mm, address);
1591         if (!pgd_present(*pgd))
1592                 return NULL;
1593         pud = pud_offset(pgd, address);
1594         if (!pud_present(*pud))
1595                 return NULL;
1596         pmd = pmd_offset(pud, address);
1597
1598         *ptl = pmd_lock(mm, pmd);
1599         if (!pmd_present(*pmd))
1600                 goto unlock;
1601         if (pmd_page(*pmd) != page)
1602                 goto unlock;
1603         /*
1604          * split_vma() may create temporary aliased mappings. There is
1605          * no risk as long as all huge pmd are found and have their
1606          * splitting bit set before __split_huge_page_refcount
1607          * runs. Finding the same huge pmd more than once during the
1608          * same rmap walk is not a problem.
1609          */
1610         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1611             pmd_trans_splitting(*pmd))
1612                 goto unlock;
1613         if (pmd_trans_huge(*pmd)) {
1614                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1615                           !pmd_trans_splitting(*pmd));
1616                 return pmd;
1617         }
1618 unlock:
1619         spin_unlock(*ptl);
1620         return NULL;
1621 }
1622
1623 static int __split_huge_page_splitting(struct page *page,
1624                                        struct vm_area_struct *vma,
1625                                        unsigned long address)
1626 {
1627         struct mm_struct *mm = vma->vm_mm;
1628         spinlock_t *ptl;
1629         pmd_t *pmd;
1630         int ret = 0;
1631         /* For mmu_notifiers */
1632         const unsigned long mmun_start = address;
1633         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1634
1635         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1636         pmd = page_check_address_pmd(page, mm, address,
1637                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1638         if (pmd) {
1639                 /*
1640                  * We can't temporarily set the pmd to null in order
1641                  * to split it, the pmd must remain marked huge at all
1642                  * times or the VM won't take the pmd_trans_huge paths
1643                  * and it won't wait on the anon_vma->root->rwsem to
1644                  * serialize against split_huge_page*.
1645                  */
1646                 pmdp_splitting_flush(vma, address, pmd);
1647
1648                 ret = 1;
1649                 spin_unlock(ptl);
1650         }
1651         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1652
1653         return ret;
1654 }
1655
1656 static void __split_huge_page_refcount(struct page *page,
1657                                        struct list_head *list)
1658 {
1659         int i;
1660         struct zone *zone = page_zone(page);
1661         struct lruvec *lruvec;
1662         int tail_count = 0;
1663
1664         /* prevent PageLRU to go away from under us, and freeze lru stats */
1665         spin_lock_irq(&zone->lru_lock);
1666         lruvec = mem_cgroup_page_lruvec(page, zone);
1667
1668         compound_lock(page);
1669         /* complete memcg works before add pages to LRU */
1670         mem_cgroup_split_huge_fixup(page);
1671
1672         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1673                 struct page *page_tail = page + i;
1674
1675                 /* tail_page->_mapcount cannot change */
1676                 BUG_ON(page_mapcount(page_tail) < 0);
1677                 tail_count += page_mapcount(page_tail);
1678                 /* check for overflow */
1679                 BUG_ON(tail_count < 0);
1680                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1681                 /*
1682                  * tail_page->_count is zero and not changing from
1683                  * under us. But get_page_unless_zero() may be running
1684                  * from under us on the tail_page. If we used
1685                  * atomic_set() below instead of atomic_add(), we
1686                  * would then run atomic_set() concurrently with
1687                  * get_page_unless_zero(), and atomic_set() is
1688                  * implemented in C not using locked ops. spin_unlock
1689                  * on x86 sometime uses locked ops because of PPro
1690                  * errata 66, 92, so unless somebody can guarantee
1691                  * atomic_set() here would be safe on all archs (and
1692                  * not only on x86), it's safer to use atomic_add().
1693                  */
1694                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1695                            &page_tail->_count);
1696
1697                 /* after clearing PageTail the gup refcount can be released */
1698                 smp_mb__after_atomic();
1699
1700                 /*
1701                  * retain hwpoison flag of the poisoned tail page:
1702                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1703                  *   by the memory-failure.
1704                  */
1705                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1706                 page_tail->flags |= (page->flags &
1707                                      ((1L << PG_referenced) |
1708                                       (1L << PG_swapbacked) |
1709                                       (1L << PG_mlocked) |
1710                                       (1L << PG_uptodate) |
1711                                       (1L << PG_active) |
1712                                       (1L << PG_unevictable)));
1713                 page_tail->flags |= (1L << PG_dirty);
1714
1715                 /* clear PageTail before overwriting first_page */
1716                 smp_wmb();
1717
1718                 /*
1719                  * __split_huge_page_splitting() already set the
1720                  * splitting bit in all pmd that could map this
1721                  * hugepage, that will ensure no CPU can alter the
1722                  * mapcount on the head page. The mapcount is only
1723                  * accounted in the head page and it has to be
1724                  * transferred to all tail pages in the below code. So
1725                  * for this code to be safe, the split the mapcount
1726                  * can't change. But that doesn't mean userland can't
1727                  * keep changing and reading the page contents while
1728                  * we transfer the mapcount, so the pmd splitting
1729                  * status is achieved setting a reserved bit in the
1730                  * pmd, not by clearing the present bit.
