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 static void khugepaged_slab_exit(void);
  71
  72 #define MM_SLOTS_HASH_BITS 10
  73 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  74
  75 static struct kmem_cache *mm_slot_cache __read_mostly;
  76
  77 /**
  78  * struct mm_slot - hash lookup from mm to mm_slot
  79  * @hash: hash collision list
  80  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  81  * @mm: the mm that this information is valid for
  82  */
  83 struct mm_slot {
  84         struct hlist_node hash;
  85         struct list_head mm_node;
  86         struct mm_struct *mm;
  87 };
  88
  89 /**
  90  * struct khugepaged_scan - cursor for scanning
  91  * @mm_head: the head of the mm list to scan
  92  * @mm_slot: the current mm_slot we are scanning
  93  * @address: the next address inside that to be scanned
  94  *
  95  * There is only the one khugepaged_scan instance of this cursor structure.
  96  */
  97 struct khugepaged_scan {
  98         struct list_head mm_head;
  99         struct mm_slot *mm_slot;
 100         unsigned long address;
 101 };
 102 static struct khugepaged_scan khugepaged_scan = {
 103         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 104 };
 105
 106
 107 static int set_recommended_min_free_kbytes(void)
 108 {
 109         struct zone *zone;
 110         int nr_zones = 0;
 111         unsigned long recommended_min;
 112
 113         /* khugepaged thread has stopped to failed to start */
 114         if (!khugepaged_thread)
 115                 return 0;
 116
 117         for_each_populated_zone(zone)
 118                 nr_zones++;
 119
 120         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 121         recommended_min = pageblock_nr_pages * nr_zones * 2;
 122
 123         /*
 124          * Make sure that on average at least two pageblocks are almost free
 125          * of another type, one for a migratetype to fall back to and a
 126          * second to avoid subsequent fallbacks of other types There are 3
 127          * MIGRATE_TYPES we care about.
 128          */
 129         recommended_min += pageblock_nr_pages * nr_zones *
 130                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 131
 132         /* don't ever allow to reserve more than 5% of the lowmem */
 133         recommended_min = min(recommended_min,
 134                               (unsigned long) nr_free_buffer_pages() / 20);
 135         recommended_min <<= (PAGE_SHIFT-10);
 136
 137         if (recommended_min > min_free_kbytes) {
 138                 if (user_min_free_kbytes >= 0)
 139                         pr_info("raising min_free_kbytes from %d to %lu "
 140                                 "to help transparent hugepage allocations\n",
 141                                 min_free_kbytes, recommended_min);
 142
 143                 min_free_kbytes = recommended_min;
 144         }
 145         setup_per_zone_wmarks();
 146         return 0;
 147 }
 148 late_initcall(set_recommended_min_free_kbytes);
 149
 150 static int start_khugepaged(void)
 151 {
 152         int err = 0;
 153         if (khugepaged_enabled()) {
 154                 if (!khugepaged_thread)
 155                         khugepaged_thread = kthread_run(khugepaged, NULL,
 156                                                         "khugepaged");
 157                 if (unlikely(IS_ERR(khugepaged_thread))) {
 158                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 159                         err = PTR_ERR(khugepaged_thread);
 160                         khugepaged_thread = NULL;
 161                 }
 162
 163                 if (!list_empty(&khugepaged_scan.mm_head))
 164                         wake_up_interruptible(&khugepaged_wait);
 165
 166                 set_recommended_min_free_kbytes();
 167         } else if (khugepaged_thread) {
 168                 kthread_stop(khugepaged_thread);
 169                 khugepaged_thread = NULL;
 170         }
 171
 172         return err;
 173 }
 174
 175 static atomic_t huge_zero_refcount;
 176 struct page *huge_zero_page __read_mostly;
 177
 178 static inline bool is_huge_zero_pmd(pmd_t pmd)
 179 {
 180         return is_huge_zero_page(pmd_page(pmd));
 181 }
 182
 183 static struct page *get_huge_zero_page(void)
 184 {
 185         struct page *zero_page;
 186 retry:
 187         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 188                 return READ_ONCE(huge_zero_page);
 189
 190         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 191                         HPAGE_PMD_ORDER);
 192         if (!zero_page) {
 193                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 194                 return NULL;
 195         }
 196         count_vm_event(THP_ZERO_PAGE_ALLOC);
 197         preempt_disable();
 198         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 199                 preempt_enable();
 200                 __free_pages(zero_page, compound_order(zero_page));
 201                 goto retry;
 202         }
 203
 204         /* We take additional reference here. It will be put back by shrinker */
 205         atomic_set(&huge_zero_refcount, 2);
 206         preempt_enable();
 207         return READ_ONCE(huge_zero_page);
 208 }
 209
 210 static void put_huge_zero_page(void)
 211 {
 212         /*
 213          * Counter should never go to zero here. Only shrinker can put
 214          * last reference.
 215          */
 216         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 217 }
 218
 219 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 220                                         struct shrink_control *sc)
 221 {
 222         /* we can free zero page only if last reference remains */
 223         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 224 }
 225
 226 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 227                                        struct shrink_control *sc)
 228 {
 229         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 230                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 231                 BUG_ON(zero_page == NULL);
 232                 __free_pages(zero_page, compound_order(zero_page));
 233                 return HPAGE_PMD_NR;
 234         }
 235
 236         return 0;
 237 }
 238
 239 static struct shrinker huge_zero_page_shrinker = {
 240         .count_objects = shrink_huge_zero_page_count,
 241         .scan_objects = shrink_huge_zero_page_scan,
 242         .seeks = DEFAULT_SEEKS,
 243 };
 244
 245 #ifdef CONFIG_SYSFS
 246
 247 static ssize_t double_flag_show(struct kobject *kobj,
 248                                 struct kobj_attribute *attr, char *buf,
 249                                 enum transparent_hugepage_flag enabled,
 250                                 enum transparent_hugepage_flag req_madv)
 251 {
 252         if (test_bit(enabled, &transparent_hugepage_flags)) {
 253                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 254                 return sprintf(buf, "[always] madvise never\n");
 255         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 256                 return sprintf(buf, "always [madvise] never\n");
 257         else
 258                 return sprintf(buf, "always madvise [never]\n");
 259 }
 260 static ssize_t double_flag_store(struct kobject *kobj,
 261                                  struct kobj_attribute *attr,
 262                                  const char *buf, size_t count,
 263                                  enum transparent_hugepage_flag enabled,
 264                                  enum transparent_hugepage_flag req_madv)
 265 {
 266         if (!memcmp("always", buf,
 267                     min(sizeof("always")-1, count))) {
 268                 set_bit(enabled, &transparent_hugepage_flags);
 269                 clear_bit(req_madv, &transparent_hugepage_flags);
 270         } else if (!memcmp("madvise", buf,
 271                            min(sizeof("madvise")-1, count))) {
 272                 clear_bit(enabled, &transparent_hugepage_flags);
 273                 set_bit(req_madv, &transparent_hugepage_flags);
 274         } else if (!memcmp("never", buf,
 275                            min(sizeof("never")-1, count))) {
 276                 clear_bit(enabled, &transparent_hugepage_flags);
 277                 clear_bit(req_madv, &transparent_hugepage_flags);
 278         } else
 279                 return -EINVAL;
 280
 281         return count;
 282 }
 283
 284 static ssize_t enabled_show(struct kobject *kobj,
 285                             struct kobj_attribute *attr, char *buf)
 286 {
 287         return double_flag_show(kobj, attr, buf,
 288                                 TRANSPARENT_HUGEPAGE_FLAG,
 289                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 290 }
 291 static ssize_t enabled_store(struct kobject *kobj,
 292                              struct kobj_attribute *attr,
 293                              const char *buf, size_t count)
 294 {
 295         ssize_t ret;
 296
 297         ret = double_flag_store(kobj, attr, buf, count,
 298                                 TRANSPARENT_HUGEPAGE_FLAG,
 299                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 300
 301         if (ret > 0) {
 302                 int err;
 303
 304                 mutex_lock(&khugepaged_mutex);
 305                 err = start_khugepaged();
 306                 mutex_unlock(&khugepaged_mutex);
 307
 308                 if (err)
 309                         ret = err;
 310         }
 311
 312         return ret;
 313 }
 314 static struct kobj_attribute enabled_attr =
 315         __ATTR(enabled, 0644, enabled_show, enabled_store);
 316
 317 static ssize_t single_flag_show(struct kobject *kobj,
 318                                 struct kobj_attribute *attr, char *buf,
 319                                 enum transparent_hugepage_flag flag)
 320 {
 321         return sprintf(buf, "%d\n",
 322                        !!test_bit(flag, &transparent_hugepage_flags));
 323 }
 324
 325 static ssize_t single_flag_store(struct kobject *kobj,
 326                                  struct kobj_attribute *attr,
 327                                  const char *buf, size_t count,
 328                                  enum transparent_hugepage_flag flag)
 329 {
 330         unsigned long value;
 331         int ret;
 332
 333         ret = kstrtoul(buf, 10, &value);
 334         if (ret < 0)
 335                 return ret;
 336         if (value > 1)
 337                 return -EINVAL;
 338
 339         if (value)
 340                 set_bit(flag, &transparent_hugepage_flags);
 341         else
 342                 clear_bit(flag, &transparent_hugepage_flags);
 343
 344         return count;
 345 }
 346
 347 /*
 348  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 349  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 350  * memory just to allocate one more hugepage.
