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/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/kthread.h>
  22 #include <linux/khugepaged.h>
  23 #include <linux/freezer.h>
  24 #include <linux/mman.h>
  25 #include <linux/pagemap.h>
  26 #include <linux/migrate.h>
  27 #include <linux/hashtable.h>
  28 #include <linux/userfaultfd_k.h>
  29 #include <linux/page_idle.h>
  30
  31 #include <asm/tlb.h>
  32 #include <asm/pgalloc.h>
  33 #include "internal.h"
  34
  35 enum scan_result {
  36         SCAN_FAIL,
  37         SCAN_SUCCEED,
  38         SCAN_PMD_NULL,
  39         SCAN_EXCEED_NONE_PTE,
  40         SCAN_PTE_NON_PRESENT,
  41         SCAN_PAGE_RO,
  42         SCAN_NO_REFERENCED_PAGE,
  43         SCAN_PAGE_NULL,
  44         SCAN_SCAN_ABORT,
  45         SCAN_PAGE_COUNT,
  46         SCAN_PAGE_LRU,
  47         SCAN_PAGE_LOCK,
  48         SCAN_PAGE_ANON,
  49         SCAN_PAGE_COMPOUND,
  50         SCAN_ANY_PROCESS,
  51         SCAN_VMA_NULL,
  52         SCAN_VMA_CHECK,
  53         SCAN_ADDRESS_RANGE,
  54         SCAN_SWAP_CACHE_PAGE,
  55         SCAN_DEL_PAGE_LRU,
  56         SCAN_ALLOC_HUGE_PAGE_FAIL,
  57         SCAN_CGROUP_CHARGE_FAIL
  58 };
  59
  60 #define CREATE_TRACE_POINTS
  61 #include <trace/events/huge_memory.h>
  62
  63 /*
  64  * By default transparent hugepage support is disabled in order that avoid
  65  * to risk increase the memory footprint of applications without a guaranteed
  66  * benefit. When transparent hugepage support is enabled, is for all mappings,
  67  * and khugepaged scans all mappings.
  68  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  69  * for all hugepage allocations.
  70  */
  71 unsigned long transparent_hugepage_flags __read_mostly =
  72 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  73         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  74 #endif
  75 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  76         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  77 #endif
  78         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  79         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  80         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  81
  82 /* default scan 8*512 pte (or vmas) every 30 second */
  83 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  84 static unsigned int khugepaged_pages_collapsed;
  85 static unsigned int khugepaged_full_scans;
  86 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  87 /* during fragmentation poll the hugepage allocator once every minute */
  88 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  89 static struct task_struct *khugepaged_thread __read_mostly;
  90 static DEFINE_MUTEX(khugepaged_mutex);
  91 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  92 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  93 /*
  94  * default collapse hugepages if there is at least one pte mapped like
  95  * it would have happened if the vma was large enough during page
  96  * fault.
  97  */
  98 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  99
 100 static int khugepaged(void *none);
 101 static int khugepaged_slab_init(void);
 102 static void khugepaged_slab_exit(void);
 103
 104 #define MM_SLOTS_HASH_BITS 10
 105 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 106
 107 static struct kmem_cache *mm_slot_cache __read_mostly;
 108
 109 /**
 110  * struct mm_slot - hash lookup from mm to mm_slot
 111  * @hash: hash collision list
 112  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 113  * @mm: the mm that this information is valid for
 114  */
 115 struct mm_slot {
 116         struct hlist_node hash;
 117         struct list_head mm_node;
 118         struct mm_struct *mm;
 119 };
 120
 121 /**
 122  * struct khugepaged_scan - cursor for scanning
 123  * @mm_head: the head of the mm list to scan
 124  * @mm_slot: the current mm_slot we are scanning
 125  * @address: the next address inside that to be scanned
 126  *
 127  * There is only the one khugepaged_scan instance of this cursor structure.
 128  */
 129 struct khugepaged_scan {
 130         struct list_head mm_head;
 131         struct mm_slot *mm_slot;
 132         unsigned long address;
 133 };
 134 static struct khugepaged_scan khugepaged_scan = {
 135         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 136 };
 137
 138 static DEFINE_SPINLOCK(split_queue_lock);
 139 static LIST_HEAD(split_queue);
 140 static unsigned long split_queue_len;
 141 static struct shrinker deferred_split_shrinker;
 142
 143 static void set_recommended_min_free_kbytes(void)
 144 {
 145         struct zone *zone;
 146         int nr_zones = 0;
 147         unsigned long recommended_min;
 148
 149         for_each_populated_zone(zone)
 150                 nr_zones++;
 151
 152         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 153         recommended_min = pageblock_nr_pages * nr_zones * 2;
 154
 155         /*
 156          * Make sure that on average at least two pageblocks are almost free
 157          * of another type, one for a migratetype to fall back to and a
 158          * second to avoid subsequent fallbacks of other types There are 3
 159          * MIGRATE_TYPES we care about.
 160          */
 161         recommended_min += pageblock_nr_pages * nr_zones *
 162                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 163
 164         /* don't ever allow to reserve more than 5% of the lowmem */
 165         recommended_min = min(recommended_min,
 166                               (unsigned long) nr_free_buffer_pages() / 20);
 167         recommended_min <<= (PAGE_SHIFT-10);
 168
 169         if (recommended_min > min_free_kbytes) {
 170                 if (user_min_free_kbytes >= 0)
 171                         pr_info("raising min_free_kbytes from %d to %lu "
 172                                 "to help transparent hugepage allocations\n",
 173                                 min_free_kbytes, recommended_min);
 174
 175                 min_free_kbytes = recommended_min;
 176         }
 177         setup_per_zone_wmarks();
 178 }
 179
 180 static int start_stop_khugepaged(void)
 181 {
 182         int err = 0;
 183         if (khugepaged_enabled()) {
 184                 if (!khugepaged_thread)
 185                         khugepaged_thread = kthread_run(khugepaged, NULL,
 186                                                         "khugepaged");
 187                 if (IS_ERR(khugepaged_thread)) {
 188                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 189                         err = PTR_ERR(khugepaged_thread);
 190                         khugepaged_thread = NULL;
 191                         goto fail;
 192                 }
 193
 194                 if (!list_empty(&khugepaged_scan.mm_head))
 195                         wake_up_interruptible(&khugepaged_wait);
 196
 197                 set_recommended_min_free_kbytes();
 198         } else if (khugepaged_thread) {
 199                 kthread_stop(khugepaged_thread);
 200                 khugepaged_thread = NULL;
 201         }
 202 fail:
 203         return err;
 204 }
 205
 206 static atomic_t huge_zero_refcount;
 207 struct page *huge_zero_page __read_mostly;
 208
 209 struct page *get_huge_zero_page(void)
 210 {
 211         struct page *zero_page;
 212 retry:
 213         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 214                 return READ_ONCE(huge_zero_page);
 215
 216         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 217                         HPAGE_PMD_ORDER);
 218         if (!zero_page) {
 219                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 220                 return NULL;
 221         }
 222         count_vm_event(THP_ZERO_PAGE_ALLOC);
 223         preempt_disable();
 224         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 225                 preempt_enable();
 226                 __free_pages(zero_page, compound_order(zero_page));
 227                 goto retry;
 228         }
 229
 230         /* We take additional reference here. It will be put back by shrinker */
 231         atomic_set(&huge_zero_refcount, 2);
 232         preempt_enable();
 233         return READ_ONCE(huge_zero_page);
 234 }
 235
 236 static void put_huge_zero_page(void)
 237 {
 238         /*
 239          * Counter should never go to zero here. Only shrinker can put
 240          * last reference.