1731                 */
1732                 page_tail->_mapcount = page->_mapcount;
1733
1734                 BUG_ON(page_tail->mapping);
1735                 page_tail->mapping = page->mapping;
1736
1737                 page_tail->index = page->index + i;
1738                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1739
1740                 BUG_ON(!PageAnon(page_tail));
1741                 BUG_ON(!PageUptodate(page_tail));
1742                 BUG_ON(!PageDirty(page_tail));
1743                 BUG_ON(!PageSwapBacked(page_tail));
1744
1745                 lru_add_page_tail(page, page_tail, lruvec, list);
1746         }
1747         atomic_sub(tail_count, &page->_count);
1748         BUG_ON(atomic_read(&page->_count) <= 0);
1749
1750         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1751
1752         ClearPageCompound(page);
1753         compound_unlock(page);
1754         spin_unlock_irq(&zone->lru_lock);
1755
1756         for (i = 1; i < HPAGE_PMD_NR; i++) {
1757                 struct page *page_tail = page + i;
1758                 BUG_ON(page_count(page_tail) <= 0);
1759                 /*
1760                  * Tail pages may be freed if there wasn't any mapping
1761                  * like if add_to_swap() is running on a lru page that
1762                  * had its mapping zapped. And freeing these pages
1763                  * requires taking the lru_lock so we do the put_page
1764                  * of the tail pages after the split is complete.
1765                  */
1766                 put_page(page_tail);
1767         }
1768
1769         /*
1770          * Only the head page (now become a regular page) is required
1771          * to be pinned by the caller.
1772          */
1773         BUG_ON(page_count(page) <= 0);
1774 }
1775
1776 static int __split_huge_page_map(struct page *page,
1777                                  struct vm_area_struct *vma,
1778                                  unsigned long address)
1779 {
1780         struct mm_struct *mm = vma->vm_mm;
1781         spinlock_t *ptl;
1782         pmd_t *pmd, _pmd;
1783         int ret = 0, i;
1784         pgtable_t pgtable;
1785         unsigned long haddr;
1786
1787         pmd = page_check_address_pmd(page, mm, address,
1788                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1789         if (pmd) {
1790                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1791                 pmd_populate(mm, &_pmd, pgtable);
1792                 if (pmd_write(*pmd))
1793                         BUG_ON(page_mapcount(page) != 1);
1794
1795                 haddr = address;
1796                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1797                         pte_t *pte, entry;
1798                         BUG_ON(PageCompound(page+i));
1799                         /*
1800                          * Note that pmd_numa is not transferred deliberately
1801                          * to avoid any possibility that pte_numa leaks to
1802                          * a PROT_NONE VMA by accident.
1803                          */
1804                         entry = mk_pte(page + i, vma->vm_page_prot);
1805                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1806                         if (!pmd_write(*pmd))
1807                                 entry = pte_wrprotect(entry);
1808                         if (!pmd_young(*pmd))
1809                                 entry = pte_mkold(entry);
1810                         pte = pte_offset_map(&_pmd, haddr);
1811                         BUG_ON(!pte_none(*pte));
1812                         set_pte_at(mm, haddr, pte, entry);
1813                         pte_unmap(pte);
1814                 }
1815
1816                 smp_wmb(); /* make pte visible before pmd */
1817                 /*
1818                  * Up to this point the pmd is present and huge and
1819                  * userland has the whole access to the hugepage
1820                  * during the split (which happens in place). If we
1821                  * overwrite the pmd with the not-huge version
1822                  * pointing to the pte here (which of course we could
1823                  * if all CPUs were bug free), userland could trigger
1824                  * a small page size TLB miss on the small sized TLB
1825                  * while the hugepage TLB entry is still established
1826                  * in the huge TLB. Some CPU doesn't like that. See
1827                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1828                  * Erratum 383 on page 93. Intel should be safe but is
1829                  * also warns that it's only safe if the permission
1830                  * and cache attributes of the two entries loaded in
1831                  * the two TLB is identical (which should be the case
1832                  * here). But it is generally safer to never allow
1833                  * small and huge TLB entries for the same virtual
1834                  * address to be loaded simultaneously. So instead of
1835                  * doing "pmd_populate(); flush_tlb_range();" we first
1836                  * mark the current pmd notpresent (atomically because
1837                  * here the pmd_trans_huge and pmd_trans_splitting
1838                  * must remain set at all times on the pmd until the
1839                  * split is complete for this pmd), then we flush the
1840                  * SMP TLB and finally we write the non-huge version
1841                  * of the pmd entry with pmd_populate.
1842                  */
1843                 pmdp_invalidate(vma, address, pmd);
1844                 pmd_populate(mm, pmd, pgtable);
1845                 ret = 1;
1846                 spin_unlock(ptl);
1847         }
1848
1849         return ret;
1850 }
1851
1852 /* must be called with anon_vma->root->rwsem held */
1853 static void __split_huge_page(struct page *page,
1854                               struct anon_vma *anon_vma,
1855                               struct list_head *list)
1856 {
1857         int mapcount, mapcount2;
1858         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1859         struct anon_vma_chain *avc;
1860
1861         BUG_ON(!PageHead(page));
1862         BUG_ON(PageTail(page));
1863
1864         mapcount = 0;
1865         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1866                 struct vm_area_struct *vma = avc->vma;
1867                 unsigned long addr = vma_address(page, vma);
1868                 BUG_ON(is_vma_temporary_stack(vma));
1869                 mapcount += __split_huge_page_splitting(page, vma, addr);
1870         }
1871         /*
1872          * It is critical that new vmas are added to the tail of the
1873          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1874          * and establishes a child pmd before
1875          * __split_huge_page_splitting() freezes the parent pmd (so if
1876          * we fail to prevent copy_huge_pmd() from running until the
1877          * whole __split_huge_page() is complete), we will still see
1878          * the newly established pmd of the child later during the
1879          * walk, to be able to set it as pmd_trans_splitting too.