 351  */
 352 static ssize_t defrag_show(struct kobject *kobj,
 353                            struct kobj_attribute *attr, char *buf)
 354 {
 355         return double_flag_show(kobj, attr, buf,
 356                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 357                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 358 }
 359 static ssize_t defrag_store(struct kobject *kobj,
 360                             struct kobj_attribute *attr,
 361                             const char *buf, size_t count)
 362 {
 363         return double_flag_store(kobj, attr, buf, count,
 364                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 365                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 366 }
 367 static struct kobj_attribute defrag_attr =
 368         __ATTR(defrag, 0644, defrag_show, defrag_store);
 369
 370 static ssize_t use_zero_page_show(struct kobject *kobj,
 371                 struct kobj_attribute *attr, char *buf)
 372 {
 373         return single_flag_show(kobj, attr, buf,
 374                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 375 }
 376 static ssize_t use_zero_page_store(struct kobject *kobj,
 377                 struct kobj_attribute *attr, const char *buf, size_t count)
 378 {
 379         return single_flag_store(kobj, attr, buf, count,
 380                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 381 }
 382 static struct kobj_attribute use_zero_page_attr =
 383         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 384 #ifdef CONFIG_DEBUG_VM
 385 static ssize_t debug_cow_show(struct kobject *kobj,
 386                                 struct kobj_attribute *attr, char *buf)
 387 {
 388         return single_flag_show(kobj, attr, buf,
 389                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 390 }
 391 static ssize_t debug_cow_store(struct kobject *kobj,
 392                                struct kobj_attribute *attr,
 393                                const char *buf, size_t count)
 394 {
 395         return single_flag_store(kobj, attr, buf, count,
 396                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 397 }
 398 static struct kobj_attribute debug_cow_attr =
 399         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 400 #endif /* CONFIG_DEBUG_VM */
 401
 402 static struct attribute *hugepage_attr[] = {
 403         &enabled_attr.attr,
 404         &defrag_attr.attr,
 405         &use_zero_page_attr.attr,
 406 #ifdef CONFIG_DEBUG_VM
 407         &debug_cow_attr.attr,
 408 #endif
 409         NULL,
 410 };
 411
 412 static struct attribute_group hugepage_attr_group = {
 413         .attrs = hugepage_attr,
 414 };
 415
 416 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 417                                          struct kobj_attribute *attr,
 418                                          char *buf)
 419 {
 420         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 421 }
 422
 423 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 424                                           struct kobj_attribute *attr,
 425                                           const char *buf, size_t count)
 426 {
 427         unsigned long msecs;
 428         int err;
 429
 430         err = kstrtoul(buf, 10, &msecs);
 431         if (err || msecs > UINT_MAX)
 432                 return -EINVAL;
 433
 434         khugepaged_scan_sleep_millisecs = msecs;
 435         wake_up_interruptible(&khugepaged_wait);
 436
 437         return count;
 438 }
 439 static struct kobj_attribute scan_sleep_millisecs_attr =
 440         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 441                scan_sleep_millisecs_store);
 442
 443 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 444                                           struct kobj_attribute *attr,
 445                                           char *buf)
 446 {
 447         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 448 }
 449
 450 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 451                                            struct kobj_attribute *attr,
 452                                            const char *buf, size_t count)
 453 {
 454         unsigned long msecs;
 455         int err;
 456
 457         err = kstrtoul(buf, 10, &msecs);
 458         if (err || msecs > UINT_MAX)
 459                 return -EINVAL;
 460
 461         khugepaged_alloc_sleep_millisecs = msecs;
 462         wake_up_interruptible(&khugepaged_wait);
 463
 464         return count;
 465 }
 466 static struct kobj_attribute alloc_sleep_millisecs_attr =
 467         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 468                alloc_sleep_millisecs_store);
 469
 470 static ssize_t pages_to_scan_show(struct kobject *kobj,
 471                                   struct kobj_attribute *attr,
 472                                   char *buf)
 473 {
 474         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 475 }
 476 static ssize_t pages_to_scan_store(struct kobject *kobj,
 477                                    struct kobj_attribute *attr,
 478                                    const char *buf, size_t count)
 479 {
 480         int err;
 481         unsigned long pages;
 482
 483         err = kstrtoul(buf, 10, &pages);
 484         if (err || !pages || pages > UINT_MAX)
 485                 return -EINVAL;
 486
 487         khugepaged_pages_to_scan = pages;
 488
 489         return count;
 490 }
 491 static struct kobj_attribute pages_to_scan_attr =
 492         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 493                pages_to_scan_store);
 494
 495 static ssize_t pages_collapsed_show(struct kobject *kobj,
 496                                     struct kobj_attribute *attr,
 497                                     char *buf)
 498 {
 499         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 500 }
 501 static struct kobj_attribute pages_collapsed_attr =
 502         __ATTR_RO(pages_collapsed);
 503
 504 static ssize_t full_scans_show(struct kobject *kobj,
 505                                struct kobj_attribute *attr,
 506                                char *buf)
 507 {
 508         return sprintf(buf, "%u\n", khugepaged_full_scans);
 509 }
 510 static struct kobj_attribute full_scans_attr =
 511         __ATTR_RO(full_scans);
 512
 513 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 514                                       struct kobj_attribute *attr, char *buf)
 515 {
 516         return single_flag_show(kobj, attr, buf,
 517                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 518 }
 519 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 520                                        struct kobj_attribute *attr,
 521                                        const char *buf, size_t count)
 522 {
 523         return single_flag_store(kobj, attr, buf, count,
 524                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 525 }
 526 static struct kobj_attribute khugepaged_defrag_attr =
 527         __ATTR(defrag, 0644, khugepaged_defrag_show,
 528                khugepaged_defrag_store);
 529
 530 /*
 531  * max_ptes_none controls if khugepaged should collapse hugepages over
 532  * any unmapped ptes in turn potentially increasing the memory
 533  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 534  * reduce the available free memory in the system as it
 535  * runs. Increasing max_ptes_none will instead potentially reduce the
 536  * free memory in the system during the khugepaged scan.
 537  */
 538 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 539                                              struct kobj_attribute *attr,
 540                                              char *buf)
 541 {
 542         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 543 }
 544 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 545                                               struct kobj_attribute *attr,
 546                                               const char *buf, size_t count)
 547 {
 548         int err;
 549         unsigned long max_ptes_none;
 550
 551         err = kstrtoul(buf, 10, &max_ptes_none);
 552         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 553                 return -EINVAL;
 554
 555         khugepaged_max_ptes_none = max_ptes_none;
 556
 557         return count;
 558 }
 559 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 560         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 561                khugepaged_max_ptes_none_store);
 562
 563 static struct attribute *khugepaged_attr[] = {
 564         &khugepaged_defrag_attr.attr,
 565         &khugepaged_max_ptes_none_attr.attr,
 566         &pages_to_scan_attr.attr,
 567         &pages_collapsed_attr.attr,
 568         &full_scans_attr.attr,
 569         &scan_sleep_millisecs_attr.attr,
 570         &alloc_sleep_millisecs_attr.attr,
 571         NULL,
 572 };
 573
 574 static struct attribute_group khugepaged_attr_group = {
 575         .attrs = khugepaged_attr,
 576         .name = "khugepaged",
 577 };
 578
 579 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 580 {
 581         int err;
 582
 583         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 584         if (unlikely(!*hugepage_kobj)) {
 585                 pr_err("failed to create transparent hugepage kobject\n");
 586                 return -ENOMEM;
 587         }
 588
 589         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 590         if (err) {
 591                 pr_err("failed to register transparent hugepage group\n");
 592                 goto delete_obj;
 593         }
 594
 595         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 596         if (err) {
 597                 pr_err("failed to register transparent hugepage group\n");
 598                 goto remove_hp_group;
 599         }
 600
 601         return 0;
 602
 603 remove_hp_group:
 604         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 605 delete_obj:
 606         kobject_put(*hugepage_kobj);
 607         return err;
 608 }
 609
 610 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 611 {
 612         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 613         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 614         kobject_put(hugepage_kobj);
 615 }
 616 #else
 617 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 618 {
 619         return 0;
 620 }
 621
 622 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 623 {
 624 }
 625 #endif /* CONFIG_SYSFS */
 626
 627 static int __init hugepage_init(void)
 628 {
 629         int err;
 630         struct kobject *hugepage_kobj;
 631
 632         if (!has_transparent_hugepage()) {
 633                 transparent_hugepage_flags = 0;
 634                 return -EINVAL;
 635         }
 636
 637         err = hugepage_init_sysfs(&hugepage_kobj);
 638         if (err)
 639                 goto err_sysfs;
 640
 641         err = khugepaged_slab_init();
 642         if (err)
 643                 goto err_slab;
 644
 645         err = register_shrinker(&huge_zero_page_shrinker);
 646         if (err)
 647                 goto err_hzp_shrinker;
 648
 649         /*
 650          * By default disable transparent hugepages on smaller systems,
 651          * where the extra memory used could hurt more than TLB overhead
 652          * is likely to save.  The admin can still enable it through /sys.
 653          */
 654         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 655                 transparent_hugepage_flags = 0;
 656
 657         err = start_khugepaged();
 658         if (err)
 659                 goto err_khugepaged;
 660
 661         return 0;
 662 err_khugepaged:
 663         unregister_shrinker(&huge_zero_page_shrinker);
 664 err_hzp_shrinker:
 665         khugepaged_slab_exit();
 666 err_slab:
 667         hugepage_exit_sysfs(hugepage_kobj);
 668 err_sysfs:
 669         return err;
 670 }
 671 subsys_initcall(hugepage_init);
 672
 673 static int __init setup_transparent_hugepage(char *str)
 674 {
 675         int ret = 0;
 676         if (!str)
 677                 goto out;
 678         if (!strcmp(str, "always")) {
 679                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 680                         &transparent_hugepage_flags);
 681                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 682                           &transparent_hugepage_flags);
 683                 ret = 1;
 684         } else if (!strcmp(str, "madvise")) {
 685                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 686                           &transparent_hugepage_flags);
 687                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 688                         &transparent_hugepage_flags);
 689                 ret = 1;
 690         } else if (!strcmp(str, "never")) {
 691                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 692                           &transparent_hugepage_flags);
 693                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 694                           &transparent_hugepage_flags);
 695                 ret = 1;
 696         }
 697 out:
 698         if (!ret)
 699                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 700         return ret;
 701 }
 702 __setup("transparent_hugepage=", setup_transparent_hugepage);
 703
 704 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 705 {
 706         if (likely(vma->vm_flags & VM_WRITE))
 707                 pmd = pmd_mkwrite(pmd);
 708         return pmd;
 709 }
 710
 711 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 712 {
 713         pmd_t entry;
 714         entry = mk_pmd(page, prot);
 715         entry = pmd_mkhuge(entry);
 716         return entry;
 717 }
 718
 719 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 720                                         struct vm_area_struct *vma,
 721                                         unsigned long haddr, pmd_t *pmd,
 722                                         struct page *page, gfp_t gfp)
 723 {
 724         struct mem_cgroup *memcg;
 725         pgtable_t pgtable;
 726         spinlock_t *ptl;
 727
 728         VM_BUG_ON_PAGE(!PageCompound(page), page);
 729
 730         if (mem_cgroup_try_charge(page, mm, gfp, &memcg))
 731                 return VM_FAULT_OOM;
 732
 733         pgtable = pte_alloc_one(mm, haddr);
 734         if (unlikely(!pgtable)) {
 735                 mem_cgroup_cancel_charge(page, memcg);
 736                 return VM_FAULT_OOM;
 737         }
 738
 739         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 740         /*
 741          * The memory barrier inside __SetPageUptodate makes sure that
 742          * clear_huge_page writes become visible before the set_pmd_at()
 743          * write.
 744          */
 745         __SetPageUptodate(page);
 746
 747         ptl = pmd_lock(mm, pmd);
 748         if (unlikely(!pmd_none(*pmd))) {
 749                 spin_unlock(ptl);
 750                 mem_cgroup_cancel_charge(page, memcg);
 751                 put_page(page);
 752                 pte_free(mm, pgtable);
 753         } else {
 754                 pmd_t entry;
 755                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 756                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 757                 page_add_new_anon_rmap(page, vma, haddr);
 758                 mem_cgroup_commit_charge(page, memcg, false);
 759                 lru_cache_add_active_or_unevictable(page, vma);
 760                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 761                 set_pmd_at(mm, haddr, pmd, entry);
 762                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 763                 atomic_long_inc(&mm->nr_ptes);
 764                 spin_unlock(ptl);
 765         }
 766
 767         return 0;
 768 }
 769
 770 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 771 {
 772         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 773 }
 774
 775 /* Caller must hold page table lock. */
 776 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 777                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 778                 struct page *zero_page)
 779 {
 780         pmd_t entry;
 781         if (!pmd_none(*pmd))
 782                 return false;
 783         entry = mk_pmd(zero_page, vma->vm_page_prot);
 784         entry = pmd_mkhuge(entry);
 785         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 786         set_pmd_at(mm, haddr, pmd, entry);
 787         atomic_long_inc(&mm->nr_ptes);
 788         return true;
 789 }
 790
 791 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 792                                unsigned long address, pmd_t *pmd,
 793                                unsigned int flags)
 794 {
 795         gfp_t gfp;
 796         struct page *page;
 797         unsigned long haddr = address & HPAGE_PMD_MASK;
 798
 799         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 800                 return VM_FAULT_FALLBACK;
 801         if (unlikely(anon_vma_prepare(vma)))
 802                 return VM_FAULT_OOM;
 803         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 804                 return VM_FAULT_OOM;
 805         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 806                         transparent_hugepage_use_zero_page()) {
 807                 spinlock_t *ptl;
 808                 pgtable_t pgtable;
 809                 struct page *zero_page;
 810                 bool set;
 811                 pgtable = pte_alloc_one(mm, haddr);
 812                 if (unlikely(!pgtable))
 813                         return VM_FAULT_OOM;
 814                 zero_page = get_huge_zero_page();
 815                 if (unlikely(!zero_page)) {
 816                         pte_free(mm, pgtable);
 817                         count_vm_event(THP_FAULT_FALLBACK);
 818                         return VM_FAULT_FALLBACK;
 819                 }
 820                 ptl = pmd_lock(mm, pmd);
 821                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 822                                 zero_page);
 823                 spin_unlock(ptl);
 824                 if (!set) {
 825                         pte_free(mm, pgtable);
 826                         put_huge_zero_page();
 827                 }
 828                 return 0;
 829         }
 830         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 831         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 832         if (unlikely(!page)) {
 833                 count_vm_event(THP_FAULT_FALLBACK);
 834                 return VM_FAULT_FALLBACK;
 835         }
 836         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page, gfp))) {
 837                 put_page(page);
 838                 count_vm_event(THP_FAULT_FALLBACK);
 839                 return VM_FAULT_FALLBACK;
 840         }
 841
 842         count_vm_event(THP_FAULT_ALLOC);
 843         return 0;
 844 }
 845
 846 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 847                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 848                   struct vm_area_struct *vma)
 849 {
 850         spinlock_t *dst_ptl, *src_ptl;
 851         struct page *src_page;
 852         pmd_t pmd;
 853         pgtable_t pgtable;
 854         int ret;
 855
 856         ret = -ENOMEM;
 857         pgtable = pte_alloc_one(dst_mm, addr);
 858         if (unlikely(!pgtable))
 859                 goto out;
 860
 861         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 862         src_ptl = pmd_lockptr(src_mm, src_pmd);
 863         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 864
 865         ret = -EAGAIN;
 866         pmd = *src_pmd;
 867         if (unlikely(!pmd_trans_huge(pmd))) {
 868                 pte_free(dst_mm, pgtable);
 869                 goto out_unlock;
 870         }
 871         /*
 872          * When page table lock is held, the huge zero pmd should not be
 873          * under splitting since we don't split the page itself, only pmd to
 874          * a page table.