 241          */
 242         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 243 }
 244
 245 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 246                                         struct shrink_control *sc)
 247 {
 248         /* we can free zero page only if last reference remains */
 249         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 250 }
 251
 252 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 253                                        struct shrink_control *sc)
 254 {
 255         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 256                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 257                 BUG_ON(zero_page == NULL);
 258                 __free_pages(zero_page, compound_order(zero_page));
 259                 return HPAGE_PMD_NR;
 260         }
 261
 262         return 0;
 263 }
 264
 265 static struct shrinker huge_zero_page_shrinker = {
 266         .count_objects = shrink_huge_zero_page_count,
 267         .scan_objects = shrink_huge_zero_page_scan,
 268         .seeks = DEFAULT_SEEKS,
 269 };
 270
 271 #ifdef CONFIG_SYSFS
 272
 273 static ssize_t double_flag_show(struct kobject *kobj,
 274                                 struct kobj_attribute *attr, char *buf,
 275                                 enum transparent_hugepage_flag enabled,
 276                                 enum transparent_hugepage_flag req_madv)
 277 {
 278         if (test_bit(enabled, &transparent_hugepage_flags)) {
 279                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 280                 return sprintf(buf, "[always] madvise never\n");
 281         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 282                 return sprintf(buf, "always [madvise] never\n");
 283         else
 284                 return sprintf(buf, "always madvise [never]\n");
 285 }
 286 static ssize_t double_flag_store(struct kobject *kobj,
 287                                  struct kobj_attribute *attr,
 288                                  const char *buf, size_t count,
 289                                  enum transparent_hugepage_flag enabled,
 290                                  enum transparent_hugepage_flag req_madv)
 291 {
 292         if (!memcmp("always", buf,
 293                     min(sizeof("always")-1, count))) {
 294                 set_bit(enabled, &transparent_hugepage_flags);
 295                 clear_bit(req_madv, &transparent_hugepage_flags);
 296         } else if (!memcmp("madvise", buf,
 297                            min(sizeof("madvise")-1, count))) {
 298                 clear_bit(enabled, &transparent_hugepage_flags);
 299                 set_bit(req_madv, &transparent_hugepage_flags);
 300         } else if (!memcmp("never", buf,
 301                            min(sizeof("never")-1, count))) {
 302                 clear_bit(enabled, &transparent_hugepage_flags);
 303                 clear_bit(req_madv, &transparent_hugepage_flags);
 304         } else
 305                 return -EINVAL;
 306
 307         return count;
 308 }
 309
 310 static ssize_t enabled_show(struct kobject *kobj,
 311                             struct kobj_attribute *attr, char *buf)
 312 {
 313         return double_flag_show(kobj, attr, buf,
 314                                 TRANSPARENT_HUGEPAGE_FLAG,
 315                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 316 }
 317 static ssize_t enabled_store(struct kobject *kobj,
 318                              struct kobj_attribute *attr,
 319                              const char *buf, size_t count)
 320 {
 321         ssize_t ret;
 322
 323         ret = double_flag_store(kobj, attr, buf, count,
 324                                 TRANSPARENT_HUGEPAGE_FLAG,
 325                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 326
 327         if (ret > 0) {
 328                 int err;
 329
 330                 mutex_lock(&khugepaged_mutex);
 331                 err = start_stop_khugepaged();
 332                 mutex_unlock(&khugepaged_mutex);
 333
 334                 if (err)
 335                         ret = err;
 336         }
 337
 338         return ret;
 339 }
 340 static struct kobj_attribute enabled_attr =
 341         __ATTR(enabled, 0644, enabled_show, enabled_store);
 342
 343 static ssize_t single_flag_show(struct kobject *kobj,
 344                                 struct kobj_attribute *attr, char *buf,
 345                                 enum transparent_hugepage_flag flag)
 346 {
 347         return sprintf(buf, "%d\n",
 348                        !!test_bit(flag, &transparent_hugepage_flags));
 349 }
 350
 351 static ssize_t single_flag_store(struct kobject *kobj,
 352                                  struct kobj_attribute *attr,
 353                                  const char *buf, size_t count,
 354                                  enum transparent_hugepage_flag flag)
 355 {
 356         unsigned long value;
 357         int ret;
 358
 359         ret = kstrtoul(buf, 10, &value);
 360         if (ret < 0)
 361                 return ret;
 362         if (value > 1)
 363                 return -EINVAL;
 364
 365         if (value)
 366                 set_bit(flag, &transparent_hugepage_flags);
 367         else
 368                 clear_bit(flag, &transparent_hugepage_flags);
 369
 370         return count;
 371 }
 372
 373 /*
 374  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 375  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 376  * memory just to allocate one more hugepage.
 377  */
 378 static ssize_t defrag_show(struct kobject *kobj,
 379                            struct kobj_attribute *attr, char *buf)
 380 {
 381         return double_flag_show(kobj, attr, buf,
 382                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 383                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 384 }
 385 static ssize_t defrag_store(struct kobject *kobj,
 386                             struct kobj_attribute *attr,
 387                             const char *buf, size_t count)
 388 {
 389         return double_flag_store(kobj, attr, buf, count,
 390                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 391                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 392 }
 393 static struct kobj_attribute defrag_attr =
 394         __ATTR(defrag, 0644, defrag_show, defrag_store);
 395
 396 static ssize_t use_zero_page_show(struct kobject *kobj,
 397                 struct kobj_attribute *attr, char *buf)
 398 {
 399         return single_flag_show(kobj, attr, buf,
 400                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 401 }
 402 static ssize_t use_zero_page_store(struct kobject *kobj,
 403                 struct kobj_attribute *attr, const char *buf, size_t count)
 404 {
 405         return single_flag_store(kobj, attr, buf, count,
 406                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 407 }
 408 static struct kobj_attribute use_zero_page_attr =
 409         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 410 #ifdef CONFIG_DEBUG_VM
 411 static ssize_t debug_cow_show(struct kobject *kobj,
 412                                 struct kobj_attribute *attr, char *buf)
 413 {
 414         return single_flag_show(kobj, attr, buf,
 415                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 416 }
 417 static ssize_t debug_cow_store(struct kobject *kobj,
 418                                struct kobj_attribute *attr,
 419                                const char *buf, size_t count)
 420 {
 421         return single_flag_store(kobj, attr, buf, count,
 422                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 423 }
 424 static struct kobj_attribute debug_cow_attr =
 425         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 426 #endif /* CONFIG_DEBUG_VM */
 427
 428 static struct attribute *hugepage_attr[] = {
 429         &enabled_attr.attr,
 430         &defrag_attr.attr,
 431         &use_zero_page_attr.attr,
 432 #ifdef CONFIG_DEBUG_VM
 433         &debug_cow_attr.attr,
 434 #endif
 435         NULL,
 436 };
 437
 438 static struct attribute_group hugepage_attr_group = {
 439         .attrs = hugepage_attr,
 440 };
 441
 442 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 443                                          struct kobj_attribute *attr,
 444                                          char *buf)
 445 {
 446         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 447 }
 448
 449 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 450                                           struct kobj_attribute *attr,
 451                                           const char *buf, size_t count)
 452 {
 453         unsigned long msecs;
 454         int err;
 455
 456         err = kstrtoul(buf, 10, &msecs);
 457         if (err || msecs > UINT_MAX)
 458                 return -EINVAL;
 459
 460         khugepaged_scan_sleep_millisecs = msecs;
 461         wake_up_interruptible(&khugepaged_wait);
 462
 463         return count;
 464 }
 465 static struct kobj_attribute scan_sleep_millisecs_attr =
 466         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 467                scan_sleep_millisecs_store);
 468
 469 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 470                                           struct kobj_attribute *attr,
 471                                           char *buf)
 472 {
 473         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 474 }
 475
 476 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 477                                            struct kobj_attribute *attr,
 478                                            const char *buf, size_t count)
 479 {
 480         unsigned long msecs;
 481         int err;
 482
 483         err = kstrtoul(buf, 10, &msecs);
 484         if (err || msecs > UINT_MAX)
 485                 return -EINVAL;
 486
 487         khugepaged_alloc_sleep_millisecs = msecs;
 488         wake_up_interruptible(&khugepaged_wait);
 489
 490         return count;
 491 }
 492 static struct kobj_attribute alloc_sleep_millisecs_attr =
 493         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 494                alloc_sleep_millisecs_store);
 495
 496 static ssize_t pages_to_scan_show(struct kobject *kobj,
 497                                   struct kobj_attribute *attr,
 498                                   char *buf)
 499 {
 500         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 501 }
 502 static ssize_t pages_to_scan_store(struct kobject *kobj,
 503                                    struct kobj_attribute *attr,
 504                                    const char *buf, size_t count)
 505 {
 506         int err;
 507         unsigned long pages;
 508
 509         err = kstrtoul(buf, 10, &pages);
 510         if (err || !pages || pages > UINT_MAX)
 511                 return -EINVAL;
 512
 513         khugepaged_pages_to_scan = pages;
 514
 515         return count;
 516 }
 517 static struct kobj_attribute pages_to_scan_attr =
 518         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 519                pages_to_scan_store);
 520
 521 static ssize_t pages_collapsed_show(struct kobject *kobj,
 522                                     struct kobj_attribute *attr,
 523                                     char *buf)
 524 {
 525         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 526 }
 527 static struct kobj_attribute pages_collapsed_attr =
 528         __ATTR_RO(pages_collapsed);
 529
 530 static ssize_t full_scans_show(struct kobject *kobj,
 531                                struct kobj_attribute *attr,
 532                                char *buf)
 533 {
 534         return sprintf(buf, "%u\n", khugepaged_full_scans);
 535 }
 536 static struct kobj_attribute full_scans_attr =
 537         __ATTR_RO(full_scans);
 538
 539 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 540                                       struct kobj_attribute *attr, char *buf)
 541 {
 542         return single_flag_show(kobj, attr, buf,
 543                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 544 }
 545 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 546                                        struct kobj_attribute *attr,
 547                                        const char *buf, size_t count)
 548 {
 549         return single_flag_store(kobj, attr, buf, count,
 550                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 551 }
 552 static struct kobj_attribute khugepaged_defrag_attr =
 553         __ATTR(defrag, 0644, khugepaged_defrag_show,
 554                khugepaged_defrag_store);
 555
 556 /*
 557  * max_ptes_none controls if khugepaged should collapse hugepages over
 558  * any unmapped ptes in turn potentially increasing the memory
 559  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 560  * reduce the available free memory in the system as it
 561  * runs. Increasing max_ptes_none will instead potentially reduce the
 562  * free memory in the system during the khugepaged scan.