1880          */
1881         if (mapcount != page_mapcount(page)) {
1882                 pr_err("mapcount %d page_mapcount %d\n",
1883                         mapcount, page_mapcount(page));
1884                 BUG();
1885         }
1886
1887         __split_huge_page_refcount(page, list);
1888
1889         mapcount2 = 0;
1890         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1891                 struct vm_area_struct *vma = avc->vma;
1892                 unsigned long addr = vma_address(page, vma);
1893                 BUG_ON(is_vma_temporary_stack(vma));
1894                 mapcount2 += __split_huge_page_map(page, vma, addr);
1895         }
1896         if (mapcount != mapcount2) {
1897                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1898                         mapcount, mapcount2, page_mapcount(page));
1899                 BUG();
1900         }
1901 }
1902
1903 /*
1904  * Split a hugepage into normal pages. This doesn't change the position of head
1905  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1906  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1907  * from the hugepage.
1908  * Return 0 if the hugepage is split successfully otherwise return 1.
1909  */
1910 int split_huge_page_to_list(struct page *page, struct list_head *list)
1911 {
1912         struct anon_vma *anon_vma;
1913         int ret = 1;
1914
1915         BUG_ON(is_huge_zero_page(page));
1916         BUG_ON(!PageAnon(page));
1917
1918         /*
1919          * The caller does not necessarily hold an mmap_sem that would prevent
1920          * the anon_vma disappearing so we first we take a reference to it
1921          * and then lock the anon_vma for write. This is similar to
1922          * page_lock_anon_vma_read except the write lock is taken to serialise
1923          * against parallel split or collapse operations.
1924          */
1925         anon_vma = page_get_anon_vma(page);
1926         if (!anon_vma)
1927                 goto out;
1928         anon_vma_lock_write(anon_vma);
1929
1930         ret = 0;
1931         if (!PageCompound(page))
1932                 goto out_unlock;
1933
1934         BUG_ON(!PageSwapBacked(page));
1935         __split_huge_page(page, anon_vma, list);
1936         count_vm_event(THP_SPLIT);
1937
1938         BUG_ON(PageCompound(page));
1939 out_unlock:
1940         anon_vma_unlock_write(anon_vma);
1941         put_anon_vma(anon_vma);
1942 out:
1943         return ret;
1944 }
1945
1946 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1947
1948 int hugepage_madvise(struct vm_area_struct *vma,
1949                      unsigned long *vm_flags, int advice)
1950 {
1951         switch (advice) {
1952         case MADV_HUGEPAGE:
1953 #ifdef CONFIG_S390
1954                 /*
1955                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1956                  * can't handle this properly after s390_enable_sie, so we simply
1957                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1958                  */
1959                 if (mm_has_pgste(vma->vm_mm))
1960                         return 0;
1961 #endif
1962                 /*
1963                  * Be somewhat over-protective like KSM for now!
1964                  */
1965                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1966                         return -EINVAL;
1967                 *vm_flags &= ~VM_NOHUGEPAGE;
1968                 *vm_flags |= VM_HUGEPAGE;
1969                 /*
1970                  * If the vma become good for khugepaged to scan,
1971                  * register it here without waiting a page fault that
1972                  * may not happen any time soon.
1973                  */
1974                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1975                         return -ENOMEM;
1976                 break;
1977         case MADV_NOHUGEPAGE:
1978                 /*
1979                  * Be somewhat over-protective like KSM for now!
1980                  */
1981                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1982                         return -EINVAL;
1983                 *vm_flags &= ~VM_HUGEPAGE;
1984                 *vm_flags |= VM_NOHUGEPAGE;
1985                 /*
1986                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1987                  * this vma even if we leave the mm registered in khugepaged if
1988                  * it got registered before VM_NOHUGEPAGE was set.
1989                  */
1990                 break;
1991         }
1992
1993         return 0;
1994 }
1995
1996 static int __init khugepaged_slab_init(void)
1997 {
1998         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1999                                           sizeof(struct mm_slot),
2000                                           __alignof__(struct mm_slot), 0, NULL);
2001         if (!mm_slot_cache)
2002                 return -ENOMEM;
2003
2004         return 0;
2005 }
2006
2007 static inline struct mm_slot *alloc_mm_slot(void)
2008 {
2009         if (!mm_slot_cache)     /* initialization failed */
2010                 return NULL;
2011         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2012 }
2013
2014 static inline void free_mm_slot(struct mm_slot *mm_slot)
2015 {
2016         kmem_cache_free(mm_slot_cache, mm_slot);
2017 }
2018
2019 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2020 {
2021         struct mm_slot *mm_slot;
2022
2023         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2024                 if (mm == mm_slot->mm)
2025                         return mm_slot;
2026
2027         return NULL;
2028 }
2029
2030 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2031                                     struct mm_slot *mm_slot)
2032 {
2033         mm_slot->mm = mm;
2034         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2035 }
2036
2037 static inline int khugepaged_test_exit(struct mm_struct *mm)
2038 {
2039         return atomic_read(&mm->mm_users) == 0;
2040 }
2041
2042 int __khugepaged_enter(struct mm_struct *mm)
2043 {
2044         struct mm_slot *mm_slot;
2045         int wakeup;
2046
2047         mm_slot = alloc_mm_slot();
2048         if (!mm_slot)
2049                 return -ENOMEM;
2050
2051         /* __khugepaged_exit() must not run from under us */
2052         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2053         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2054                 free_mm_slot(mm_slot);
2055                 return 0;
2056         }
2057
2058         spin_lock(&khugepaged_mm_lock);
2059         insert_to_mm_slots_hash(mm, mm_slot);
2060         /*
2061          * Insert just behind the scanning cursor, to let the area settle
2062          * down a little.
2063          */
2064         wakeup = list_empty(&khugepaged_scan.mm_head);
2065         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2066         spin_unlock(&khugepaged_mm_lock);
2067
2068         atomic_inc(&mm->mm_count);
2069         if (wakeup)
2070                 wake_up_interruptible(&khugepaged_wait);
2071
2072         return 0;
2073 }
2074
2075 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2076                                unsigned long vm_flags)
2077 {
2078         unsigned long hstart, hend;
2079         if (!vma->anon_vma)
2080                 /*
2081                  * Not yet faulted in so we will register later in the
2082                  * page fault if needed.