 875          */
 876         if (is_huge_zero_pmd(pmd)) {
 877                 struct page *zero_page;
 878                 bool set;
 879                 /*
 880                  * get_huge_zero_page() will never allocate a new page here,
 881                  * since we already have a zero page to copy. It just takes a
 882                  * reference.
 883                  */
 884                 zero_page = get_huge_zero_page();
 885                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 886                                 zero_page);
 887                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 888                 ret = 0;
 889                 goto out_unlock;
 890         }
 891
 892         if (unlikely(pmd_trans_splitting(pmd))) {
 893                 /* split huge page running from under us */
 894                 spin_unlock(src_ptl);
 895                 spin_unlock(dst_ptl);
 896                 pte_free(dst_mm, pgtable);
 897
 898                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 899                 goto out;
 900         }
 901         src_page = pmd_page(pmd);
 902         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 903         get_page(src_page);
 904         page_dup_rmap(src_page);
 905         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 906
 907         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 908         pmd = pmd_mkold(pmd_wrprotect(pmd));
 909         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 910         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 911         atomic_long_inc(&dst_mm->nr_ptes);
 912
 913         ret = 0;
 914 out_unlock:
 915         spin_unlock(src_ptl);
 916         spin_unlock(dst_ptl);
 917 out:
 918         return ret;
 919 }
 920
 921 void huge_pmd_set_accessed(struct mm_struct *mm,
 922                            struct vm_area_struct *vma,
 923                            unsigned long address,
 924                            pmd_t *pmd, pmd_t orig_pmd,
 925                            int dirty)
 926 {
 927         spinlock_t *ptl;
 928         pmd_t entry;
 929         unsigned long haddr;
 930
 931         ptl = pmd_lock(mm, pmd);
 932         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 933                 goto unlock;
 934
 935         entry = pmd_mkyoung(orig_pmd);
 936         haddr = address & HPAGE_PMD_MASK;
 937         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 938                 update_mmu_cache_pmd(vma, address, pmd);
 939
 940 unlock:
 941         spin_unlock(ptl);
 942 }
 943
 944 /*
 945  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 946  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 947  * the source page gets split and a tail freed before copy completes.
 948  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 949  */
 950 static void get_user_huge_page(struct page *page)
 951 {
 952         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 953                 struct page *endpage = page + HPAGE_PMD_NR;
 954
 955                 atomic_add(HPAGE_PMD_NR, &page->_count);
 956                 while (++page < endpage)
 957                         get_huge_page_tail(page);
 958         } else {
 959                 get_page(page);
 960         }
 961 }
 962
 963 static void put_user_huge_page(struct page *page)
 964 {
 965         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 966                 struct page *endpage = page + HPAGE_PMD_NR;
 967
 968                 while (page < endpage)
 969                         put_page(page++);
 970         } else {
 971                 put_page(page);
 972         }
 973 }
 974
 975 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 976                                         struct vm_area_struct *vma,
 977                                         unsigned long address,
 978                                         pmd_t *pmd, pmd_t orig_pmd,
 979                                         struct page *page,
 980                                         unsigned long haddr)
 981 {
 982         struct mem_cgroup *memcg;
 983         spinlock_t *ptl;
 984         pgtable_t pgtable;
 985         pmd_t _pmd;
 986         int ret = 0, i;
 987         struct page **pages;
 988         unsigned long mmun_start;       /* For mmu_notifiers */
 989         unsigned long mmun_end;         /* For mmu_notifiers */
 990
 991         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 992                         GFP_KERNEL);
 993         if (unlikely(!pages)) {
 994                 ret |= VM_FAULT_OOM;
 995                 goto out;
 996         }
 997
 998         for (i = 0; i < HPAGE_PMD_NR; i++) {
 999                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1000                                                __GFP_OTHER_NODE,
1001                                                vma, address, page_to_nid(page));
1002                 if (unlikely(!pages[i] ||
1003                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1004                                                    &memcg))) {
1005                         if (pages[i])
1006                                 put_page(pages[i]);
1007                         while (--i >= 0) {
1008                                 memcg = (void *)page_private(pages[i]);
1009                                 set_page_private(pages[i], 0);
1010                                 mem_cgroup_cancel_charge(pages[i], memcg);
1011                                 put_page(pages[i]);
1012                         }
1013                         kfree(pages);
1014                         ret |= VM_FAULT_OOM;
1015                         goto out;
1016                 }
1017                 set_page_private(pages[i], (unsigned long)memcg);
1018         }
1019
1020         for (i = 0; i < HPAGE_PMD_NR; i++) {
1021                 copy_user_highpage(pages[i], page + i,
1022                                    haddr + PAGE_SIZE * i, vma);
1023                 __SetPageUptodate(pages[i]);
1024                 cond_resched();
1025         }
1026
1027         mmun_start = haddr;
1028         mmun_end   = haddr + HPAGE_PMD_SIZE;
1029         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1030
1031         ptl = pmd_lock(mm, pmd);
1032         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1033                 goto out_free_pages;
1034         VM_BUG_ON_PAGE(!PageHead(page), page);
1035
1036         pmdp_clear_flush_notify(vma, haddr, pmd);
1037         /* leave pmd empty until pte is filled */
1038
1039         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1040         pmd_populate(mm, &_pmd, pgtable);
1041
1042         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1043                 pte_t *pte, entry;
1044                 entry = mk_pte(pages[i], vma->vm_page_prot);
1045                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1046                 memcg = (void *)page_private(pages[i]);
1047                 set_page_private(pages[i], 0);
1048                 page_add_new_anon_rmap(pages[i], vma, haddr);
1049                 mem_cgroup_commit_charge(pages[i], memcg, false);
1050                 lru_cache_add_active_or_unevictable(pages[i], vma);
1051                 pte = pte_offset_map(&_pmd, haddr);
1052                 VM_BUG_ON(!pte_none(*pte));
1053                 set_pte_at(mm, haddr, pte, entry);
1054                 pte_unmap(pte);
1055         }
1056         kfree(pages);
1057
1058         smp_wmb(); /* make pte visible before pmd */
1059         pmd_populate(mm, pmd, pgtable);
1060         page_remove_rmap(page);
1061         spin_unlock(ptl);
1062
1063         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1064
1065         ret |= VM_FAULT_WRITE;
1066         put_page(page);
1067
1068 out:
1069         return ret;
1070
1071 out_free_pages:
1072         spin_unlock(ptl);
1073         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1074         for (i = 0; i < HPAGE_PMD_NR; i++) {
1075                 memcg = (void *)page_private(pages[i]);
1076                 set_page_private(pages[i], 0);
1077                 mem_cgroup_cancel_charge(pages[i], memcg);
1078                 put_page(pages[i]);
1079         }
1080         kfree(pages);
1081         goto out;
1082 }
1083
1084 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1085                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1086 {
1087         spinlock_t *ptl;
1088         int ret = 0;
1089         struct page *page = NULL, *new_page;
1090         struct mem_cgroup *memcg;
1091         unsigned long haddr;
1092         unsigned long mmun_start;       /* For mmu_notifiers */
1093         unsigned long mmun_end;         /* For mmu_notifiers */
1094         gfp_t huge_gfp;                 /* for allocation and charge */
1095
1096         ptl = pmd_lockptr(mm, pmd);
1097         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1098         haddr = address & HPAGE_PMD_MASK;
1099         if (is_huge_zero_pmd(orig_pmd))
1100                 goto alloc;
1101         spin_lock(ptl);
1102         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1103                 goto out_unlock;
1104
1105         page = pmd_page(orig_pmd);
1106         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1107         if (page_mapcount(page) == 1) {
1108                 pmd_t entry;
1109                 entry = pmd_mkyoung(orig_pmd);
1110                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1111                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1112                         update_mmu_cache_pmd(vma, address, pmd);
1113                 ret |= VM_FAULT_WRITE;
1114                 goto out_unlock;
1115         }
1116         get_user_huge_page(page);
1117         spin_unlock(ptl);
1118 alloc:
1119         if (transparent_hugepage_enabled(vma) &&
1120             !transparent_hugepage_debug_cow()) {
1121                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1122                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1123         } else
1124                 new_page = NULL;
1125
1126         if (unlikely(!new_page)) {
1127                 if (!page) {
1128                         split_huge_page_pmd(vma, address, pmd);
1129                         ret |= VM_FAULT_FALLBACK;
1130                 } else {
1131                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1132                                         pmd, orig_pmd, page, haddr);
1133                         if (ret & VM_FAULT_OOM) {
1134                                 split_huge_page(page);
1135                                 ret |= VM_FAULT_FALLBACK;
1136                         }
1137                         put_user_huge_page(page);
1138                 }
1139                 count_vm_event(THP_FAULT_FALLBACK);
1140                 goto out;
1141         }
1142
1143         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1144                 put_page(new_page);
1145                 if (page) {
1146                         split_huge_page(page);
1147                         put_user_huge_page(page);
1148                 } else
1149                         split_huge_page_pmd(vma, address, pmd);
1150                 ret |= VM_FAULT_FALLBACK;
1151                 count_vm_event(THP_FAULT_FALLBACK);
1152                 goto out;
1153         }
1154
1155         count_vm_event(THP_FAULT_ALLOC);
1156
1157         if (!page)
1158                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1159         else
1160                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1161         __SetPageUptodate(new_page);
1162
1163         mmun_start = haddr;
1164         mmun_end   = haddr + HPAGE_PMD_SIZE;
1165         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1166
1167         spin_lock(ptl);
1168         if (page)
1169                 put_user_huge_page(page);
1170         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1171                 spin_unlock(ptl);
1172                 mem_cgroup_cancel_charge(new_page, memcg);
1173                 put_page(new_page);
1174                 goto out_mn;
1175         } else {
1176                 pmd_t entry;
1177                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1178                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1179                 pmdp_clear_flush_notify(vma, haddr, pmd);
1180                 page_add_new_anon_rmap(new_page, vma, haddr);
1181                 mem_cgroup_commit_charge(new_page, memcg, false);
1182                 lru_cache_add_active_or_unevictable(new_page, vma);
1183                 set_pmd_at(mm, haddr, pmd, entry);
1184                 update_mmu_cache_pmd(vma, address, pmd);
1185                 if (!page) {
1186                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1187                         put_huge_zero_page();
1188                 } else {
1189                         VM_BUG_ON_PAGE(!PageHead(page), page);
1190                         page_remove_rmap(page);
1191                         put_page(page);
1192                 }
1193                 ret |= VM_FAULT_WRITE;
1194         }
1195         spin_unlock(ptl);
1196 out_mn:
1197         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1198 out:
1199         return ret;
1200 out_unlock:
1201         spin_unlock(ptl);
1202         return ret;
1203 }
1204
1205 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1206                                    unsigned long addr,
1207                                    pmd_t *pmd,
1208                                    unsigned int flags)
1209 {
1210         struct mm_struct *mm = vma->vm_mm;
1211         struct page *page = NULL;
1212
1213         assert_spin_locked(pmd_lockptr(mm, pmd));
1214
1215         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1216                 goto out;
1217
1218         /* Avoid dumping huge zero page */
1219         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1220                 return ERR_PTR(-EFAULT);
1221
1222         /* Full NUMA hinting faults to serialise migration in fault paths */
1223         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1224                 goto out;
1225
1226         page = pmd_page(*pmd);
1227         VM_BUG_ON_PAGE(!PageHead(page), page);
1228         if (flags & FOLL_TOUCH) {
1229                 pmd_t _pmd;
1230                 /*
1231                  * We should set the dirty bit only for FOLL_WRITE but
1232                  * for now the dirty bit in the pmd is meaningless.