 563  */
 564 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 565                                              struct kobj_attribute *attr,
 566                                              char *buf)
 567 {
 568         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 569 }
 570 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 571                                               struct kobj_attribute *attr,
 572                                               const char *buf, size_t count)
 573 {
 574         int err;
 575         unsigned long max_ptes_none;
 576
 577         err = kstrtoul(buf, 10, &max_ptes_none);
 578         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 579                 return -EINVAL;
 580
 581         khugepaged_max_ptes_none = max_ptes_none;
 582
 583         return count;
 584 }
 585 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 586         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 587                khugepaged_max_ptes_none_store);
 588
 589 static struct attribute *khugepaged_attr[] = {
 590         &khugepaged_defrag_attr.attr,
 591         &khugepaged_max_ptes_none_attr.attr,
 592         &pages_to_scan_attr.attr,
 593         &pages_collapsed_attr.attr,
 594         &full_scans_attr.attr,
 595         &scan_sleep_millisecs_attr.attr,
 596         &alloc_sleep_millisecs_attr.attr,
 597         NULL,
 598 };
 599
 600 static struct attribute_group khugepaged_attr_group = {
 601         .attrs = khugepaged_attr,
 602         .name = "khugepaged",
 603 };
 604
 605 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 606 {
 607         int err;
 608
 609         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 610         if (unlikely(!*hugepage_kobj)) {
 611                 pr_err("failed to create transparent hugepage kobject\n");
 612                 return -ENOMEM;
 613         }
 614
 615         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 616         if (err) {
 617                 pr_err("failed to register transparent hugepage group\n");
 618                 goto delete_obj;
 619         }
 620
 621         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 622         if (err) {
 623                 pr_err("failed to register transparent hugepage group\n");
 624                 goto remove_hp_group;
 625         }
 626
 627         return 0;
 628
 629 remove_hp_group:
 630         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 631 delete_obj:
 632         kobject_put(*hugepage_kobj);
 633         return err;
 634 }
 635
 636 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 637 {
 638         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 639         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 640         kobject_put(hugepage_kobj);
 641 }
 642 #else
 643 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 644 {
 645         return 0;
 646 }
 647
 648 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 649 {
 650 }
 651 #endif /* CONFIG_SYSFS */
 652
 653 static int __init hugepage_init(void)
 654 {
 655         int err;
 656         struct kobject *hugepage_kobj;
 657
 658         if (!has_transparent_hugepage()) {
 659                 transparent_hugepage_flags = 0;
 660                 return -EINVAL;
 661         }
 662
 663         err = hugepage_init_sysfs(&hugepage_kobj);
 664         if (err)
 665                 goto err_sysfs;
 666
 667         err = khugepaged_slab_init();
 668         if (err)
 669                 goto err_slab;
 670
 671         err = register_shrinker(&huge_zero_page_shrinker);
 672         if (err)
 673                 goto err_hzp_shrinker;
 674         err = register_shrinker(&deferred_split_shrinker);
 675         if (err)
 676                 goto err_split_shrinker;
 677
 678         /*
 679          * By default disable transparent hugepages on smaller systems,
 680          * where the extra memory used could hurt more than TLB overhead
 681          * is likely to save.  The admin can still enable it through /sys.
 682          */
 683         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 684                 transparent_hugepage_flags = 0;
 685                 return 0;
 686         }
 687
 688         err = start_stop_khugepaged();
 689         if (err)
 690                 goto err_khugepaged;
 691
 692         return 0;
 693 err_khugepaged:
 694         unregister_shrinker(&deferred_split_shrinker);
 695 err_split_shrinker:
 696         unregister_shrinker(&huge_zero_page_shrinker);
 697 err_hzp_shrinker:
 698         khugepaged_slab_exit();
 699 err_slab:
 700         hugepage_exit_sysfs(hugepage_kobj);
 701 err_sysfs:
 702         return err;
 703 }
 704 subsys_initcall(hugepage_init);
 705
 706 static int __init setup_transparent_hugepage(char *str)
 707 {
 708         int ret = 0;
 709         if (!str)
 710                 goto out;
 711         if (!strcmp(str, "always")) {
 712                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 713                         &transparent_hugepage_flags);
 714                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 715                           &transparent_hugepage_flags);
 716                 ret = 1;
 717         } else if (!strcmp(str, "madvise")) {
 718                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 719                           &transparent_hugepage_flags);
 720                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 721                         &transparent_hugepage_flags);
 722                 ret = 1;
 723         } else if (!strcmp(str, "never")) {
 724                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 725                           &transparent_hugepage_flags);
 726                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 727                           &transparent_hugepage_flags);
 728                 ret = 1;
 729         }
 730 out:
 731         if (!ret)
 732                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 733         return ret;
 734 }
 735 __setup("transparent_hugepage=", setup_transparent_hugepage);
 736
 737 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 738 {
 739         if (likely(vma->vm_flags & VM_WRITE))
 740                 pmd = pmd_mkwrite(pmd);
 741         return pmd;
 742 }
 743
 744 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 745 {
 746         pmd_t entry;
 747         entry = mk_pmd(page, prot);
 748         entry = pmd_mkhuge(entry);
 749         return entry;
 750 }
 751
 752 static inline struct list_head *page_deferred_list(struct page *page)
 753 {
 754         /*
 755          * ->lru in the tail pages is occupied by compound_head.
 756          * Let's use ->mapping + ->index in the second tail page as list_head.
 757          */
 758         return (struct list_head *)&page[2].mapping;
 759 }
 760
 761 void prep_transhuge_page(struct page *page)
 762 {
 763         /*
 764          * we use page->mapping and page->indexlru in second tail page
 765          * as list_head: assuming THP order >= 2
 766          */
 767         BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 768
 769         INIT_LIST_HEAD(page_deferred_list(page));
 770         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 771 }
 772
 773 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 774                                         struct vm_area_struct *vma,
 775                                         unsigned long address, pmd_t *pmd,
 776                                         struct page *page, gfp_t gfp,
 777                                         unsigned int flags)
 778 {
 779         struct mem_cgroup *memcg;
 780         pgtable_t pgtable;
 781         spinlock_t *ptl;
 782         unsigned long haddr = address & HPAGE_PMD_MASK;
 783
 784         VM_BUG_ON_PAGE(!PageCompound(page), page);
 785
 786         if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
 787                 put_page(page);
 788                 count_vm_event(THP_FAULT_FALLBACK);
 789                 return VM_FAULT_FALLBACK;
 790         }
 791
 792         pgtable = pte_alloc_one(mm, haddr);
 793         if (unlikely(!pgtable)) {
 794                 mem_cgroup_cancel_charge(page, memcg, true);
 795                 put_page(page);
 796                 return VM_FAULT_OOM;
 797         }
 798
 799         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 800         /*
 801          * The memory barrier inside __SetPageUptodate makes sure that
 802          * clear_huge_page writes become visible before the set_pmd_at()
 803          * write.