2083                  */
2084                 return 0;
2085         if (vma->vm_ops)
2086                 /* khugepaged not yet working on file or special mappings */
2087                 return 0;
2088         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2089         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2090         hend = vma->vm_end & HPAGE_PMD_MASK;
2091         if (hstart < hend)
2092                 return khugepaged_enter(vma, vm_flags);
2093         return 0;
2094 }
2095
2096 void __khugepaged_exit(struct mm_struct *mm)
2097 {
2098         struct mm_slot *mm_slot;
2099         int free = 0;
2100
2101         spin_lock(&khugepaged_mm_lock);
2102         mm_slot = get_mm_slot(mm);
2103         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2104                 hash_del(&mm_slot->hash);
2105                 list_del(&mm_slot->mm_node);
2106                 free = 1;
2107         }
2108         spin_unlock(&khugepaged_mm_lock);
2109
2110         if (free) {
2111                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2112                 free_mm_slot(mm_slot);
2113                 mmdrop(mm);
2114         } else if (mm_slot) {
2115                 /*
2116                  * This is required to serialize against
2117                  * khugepaged_test_exit() (which is guaranteed to run
2118                  * under mmap sem read mode). Stop here (after we
2119                  * return all pagetables will be destroyed) until
2120                  * khugepaged has finished working on the pagetables
2121                  * under the mmap_sem.
2122                  */
2123                 down_write(&mm->mmap_sem);
2124                 up_write(&mm->mmap_sem);
2125         }
2126 }
2127
2128 static void release_pte_page(struct page *page)
2129 {
2130         /* 0 stands for page_is_file_cache(page) == false */
2131         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2132         unlock_page(page);
2133         putback_lru_page(page);
2134 }
2135
2136 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2137 {
2138         while (--_pte >= pte) {
2139                 pte_t pteval = *_pte;
2140                 if (!pte_none(pteval))
2141                         release_pte_page(pte_page(pteval));
2142         }
2143 }
2144
2145 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2146                                         unsigned long address,
2147                                         pte_t *pte)
2148 {
2149         struct page *page;
2150         pte_t *_pte;
2151         int referenced = 0, none = 0;
2152         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2153              _pte++, address += PAGE_SIZE) {
2154                 pte_t pteval = *_pte;
2155                 if (pte_none(pteval)) {
2156                         if (++none <= khugepaged_max_ptes_none)
2157                                 continue;
2158                         else
2159                                 goto out;
2160                 }
2161                 if (!pte_present(pteval) || !pte_write(pteval))
2162                         goto out;
2163                 page = vm_normal_page(vma, address, pteval);
2164                 if (unlikely(!page))
2165                         goto out;
2166
2167                 VM_BUG_ON_PAGE(PageCompound(page), page);
2168                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2169                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2170
2171                 /* cannot use mapcount: can't collapse if there's a gup pin */
2172                 if (page_count(page) != 1)
2173                         goto out;
2174                 /*
2175                  * We can do it before isolate_lru_page because the
2176                  * page can't be freed from under us. NOTE: PG_lock
2177                  * is needed to serialize against split_huge_page
2178                  * when invoked from the VM.
2179                  */
2180                 if (!trylock_page(page))
2181                         goto out;
2182                 /*
2183                  * Isolate the page to avoid collapsing an hugepage
2184                  * currently in use by the VM.
2185                  */
2186                 if (isolate_lru_page(page)) {
2187                         unlock_page(page);
2188                         goto out;
2189                 }
2190                 /* 0 stands for page_is_file_cache(page) == false */
2191                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2192                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2193                 VM_BUG_ON_PAGE(PageLRU(page), page);
2194
2195                 /* If there is no mapped pte young don't collapse the page */
2196                 if (pte_young(pteval) || PageReferenced(page) ||
2197                     mmu_notifier_test_young(vma->vm_mm, address))
2198                         referenced = 1;
2199         }
2200         if (likely(referenced))
2201                 return 1;
2202 out:
2203         release_pte_pages(pte, _pte);
2204         return 0;
2205 }
2206
2207 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2208                                       struct vm_area_struct *vma,
2209                                       unsigned long address,
2210                                       spinlock_t *ptl)
2211 {
2212         pte_t *_pte;
2213         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2214                 pte_t pteval = *_pte;
2215                 struct page *src_page;
2216
2217                 if (pte_none(pteval)) {
2218                         clear_user_highpage(page, address);
2219                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2220                 } else {
2221                         src_page = pte_page(pteval);
2222                         copy_user_highpage(page, src_page, address, vma);
2223                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2224                         release_pte_page(src_page);
2225                         /*
2226                          * ptl mostly unnecessary, but preempt has to
2227                          * be disabled to update the per-cpu stats
2228                          * inside page_remove_rmap().
2229                          */
2230                         spin_lock(ptl);
2231                         /*
2232                          * paravirt calls inside pte_clear here are
2233                          * superfluous.
2234                          */
2235                         pte_clear(vma->vm_mm, address, _pte);
2236                         page_remove_rmap(src_page);
2237                         spin_unlock(ptl);
2238                         free_page_and_swap_cache(src_page);
2239                 }
2240
2241                 address += PAGE_SIZE;
2242                 page++;
2243         }
2244 }
2245
2246 static void khugepaged_alloc_sleep(void)
2247 {
2248         wait_event_freezable_timeout(khugepaged_wait, false,
2249                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2250 }
2251
2252 static int khugepaged_node_load[MAX_NUMNODES];
2253
2254 static bool khugepaged_scan_abort(int nid)
2255 {
2256         int i;
2257
2258         /*
2259          * If zone_reclaim_mode is disabled, then no extra effort is made to
2260          * allocate memory locally.