1233                  * And if the dirty bit will become meaningful and
1234                  * we'll only set it with FOLL_WRITE, an atomic
1235                  * set_bit will be required on the pmd to set the
1236                  * young bit, instead of the current set_pmd_at.
1237                  */
1238                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1239                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1240                                           pmd, _pmd,  1))
1241                         update_mmu_cache_pmd(vma, addr, pmd);
1242         }
1243         if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1244                 if (page->mapping && trylock_page(page)) {
1245                         lru_add_drain();
1246                         if (page->mapping)
1247                                 mlock_vma_page(page);
1248                         unlock_page(page);
1249                 }
1250         }
1251         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1252         VM_BUG_ON_PAGE(!PageCompound(page), page);
1253         if (flags & FOLL_GET)
1254                 get_page_foll(page);
1255
1256 out:
1257         return page;
1258 }
1259
1260 /* NUMA hinting page fault entry point for trans huge pmds */
1261 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1262                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1263 {
1264         spinlock_t *ptl;
1265         struct anon_vma *anon_vma = NULL;
1266         struct page *page;
1267         unsigned long haddr = addr & HPAGE_PMD_MASK;
1268         int page_nid = -1, this_nid = numa_node_id();
1269         int target_nid, last_cpupid = -1;
1270         bool page_locked;
1271         bool migrated = false;
1272         bool was_writable;
1273         int flags = 0;
1274
1275         /* A PROT_NONE fault should not end up here */
1276         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1277
1278         ptl = pmd_lock(mm, pmdp);
1279         if (unlikely(!pmd_same(pmd, *pmdp)))
1280                 goto out_unlock;
1281
1282         /*
1283          * If there are potential migrations, wait for completion and retry
1284          * without disrupting NUMA hinting information. Do not relock and
1285          * check_same as the page may no longer be mapped.
1286          */
1287         if (unlikely(pmd_trans_migrating(*pmdp))) {
1288                 page = pmd_page(*pmdp);
1289                 spin_unlock(ptl);
1290                 wait_on_page_locked(page);
1291                 goto out;
1292         }
1293
1294         page = pmd_page(pmd);
1295         BUG_ON(is_huge_zero_page(page));
1296         page_nid = page_to_nid(page);
1297         last_cpupid = page_cpupid_last(page);
1298         count_vm_numa_event(NUMA_HINT_FAULTS);
1299         if (page_nid == this_nid) {
1300                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1301                 flags |= TNF_FAULT_LOCAL;
1302         }
1303
1304         /* See similar comment in do_numa_page for explanation */
1305         if (!(vma->vm_flags & VM_WRITE))
1306                 flags |= TNF_NO_GROUP;
1307
1308         /*
1309          * Acquire the page lock to serialise THP migrations but avoid dropping
1310          * page_table_lock if at all possible
1311          */
1312         page_locked = trylock_page(page);
1313         target_nid = mpol_misplaced(page, vma, haddr);
1314         if (target_nid == -1) {
1315                 /* If the page was locked, there are no parallel migrations */
1316                 if (page_locked)
1317                         goto clear_pmdnuma;
1318         }
1319
1320         /* Migration could have started since the pmd_trans_migrating check */
1321         if (!page_locked) {
1322                 spin_unlock(ptl);
1323                 wait_on_page_locked(page);
1324                 page_nid = -1;
1325                 goto out;
1326         }
1327
1328         /*
1329          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1330          * to serialises splits
1331          */
1332         get_page(page);
1333         spin_unlock(ptl);
1334         anon_vma = page_lock_anon_vma_read(page);
1335
1336         /* Confirm the PMD did not change while page_table_lock was released */
1337         spin_lock(ptl);
1338         if (unlikely(!pmd_same(pmd, *pmdp))) {
1339                 unlock_page(page);
1340                 put_page(page);
1341                 page_nid = -1;
1342                 goto out_unlock;
1343         }
1344
1345         /* Bail if we fail to protect against THP splits for any reason */
1346         if (unlikely(!anon_vma)) {
1347                 put_page(page);
1348                 page_nid = -1;
1349                 goto clear_pmdnuma;
1350         }
1351
1352         /*
1353          * Migrate the THP to the requested node, returns with page unlocked
1354          * and access rights restored.
1355          */
1356         spin_unlock(ptl);
1357         migrated = migrate_misplaced_transhuge_page(mm, vma,
1358                                 pmdp, pmd, addr, page, target_nid);
1359         if (migrated) {
1360                 flags |= TNF_MIGRATED;
1361                 page_nid = target_nid;
1362         } else
1363                 flags |= TNF_MIGRATE_FAIL;
1364
1365         goto out;
1366 clear_pmdnuma:
1367         BUG_ON(!PageLocked(page));
1368         was_writable = pmd_write(pmd);
1369         pmd = pmd_modify(pmd, vma->vm_page_prot);
1370         pmd = pmd_mkyoung(pmd);
1371         if (was_writable)
1372                 pmd = pmd_mkwrite(pmd);
1373         set_pmd_at(mm, haddr, pmdp, pmd);
1374         update_mmu_cache_pmd(vma, addr, pmdp);
1375         unlock_page(page);
1376 out_unlock:
1377         spin_unlock(ptl);
1378
1379 out:
1380         if (anon_vma)
1381                 page_unlock_anon_vma_read(anon_vma);
1382
1383         if (page_nid != -1)
1384                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1385
1386         return 0;
1387 }
1388
1389 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1390                  pmd_t *pmd, unsigned long addr)
1391 {
1392         spinlock_t *ptl;
1393         int ret = 0;
1394
1395         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1396                 struct page *page;
1397                 pgtable_t pgtable;
1398                 pmd_t orig_pmd;
1399                 /*
1400                  * For architectures like ppc64 we look at deposited pgtable
1401                  * when calling pmdp_get_and_clear. So do the
1402                  * pgtable_trans_huge_withdraw after finishing pmdp related
1403                  * operations.
1404                  */
1405                 orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd,
1406                                                    tlb->fullmm);
1407                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1408                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1409                 if (is_huge_zero_pmd(orig_pmd)) {
1410                         atomic_long_dec(&tlb->mm->nr_ptes);
1411                         spin_unlock(ptl);
1412                         put_huge_zero_page();
1413                 } else {
1414                         page = pmd_page(orig_pmd);
1415                         page_remove_rmap(page);
1416                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1417                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1418                         VM_BUG_ON_PAGE(!PageHead(page), page);
1419                         atomic_long_dec(&tlb->mm->nr_ptes);
1420                         spin_unlock(ptl);
1421                         tlb_remove_page(tlb, page);
1422                 }
1423                 pte_free(tlb->mm, pgtable);
1424                 ret = 1;
1425         }
1426         return ret;
1427 }
1428
1429 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1430                   unsigned long old_addr,
1431                   unsigned long new_addr, unsigned long old_end,
1432                   pmd_t *old_pmd, pmd_t *new_pmd)
1433 {
1434         spinlock_t *old_ptl, *new_ptl;
1435         int ret = 0;
1436         pmd_t pmd;
1437
1438         struct mm_struct *mm = vma->vm_mm;
1439
1440         if ((old_addr & ~HPAGE_PMD_MASK) ||
1441             (new_addr & ~HPAGE_PMD_MASK) ||
1442             old_end - old_addr < HPAGE_PMD_SIZE ||
1443             (new_vma->vm_flags & VM_NOHUGEPAGE))
1444                 goto out;
1445
1446         /*
1447          * The destination pmd shouldn't be established, free_pgtables()
1448          * should have release it.
1449          */
1450         if (WARN_ON(!pmd_none(*new_pmd))) {
1451                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1452                 goto out;
1453         }
1454
1455         /*
1456          * We don't have to worry about the ordering of src and dst
1457          * ptlocks because exclusive mmap_sem prevents deadlock.
1458          */
1459         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1460         if (ret == 1) {
1461                 new_ptl = pmd_lockptr(mm, new_pmd);
1462                 if (new_ptl != old_ptl)
1463                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1464                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1465                 VM_BUG_ON(!pmd_none(*new_pmd));
1466
1467                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1468                         pgtable_t pgtable;
1469                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1470                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1471                 }
1472                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1473                 if (new_ptl != old_ptl)
1474                         spin_unlock(new_ptl);
1475                 spin_unlock(old_ptl);
1476         }
1477 out:
1478         return ret;
1479 }
1480
1481 /*
1482  * Returns
1483  *  - 0 if PMD could not be locked
1484  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1485  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1486  */
1487 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1488                 unsigned long addr, pgprot_t newprot, int prot_numa)
1489 {
1490         struct mm_struct *mm = vma->vm_mm;
1491         spinlock_t *ptl;
1492         int ret = 0;
1493
1494         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1495                 pmd_t entry;
1496                 bool preserve_write = prot_numa && pmd_write(*pmd);
1497                 ret = 1;
1498
1499                 /*
1500                  * Avoid trapping faults against the zero page. The read-only
1501                  * data is likely to be read-cached on the local CPU and
1502                  * local/remote hits to the zero page are not interesting.
1503                  */
1504                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1505                         spin_unlock(ptl);
1506                         return ret;
1507                 }
1508
1509                 if (!prot_numa || !pmd_protnone(*pmd)) {
1510                         entry = pmdp_get_and_clear_notify(mm, addr, pmd);
1511                         entry = pmd_modify(entry, newprot);
1512                         if (preserve_write)
1513                                 entry = pmd_mkwrite(entry);
1514                         ret = HPAGE_PMD_NR;
1515                         set_pmd_at(mm, addr, pmd, entry);
1516                         BUG_ON(!preserve_write && pmd_write(entry));
1517                 }
1518                 spin_unlock(ptl);
1519         }
1520
1521         return ret;
1522 }
1523
1524 /*
1525  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1526  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1527  *
1528  * Note that if it returns 1, this routine returns without unlocking page
1529  * table locks. So callers must unlock them.
1530  */
1531 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1532                 spinlock_t **ptl)
1533 {
1534         *ptl = pmd_lock(vma->vm_mm, pmd);
1535         if (likely(pmd_trans_huge(*pmd))) {
1536                 if (unlikely(pmd_trans_splitting(*pmd))) {
1537                         spin_unlock(*ptl);
1538                         wait_split_huge_page(vma->anon_vma, pmd);
1539                         return -1;
1540                 } else {
1541                         /* Thp mapped by 'pmd' is stable, so we can
1542                          * handle it as it is. */
1543                         return 1;
1544                 }
1545         }
1546         spin_unlock(*ptl);
1547         return 0;
1548 }
1549
1550 /*
1551  * This function returns whether a given @page is mapped onto the @address
1552  * in the virtual space of @mm.