 804          */
 805         __SetPageUptodate(page);
 806
 807         ptl = pmd_lock(mm, pmd);
 808         if (unlikely(!pmd_none(*pmd))) {
 809                 spin_unlock(ptl);
 810                 mem_cgroup_cancel_charge(page, memcg, true);
 811                 put_page(page);
 812                 pte_free(mm, pgtable);
 813         } else {
 814                 pmd_t entry;
 815
 816                 /* Deliver the page fault to userland */
 817                 if (userfaultfd_missing(vma)) {
 818                         int ret;
 819
 820                         spin_unlock(ptl);
 821                         mem_cgroup_cancel_charge(page, memcg, true);
 822                         put_page(page);
 823                         pte_free(mm, pgtable);
 824                         ret = handle_userfault(vma, address, flags,
 825                                                VM_UFFD_MISSING);
 826                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 827                         return ret;
 828                 }
 829
 830                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 831                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 832                 page_add_new_anon_rmap(page, vma, haddr, true);
 833                 mem_cgroup_commit_charge(page, memcg, false, true);
 834                 lru_cache_add_active_or_unevictable(page, vma);
 835                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 836                 set_pmd_at(mm, haddr, pmd, entry);
 837                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 838                 atomic_long_inc(&mm->nr_ptes);
 839                 spin_unlock(ptl);
 840                 count_vm_event(THP_FAULT_ALLOC);
 841         }
 842
 843         return 0;
 844 }
 845
 846 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 847 {
 848         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 849 }
 850
 851 /* Caller must hold page table lock. */
 852 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 853                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 854                 struct page *zero_page)
 855 {
 856         pmd_t entry;
 857         if (!pmd_none(*pmd))
 858                 return false;
 859         entry = mk_pmd(zero_page, vma->vm_page_prot);
 860         entry = pmd_mkhuge(entry);
 861         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 862         set_pmd_at(mm, haddr, pmd, entry);
 863         atomic_long_inc(&mm->nr_ptes);
 864         return true;
 865 }
 866
 867 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 868                                unsigned long address, pmd_t *pmd,
 869                                unsigned int flags)
 870 {
 871         gfp_t gfp;
 872         struct page *page;
 873         unsigned long haddr = address & HPAGE_PMD_MASK;
 874
 875         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 876                 return VM_FAULT_FALLBACK;
 877         if (unlikely(anon_vma_prepare(vma)))
 878                 return VM_FAULT_OOM;
 879         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 880                 return VM_FAULT_OOM;
 881         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 882                         transparent_hugepage_use_zero_page()) {
 883                 spinlock_t *ptl;
 884                 pgtable_t pgtable;
 885                 struct page *zero_page;
 886                 bool set;
 887                 int ret;
 888                 pgtable = pte_alloc_one(mm, haddr);
 889                 if (unlikely(!pgtable))
 890                         return VM_FAULT_OOM;
 891                 zero_page = get_huge_zero_page();
 892                 if (unlikely(!zero_page)) {
 893                         pte_free(mm, pgtable);
 894                         count_vm_event(THP_FAULT_FALLBACK);
 895                         return VM_FAULT_FALLBACK;
 896                 }
 897                 ptl = pmd_lock(mm, pmd);
 898                 ret = 0;
 899                 set = false;
 900                 if (pmd_none(*pmd)) {
 901                         if (userfaultfd_missing(vma)) {
 902                                 spin_unlock(ptl);
 903                                 ret = handle_userfault(vma, address, flags,
 904                                                        VM_UFFD_MISSING);
 905                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 906                         } else {
 907                                 set_huge_zero_page(pgtable, mm, vma,
 908                                                    haddr, pmd,
 909                                                    zero_page);
 910                                 spin_unlock(ptl);
 911                                 set = true;
 912                         }
 913                 } else
 914                         spin_unlock(ptl);
 915                 if (!set) {
 916                         pte_free(mm, pgtable);
 917                         put_huge_zero_page();
 918                 }
 919                 return ret;
 920         }
 921         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 922         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 923         if (unlikely(!page)) {
 924                 count_vm_event(THP_FAULT_FALLBACK);
 925                 return VM_FAULT_FALLBACK;
 926         }
 927         prep_transhuge_page(page);
 928         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 929                                             flags);
 930 }
 931
 932 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 933                 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
 934 {
 935         struct mm_struct *mm = vma->vm_mm;
 936         pmd_t entry;
 937         spinlock_t *ptl;
 938
 939         ptl = pmd_lock(mm, pmd);
 940         if (pmd_none(*pmd)) {
 941                 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
 942                 if (write) {
 943                         entry = pmd_mkyoung(pmd_mkdirty(entry));
 944                         entry = maybe_pmd_mkwrite(entry, vma);
 945                 }
 946                 set_pmd_at(mm, addr, pmd, entry);
 947                 update_mmu_cache_pmd(vma, addr, pmd);
 948         }
 949         spin_unlock(ptl);
 950 }
 951
 952 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 953                         pmd_t *pmd, unsigned long pfn, bool write)
 954 {
 955         pgprot_t pgprot = vma->vm_page_prot;
 956         /*
 957          * If we had pmd_special, we could avoid all these restrictions,
 958          * but we need to be consistent with PTEs and architectures that
 959          * can't support a 'special' bit.
 960          */
 961         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 962         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 963                                                 (VM_PFNMAP|VM_MIXEDMAP));
 964         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 965         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
 966
 967         if (addr < vma->vm_start || addr >= vma->vm_end)
 968                 return VM_FAULT_SIGBUS;
 969         if (track_pfn_insert(vma, &pgprot, pfn))
 970                 return VM_FAULT_SIGBUS;
 971         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 972         return VM_FAULT_NOPAGE;
 973 }
 974
 975 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 976                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 977                   struct vm_area_struct *vma)
 978 {
 979         spinlock_t *dst_ptl, *src_ptl;
 980         struct page *src_page;
 981         pmd_t pmd;
 982         pgtable_t pgtable;
 983         int ret;
 984
 985         ret = -ENOMEM;
 986         pgtable = pte_alloc_one(dst_mm, addr);
 987         if (unlikely(!pgtable))
 988                 goto out;
 989
 990         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 991         src_ptl = pmd_lockptr(src_mm, src_pmd);
 992         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 993
 994         ret = -EAGAIN;
 995         pmd = *src_pmd;
 996         if (unlikely(!pmd_trans_huge(pmd))) {
 997                 pte_free(dst_mm, pgtable);
 998                 goto out_unlock;
 999         }
1000         /*
1001          * When page table lock is held, the huge zero pmd should not be
1002          * under splitting since we don't split the page itself, only pmd to
1003          * a page table.
1004          */
1005         if (is_huge_zero_pmd(pmd)) {
1006                 struct page *zero_page;
1007                 /*
1008                  * get_huge_zero_page() will never allocate a new page here,
1009                  * since we already have a zero page to copy. It just takes a
1010                  * reference.
1011                  */
1012                 zero_page = get_huge_zero_page();
1013                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1014                                 zero_page);
1015                 ret = 0;
1016                 goto out_unlock;
1017         }
1018
1019         src_page = pmd_page(pmd);
1020         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1021         get_page(src_page);
1022         page_dup_rmap(src_page, true);
1023         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1024
1025         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1026         pmd = pmd_mkold(pmd_wrprotect(pmd));
1027         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1028         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1029         atomic_long_inc(&dst_mm->nr_ptes);
1030
1031         ret = 0;
1032 out_unlock:
1033         spin_unlock(src_ptl);
1034         spin_unlock(dst_ptl);
1035 out:
1036         return ret;
1037 }
1038
1039 void huge_pmd_set_accessed(struct mm_struct *mm,
1040                            struct vm_area_struct *vma,
1041                            unsigned long address,
1042                            pmd_t *pmd, pmd_t orig_pmd,
1043                            int dirty)
1044 {
1045         spinlock_t *ptl;
1046         pmd_t entry;
1047         unsigned long haddr;
1048
1049         ptl = pmd_lock(mm, pmd);
1050         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1051                 goto unlock;
1052
1053         entry = pmd_mkyoung(orig_pmd);
1054         haddr = address & HPAGE_PMD_MASK;
1055         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1056                 update_mmu_cache_pmd(vma, address, pmd);
1057
1058 unlock:
1059         spin_unlock(ptl);
1060 }
1061
1062 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1063                                         struct vm_area_struct *vma,
1064                                         unsigned long address,
1065                                         pmd_t *pmd, pmd_t orig_pmd,
1066                                         struct page *page,
1067                                         unsigned long haddr)
1068 {
1069         struct mem_cgroup *memcg;
1070         spinlock_t *ptl;
1071         pgtable_t pgtable;
1072         pmd_t _pmd;
1073         int ret = 0, i;
1074         struct page **pages;
1075         unsigned long mmun_start;       /* For mmu_notifiers */
1076         unsigned long mmun_end;         /* For mmu_notifiers */
1077
1078         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1079                         GFP_KERNEL);
1080         if (unlikely(!pages)) {
1081                 ret |= VM_FAULT_OOM;
1082                 goto out;
1083         }
1084
1085         for (i = 0; i < HPAGE_PMD_NR; i++) {
1086                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1087                                                __GFP_OTHER_NODE,
1088                                                vma, address, page_to_nid(page));
1089                 if (unlikely(!pages[i] ||
1090                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1091                                                    &memcg, false))) {
1092                         if (pages[i])
1093                                 put_page(pages[i]);
1094                         while (--i >= 0) {
1095                                 memcg = (void *)page_private(pages[i]);
1096                                 set_page_private(pages[i], 0);
1097                                 mem_cgroup_cancel_charge(pages[i], memcg,
1098                                                 false);
1099                                 put_page(pages[i]);
1100                         }
1101                         kfree(pages);
1102                         ret |= VM_FAULT_OOM;
1103                         goto out;
1104                 }
1105                 set_page_private(pages[i], (unsigned long)memcg);
1106         }
1107
1108         for (i = 0; i < HPAGE_PMD_NR; i++) {
1109                 copy_user_highpage(pages[i], page + i,
1110                                    haddr + PAGE_SIZE * i, vma);
1111                 __SetPageUptodate(pages[i]);
1112                 cond_resched();
1113         }
1114
1115         mmun_start = haddr;
1116         mmun_end   = haddr + HPAGE_PMD_SIZE;
1117         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1118
1119         ptl = pmd_lock(mm, pmd);
1120         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1121                 goto out_free_pages;
1122         VM_BUG_ON_PAGE(!PageHead(page), page);
1123
1124         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1125         /* leave pmd empty until pte is filled */
1126
1127         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1128         pmd_populate(mm, &_pmd, pgtable);
1129
1130         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1131                 pte_t *pte, entry;
1132                 entry = mk_pte(pages[i], vma->vm_page_prot);
1133                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1134                 memcg = (void *)page_private(pages[i]);
1135                 set_page_private(pages[i], 0);
1136                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1137                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1138                 lru_cache_add_active_or_unevictable(pages[i], vma);
1139                 pte = pte_offset_map(&_pmd, haddr);
1140                 VM_BUG_ON(!pte_none(*pte));
1141                 set_pte_at(mm, haddr, pte, entry);
1142                 pte_unmap(pte);
1143         }
1144         kfree(pages);
1145
1146         smp_wmb(); /* make pte visible before pmd */
1147         pmd_populate(mm, pmd, pgtable);
1148         page_remove_rmap(page, true);
1149         spin_unlock(ptl);
1150
1151         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1152
1153         ret |= VM_FAULT_WRITE;
1154         put_page(page);
1155
1156 out:
1157         return ret;
1158
1159 out_free_pages:
1160         spin_unlock(ptl);
1161         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1162         for (i = 0; i < HPAGE_PMD_NR; i++) {
1163                 memcg = (void *)page_private(pages[i]);
1164                 set_page_private(pages[i], 0);
1165                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1166                 put_page(pages[i]);
1167         }
1168         kfree(pages);
1169         goto out;
1170 }
1171
1172 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1173                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1174 {
1175         spinlock_t *ptl;
1176         int ret = 0;
1177         struct page *page = NULL, *new_page;
1178         struct mem_cgroup *memcg;
1179         unsigned long haddr;
1180         unsigned long mmun_start;       /* For mmu_notifiers */
1181         unsigned long mmun_end;         /* For mmu_notifiers */
1182         gfp_t huge_gfp;                 /* for allocation and charge */
1183
1184         ptl = pmd_lockptr(mm, pmd);
1185         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1186         haddr = address & HPAGE_PMD_MASK;
1187         if (is_huge_zero_pmd(orig_pmd))
1188                 goto alloc;
1189         spin_lock(ptl);
1190         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1191                 goto out_unlock;
1192
1193         page = pmd_page(orig_pmd);
1194         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1195         /*
1196          * We can only reuse the page if nobody else maps the huge page or it's
1197          * part. We can do it by checking page_mapcount() on each sub-page, but
1198          * it's expensive.