2261          */
2262         if (!zone_reclaim_mode)
2263                 return false;
2264
2265         /* If there is a count for this node already, it must be acceptable */
2266         if (khugepaged_node_load[nid])
2267                 return false;
2268
2269         for (i = 0; i < MAX_NUMNODES; i++) {
2270                 if (!khugepaged_node_load[i])
2271                         continue;
2272                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2273                         return true;
2274         }
2275         return false;
2276 }
2277
2278 #ifdef CONFIG_NUMA
2279 static int khugepaged_find_target_node(void)
2280 {
2281         static int last_khugepaged_target_node = NUMA_NO_NODE;
2282         int nid, target_node = 0, max_value = 0;
2283
2284         /* find first node with max normal pages hit */
2285         for (nid = 0; nid < MAX_NUMNODES; nid++)
2286                 if (khugepaged_node_load[nid] > max_value) {
2287                         max_value = khugepaged_node_load[nid];
2288                         target_node = nid;
2289                 }
2290
2291         /* do some balance if several nodes have the same hit record */
2292         if (target_node <= last_khugepaged_target_node)
2293                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2294                                 nid++)
2295                         if (max_value == khugepaged_node_load[nid]) {
2296                                 target_node = nid;
2297                                 break;
2298                         }
2299
2300         last_khugepaged_target_node = target_node;
2301         return target_node;
2302 }
2303
2304 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2305 {
2306         if (IS_ERR(*hpage)) {
2307                 if (!*wait)
2308                         return false;
2309
2310                 *wait = false;
2311                 *hpage = NULL;
2312                 khugepaged_alloc_sleep();
2313         } else if (*hpage) {
2314                 put_page(*hpage);
2315                 *hpage = NULL;
2316         }
2317
2318         return true;
2319 }
2320
2321 static struct page
2322 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2323                        struct vm_area_struct *vma, unsigned long address,
2324                        int node)
2325 {
2326         VM_BUG_ON_PAGE(*hpage, *hpage);
2327
2328         /*
2329          * Before allocating the hugepage, release the mmap_sem read lock.
2330          * The allocation can take potentially a long time if it involves
2331          * sync compaction, and we do not need to hold the mmap_sem during
2332          * that. We will recheck the vma after taking it again in write mode.
2333          */
2334         up_read(&mm->mmap_sem);
2335
2336         *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2337                 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2338         if (unlikely(!*hpage)) {
2339                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2340                 *hpage = ERR_PTR(-ENOMEM);
2341                 return NULL;
2342         }
2343
2344         count_vm_event(THP_COLLAPSE_ALLOC);
2345         return *hpage;
2346 }
2347 #else
2348 static int khugepaged_find_target_node(void)
2349 {
2350         return 0;
2351 }
2352
2353 static inline struct page *alloc_hugepage(int defrag)
2354 {
2355         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2356                            HPAGE_PMD_ORDER);
2357 }
2358
2359 static struct page *khugepaged_alloc_hugepage(bool *wait)
2360 {
2361         struct page *hpage;
2362
2363         do {
2364                 hpage = alloc_hugepage(khugepaged_defrag());
2365                 if (!hpage) {
2366                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2367                         if (!*wait)
2368                                 return NULL;
2369
2370                         *wait = false;
2371                         khugepaged_alloc_sleep();
2372                 } else
2373                         count_vm_event(THP_COLLAPSE_ALLOC);
2374         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2375
2376         return hpage;
2377 }
2378
2379 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2380 {
2381         if (!*hpage)
2382                 *hpage = khugepaged_alloc_hugepage(wait);
2383
2384         if (unlikely(!*hpage))
2385                 return false;
2386
2387         return true;
2388 }
2389
2390 static struct page
2391 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2392                        struct vm_area_struct *vma, unsigned long address,
2393                        int node)
2394 {
2395         up_read(&mm->mmap_sem);
2396         VM_BUG_ON(!*hpage);
2397         return  *hpage;
2398 }
2399 #endif
2400
2401 static bool hugepage_vma_check(struct vm_area_struct *vma)
2402 {
2403         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2404             (vma->vm_flags & VM_NOHUGEPAGE))
2405                 return false;
2406
2407         if (!vma->anon_vma || vma->vm_ops)
2408                 return false;
2409         if (is_vma_temporary_stack(vma))
2410                 return false;
2411         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2412         return true;
2413 }
2414
2415 static void collapse_huge_page(struct mm_struct *mm,
2416                                    unsigned long address,
2417                                    struct page **hpage,
2418                                    struct vm_area_struct *vma,
2419                                    int node)
2420 {
2421         pmd_t *pmd, _pmd;
2422         pte_t *pte;
2423         pgtable_t pgtable;
2424         struct page *new_page;
2425         spinlock_t *pmd_ptl, *pte_ptl;
2426         int isolated;
2427         unsigned long hstart, hend;
2428         struct mem_cgroup *memcg;
2429         unsigned long mmun_start;       /* For mmu_notifiers */
2430         unsigned long mmun_end;         /* For mmu_notifiers */
2431
2432         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2433
2434         /* release the mmap_sem read lock. */
2435         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2436         if (!new_page)
2437                 return;
2438
2439         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2440                                            GFP_TRANSHUGE, &memcg)))
2441                 return;
2442
2443         /*
2444          * Prevent all access to pagetables with the exception of
2445          * gup_fast later hanlded by the ptep_clear_flush and the VM
2446          * handled by the anon_vma lock + PG_lock.