1553  *
1554  * When it's true, this function returns *pmd with holding the page table lock
1555  * and passing it back to the caller via @ptl.
1556  * If it's false, returns NULL without holding the page table lock.
1557  */
1558 pmd_t *page_check_address_pmd(struct page *page,
1559                               struct mm_struct *mm,
1560                               unsigned long address,
1561                               enum page_check_address_pmd_flag flag,
1562                               spinlock_t **ptl)
1563 {
1564         pgd_t *pgd;
1565         pud_t *pud;
1566         pmd_t *pmd;
1567
1568         if (address & ~HPAGE_PMD_MASK)
1569                 return NULL;
1570
1571         pgd = pgd_offset(mm, address);
1572         if (!pgd_present(*pgd))
1573                 return NULL;
1574         pud = pud_offset(pgd, address);
1575         if (!pud_present(*pud))
1576                 return NULL;
1577         pmd = pmd_offset(pud, address);
1578
1579         *ptl = pmd_lock(mm, pmd);
1580         if (!pmd_present(*pmd))
1581                 goto unlock;
1582         if (pmd_page(*pmd) != page)
1583                 goto unlock;
1584         /*
1585          * split_vma() may create temporary aliased mappings. There is
1586          * no risk as long as all huge pmd are found and have their
1587          * splitting bit set before __split_huge_page_refcount
1588          * runs. Finding the same huge pmd more than once during the
1589          * same rmap walk is not a problem.
1590          */
1591         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1592             pmd_trans_splitting(*pmd))
1593                 goto unlock;
1594         if (pmd_trans_huge(*pmd)) {
1595                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1596                           !pmd_trans_splitting(*pmd));
1597                 return pmd;
1598         }
1599 unlock:
1600         spin_unlock(*ptl);
1601         return NULL;
1602 }
1603
1604 static int __split_huge_page_splitting(struct page *page,
1605                                        struct vm_area_struct *vma,
1606                                        unsigned long address)
1607 {
1608         struct mm_struct *mm = vma->vm_mm;
1609         spinlock_t *ptl;
1610         pmd_t *pmd;
1611         int ret = 0;
1612         /* For mmu_notifiers */
1613         const unsigned long mmun_start = address;
1614         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1615
1616         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1617         pmd = page_check_address_pmd(page, mm, address,
1618                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1619         if (pmd) {
1620                 /*
1621                  * We can't temporarily set the pmd to null in order
1622                  * to split it, the pmd must remain marked huge at all
1623                  * times or the VM won't take the pmd_trans_huge paths
1624                  * and it won't wait on the anon_vma->root->rwsem to
1625                  * serialize against split_huge_page*.
1626                  */
1627                 pmdp_splitting_flush(vma, address, pmd);
1628
1629                 ret = 1;
1630                 spin_unlock(ptl);
1631         }
1632         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1633
1634         return ret;
1635 }
1636
1637 static void __split_huge_page_refcount(struct page *page,
1638                                        struct list_head *list)
1639 {
1640         int i;
1641         struct zone *zone = page_zone(page);
1642         struct lruvec *lruvec;
1643         int tail_count = 0;
1644
1645         /* prevent PageLRU to go away from under us, and freeze lru stats */
1646         spin_lock_irq(&zone->lru_lock);
1647         lruvec = mem_cgroup_page_lruvec(page, zone);
1648
1649         compound_lock(page);
1650         /* complete memcg works before add pages to LRU */
1651         mem_cgroup_split_huge_fixup(page);
1652
1653         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1654                 struct page *page_tail = page + i;
1655
1656                 /* tail_page->_mapcount cannot change */
1657                 BUG_ON(page_mapcount(page_tail) < 0);
1658                 tail_count += page_mapcount(page_tail);
1659                 /* check for overflow */
1660                 BUG_ON(tail_count < 0);
1661                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1662                 /*
1663                  * tail_page->_count is zero and not changing from
1664                  * under us. But get_page_unless_zero() may be running
1665                  * from under us on the tail_page. If we used
1666                  * atomic_set() below instead of atomic_add(), we
1667                  * would then run atomic_set() concurrently with
1668                  * get_page_unless_zero(), and atomic_set() is
1669                  * implemented in C not using locked ops. spin_unlock
1670                  * on x86 sometime uses locked ops because of PPro
1671                  * errata 66, 92, so unless somebody can guarantee
1672                  * atomic_set() here would be safe on all archs (and
1673                  * not only on x86), it's safer to use atomic_add().
1674                  */
1675                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1676                            &page_tail->_count);
1677
1678                 /* after clearing PageTail the gup refcount can be released */
1679                 smp_mb__after_atomic();
1680
1681                 /*
1682                  * retain hwpoison flag of the poisoned tail page:
1683                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1684                  *   by the memory-failure.
1685                  */
1686                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1687                 page_tail->flags |= (page->flags &
1688                                      ((1L << PG_referenced) |
1689                                       (1L << PG_swapbacked) |
1690                                       (1L << PG_mlocked) |
1691                                       (1L << PG_uptodate) |
1692                                       (1L << PG_active) |
1693                                       (1L << PG_unevictable)));
1694                 page_tail->flags |= (1L << PG_dirty);
1695
1696                 /* clear PageTail before overwriting first_page */
1697                 smp_wmb();
1698
1699                 /*
1700                  * __split_huge_page_splitting() already set the
1701                  * splitting bit in all pmd that could map this
1702                  * hugepage, that will ensure no CPU can alter the
1703                  * mapcount on the head page. The mapcount is only
1704                  * accounted in the head page and it has to be
1705                  * transferred to all tail pages in the below code. So
1706                  * for this code to be safe, the split the mapcount
1707                  * can't change. But that doesn't mean userland can't
1708                  * keep changing and reading the page contents while
1709                  * we transfer the mapcount, so the pmd splitting
1710                  * status is achieved setting a reserved bit in the
1711                  * pmd, not by clearing the present bit.
1712                 */
1713                 page_tail->_mapcount = page->_mapcount;
1714
1715                 BUG_ON(page_tail->mapping);
1716                 page_tail->mapping = page->mapping;
1717
1718                 page_tail->index = page->index + i;
1719                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1720
1721                 BUG_ON(!PageAnon(page_tail));
1722                 BUG_ON(!PageUptodate(page_tail));
1723                 BUG_ON(!PageDirty(page_tail));
1724                 BUG_ON(!PageSwapBacked(page_tail));
1725
1726                 lru_add_page_tail(page, page_tail, lruvec, list);
1727         }
1728         atomic_sub(tail_count, &page->_count);
1729         BUG_ON(atomic_read(&page->_count) <= 0);
1730
1731         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1732
1733         ClearPageCompound(page);
1734         compound_unlock(page);
1735         spin_unlock_irq(&zone->lru_lock);
1736
1737         for (i = 1; i < HPAGE_PMD_NR; i++) {
1738                 struct page *page_tail = page + i;
1739                 BUG_ON(page_count(page_tail) <= 0);
1740                 /*
1741                  * Tail pages may be freed if there wasn't any mapping
1742                  * like if add_to_swap() is running on a lru page that
1743                  * had its mapping zapped. And freeing these pages
1744                  * requires taking the lru_lock so we do the put_page
1745                  * of the tail pages after the split is complete.
1746                  */
1747                 put_page(page_tail);
1748         }
1749
1750         /*
1751          * Only the head page (now become a regular page) is required
1752          * to be pinned by the caller.
1753          */
1754         BUG_ON(page_count(page) <= 0);
1755 }
1756
1757 static int __split_huge_page_map(struct page *page,
1758                                  struct vm_area_struct *vma,
1759                                  unsigned long address)
1760 {
1761         struct mm_struct *mm = vma->vm_mm;
1762         spinlock_t *ptl;
1763         pmd_t *pmd, _pmd;
1764         int ret = 0, i;
1765         pgtable_t pgtable;
1766         unsigned long haddr;
1767
1768         pmd = page_check_address_pmd(page, mm, address,
1769                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1770         if (pmd) {
1771                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1772                 pmd_populate(mm, &_pmd, pgtable);
1773                 if (pmd_write(*pmd))
1774                         BUG_ON(page_mapcount(page) != 1);
1775
1776                 haddr = address;
1777                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1778                         pte_t *pte, entry;
1779                         BUG_ON(PageCompound(page+i));
1780                         /*
1781                          * Note that NUMA hinting access restrictions are not
1782                          * transferred to avoid any possibility of altering
1783                          * permissions across VMAs.
1784                          */
1785                         entry = mk_pte(page + i, vma->vm_page_prot);
1786                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1787                         if (!pmd_write(*pmd))
1788                                 entry = pte_wrprotect(entry);
1789                         if (!pmd_young(*pmd))
1790                                 entry = pte_mkold(entry);
1791                         pte = pte_offset_map(&_pmd, haddr);
1792                         BUG_ON(!pte_none(*pte));
1793                         set_pte_at(mm, haddr, pte, entry);
1794                         pte_unmap(pte);
1795                 }
1796
1797                 smp_wmb(); /* make pte visible before pmd */
1798                 /*
1799                  * Up to this point the pmd is present and huge and
1800                  * userland has the whole access to the hugepage
1801                  * during the split (which happens in place). If we
1802                  * overwrite the pmd with the not-huge version
1803                  * pointing to the pte here (which of course we could
1804                  * if all CPUs were bug free), userland could trigger
1805                  * a small page size TLB miss on the small sized TLB
1806                  * while the hugepage TLB entry is still established
1807                  * in the huge TLB. Some CPU doesn't like that. See
1808                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1809                  * Erratum 383 on page 93. Intel should be safe but is
1810                  * also warns that it's only safe if the permission
1811                  * and cache attributes of the two entries loaded in
1812                  * the two TLB is identical (which should be the case
1813                  * here). But it is generally safer to never allow
1814                  * small and huge TLB entries for the same virtual
1815                  * address to be loaded simultaneously. So instead of
1816                  * doing "pmd_populate(); flush_tlb_range();" we first
1817                  * mark the current pmd notpresent (atomically because
1818                  * here the pmd_trans_huge and pmd_trans_splitting
1819                  * must remain set at all times on the pmd until the
1820                  * split is complete for this pmd), then we flush the
1821                  * SMP TLB and finally we write the non-huge version
1822                  * of the pmd entry with pmd_populate.
1823                  */
1824                 pmdp_invalidate(vma, address, pmd);
1825                 pmd_populate(mm, pmd, pgtable);
1826                 ret = 1;
1827                 spin_unlock(ptl);
1828         }
1829
1830         return ret;
1831 }
1832
1833 /* must be called with anon_vma->root->rwsem held */
1834 static void __split_huge_page(struct page *page,
1835                               struct anon_vma *anon_vma,
1836                               struct list_head *list)
1837 {
1838         int mapcount, mapcount2;
1839         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1840         struct anon_vma_chain *avc;
1841
1842         BUG_ON(!PageHead(page));
1843         BUG_ON(PageTail(page));
1844
1845         mapcount = 0;
1846         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1847                 struct vm_area_struct *vma = avc->vma;
1848                 unsigned long addr = vma_address(page, vma);
1849                 BUG_ON(is_vma_temporary_stack(vma));
1850                 mapcount += __split_huge_page_splitting(page, vma, addr);
1851         }
1852         /*
1853          * It is critical that new vmas are added to the tail of the
1854          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1855          * and establishes a child pmd before
1856          * __split_huge_page_splitting() freezes the parent pmd (so if
1857          * we fail to prevent copy_huge_pmd() from running until the
1858          * whole __split_huge_page() is complete), we will still see
1859          * the newly established pmd of the child later during the
1860          * walk, to be able to set it as pmd_trans_splitting too.