1199          * The cheaper way is to check page_count() to be equal 1: every
1200          * mapcount takes page reference reference, so this way we can
1201          * guarantee, that the PMD is the only mapping.
1202          * This can give false negative if somebody pinned the page, but that's
1203          * fine.
1204          */
1205         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1206                 pmd_t entry;
1207                 entry = pmd_mkyoung(orig_pmd);
1208                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1209                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1210                         update_mmu_cache_pmd(vma, address, pmd);
1211                 ret |= VM_FAULT_WRITE;
1212                 goto out_unlock;
1213         }
1214         get_page(page);
1215         spin_unlock(ptl);
1216 alloc:
1217         if (transparent_hugepage_enabled(vma) &&
1218             !transparent_hugepage_debug_cow()) {
1219                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1220                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1221         } else
1222                 new_page = NULL;
1223
1224         if (likely(new_page)) {
1225                 prep_transhuge_page(new_page);
1226         } else {
1227                 if (!page) {
1228                         split_huge_pmd(vma, pmd, address);
1229                         ret |= VM_FAULT_FALLBACK;
1230                 } else {
1231                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1232                                         pmd, orig_pmd, page, haddr);
1233                         if (ret & VM_FAULT_OOM) {
1234                                 split_huge_pmd(vma, pmd, address);
1235                                 ret |= VM_FAULT_FALLBACK;
1236                         }
1237                         put_page(page);
1238                 }
1239                 count_vm_event(THP_FAULT_FALLBACK);
1240                 goto out;
1241         }
1242
1243         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1244                                            true))) {
1245                 put_page(new_page);
1246                 if (page) {
1247                         split_huge_pmd(vma, pmd, address);
1248                         put_page(page);
1249                 } else
1250                         split_huge_pmd(vma, pmd, address);
1251                 ret |= VM_FAULT_FALLBACK;
1252                 count_vm_event(THP_FAULT_FALLBACK);
1253                 goto out;
1254         }
1255
1256         count_vm_event(THP_FAULT_ALLOC);
1257
1258         if (!page)
1259                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1260         else
1261                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1262         __SetPageUptodate(new_page);
1263
1264         mmun_start = haddr;
1265         mmun_end   = haddr + HPAGE_PMD_SIZE;
1266         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1267
1268         spin_lock(ptl);
1269         if (page)
1270                 put_page(page);
1271         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1272                 spin_unlock(ptl);
1273                 mem_cgroup_cancel_charge(new_page, memcg, true);
1274                 put_page(new_page);
1275                 goto out_mn;
1276         } else {
1277                 pmd_t entry;
1278                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1279                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1280                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1281                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1282                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1283                 lru_cache_add_active_or_unevictable(new_page, vma);
1284                 set_pmd_at(mm, haddr, pmd, entry);
1285                 update_mmu_cache_pmd(vma, address, pmd);
1286                 if (!page) {
1287                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1288                         put_huge_zero_page();
1289                 } else {
1290                         VM_BUG_ON_PAGE(!PageHead(page), page);
1291                         page_remove_rmap(page, true);
1292                         put_page(page);
1293                 }
1294                 ret |= VM_FAULT_WRITE;
1295         }
1296         spin_unlock(ptl);
1297 out_mn:
1298         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1299 out:
1300         return ret;
1301 out_unlock:
1302         spin_unlock(ptl);
1303         return ret;
1304 }
1305
1306 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1307                                    unsigned long addr,
1308                                    pmd_t *pmd,
1309                                    unsigned int flags)
1310 {
1311         struct mm_struct *mm = vma->vm_mm;
1312         struct page *page = NULL;
1313
1314         assert_spin_locked(pmd_lockptr(mm, pmd));
1315
1316         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1317                 goto out;
1318
1319         /* Avoid dumping huge zero page */
1320         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1321                 return ERR_PTR(-EFAULT);
1322
1323         /* Full NUMA hinting faults to serialise migration in fault paths */
1324         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1325                 goto out;
1326
1327         page = pmd_page(*pmd);
1328         VM_BUG_ON_PAGE(!PageHead(page), page);
1329         if (flags & FOLL_TOUCH) {
1330                 pmd_t _pmd;
1331                 /*
1332                  * We should set the dirty bit only for FOLL_WRITE but
1333                  * for now the dirty bit in the pmd is meaningless.
1334                  * And if the dirty bit will become meaningful and
1335                  * we'll only set it with FOLL_WRITE, an atomic
1336                  * set_bit will be required on the pmd to set the
1337                  * young bit, instead of the current set_pmd_at.
1338                  */
1339                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1340                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1341                                           pmd, _pmd,  1))
1342                         update_mmu_cache_pmd(vma, addr, pmd);
1343         }
1344         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1345                 /*
1346                  * We don't mlock() pte-mapped THPs. This way we can avoid
1347                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1348                  *
1349                  * In most cases the pmd is the only mapping of the page as we
1350                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1351                  * writable private mappings in populate_vma_page_range().
1352                  *
1353                  * The only scenario when we have the page shared here is if we
1354                  * mlocking read-only mapping shared over fork(). We skip
1355                  * mlocking such pages.
1356                  */
1357                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1358                                 page->mapping && trylock_page(page)) {
1359                         lru_add_drain();
1360                         if (page->mapping)
1361                                 mlock_vma_page(page);
1362                         unlock_page(page);
1363                 }
1364         }
1365         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1366         VM_BUG_ON_PAGE(!PageCompound(page), page);
1367         if (flags & FOLL_GET)
1368                 get_page(page);
1369
1370 out:
1371         return page;
1372 }
1373
1374 /* NUMA hinting page fault entry point for trans huge pmds */
1375 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1376                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1377 {
1378         spinlock_t *ptl;
1379         struct anon_vma *anon_vma = NULL;
1380         struct page *page;
1381         unsigned long haddr = addr & HPAGE_PMD_MASK;
1382         int page_nid = -1, this_nid = numa_node_id();
1383         int target_nid, last_cpupid = -1;
1384         bool page_locked;
1385         bool migrated = false;
1386         bool was_writable;
1387         int flags = 0;
1388
1389         /* A PROT_NONE fault should not end up here */
1390         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1391
1392         ptl = pmd_lock(mm, pmdp);
1393         if (unlikely(!pmd_same(pmd, *pmdp)))
1394                 goto out_unlock;
1395
1396         /*
1397          * If there are potential migrations, wait for completion and retry
1398          * without disrupting NUMA hinting information. Do not relock and
1399          * check_same as the page may no longer be mapped.
1400          */
1401         if (unlikely(pmd_trans_migrating(*pmdp))) {
1402                 page = pmd_page(*pmdp);
1403                 spin_unlock(ptl);
1404                 wait_on_page_locked(page);
1405                 goto out;
1406         }
1407
1408         page = pmd_page(pmd);
1409         BUG_ON(is_huge_zero_page(page));
1410         page_nid = page_to_nid(page);
1411         last_cpupid = page_cpupid_last(page);
1412         count_vm_numa_event(NUMA_HINT_FAULTS);
1413         if (page_nid == this_nid) {
1414                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1415                 flags |= TNF_FAULT_LOCAL;
1416         }
1417
1418         /* See similar comment in do_numa_page for explanation */
1419         if (!(vma->vm_flags & VM_WRITE))
1420                 flags |= TNF_NO_GROUP;
1421
1422         /*
1423          * Acquire the page lock to serialise THP migrations but avoid dropping
1424          * page_table_lock if at all possible
1425          */
1426         page_locked = trylock_page(page);
1427         target_nid = mpol_misplaced(page, vma, haddr);
1428         if (target_nid == -1) {
1429                 /* If the page was locked, there are no parallel migrations */
1430                 if (page_locked)
1431                         goto clear_pmdnuma;
1432         }
1433
1434         /* Migration could have started since the pmd_trans_migrating check */
1435         if (!page_locked) {
1436                 spin_unlock(ptl);
1437                 wait_on_page_locked(page);
1438                 page_nid = -1;
1439                 goto out;
1440         }
1441
1442         /*
1443          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1444          * to serialises splits
1445          */
1446         get_page(page);
1447         spin_unlock(ptl);
1448         anon_vma = page_lock_anon_vma_read(page);
1449
1450         /* Confirm the PMD did not change while page_table_lock was released */
1451         spin_lock(ptl);
1452         if (unlikely(!pmd_same(pmd, *pmdp))) {
1453                 unlock_page(page);
1454                 put_page(page);
1455                 page_nid = -1;
1456                 goto out_unlock;
1457         }
1458
1459         /* Bail if we fail to protect against THP splits for any reason */
1460         if (unlikely(!anon_vma)) {
1461                 put_page(page);
1462                 page_nid = -1;
1463                 goto clear_pmdnuma;
1464         }
1465
1466         /*
1467          * Migrate the THP to the requested node, returns with page unlocked
1468          * and access rights restored.