2447          */
2448         down_write(&mm->mmap_sem);
2449         if (unlikely(khugepaged_test_exit(mm)))
2450                 goto out;
2451
2452         vma = find_vma(mm, address);
2453         if (!vma)
2454                 goto out;
2455         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2456         hend = vma->vm_end & HPAGE_PMD_MASK;
2457         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2458                 goto out;
2459         if (!hugepage_vma_check(vma))
2460                 goto out;
2461         pmd = mm_find_pmd(mm, address);
2462         if (!pmd)
2463                 goto out;
2464
2465         anon_vma_lock_write(vma->anon_vma);
2466
2467         pte = pte_offset_map(pmd, address);
2468         pte_ptl = pte_lockptr(mm, pmd);
2469
2470         mmun_start = address;
2471         mmun_end   = address + HPAGE_PMD_SIZE;
2472         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2473         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2474         /*
2475          * After this gup_fast can't run anymore. This also removes
2476          * any huge TLB entry from the CPU so we won't allow
2477          * huge and small TLB entries for the same virtual address
2478          * to avoid the risk of CPU bugs in that area.
2479          */
2480         _pmd = pmdp_clear_flush(vma, address, pmd);
2481         spin_unlock(pmd_ptl);
2482         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2483
2484         spin_lock(pte_ptl);
2485         isolated = __collapse_huge_page_isolate(vma, address, pte);
2486         spin_unlock(pte_ptl);
2487
2488         if (unlikely(!isolated)) {
2489                 pte_unmap(pte);
2490                 spin_lock(pmd_ptl);
2491                 BUG_ON(!pmd_none(*pmd));
2492                 /*
2493                  * We can only use set_pmd_at when establishing
2494                  * hugepmds and never for establishing regular pmds that
2495                  * points to regular pagetables. Use pmd_populate for that
2496                  */
2497                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2498                 spin_unlock(pmd_ptl);
2499                 anon_vma_unlock_write(vma->anon_vma);
2500                 goto out;
2501         }
2502
2503         /*
2504          * All pages are isolated and locked so anon_vma rmap
2505          * can't run anymore.
2506          */
2507         anon_vma_unlock_write(vma->anon_vma);
2508
2509         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2510         pte_unmap(pte);
2511         __SetPageUptodate(new_page);
2512         pgtable = pmd_pgtable(_pmd);
2513
2514         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2515         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2516
2517         /*
2518          * spin_lock() below is not the equivalent of smp_wmb(), so
2519          * this is needed to avoid the copy_huge_page writes to become
2520          * visible after the set_pmd_at() write.
2521          */
2522         smp_wmb();
2523
2524         spin_lock(pmd_ptl);
2525         BUG_ON(!pmd_none(*pmd));
2526         page_add_new_anon_rmap(new_page, vma, address);
2527         mem_cgroup_commit_charge(new_page, memcg, false);
2528         lru_cache_add_active_or_unevictable(new_page, vma);
2529         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2530         set_pmd_at(mm, address, pmd, _pmd);
2531         update_mmu_cache_pmd(vma, address, pmd);
2532         spin_unlock(pmd_ptl);
2533
2534         *hpage = NULL;
2535
2536         khugepaged_pages_collapsed++;
2537 out_up_write:
2538         up_write(&mm->mmap_sem);
2539         return;
2540
2541 out:
2542         mem_cgroup_cancel_charge(new_page, memcg);
2543         goto out_up_write;
2544 }
2545
2546 static int khugepaged_scan_pmd(struct mm_struct *mm,
2547                                struct vm_area_struct *vma,
2548                                unsigned long address,
2549                                struct page **hpage)
2550 {
2551         pmd_t *pmd;
2552         pte_t *pte, *_pte;
2553         int ret = 0, referenced = 0, none = 0;
2554         struct page *page;
2555         unsigned long _address;
2556         spinlock_t *ptl;
2557         int node = NUMA_NO_NODE;
2558
2559         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2560
2561         pmd = mm_find_pmd(mm, address);
2562         if (!pmd)
2563                 goto out;
2564
2565         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2566         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2567         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2568              _pte++, _address += PAGE_SIZE) {
2569                 pte_t pteval = *_pte;
2570                 if (pte_none(pteval)) {
2571                         if (++none <= khugepaged_max_ptes_none)
2572                                 continue;
2573                         else
2574                                 goto out_unmap;
2575                 }
2576                 if (!pte_present(pteval) || !pte_write(pteval))
2577                         goto out_unmap;
2578                 page = vm_normal_page(vma, _address, pteval);
2579                 if (unlikely(!page))
2580                         goto out_unmap;
2581                 /*
2582                  * Record which node the original page is from and save this
2583                  * information to khugepaged_node_load[].
2584                  * Khupaged will allocate hugepage from the node has the max
2585                  * hit record.
2586                  */
2587                 node = page_to_nid(page);
2588                 if (khugepaged_scan_abort(node))
2589                         goto out_unmap;
2590                 khugepaged_node_load[node]++;
2591                 VM_BUG_ON_PAGE(PageCompound(page), page);
2592                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2593                         goto out_unmap;
2594                 /* cannot use mapcount: can't collapse if there's a gup pin */
2595                 if (page_count(page) != 1)
2596                         goto out_unmap;
2597                 if (pte_young(pteval) || PageReferenced(page) ||
2598                     mmu_notifier_test_young(vma->vm_mm, address))
2599                         referenced = 1;
2600         }
2601         if (referenced)
2602                 ret = 1;
2603 out_unmap:
2604         pte_unmap_unlock(pte, ptl);
2605         if (ret) {
2606                 node = khugepaged_find_target_node();
2607                 /* collapse_huge_page will return with the mmap_sem released */
2608                 collapse_huge_page(mm, address, hpage, vma, node);
2609         }
2610 out:
2611         return ret;
2612 }
2613
2614 static void collect_mm_slot(struct mm_slot *mm_slot)
2615 {
2616         struct mm_struct *mm = mm_slot->mm;
2617
2618         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2619
2620         if (khugepaged_test_exit(mm)) {
2621                 /* free mm_slot */
2622                 hash_del(&mm_slot->hash);
2623                 list_del(&mm_slot->mm_node);
2624
2625                 /*
2626                  * Not strictly needed because the mm exited already.