1861          */
1862         if (mapcount != page_mapcount(page)) {
1863                 pr_err("mapcount %d page_mapcount %d\n",
1864                         mapcount, page_mapcount(page));
1865                 BUG();
1866         }
1867
1868         __split_huge_page_refcount(page, list);
1869
1870         mapcount2 = 0;
1871         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1872                 struct vm_area_struct *vma = avc->vma;
1873                 unsigned long addr = vma_address(page, vma);
1874                 BUG_ON(is_vma_temporary_stack(vma));
1875                 mapcount2 += __split_huge_page_map(page, vma, addr);
1876         }
1877         if (mapcount != mapcount2) {
1878                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1879                         mapcount, mapcount2, page_mapcount(page));
1880                 BUG();
1881         }
1882 }
1883
1884 /*
1885  * Split a hugepage into normal pages. This doesn't change the position of head
1886  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1887  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1888  * from the hugepage.
1889  * Return 0 if the hugepage is split successfully otherwise return 1.
1890  */
1891 int split_huge_page_to_list(struct page *page, struct list_head *list)
1892 {
1893         struct anon_vma *anon_vma;
1894         int ret = 1;
1895
1896         BUG_ON(is_huge_zero_page(page));
1897         BUG_ON(!PageAnon(page));
1898
1899         /*
1900          * The caller does not necessarily hold an mmap_sem that would prevent
1901          * the anon_vma disappearing so we first we take a reference to it
1902          * and then lock the anon_vma for write. This is similar to
1903          * page_lock_anon_vma_read except the write lock is taken to serialise
1904          * against parallel split or collapse operations.
1905          */
1906         anon_vma = page_get_anon_vma(page);
1907         if (!anon_vma)
1908                 goto out;
1909         anon_vma_lock_write(anon_vma);
1910
1911         ret = 0;
1912         if (!PageCompound(page))
1913                 goto out_unlock;
1914
1915         BUG_ON(!PageSwapBacked(page));
1916         __split_huge_page(page, anon_vma, list);
1917         count_vm_event(THP_SPLIT);
1918
1919         BUG_ON(PageCompound(page));
1920 out_unlock:
1921         anon_vma_unlock_write(anon_vma);
1922         put_anon_vma(anon_vma);
1923 out:
1924         return ret;
1925 }
1926
1927 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1928
1929 int hugepage_madvise(struct vm_area_struct *vma,
1930                      unsigned long *vm_flags, int advice)
1931 {
1932         switch (advice) {
1933         case MADV_HUGEPAGE:
1934 #ifdef CONFIG_S390
1935                 /*
1936                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1937                  * can't handle this properly after s390_enable_sie, so we simply
1938                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1939                  */
1940                 if (mm_has_pgste(vma->vm_mm))
1941                         return 0;
1942 #endif
1943                 /*
1944                  * Be somewhat over-protective like KSM for now!
1945                  */
1946                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1947                         return -EINVAL;
1948                 *vm_flags &= ~VM_NOHUGEPAGE;
1949                 *vm_flags |= VM_HUGEPAGE;
1950                 /*
1951                  * If the vma become good for khugepaged to scan,
1952                  * register it here without waiting a page fault that
1953                  * may not happen any time soon.
1954                  */
1955                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1956                         return -ENOMEM;
1957                 break;
1958         case MADV_NOHUGEPAGE:
1959                 /*
1960                  * Be somewhat over-protective like KSM for now!
1961                  */
1962                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1963                         return -EINVAL;
1964                 *vm_flags &= ~VM_HUGEPAGE;
1965                 *vm_flags |= VM_NOHUGEPAGE;
1966                 /*
1967                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1968                  * this vma even if we leave the mm registered in khugepaged if
1969                  * it got registered before VM_NOHUGEPAGE was set.
1970                  */
1971                 break;
1972         }
1973
1974         return 0;
1975 }
1976
1977 static int __init khugepaged_slab_init(void)
1978 {
1979         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1980                                           sizeof(struct mm_slot),
1981                                           __alignof__(struct mm_slot), 0, NULL);
1982         if (!mm_slot_cache)
1983                 return -ENOMEM;
1984
1985         return 0;
1986 }
1987
1988 static void __init khugepaged_slab_exit(void)
1989 {
1990         kmem_cache_destroy(mm_slot_cache);
1991 }
1992
1993 static inline struct mm_slot *alloc_mm_slot(void)
1994 {
1995         if (!mm_slot_cache)     /* initialization failed */
1996                 return NULL;
1997         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1998 }
1999
2000 static inline void free_mm_slot(struct mm_slot *mm_slot)
2001 {
2002         kmem_cache_free(mm_slot_cache, mm_slot);
2003 }
2004
2005 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2006 {
2007         struct mm_slot *mm_slot;
2008
2009         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2010                 if (mm == mm_slot->mm)
2011                         return mm_slot;
2012
2013         return NULL;
2014 }
2015
2016 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2017                                     struct mm_slot *mm_slot)
2018 {
2019         mm_slot->mm = mm;
2020         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2021 }
2022
2023 static inline int khugepaged_test_exit(struct mm_struct *mm)
2024 {
2025         return atomic_read(&mm->mm_users) == 0;
2026 }
2027
2028 int __khugepaged_enter(struct mm_struct *mm)
2029 {
2030         struct mm_slot *mm_slot;
2031         int wakeup;
2032
2033         mm_slot = alloc_mm_slot();
2034         if (!mm_slot)
2035                 return -ENOMEM;
2036
2037         /* __khugepaged_exit() must not run from under us */
2038         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2039         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2040                 free_mm_slot(mm_slot);
2041                 return 0;
2042         }
2043
2044         spin_lock(&khugepaged_mm_lock);
2045         insert_to_mm_slots_hash(mm, mm_slot);
2046         /*
2047          * Insert just behind the scanning cursor, to let the area settle
2048          * down a little.
2049          */
2050         wakeup = list_empty(&khugepaged_scan.mm_head);
2051         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2052         spin_unlock(&khugepaged_mm_lock);
2053
2054         atomic_inc(&mm->mm_count);
2055         if (wakeup)
2056                 wake_up_interruptible(&khugepaged_wait);
2057
2058         return 0;
2059 }
2060
2061 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2062                                unsigned long vm_flags)
2063 {
2064         unsigned long hstart, hend;
2065         if (!vma->anon_vma)
2066                 /*
2067                  * Not yet faulted in so we will register later in the
2068                  * page fault if needed.
2069                  */
2070                 return 0;
2071         if (vma->vm_ops)
2072                 /* khugepaged not yet working on file or special mappings */
2073                 return 0;
2074         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2075         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2076         hend = vma->vm_end & HPAGE_PMD_MASK;
2077         if (hstart < hend)
2078                 return khugepaged_enter(vma, vm_flags);
2079         return 0;
2080 }
2081
2082 void __khugepaged_exit(struct mm_struct *mm)
2083 {
2084         struct mm_slot *mm_slot;
2085         int free = 0;
2086
2087         spin_lock(&khugepaged_mm_lock);
2088         mm_slot = get_mm_slot(mm);
2089         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2090                 hash_del(&mm_slot->hash);
2091                 list_del(&mm_slot->mm_node);
2092                 free = 1;
2093         }
2094         spin_unlock(&khugepaged_mm_lock);
2095
2096         if (free) {
2097                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2098                 free_mm_slot(mm_slot);
2099                 mmdrop(mm);
2100         } else if (mm_slot) {
2101                 /*
2102                  * This is required to serialize against
2103                  * khugepaged_test_exit() (which is guaranteed to run
2104                  * under mmap sem read mode). Stop here (after we
2105                  * return all pagetables will be destroyed) until
2106                  * khugepaged has finished working on the pagetables
2107                  * under the mmap_sem.
2108                  */
2109                 down_write(&mm->mmap_sem);
2110                 up_write(&mm->mmap_sem);
2111         }
2112 }
2113
2114 static void release_pte_page(struct page *page)
2115 {
2116         /* 0 stands for page_is_file_cache(page) == false */
2117         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2118         unlock_page(page);
2119         putback_lru_page(page);
2120 }
2121
2122 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2123 {
2124         while (--_pte >= pte) {
2125                 pte_t pteval = *_pte;
2126                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2127                         release_pte_page(pte_page(pteval));
2128         }
2129 }
2130
2131 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2132                                         unsigned long address,
2133                                         pte_t *pte)
2134 {
2135         struct page *page;
2136         pte_t *_pte;
2137         int none_or_zero = 0;
2138         bool referenced = false, writable = false;
2139         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2140              _pte++, address += PAGE_SIZE) {
2141                 pte_t pteval = *_pte;
2142                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2143                         if (++none_or_zero <= khugepaged_max_ptes_none)
2144                                 continue;
2145                         else
2146                                 goto out;
2147                 }
2148                 if (!pte_present(pteval))
2149                         goto out;
2150                 page = vm_normal_page(vma, address, pteval);
2151                 if (unlikely(!page))
2152                         goto out;
2153
2154                 VM_BUG_ON_PAGE(PageCompound(page), page);
2155                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2156                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2157
2158                 /*
2159                  * We can do it before isolate_lru_page because the
2160                  * page can't be freed from under us. NOTE: PG_lock
2161                  * is needed to serialize against split_huge_page
2162                  * when invoked from the VM.
2163                  */
2164                 if (!trylock_page(page))
2165                         goto out;
2166
2167                 /*
2168                  * cannot use mapcount: can't collapse if there's a gup pin.
2169                  * The page must only be referenced by the scanned process
2170                  * and page swap cache.
2171                  */
2172                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2173                         unlock_page(page);
2174                         goto out;
2175                 }
2176                 if (pte_write(pteval)) {
2177                         writable = true;
2178                 } else {
2179                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2180                                 unlock_page(page);
2181                                 goto out;
2182                         }
2183                         /*
2184                          * Page is not in the swap cache. It can be collapsed
2185                          * into a THP.
2186                          */
2187                 }
2188
2189                 /*
2190                  * Isolate the page to avoid collapsing an hugepage
2191                  * currently in use by the VM.
2192                  */
2193                 if (isolate_lru_page(page)) {
2194                         unlock_page(page);
2195                         goto out;
2196                 }
2197                 /* 0 stands for page_is_file_cache(page) == false */
2198                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2199                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2200                 VM_BUG_ON_PAGE(PageLRU(page), page);
2201
2202                 /* If there is no mapped pte young don't collapse the page */
2203                 if (pte_young(pteval) || PageReferenced(page) ||
2204                     mmu_notifier_test_young(vma->vm_mm, address))
2205                         referenced = true;
2206         }
2207         if (likely(referenced && writable))
2208                 return 1;
2209 out:
2210         release_pte_pages(pte, _pte);
2211         return 0;
2212 }
2213
2214 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2215                                       struct vm_area_struct *vma,
2216                                       unsigned long address,
2217                                       spinlock_t *ptl)
2218 {
2219         pte_t *_pte;
2220         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2221                 pte_t pteval = *_pte;
2222                 struct page *src_page;
2223
2224                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2225                         clear_user_highpage(page, address);
2226                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2227                         if (is_zero_pfn(pte_pfn(pteval))) {
2228                                 /*
2229                                  * ptl mostly unnecessary.
2230                                  */
2231                                 spin_lock(ptl);
2232                                 /*
2233                                  * paravirt calls inside pte_clear here are
2234                                  * superfluous.
2235                                  */
2236                                 pte_clear(vma->vm_mm, address, _pte);
2237                                 spin_unlock(ptl);
2238                         }
2239                 } else {
2240                         src_page = pte_page(pteval);
2241                         copy_user_highpage(page, src_page, address, vma);
2242                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2243                         release_pte_page(src_page);
2244                         /*
2245                          * ptl mostly unnecessary, but preempt has to
2246                          * be disabled to update the per-cpu stats
2247                          * inside page_remove_rmap().
2248                          */
2249                         spin_lock(ptl);
2250                         /*
2251                          * paravirt calls inside pte_clear here are
2252                          * superfluous.