1469          */
1470         spin_unlock(ptl);
1471         migrated = migrate_misplaced_transhuge_page(mm, vma,
1472                                 pmdp, pmd, addr, page, target_nid);
1473         if (migrated) {
1474                 flags |= TNF_MIGRATED;
1475                 page_nid = target_nid;
1476         } else
1477                 flags |= TNF_MIGRATE_FAIL;
1478
1479         goto out;
1480 clear_pmdnuma:
1481         BUG_ON(!PageLocked(page));
1482         was_writable = pmd_write(pmd);
1483         pmd = pmd_modify(pmd, vma->vm_page_prot);
1484         pmd = pmd_mkyoung(pmd);
1485         if (was_writable)
1486                 pmd = pmd_mkwrite(pmd);
1487         set_pmd_at(mm, haddr, pmdp, pmd);
1488         update_mmu_cache_pmd(vma, addr, pmdp);
1489         unlock_page(page);
1490 out_unlock:
1491         spin_unlock(ptl);
1492
1493 out:
1494         if (anon_vma)
1495                 page_unlock_anon_vma_read(anon_vma);
1496
1497         if (page_nid != -1)
1498                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1499
1500         return 0;
1501 }
1502
1503 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1504                  pmd_t *pmd, unsigned long addr)
1505 {
1506         pmd_t orig_pmd;
1507         spinlock_t *ptl;
1508
1509         if (!__pmd_trans_huge_lock(pmd, vma, &ptl))
1510                 return 0;
1511         /*
1512          * For architectures like ppc64 we look at deposited pgtable
1513          * when calling pmdp_huge_get_and_clear. So do the
1514          * pgtable_trans_huge_withdraw after finishing pmdp related
1515          * operations.
1516          */
1517         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1518                         tlb->fullmm);
1519         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1520         if (vma_is_dax(vma)) {
1521                 spin_unlock(ptl);
1522                 if (is_huge_zero_pmd(orig_pmd))
1523                         put_huge_zero_page();
1524         } else if (is_huge_zero_pmd(orig_pmd)) {
1525                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1526                 atomic_long_dec(&tlb->mm->nr_ptes);
1527                 spin_unlock(ptl);
1528                 put_huge_zero_page();
1529         } else {
1530                 struct page *page = pmd_page(orig_pmd);
1531                 page_remove_rmap(page, true);
1532                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1533                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1534                 VM_BUG_ON_PAGE(!PageHead(page), page);
1535                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1536                 atomic_long_dec(&tlb->mm->nr_ptes);
1537                 spin_unlock(ptl);
1538                 tlb_remove_page(tlb, page);
1539         }
1540         return 1;
1541 }
1542
1543 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1544                   unsigned long old_addr,
1545                   unsigned long new_addr, unsigned long old_end,
1546                   pmd_t *old_pmd, pmd_t *new_pmd)
1547 {
1548         spinlock_t *old_ptl, *new_ptl;
1549         pmd_t pmd;
1550
1551         struct mm_struct *mm = vma->vm_mm;
1552
1553         if ((old_addr & ~HPAGE_PMD_MASK) ||
1554             (new_addr & ~HPAGE_PMD_MASK) ||
1555             old_end - old_addr < HPAGE_PMD_SIZE ||
1556             (new_vma->vm_flags & VM_NOHUGEPAGE))
1557                 return false;
1558
1559         /*
1560          * The destination pmd shouldn't be established, free_pgtables()
1561          * should have release it.
1562          */
1563         if (WARN_ON(!pmd_none(*new_pmd))) {
1564                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1565                 return false;
1566         }
1567
1568         /*
1569          * We don't have to worry about the ordering of src and dst
1570          * ptlocks because exclusive mmap_sem prevents deadlock.
1571          */
1572         if (__pmd_trans_huge_lock(old_pmd, vma, &old_ptl)) {
1573                 new_ptl = pmd_lockptr(mm, new_pmd);
1574                 if (new_ptl != old_ptl)
1575                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1576                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1577                 VM_BUG_ON(!pmd_none(*new_pmd));
1578
1579                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1580                         pgtable_t pgtable;
1581                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1582                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1583                 }
1584                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1585                 if (new_ptl != old_ptl)
1586                         spin_unlock(new_ptl);
1587                 spin_unlock(old_ptl);
1588                 return true;
1589         }
1590         return false;
1591 }
1592
1593 /*
1594  * Returns
1595  *  - 0 if PMD could not be locked
1596  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1597  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1598  */
1599 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1600                 unsigned long addr, pgprot_t newprot, int prot_numa)
1601 {
1602         struct mm_struct *mm = vma->vm_mm;
1603         spinlock_t *ptl;
1604         int ret = 0;
1605
1606         if (__pmd_trans_huge_lock(pmd, vma, &ptl)) {
1607                 pmd_t entry;
1608                 bool preserve_write = prot_numa && pmd_write(*pmd);
1609                 ret = 1;
1610
1611                 /*
1612                  * Avoid trapping faults against the zero page. The read-only
1613                  * data is likely to be read-cached on the local CPU and
1614                  * local/remote hits to the zero page are not interesting.
1615                  */
1616                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1617                         spin_unlock(ptl);
1618                         return ret;
1619                 }
1620
1621                 if (!prot_numa || !pmd_protnone(*pmd)) {
1622                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1623                         entry = pmd_modify(entry, newprot);
1624                         if (preserve_write)
1625                                 entry = pmd_mkwrite(entry);
1626                         ret = HPAGE_PMD_NR;
1627                         set_pmd_at(mm, addr, pmd, entry);
1628                         BUG_ON(!preserve_write && pmd_write(entry));
1629                 }
1630                 spin_unlock(ptl);
1631         }
1632
1633         return ret;
1634 }
1635
1636 /*
1637  * Returns true if a given pmd maps a thp, false otherwise.
1638  *
1639  * Note that if it returns true, this routine returns without unlocking page
1640  * table lock. So callers must unlock it.
1641  */
1642 bool __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1643                 spinlock_t **ptl)
1644 {
1645         *ptl = pmd_lock(vma->vm_mm, pmd);
1646         if (likely(pmd_trans_huge(*pmd)))
1647                 return true;
1648         spin_unlock(*ptl);
1649         return false;
1650 }
1651
1652 /*
1653  * This function returns whether a given @page is mapped onto the @address
1654  * in the virtual space of @mm.
1655  *
1656  * When it's true, this function returns *pmd with holding the page table lock
1657  * and passing it back to the caller via @ptl.
1658  * If it's false, returns NULL without holding the page table lock.
1659  */
1660 pmd_t *page_check_address_pmd(struct page *page,
1661                               struct mm_struct *mm,
1662                               unsigned long address,
1663                               spinlock_t **ptl)
1664 {
1665         pgd_t *pgd;
1666         pud_t *pud;
1667         pmd_t *pmd;
1668
1669         if (address & ~HPAGE_PMD_MASK)
1670                 return NULL;
1671
1672         pgd = pgd_offset(mm, address);
1673         if (!pgd_present(*pgd))
1674                 return NULL;
1675         pud = pud_offset(pgd, address);
1676         if (!pud_present(*pud))
1677                 return NULL;
1678         pmd = pmd_offset(pud, address);
1679
1680         *ptl = pmd_lock(mm, pmd);
1681         if (!pmd_present(*pmd))
1682                 goto unlock;
1683         if (pmd_page(*pmd) != page)
1684                 goto unlock;
1685         if (pmd_trans_huge(*pmd))
1686                 return pmd;
1687 unlock:
1688         spin_unlock(*ptl);
1689         return NULL;
1690 }
1691
1692 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1693
1694 int hugepage_madvise(struct vm_area_struct *vma,
1695                      unsigned long *vm_flags, int advice)
1696 {
1697         switch (advice) {
1698         case MADV_HUGEPAGE:
1699 #ifdef CONFIG_S390
1700                 /*
1701                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1702                  * can't handle this properly after s390_enable_sie, so we simply
1703                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1704                  */
1705                 if (mm_has_pgste(vma->vm_mm))
1706                         return 0;
1707 #endif
1708                 /*
1709                  * Be somewhat over-protective like KSM for now!