2627                  *
2628                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2629                  */
2630
2631                 /* khugepaged_mm_lock actually not necessary for the below */
2632                 free_mm_slot(mm_slot);
2633                 mmdrop(mm);
2634         }
2635 }
2636
2637 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2638                                             struct page **hpage)
2639         __releases(&khugepaged_mm_lock)
2640         __acquires(&khugepaged_mm_lock)
2641 {
2642         struct mm_slot *mm_slot;
2643         struct mm_struct *mm;
2644         struct vm_area_struct *vma;
2645         int progress = 0;
2646
2647         VM_BUG_ON(!pages);
2648         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2649
2650         if (khugepaged_scan.mm_slot)
2651                 mm_slot = khugepaged_scan.mm_slot;
2652         else {
2653                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2654                                      struct mm_slot, mm_node);
2655                 khugepaged_scan.address = 0;
2656                 khugepaged_scan.mm_slot = mm_slot;
2657         }
2658         spin_unlock(&khugepaged_mm_lock);
2659
2660         mm = mm_slot->mm;
2661         down_read(&mm->mmap_sem);
2662         if (unlikely(khugepaged_test_exit(mm)))
2663                 vma = NULL;
2664         else
2665                 vma = find_vma(mm, khugepaged_scan.address);
2666
2667         progress++;
2668         for (; vma; vma = vma->vm_next) {
2669                 unsigned long hstart, hend;
2670
2671                 cond_resched();
2672                 if (unlikely(khugepaged_test_exit(mm))) {
2673                         progress++;
2674                         break;
2675                 }
2676                 if (!hugepage_vma_check(vma)) {
2677 skip:
2678                         progress++;
2679                         continue;
2680                 }
2681                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2682                 hend = vma->vm_end & HPAGE_PMD_MASK;
2683                 if (hstart >= hend)
2684                         goto skip;
2685                 if (khugepaged_scan.address > hend)
2686                         goto skip;
2687                 if (khugepaged_scan.address < hstart)
2688                         khugepaged_scan.address = hstart;
2689                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2690
2691                 while (khugepaged_scan.address < hend) {
2692                         int ret;
2693                         cond_resched();
2694                         if (unlikely(khugepaged_test_exit(mm)))
2695                                 goto breakouterloop;
2696
2697                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2698                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2699                                   hend);
2700                         ret = khugepaged_scan_pmd(mm, vma,
2701                                                   khugepaged_scan.address,
2702                                                   hpage);
2703                         /* move to next address */
2704                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2705                         progress += HPAGE_PMD_NR;
2706                         if (ret)
2707                                 /* we released mmap_sem so break loop */
2708                                 goto breakouterloop_mmap_sem;
2709                         if (progress >= pages)
2710                                 goto breakouterloop;
2711                 }
2712         }
2713 breakouterloop:
2714         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2715 breakouterloop_mmap_sem:
2716
2717         spin_lock(&khugepaged_mm_lock);
2718         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2719         /*
2720          * Release the current mm_slot if this mm is about to die, or
2721          * if we scanned all vmas of this mm.
2722          */
2723         if (khugepaged_test_exit(mm) || !vma) {
2724                 /*
2725                  * Make sure that if mm_users is reaching zero while
2726                  * khugepaged runs here, khugepaged_exit will find
2727                  * mm_slot not pointing to the exiting mm.
2728                  */
2729                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2730                         khugepaged_scan.mm_slot = list_entry(
2731                                 mm_slot->mm_node.next,
2732                                 struct mm_slot, mm_node);
2733                         khugepaged_scan.address = 0;
2734                 } else {
2735                         khugepaged_scan.mm_slot = NULL;
2736                         khugepaged_full_scans++;
2737                 }
2738
2739                 collect_mm_slot(mm_slot);
2740         }
2741
2742         return progress;
2743 }
2744
2745 static int khugepaged_has_work(void)
2746 {
2747         return !list_empty(&khugepaged_scan.mm_head) &&
2748                 khugepaged_enabled();
2749 }
2750
2751 static int khugepaged_wait_event(void)
2752 {
2753         return !list_empty(&khugepaged_scan.mm_head) ||
2754                 kthread_should_stop();
2755 }
2756
2757 static void khugepaged_do_scan(void)
2758 {
2759         struct page *hpage = NULL;
2760         unsigned int progress = 0, pass_through_head = 0;
2761         unsigned int pages = khugepaged_pages_to_scan;
2762         bool wait = true;
2763
2764         barrier(); /* write khugepaged_pages_to_scan to local stack */
2765
2766         while (progress < pages) {
2767                 if (!khugepaged_prealloc_page(&hpage, &wait))
2768                         break;
2769
2770                 cond_resched();
2771
2772                 if (unlikely(kthread_should_stop() || freezing(current)))
2773                         break;
2774
2775                 spin_lock(&khugepaged_mm_lock);
2776                 if (!khugepaged_scan.mm_slot)
2777                         pass_through_head++;
2778                 if (khugepaged_has_work() &&
2779                     pass_through_head < 2)
2780                         progress += khugepaged_scan_mm_slot(pages - progress,
2781                                                             &hpage);
2782                 else
2783                         progress = pages;
2784                 spin_unlock(&khugepaged_mm_lock);
2785         }
2786
2787         if (!IS_ERR_OR_NULL(hpage))
2788                 put_page(hpage);
2789 }
2790
2791 static void khugepaged_wait_work(void)
2792 {
2793         try_to_freeze();
2794
2795         if (khugepaged_has_work()) {
2796                 if (!khugepaged_scan_sleep_millisecs)
2797                         return;
2798
2799                 wait_event_freezable_timeout(khugepaged_wait,