2253                          */
2254                         pte_clear(vma->vm_mm, address, _pte);
2255                         page_remove_rmap(src_page);
2256                         spin_unlock(ptl);
2257                         free_page_and_swap_cache(src_page);
2258                 }
2259
2260                 address += PAGE_SIZE;
2261                 page++;
2262         }
2263 }
2264
2265 static void khugepaged_alloc_sleep(void)
2266 {
2267         wait_event_freezable_timeout(khugepaged_wait, false,
2268                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2269 }
2270
2271 static int khugepaged_node_load[MAX_NUMNODES];
2272
2273 static bool khugepaged_scan_abort(int nid)
2274 {
2275         int i;
2276
2277         /*
2278          * If zone_reclaim_mode is disabled, then no extra effort is made to
2279          * allocate memory locally.
2280          */
2281         if (!zone_reclaim_mode)
2282                 return false;
2283
2284         /* If there is a count for this node already, it must be acceptable */
2285         if (khugepaged_node_load[nid])
2286                 return false;
2287
2288         for (i = 0; i < MAX_NUMNODES; i++) {
2289                 if (!khugepaged_node_load[i])
2290                         continue;
2291                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2292                         return true;
2293         }
2294         return false;
2295 }
2296
2297 #ifdef CONFIG_NUMA
2298 static int khugepaged_find_target_node(void)
2299 {
2300         static int last_khugepaged_target_node = NUMA_NO_NODE;
2301         int nid, target_node = 0, max_value = 0;
2302
2303         /* find first node with max normal pages hit */
2304         for (nid = 0; nid < MAX_NUMNODES; nid++)
2305                 if (khugepaged_node_load[nid] > max_value) {
2306                         max_value = khugepaged_node_load[nid];
2307                         target_node = nid;
2308                 }
2309
2310         /* do some balance if several nodes have the same hit record */
2311         if (target_node <= last_khugepaged_target_node)
2312                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2313                                 nid++)
2314                         if (max_value == khugepaged_node_load[nid]) {
2315                                 target_node = nid;
2316                                 break;
2317                         }
2318
2319         last_khugepaged_target_node = target_node;
2320         return target_node;
2321 }
2322
2323 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2324 {
2325         if (IS_ERR(*hpage)) {
2326                 if (!*wait)
2327                         return false;
2328
2329                 *wait = false;
2330                 *hpage = NULL;
2331                 khugepaged_alloc_sleep();
2332         } else if (*hpage) {
2333                 put_page(*hpage);
2334                 *hpage = NULL;
2335         }
2336
2337         return true;
2338 }
2339
2340 static struct page *
2341 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2342                        struct vm_area_struct *vma, unsigned long address,
2343                        int node)
2344 {
2345         VM_BUG_ON_PAGE(*hpage, *hpage);
2346
2347         /*
2348          * Before allocating the hugepage, release the mmap_sem read lock.
2349          * The allocation can take potentially a long time if it involves
2350          * sync compaction, and we do not need to hold the mmap_sem during
2351          * that. We will recheck the vma after taking it again in write mode.
2352          */
2353         up_read(&mm->mmap_sem);
2354
2355         *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2356         if (unlikely(!*hpage)) {
2357                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2358                 *hpage = ERR_PTR(-ENOMEM);
2359                 return NULL;
2360         }
2361
2362         count_vm_event(THP_COLLAPSE_ALLOC);
2363         return *hpage;
2364 }
2365 #else
2366 static int khugepaged_find_target_node(void)
2367 {
2368         return 0;
2369 }
2370
2371 static inline struct page *alloc_hugepage(int defrag)
2372 {
2373         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2374                            HPAGE_PMD_ORDER);
2375 }
2376
2377 static struct page *khugepaged_alloc_hugepage(bool *wait)
2378 {
2379         struct page *hpage;
2380
2381         do {
2382                 hpage = alloc_hugepage(khugepaged_defrag());
2383                 if (!hpage) {
2384                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2385                         if (!*wait)
2386                                 return NULL;
2387
2388                         *wait = false;
2389                         khugepaged_alloc_sleep();
2390                 } else
2391                         count_vm_event(THP_COLLAPSE_ALLOC);
2392         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2393
2394         return hpage;
2395 }
2396
2397 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2398 {
2399         if (!*hpage)
2400                 *hpage = khugepaged_alloc_hugepage(wait);
2401
2402         if (unlikely(!*hpage))
2403                 return false;
2404
2405         return true;
2406 }
2407
2408 static struct page *
2409 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2410                        struct vm_area_struct *vma, unsigned long address,
2411                        int node)
2412 {
2413         up_read(&mm->mmap_sem);
2414         VM_BUG_ON(!*hpage);
2415
2416         return  *hpage;
2417 }
2418 #endif
2419
2420 static bool hugepage_vma_check(struct vm_area_struct *vma)
2421 {
2422         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2423             (vma->vm_flags & VM_NOHUGEPAGE))
2424                 return false;
2425
2426         if (!vma->anon_vma || vma->vm_ops)
2427                 return false;
2428         if (is_vma_temporary_stack(vma))
2429                 return false;
2430         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2431         return true;
2432 }
2433
2434 static void collapse_huge_page(struct mm_struct *mm,
2435                                    unsigned long address,
2436                                    struct page **hpage,
2437                                    struct vm_area_struct *vma,
2438                                    int node)
2439 {
2440         pmd_t *pmd, _pmd;
2441         pte_t *pte;
2442         pgtable_t pgtable;
2443         struct page *new_page;
2444         spinlock_t *pmd_ptl, *pte_ptl;
2445         int isolated;
2446         unsigned long hstart, hend;
2447         struct mem_cgroup *memcg;
2448         unsigned long mmun_start;       /* For mmu_notifiers */
2449         unsigned long mmun_end;         /* For mmu_notifiers */
2450         gfp_t gfp;
2451
2452         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2453
2454         /* Only allocate from the target node */
2455         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2456                 __GFP_THISNODE;
2457
2458         /* release the mmap_sem read lock. */
2459         new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2460         if (!new_page)
2461                 return;
2462
2463         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2464                                            gfp, &memcg)))
2465                 return;
2466
2467         /*
2468          * Prevent all access to pagetables with the exception of
2469          * gup_fast later hanlded by the ptep_clear_flush and the VM
2470          * handled by the anon_vma lock + PG_lock.
2471          */
2472         down_write(&mm->mmap_sem);
2473         if (unlikely(khugepaged_test_exit(mm)))
2474                 goto out;
2475
2476         vma = find_vma(mm, address);
2477         if (!vma)
2478                 goto out;
2479         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2480         hend = vma->vm_end & HPAGE_PMD_MASK;
2481         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2482                 goto out;
2483         if (!hugepage_vma_check(vma))
2484                 goto out;
2485         pmd = mm_find_pmd(mm, address);
2486         if (!pmd)
2487                 goto out;
2488
2489         anon_vma_lock_write(vma->anon_vma);
2490
2491         pte = pte_offset_map(pmd, address);
2492         pte_ptl = pte_lockptr(mm, pmd);
2493
2494         mmun_start = address;
2495         mmun_end   = address + HPAGE_PMD_SIZE;
2496         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2497         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2498         /*
2499          * After this gup_fast can't run anymore. This also removes
2500          * any huge TLB entry from the CPU so we won't allow
2501          * huge and small TLB entries for the same virtual address
2502          * to avoid the risk of CPU bugs in that area.
2503          */
2504         _pmd = pmdp_clear_flush(vma, address, pmd);
2505         spin_unlock(pmd_ptl);
2506         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2507
2508         spin_lock(pte_ptl);
2509         isolated = __collapse_huge_page_isolate(vma, address, pte);
2510         spin_unlock(pte_ptl);
2511
2512         if (unlikely(!isolated)) {
2513                 pte_unmap(pte);
2514                 spin_lock(pmd_ptl);
2515                 BUG_ON(!pmd_none(*pmd));
2516                 /*
2517                  * We can only use set_pmd_at when establishing
2518                  * hugepmds and never for establishing regular pmds that
2519                  * points to regular pagetables. Use pmd_populate for that
2520                  */
2521                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2522                 spin_unlock(pmd_ptl);
2523                 anon_vma_unlock_write(vma->anon_vma);
2524                 goto out;
2525         }
2526
2527         /*
2528          * All pages are isolated and locked so anon_vma rmap
2529          * can't run anymore.
2530          */
2531         anon_vma_unlock_write(vma->anon_vma);
2532
2533         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2534         pte_unmap(pte);
2535         __SetPageUptodate(new_page);
2536         pgtable = pmd_pgtable(_pmd);
2537
2538         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2539         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2540
2541         /*
2542          * spin_lock() below is not the equivalent of smp_wmb(), so
2543          * this is needed to avoid the copy_huge_page writes to become
2544          * visible after the set_pmd_at() write.
2545          */
2546         smp_wmb();
2547
2548         spin_lock(pmd_ptl);
2549         BUG_ON(!pmd_none(*pmd));
2550         page_add_new_anon_rmap(new_page, vma, address);
2551         mem_cgroup_commit_charge(new_page, memcg, false);
2552         lru_cache_add_active_or_unevictable(new_page, vma);
2553         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2554         set_pmd_at(mm, address, pmd, _pmd);
2555         update_mmu_cache_pmd(vma, address, pmd);
2556         spin_unlock(pmd_ptl);
2557
2558         *hpage = NULL;
2559
2560         khugepaged_pages_collapsed++;
2561 out_up_write:
2562         up_write(&mm->mmap_sem);
2563         return;
2564
2565 out:
2566         mem_cgroup_cancel_charge(new_page, memcg);
2567         goto out_up_write;
2568 }
2569
2570 static int khugepaged_scan_pmd(struct mm_struct *mm,
2571                                struct vm_area_struct *vma,
2572                                unsigned long address,
2573                                struct page **hpage)
2574 {
2575         pmd_t *pmd;
2576         pte_t *pte, *_pte;
2577         int ret = 0, none_or_zero = 0;
2578         struct page *page;
2579         unsigned long _address;
2580         spinlock_t *ptl;
2581         int node = NUMA_NO_NODE;
2582         bool writable = false, referenced = false;
2583
2584         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2585
2586         pmd = mm_find_pmd(mm, address);
2587         if (!pmd)
2588                 goto out;
2589
2590         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2591         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2592         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2593              _pte++, _address += PAGE_SIZE) {
2594                 pte_t pteval = *_pte;
2595                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2596                         if (++none_or_zero <= khugepaged_max_ptes_none)
2597                                 continue;
2598                         else
2599                                 goto out_unmap;
2600                 }
2601                 if (!pte_present(pteval))
2602                         goto out_unmap;
2603                 if (pte_write(pteval))
2604                         writable = true;
2605
2606                 page = vm_normal_page(vma, _address, pteval);
2607                 if (unlikely(!page))
2608                         goto out_unmap;
2609                 /*
2610                  * Record which node the original page is from and save this
2611                  * information to khugepaged_node_load[].
2612                  * Khupaged will allocate hugepage from the node has the max
2613                  * hit record.
2614                  */
2615                 node = page_to_nid(page);
2616                 if (khugepaged_scan_abort(node))
2617                         goto out_unmap;
2618                 khugepaged_node_load[node]++;
2619                 VM_BUG_ON_PAGE(PageCompound(page), page);
2620                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2621                         goto out_unmap;
2622                 /*
2623                  * cannot use mapcount: can't collapse if there's a gup pin.
2624                  * The page must only be referenced by the scanned process
2625                  * and page swap cache.