1710                  */
1711                 if (*vm_flags & VM_NO_THP)
1712                         return -EINVAL;
1713                 *vm_flags &= ~VM_NOHUGEPAGE;
1714                 *vm_flags |= VM_HUGEPAGE;
1715                 /*
1716                  * If the vma become good for khugepaged to scan,
1717                  * register it here without waiting a page fault that
1718                  * may not happen any time soon.
1719                  */
1720                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1721                         return -ENOMEM;
1722                 break;
1723         case MADV_NOHUGEPAGE:
1724                 /*
1725                  * Be somewhat over-protective like KSM for now!
1726                  */
1727                 if (*vm_flags & VM_NO_THP)
1728                         return -EINVAL;
1729                 *vm_flags &= ~VM_HUGEPAGE;
1730                 *vm_flags |= VM_NOHUGEPAGE;
1731                 /*
1732                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1733                  * this vma even if we leave the mm registered in khugepaged if
1734                  * it got registered before VM_NOHUGEPAGE was set.
1735                  */
1736                 break;
1737         }
1738
1739         return 0;
1740 }
1741
1742 static int __init khugepaged_slab_init(void)
1743 {
1744         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1745                                           sizeof(struct mm_slot),
1746                                           __alignof__(struct mm_slot), 0, NULL);
1747         if (!mm_slot_cache)
1748                 return -ENOMEM;
1749
1750         return 0;
1751 }
1752
1753 static void __init khugepaged_slab_exit(void)
1754 {
1755         kmem_cache_destroy(mm_slot_cache);
1756 }
1757
1758 static inline struct mm_slot *alloc_mm_slot(void)
1759 {
1760         if (!mm_slot_cache)     /* initialization failed */
1761                 return NULL;
1762         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1763 }
1764
1765 static inline void free_mm_slot(struct mm_slot *mm_slot)
1766 {
1767         kmem_cache_free(mm_slot_cache, mm_slot);
1768 }
1769
1770 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1771 {
1772         struct mm_slot *mm_slot;
1773
1774         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1775                 if (mm == mm_slot->mm)
1776                         return mm_slot;
1777
1778         return NULL;
1779 }
1780
1781 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1782                                     struct mm_slot *mm_slot)
1783 {
1784         mm_slot->mm = mm;
1785         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1786 }
1787
1788 static inline int khugepaged_test_exit(struct mm_struct *mm)
1789 {
1790         return atomic_read(&mm->mm_users) == 0;
1791 }
1792
1793 int __khugepaged_enter(struct mm_struct *mm)
1794 {
1795         struct mm_slot *mm_slot;
1796         int wakeup;
1797
1798         mm_slot = alloc_mm_slot();
1799         if (!mm_slot)
1800                 return -ENOMEM;
1801
1802         /* __khugepaged_exit() must not run from under us */
1803         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1804         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1805                 free_mm_slot(mm_slot);
1806                 return 0;
1807         }
1808
1809         spin_lock(&khugepaged_mm_lock);
1810         insert_to_mm_slots_hash(mm, mm_slot);
1811         /*
1812          * Insert just behind the scanning cursor, to let the area settle
1813          * down a little.
1814          */
1815         wakeup = list_empty(&khugepaged_scan.mm_head);
1816         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1817         spin_unlock(&khugepaged_mm_lock);
1818
1819         atomic_inc(&mm->mm_count);
1820         if (wakeup)
1821                 wake_up_interruptible(&khugepaged_wait);
1822
1823         return 0;
1824 }
1825
1826 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1827                                unsigned long vm_flags)
1828 {
1829         unsigned long hstart, hend;
1830         if (!vma->anon_vma)
1831                 /*
1832                  * Not yet faulted in so we will register later in the
1833                  * page fault if needed.
1834                  */
1835                 return 0;
1836         if (vma->vm_ops)
1837                 /* khugepaged not yet working on file or special mappings */
1838                 return 0;
1839         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1840         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1841         hend = vma->vm_end & HPAGE_PMD_MASK;
1842         if (hstart < hend)
1843                 return khugepaged_enter(vma, vm_flags);
1844         return 0;
1845 }
1846
1847 void __khugepaged_exit(struct mm_struct *mm)
1848 {
1849         struct mm_slot *mm_slot;
1850         int free = 0;
1851
1852         spin_lock(&khugepaged_mm_lock);
1853         mm_slot = get_mm_slot(mm);
1854         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1855                 hash_del(&mm_slot->hash);
1856                 list_del(&mm_slot->mm_node);
1857                 free = 1;
1858         }
1859         spin_unlock(&khugepaged_mm_lock);
1860
1861         if (free) {
1862                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1863                 free_mm_slot(mm_slot);
1864                 mmdrop(mm);
1865         } else if (mm_slot) {
1866                 /*
1867                  * This is required to serialize against
1868                  * khugepaged_test_exit() (which is guaranteed to run
1869                  * under mmap sem read mode). Stop here (after we
1870                  * return all pagetables will be destroyed) until
1871                  * khugepaged has finished working on the pagetables
1872                  * under the mmap_sem.
1873                  */
1874                 down_write(&mm->mmap_sem);
1875                 up_write(&mm->mmap_sem);
1876         }
1877 }
1878
1879 static void release_pte_page(struct page *page)
1880 {
1881         /* 0 stands for page_is_file_cache(page) == false */
1882         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1883         unlock_page(page);
1884         putback_lru_page(page);
1885 }
1886
1887 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1888 {
1889         while (--_pte >= pte) {
1890                 pte_t pteval = *_pte;
1891                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1892                         release_pte_page(pte_page(pteval));
1893         }
1894 }
1895
1896 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1897                                         unsigned long address,
1898                                         pte_t *pte)
1899 {
1900         struct page *page = NULL;
1901         pte_t *_pte;
1902         int none_or_zero = 0, result = 0;
1903         bool referenced = false, writable = false;
1904
1905         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1906              _pte++, address += PAGE_SIZE) {
1907                 pte_t pteval = *_pte;
1908                 if (pte_none(pteval) || (pte_present(pteval) &&
1909                                 is_zero_pfn(pte_pfn(pteval)))) {
1910                         if (!userfaultfd_armed(vma) &&
1911                             ++none_or_zero <= khugepaged_max_ptes_none) {
1912                                 continue;
1913                         } else {
1914                                 result = SCAN_EXCEED_NONE_PTE;
1915                                 goto out;
1916                         }
1917                 }
1918                 if (!pte_present(pteval)) {
1919                         result = SCAN_PTE_NON_PRESENT;
1920                         goto out;
1921                 }
1922                 page = vm_normal_page(vma, address, pteval);
1923                 if (unlikely(!page)) {
1924                         result = SCAN_PAGE_NULL;
1925                         goto out;
1926                 }
1927
1928                 VM_BUG_ON_PAGE(PageCompound(page), page);
1929                 VM_BUG_ON_PAGE(!PageAnon(page), page);
1930                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1931
1932                 /*
1933                  * We can do it before isolate_lru_page because the
1934                  * page can't be freed from under us. NOTE: PG_lock
1935                  * is needed to serialize against split_huge_page
1936                  * when invoked from the VM.
1937                  */
1938                 if (!trylock_page(page)) {
1939                         result = SCAN_PAGE_LOCK;
1940                         goto out;
1941                 }
1942
1943                 /*
1944                  * cannot use mapcount: can't collapse if there's a gup pin.
1945                  * The page must only be referenced by the scanned process
1946                  * and page swap cache.
1947                  */
1948                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1949                         unlock_page(page);
1950                         result = SCAN_PAGE_COUNT;
1951                         goto out;
1952                 }
1953                 if (pte_write(pteval)) {
1954                         writable = true;
1955                 } else {
1956                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
1957                                 unlock_page(page);
1958                                 result = SCAN_SWAP_CACHE_PAGE;
1959                                 goto out;
1960                         }
1961                         /*
1962                          * Page is not in the swap cache. It can be collapsed
1963                          * into a THP.
1964                          */
1965                 }
1966
1967                 /*
1968                  * Isolate the page to avoid collapsing an hugepage
1969                  * currently in use by the VM.
1970                  */
1971                 if (isolate_lru_page(page)) {
1972                         unlock_page(page);
1973                         result = SCAN_DEL_PAGE_LRU;
1974                         goto out;
1975                 }
1976                 /* 0 stands for page_is_file_cache(page) == false */
1977                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1978                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1979                 VM_BUG_ON_PAGE(PageLRU(page), page);
1980
1981                 /* If there is no mapped pte young don't collapse the page */
1982                 if (pte_young(pteval) ||
1983                     page_is_young(page) || PageReferenced(page) ||
1984                     mmu_notifier_test_young(vma->vm_mm, address))
1985                         referenced = true;
1986         }
1987         if (likely(writable)) {
1988                 if (likely(referenced)) {
1989                         result = SCAN_SUCCEED;
1990                         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
1991                                                             referenced, writable, result);
1992                         return 1;
1993                 }
1994         } else {
1995                 result = SCAN_PAGE_RO;
1996         }
1997
1998 out:
1999         release_pte_pages(pte, _pte);
2000         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
2001                                             referenced, writable, result);
2002         return 0;
2003 }
2004
2005 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2006                                       struct vm_area_struct *vma,
2007                                       unsigned long address,
2008                                       spinlock_t *ptl)
2009 {
2010         pte_t *_pte;
2011         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2012                 pte_t pteval = *_pte;
2013                 struct page *src_page;
2014
2015                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2016                         clear_user_highpage(page, address);
2017                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2018                         if (is_zero_pfn(pte_pfn(pteval))) {
2019                                 /*
2020                                  * ptl mostly unnecessary.