2800                                              kthread_should_stop(),
2801                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2802                 return;
2803         }
2804
2805         if (khugepaged_enabled())
2806                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2807 }
2808
2809 static int khugepaged(void *none)
2810 {
2811         struct mm_slot *mm_slot;
2812
2813         set_freezable();
2814         set_user_nice(current, MAX_NICE);
2815
2816         while (!kthread_should_stop()) {
2817                 khugepaged_do_scan();
2818                 khugepaged_wait_work();
2819         }
2820
2821         spin_lock(&khugepaged_mm_lock);
2822         mm_slot = khugepaged_scan.mm_slot;
2823         khugepaged_scan.mm_slot = NULL;
2824         if (mm_slot)
2825                 collect_mm_slot(mm_slot);
2826         spin_unlock(&khugepaged_mm_lock);
2827         return 0;
2828 }
2829
2830 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2831                 unsigned long haddr, pmd_t *pmd)
2832 {
2833         struct mm_struct *mm = vma->vm_mm;
2834         pgtable_t pgtable;
2835         pmd_t _pmd;
2836         int i;
2837
2838         pmdp_clear_flush_notify(vma, haddr, pmd);
2839         /* leave pmd empty until pte is filled */
2840
2841         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2842         pmd_populate(mm, &_pmd, pgtable);
2843
2844         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2845                 pte_t *pte, entry;
2846                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2847                 entry = pte_mkspecial(entry);
2848                 pte = pte_offset_map(&_pmd, haddr);
2849                 VM_BUG_ON(!pte_none(*pte));
2850                 set_pte_at(mm, haddr, pte, entry);
2851                 pte_unmap(pte);
2852         }
2853         smp_wmb(); /* make pte visible before pmd */
2854         pmd_populate(mm, pmd, pgtable);
2855         put_huge_zero_page();
2856 }
2857
2858 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2859                 pmd_t *pmd)
2860 {
2861         spinlock_t *ptl;
2862         struct page *page;
2863         struct mm_struct *mm = vma->vm_mm;
2864         unsigned long haddr = address & HPAGE_PMD_MASK;
2865         unsigned long mmun_start;       /* For mmu_notifiers */
2866         unsigned long mmun_end;         /* For mmu_notifiers */
2867
2868         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2869
2870         mmun_start = haddr;
2871         mmun_end   = haddr + HPAGE_PMD_SIZE;
2872 again:
2873         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2874         ptl = pmd_lock(mm, pmd);
2875         if (unlikely(!pmd_trans_huge(*pmd))) {
2876                 spin_unlock(ptl);
2877                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2878                 return;
2879         }
2880         if (is_huge_zero_pmd(*pmd)) {
2881                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2882                 spin_unlock(ptl);
2883                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2884                 return;
2885         }
2886         page = pmd_page(*pmd);
2887         VM_BUG_ON_PAGE(!page_count(page), page);
2888         get_page(page);
2889         spin_unlock(ptl);
2890         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2891
2892         split_huge_page(page);
2893
2894         put_page(page);
2895
2896         /*
2897          * We don't always have down_write of mmap_sem here: a racing
2898          * do_huge_pmd_wp_page() might have copied-on-write to another
2899          * huge page before our split_huge_page() got the anon_vma lock.
2900          */
2901         if (unlikely(pmd_trans_huge(*pmd)))
2902                 goto again;
2903 }
2904
2905 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2906                 pmd_t *pmd)
2907 {
2908         struct vm_area_struct *vma;
2909
2910         vma = find_vma(mm, address);
2911         BUG_ON(vma == NULL);
2912         split_huge_page_pmd(vma, address, pmd);
2913 }
2914
2915 static void split_huge_page_address(struct mm_struct *mm,
2916                                     unsigned long address)
2917 {
2918         pgd_t *pgd;
2919         pud_t *pud;
2920         pmd_t *pmd;
2921
2922         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2923
2924         pgd = pgd_offset(mm, address);
2925         if (!pgd_present(*pgd))
2926                 return;
2927
2928         pud = pud_offset(pgd, address);
2929         if (!pud_present(*pud))
2930                 return;
2931
2932         pmd = pmd_offset(pud, address);
2933         if (!pmd_present(*pmd))
2934                 return;
2935         /*
2936          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2937          * materialize from under us.
2938          */
2939         split_huge_page_pmd_mm(mm, address, pmd);
2940 }
2941
2942 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2943                              unsigned long start,
2944                              unsigned long end,
2945                              long adjust_next)
2946 {
2947         /*
2948          * If the new start address isn't hpage aligned and it could
2949          * previously contain an hugepage: check if we need to split
2950          * an huge pmd.
2951          */
2952         if (start & ~HPAGE_PMD_MASK &&
2953             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2954             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2955                 split_huge_page_address(vma->vm_mm, start);
2956
2957         /*
2958          * If the new end address isn't hpage aligned and it could
2959          * previously contain an hugepage: check if we need to split
2960          * an huge pmd.
2961          */
2962         if (end & ~HPAGE_PMD_MASK &&
2963             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2964             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2965                 split_huge_page_address(vma->vm_mm, end);
2966
2967         /*
2968          * If we're also updating the vma->vm_next->vm_start, if the new
2969          * vm_next->vm_start isn't page aligned and it could previously
2970          * contain an hugepage: check if we need to split an huge pmd.
2971          */
2972         if (adjust_next > 0) {
2973                 struct vm_area_struct *next = vma->vm_next;
2974                 unsigned long nstart = next->vm_start;
2975                 nstart += adjust_next << PAGE_SHIFT;
2976                 if (nstart & ~HPAGE_PMD_MASK &&
2977                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2978                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2979                         split_huge_page_address(next->vm_mm, nstart);
2980         }
2981 }