2626                  */
2627                 if (page_count(page) != 1 + !!PageSwapCache(page))
2628                         goto out_unmap;
2629                 if (pte_young(pteval) || PageReferenced(page) ||
2630                     mmu_notifier_test_young(vma->vm_mm, address))
2631                         referenced = true;
2632         }
2633         if (referenced && writable)
2634                 ret = 1;
2635 out_unmap:
2636         pte_unmap_unlock(pte, ptl);
2637         if (ret) {
2638                 node = khugepaged_find_target_node();
2639                 /* collapse_huge_page will return with the mmap_sem released */
2640                 collapse_huge_page(mm, address, hpage, vma, node);
2641         }
2642 out:
2643         return ret;
2644 }
2645
2646 static void collect_mm_slot(struct mm_slot *mm_slot)
2647 {
2648         struct mm_struct *mm = mm_slot->mm;
2649
2650         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2651
2652         if (khugepaged_test_exit(mm)) {
2653                 /* free mm_slot */
2654                 hash_del(&mm_slot->hash);
2655                 list_del(&mm_slot->mm_node);
2656
2657                 /*
2658                  * Not strictly needed because the mm exited already.
2659                  *
2660                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2661                  */
2662
2663                 /* khugepaged_mm_lock actually not necessary for the below */
2664                 free_mm_slot(mm_slot);
2665                 mmdrop(mm);
2666         }
2667 }
2668
2669 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2670                                             struct page **hpage)
2671         __releases(&khugepaged_mm_lock)
2672         __acquires(&khugepaged_mm_lock)
2673 {
2674         struct mm_slot *mm_slot;
2675         struct mm_struct *mm;
2676         struct vm_area_struct *vma;
2677         int progress = 0;
2678
2679         VM_BUG_ON(!pages);
2680         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2681
2682         if (khugepaged_scan.mm_slot)
2683                 mm_slot = khugepaged_scan.mm_slot;
2684         else {
2685                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2686                                      struct mm_slot, mm_node);
2687                 khugepaged_scan.address = 0;
2688                 khugepaged_scan.mm_slot = mm_slot;
2689         }
2690         spin_unlock(&khugepaged_mm_lock);
2691
2692         mm = mm_slot->mm;
2693         down_read(&mm->mmap_sem);
2694         if (unlikely(khugepaged_test_exit(mm)))
2695                 vma = NULL;
2696         else
2697                 vma = find_vma(mm, khugepaged_scan.address);
2698
2699         progress++;
2700         for (; vma; vma = vma->vm_next) {
2701                 unsigned long hstart, hend;
2702
2703                 cond_resched();
2704                 if (unlikely(khugepaged_test_exit(mm))) {
2705                         progress++;
2706                         break;
2707                 }
2708                 if (!hugepage_vma_check(vma)) {
2709 skip:
2710                         progress++;
2711                         continue;
2712                 }
2713                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2714                 hend = vma->vm_end & HPAGE_PMD_MASK;
2715                 if (hstart >= hend)
2716                         goto skip;
2717                 if (khugepaged_scan.address > hend)
2718                         goto skip;
2719                 if (khugepaged_scan.address < hstart)
2720                         khugepaged_scan.address = hstart;
2721                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2722
2723                 while (khugepaged_scan.address < hend) {
2724                         int ret;
2725                         cond_resched();
2726                         if (unlikely(khugepaged_test_exit(mm)))
2727                                 goto breakouterloop;
2728
2729                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2730                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2731                                   hend);
2732                         ret = khugepaged_scan_pmd(mm, vma,
2733                                                   khugepaged_scan.address,
2734                                                   hpage);
2735                         /* move to next address */
2736                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2737                         progress += HPAGE_PMD_NR;
2738                         if (ret)
2739                                 /* we released mmap_sem so break loop */
2740                                 goto breakouterloop_mmap_sem;
2741                         if (progress >= pages)
2742                                 goto breakouterloop;
2743                 }
2744         }
2745 breakouterloop:
2746         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2747 breakouterloop_mmap_sem:
2748
2749         spin_lock(&khugepaged_mm_lock);
2750         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2751         /*
2752          * Release the current mm_slot if this mm is about to die, or
2753          * if we scanned all vmas of this mm.
2754          */
2755         if (khugepaged_test_exit(mm) || !vma) {
2756                 /*
2757                  * Make sure that if mm_users is reaching zero while
2758                  * khugepaged runs here, khugepaged_exit will find
2759                  * mm_slot not pointing to the exiting mm.
2760                  */
2761                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2762                         khugepaged_scan.mm_slot = list_entry(
2763                                 mm_slot->mm_node.next,
2764                                 struct mm_slot, mm_node);
2765                         khugepaged_scan.address = 0;
2766                 } else {
2767                         khugepaged_scan.mm_slot = NULL;
2768                         khugepaged_full_scans++;
2769                 }
2770
2771                 collect_mm_slot(mm_slot);
2772         }
2773
2774         return progress;
2775 }
2776
2777 static int khugepaged_has_work(void)
2778 {
2779         return !list_empty(&khugepaged_scan.mm_head) &&
2780                 khugepaged_enabled();
2781 }
2782
2783 static int khugepaged_wait_event(void)
2784 {
2785         return !list_empty(&khugepaged_scan.mm_head) ||
2786                 kthread_should_stop();
2787 }
2788
2789 static void khugepaged_do_scan(void)
2790 {
2791         struct page *hpage = NULL;
2792         unsigned int progress = 0, pass_through_head = 0;
2793         unsigned int pages = khugepaged_pages_to_scan;
2794         bool wait = true;
2795
2796         barrier(); /* write khugepaged_pages_to_scan to local stack */
2797
2798         while (progress < pages) {
2799                 if (!khugepaged_prealloc_page(&hpage, &wait))
2800                         break;
2801
2802                 cond_resched();
2803
2804                 if (unlikely(kthread_should_stop() || freezing(current)))
2805                         break;
2806
2807                 spin_lock(&khugepaged_mm_lock);
2808                 if (!khugepaged_scan.mm_slot)
2809                         pass_through_head++;
2810                 if (khugepaged_has_work() &&
2811                     pass_through_head < 2)
2812                         progress += khugepaged_scan_mm_slot(pages - progress,
2813                                                             &hpage);
2814                 else
2815                         progress = pages;
2816                 spin_unlock(&khugepaged_mm_lock);
2817         }
2818
2819         if (!IS_ERR_OR_NULL(hpage))
2820                 put_page(hpage);
2821 }
2822
2823 static void khugepaged_wait_work(void)
2824 {
2825         try_to_freeze();
2826
2827         if (khugepaged_has_work()) {
2828                 if (!khugepaged_scan_sleep_millisecs)
2829                         return;
2830
2831                 wait_event_freezable_timeout(khugepaged_wait,
2832                                              kthread_should_stop(),
2833                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2834                 return;
2835         }
2836
2837         if (khugepaged_enabled())
2838                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2839 }
2840
2841 static int khugepaged(void *none)
2842 {
2843         struct mm_slot *mm_slot;
2844
2845         set_freezable();
2846         set_user_nice(current, MAX_NICE);
2847
2848         while (!kthread_should_stop()) {
2849                 khugepaged_do_scan();
2850                 khugepaged_wait_work();
2851         }
2852
2853         spin_lock(&khugepaged_mm_lock);
2854         mm_slot = khugepaged_scan.mm_slot;
2855         khugepaged_scan.mm_slot = NULL;
2856         if (mm_slot)
2857                 collect_mm_slot(mm_slot);
2858         spin_unlock(&khugepaged_mm_lock);
2859         return 0;
2860 }
2861
2862 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2863                 unsigned long haddr, pmd_t *pmd)
2864 {
2865         struct mm_struct *mm = vma->vm_mm;
2866         pgtable_t pgtable;
2867         pmd_t _pmd;
2868         int i;
2869
2870         pmdp_clear_flush_notify(vma, haddr, pmd);
2871         /* leave pmd empty until pte is filled */
2872
2873         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2874         pmd_populate(mm, &_pmd, pgtable);
2875
2876         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2877                 pte_t *pte, entry;
2878                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2879                 entry = pte_mkspecial(entry);
2880                 pte = pte_offset_map(&_pmd, haddr);
2881                 VM_BUG_ON(!pte_none(*pte));
2882                 set_pte_at(mm, haddr, pte, entry);
2883                 pte_unmap(pte);
2884         }
2885         smp_wmb(); /* make pte visible before pmd */
2886         pmd_populate(mm, pmd, pgtable);
2887         put_huge_zero_page();
2888 }
2889
2890 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2891                 pmd_t *pmd)
2892 {
2893         spinlock_t *ptl;
2894         struct page *page;
2895         struct mm_struct *mm = vma->vm_mm;
2896         unsigned long haddr = address & HPAGE_PMD_MASK;
2897         unsigned long mmun_start;       /* For mmu_notifiers */
2898         unsigned long mmun_end;         /* For mmu_notifiers */
2899
2900         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2901
2902         mmun_start = haddr;
2903         mmun_end   = haddr + HPAGE_PMD_SIZE;
2904 again:
2905         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2906         ptl = pmd_lock(mm, pmd);
2907         if (unlikely(!pmd_trans_huge(*pmd))) {
2908                 spin_unlock(ptl);
2909                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2910                 return;
2911         }
2912         if (is_huge_zero_pmd(*pmd)) {
2913                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2914                 spin_unlock(ptl);
2915                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2916                 return;
2917         }
2918         page = pmd_page(*pmd);
2919         VM_BUG_ON_PAGE(!page_count(page), page);
2920         get_page(page);
2921         spin_unlock(ptl);
2922         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2923
2924         split_huge_page(page);
2925
2926         put_page(page);
2927
2928         /*
2929          * We don't always have down_write of mmap_sem here: a racing
2930          * do_huge_pmd_wp_page() might have copied-on-write to another
2931          * huge page before our split_huge_page() got the anon_vma lock.
2932          */
2933         if (unlikely(pmd_trans_huge(*pmd)))
2934                 goto again;
2935 }
2936
2937 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2938                 pmd_t *pmd)
2939 {
2940         struct vm_area_struct *vma;
2941
2942         vma = find_vma(mm, address);
2943         BUG_ON(vma == NULL);
2944         split_huge_page_pmd(vma, address, pmd);
2945 }
2946
2947 static void split_huge_page_address(struct mm_struct *mm,
2948                                     unsigned long address)
2949 {
2950         pgd_t *pgd;
2951         pud_t *pud;
2952         pmd_t *pmd;
2953
2954         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2955
2956         pgd = pgd_offset(mm, address);
2957         if (!pgd_present(*pgd))
2958                 return;
2959
2960         pud = pud_offset(pgd, address);
2961         if (!pud_present(*pud))
2962                 return;
2963
2964         pmd = pmd_offset(pud, address);
2965         if (!pmd_present(*pmd))
2966                 return;
2967         /*
2968          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2969          * materialize from under us.
2970          */
2971         split_huge_page_pmd_mm(mm, address, pmd);
2972 }
2973
2974 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2975                              unsigned long start,
2976                              unsigned long end,
2977                              long adjust_next)
2978 {
2979         /*
2980          * If the new start address isn't hpage aligned and it could
2981          * previously contain an hugepage: check if we need to split
2982          * an huge pmd.
2983          */
2984         if (start & ~HPAGE_PMD_MASK &&
2985             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2986             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2987                 split_huge_page_address(vma->vm_mm, start);
2988
2989         /*
2990          * If the new end address isn't hpage aligned and it could
2991          * previously contain an hugepage: check if we need to split
2992          * an huge pmd.
2993          */
2994         if (end & ~HPAGE_PMD_MASK &&
2995             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2996             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2997                 split_huge_page_address(vma->vm_mm, end);
2998
2999         /*
3000          * If we're also updating the vma->vm_next->vm_start, if the new
3001          * vm_next->vm_start isn't page aligned and it could previously
3002          * contain an hugepage: check if we need to split an huge pmd.
3003          */
3004         if (adjust_next > 0) {
3005                 struct vm_area_struct *next = vma->vm_next;
3006                 unsigned long nstart = next->vm_start;
3007                 nstart += adjust_next << PAGE_SHIFT;
3008                 if (nstart & ~HPAGE_PMD_MASK &&
3009                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3010                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3011                         split_huge_page_address(next->vm_mm, nstart);
3012         }
3013 }