2021                                  */
2022                                 spin_lock(ptl);
2023                                 /*
2024                                  * paravirt calls inside pte_clear here are
2025                                  * superfluous.
2026                                  */
2027                                 pte_clear(vma->vm_mm, address, _pte);
2028                                 spin_unlock(ptl);
2029                         }
2030                 } else {
2031                         src_page = pte_page(pteval);
2032                         copy_user_highpage(page, src_page, address, vma);
2033                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2034                         release_pte_page(src_page);
2035                         /*
2036                          * ptl mostly unnecessary, but preempt has to
2037                          * be disabled to update the per-cpu stats
2038                          * inside page_remove_rmap().
2039                          */
2040                         spin_lock(ptl);
2041                         /*
2042                          * paravirt calls inside pte_clear here are
2043                          * superfluous.
2044                          */
2045                         pte_clear(vma->vm_mm, address, _pte);
2046                         page_remove_rmap(src_page, false);
2047                         spin_unlock(ptl);
2048                         free_page_and_swap_cache(src_page);
2049                 }
2050
2051                 address += PAGE_SIZE;
2052                 page++;
2053         }
2054 }
2055
2056 static void khugepaged_alloc_sleep(void)
2057 {
2058         DEFINE_WAIT(wait);
2059
2060         add_wait_queue(&khugepaged_wait, &wait);
2061         freezable_schedule_timeout_interruptible(
2062                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2063         remove_wait_queue(&khugepaged_wait, &wait);
2064 }
2065
2066 static int khugepaged_node_load[MAX_NUMNODES];
2067
2068 static bool khugepaged_scan_abort(int nid)
2069 {
2070         int i;
2071
2072         /*
2073          * If zone_reclaim_mode is disabled, then no extra effort is made to
2074          * allocate memory locally.
2075          */
2076         if (!zone_reclaim_mode)
2077                 return false;
2078
2079         /* If there is a count for this node already, it must be acceptable */
2080         if (khugepaged_node_load[nid])
2081                 return false;
2082
2083         for (i = 0; i < MAX_NUMNODES; i++) {
2084                 if (!khugepaged_node_load[i])
2085                         continue;
2086                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2087                         return true;
2088         }
2089         return false;
2090 }
2091
2092 #ifdef CONFIG_NUMA
2093 static int khugepaged_find_target_node(void)
2094 {
2095         static int last_khugepaged_target_node = NUMA_NO_NODE;
2096         int nid, target_node = 0, max_value = 0;
2097
2098         /* find first node with max normal pages hit */
2099         for (nid = 0; nid < MAX_NUMNODES; nid++)
2100                 if (khugepaged_node_load[nid] > max_value) {
2101                         max_value = khugepaged_node_load[nid];
2102                         target_node = nid;
2103                 }
2104
2105         /* do some balance if several nodes have the same hit record */
2106         if (target_node <= last_khugepaged_target_node)
2107                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2108                                 nid++)
2109                         if (max_value == khugepaged_node_load[nid]) {
2110                                 target_node = nid;
2111                                 break;
2112                         }
2113
2114         last_khugepaged_target_node = target_node;
2115         return target_node;
2116 }
2117
2118 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2119 {
2120         if (IS_ERR(*hpage)) {
2121                 if (!*wait)
2122                         return false;
2123
2124                 *wait = false;
2125                 *hpage = NULL;
2126                 khugepaged_alloc_sleep();
2127         } else if (*hpage) {
2128                 put_page(*hpage);
2129                 *hpage = NULL;
2130         }
2131
2132         return true;
2133 }
2134
2135 static struct page *
2136 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2137                        unsigned long address, int node)
2138 {
2139         VM_BUG_ON_PAGE(*hpage, *hpage);
2140
2141         /*
2142          * Before allocating the hugepage, release the mmap_sem read lock.
2143          * The allocation can take potentially a long time if it involves
2144          * sync compaction, and we do not need to hold the mmap_sem during
2145          * that. We will recheck the vma after taking it again in write mode.
2146          */
2147         up_read(&mm->mmap_sem);
2148
2149         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2150         if (unlikely(!*hpage)) {
2151                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2152                 *hpage = ERR_PTR(-ENOMEM);
2153                 return NULL;
2154         }
2155
2156         prep_transhuge_page(*hpage);
2157         count_vm_event(THP_COLLAPSE_ALLOC);
2158         return *hpage;
2159 }
2160 #else
2161 static int khugepaged_find_target_node(void)
2162 {
2163         return 0;
2164 }
2165
2166 static inline struct page *alloc_hugepage(int defrag)
2167 {
2168         struct page *page;
2169
2170         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2171         if (page)
2172                 prep_transhuge_page(page);
2173         return page;
2174 }
2175
2176 static struct page *khugepaged_alloc_hugepage(bool *wait)
2177 {
2178         struct page *hpage;
2179
2180         do {
2181                 hpage = alloc_hugepage(khugepaged_defrag());
2182                 if (!hpage) {
2183                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2184                         if (!*wait)
2185                                 return NULL;
2186
2187                         *wait = false;
2188                         khugepaged_alloc_sleep();
2189                 } else
2190                         count_vm_event(THP_COLLAPSE_ALLOC);
2191         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2192
2193         return hpage;
2194 }
2195
2196 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2197 {
2198         if (!*hpage)
2199                 *hpage = khugepaged_alloc_hugepage(wait);
2200
2201         if (unlikely(!*hpage))
2202                 return false;
2203
2204         return true;
2205 }
2206
2207 static struct page *
2208 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2209                        unsigned long address, int node)
2210 {
2211         up_read(&mm->mmap_sem);
2212         VM_BUG_ON(!*hpage);
2213
2214         return  *hpage;
2215 }
2216 #endif
2217
2218 static bool hugepage_vma_check(struct vm_area_struct *vma)
2219 {
2220         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2221             (vma->vm_flags & VM_NOHUGEPAGE))
2222                 return false;
2223         if (!vma->anon_vma || vma->vm_ops)
2224                 return false;
2225         if (is_vma_temporary_stack(vma))
2226                 return false;
2227         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2228         return true;
2229 }
2230
2231 static void collapse_huge_page(struct mm_struct *mm,
2232                                    unsigned long address,
2233                                    struct page **hpage,
2234                                    struct vm_area_struct *vma,
2235                                    int node)
2236 {
2237         pmd_t *pmd, _pmd;
2238         pte_t *pte;
2239         pgtable_t pgtable;
2240         struct page *new_page;
2241         spinlock_t *pmd_ptl, *pte_ptl;
2242         int isolated, result = 0;
2243         unsigned long hstart, hend;
2244         struct mem_cgroup *memcg;
2245         unsigned long mmun_start;       /* For mmu_notifiers */
2246         unsigned long mmun_end;         /* For mmu_notifiers */
2247         gfp_t gfp;
2248
2249         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2250
2251         /* Only allocate from the target node */
2252         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2253                 __GFP_THISNODE;
2254
2255         /* release the mmap_sem read lock. */
2256         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2257         if (!new_page) {
2258                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2259                 goto out_nolock;
2260         }
2261
2262         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2263                 result = SCAN_CGROUP_CHARGE_FAIL;
2264                 goto out_nolock;
2265         }
2266
2267         /*
2268          * Prevent all access to pagetables with the exception of
2269          * gup_fast later hanlded by the ptep_clear_flush and the VM
2270          * handled by the anon_vma lock + PG_lock.
2271          */
2272         down_write(&mm->mmap_sem);
2273         if (unlikely(khugepaged_test_exit(mm))) {
2274                 result = SCAN_ANY_PROCESS;
2275                 goto out;
2276         }
2277
2278         vma = find_vma(mm, address);
2279         if (!vma) {
2280                 result = SCAN_VMA_NULL;
2281                 goto out;
2282         }
2283         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2284         hend = vma->vm_end & HPAGE_PMD_MASK;
2285         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2286                 result = SCAN_ADDRESS_RANGE;
2287                 goto out;
2288         }
2289         if (!hugepage_vma_check(vma)) {
2290                 result = SCAN_VMA_CHECK;
2291                 goto out;
2292         }
2293         pmd = mm_find_pmd(mm, address);
2294         if (!pmd) {
2295                 result = SCAN_PMD_NULL;
2296                 goto out;
2297         }
2298
2299         anon_vma_lock_write(vma->anon_vma);
2300
2301         pte = pte_offset_map(pmd, address);
2302         pte_ptl = pte_lockptr(mm, pmd);
2303
2304         mmun_start = address;
2305         mmun_end   = address + HPAGE_PMD_SIZE;
2306         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2307         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2308         /*
2309          * After this gup_fast can't run anymore. This also removes
2310          * any huge TLB entry from the CPU so we won't allow
2311          * huge and small TLB entries for the same virtual address
2312          * to avoid the risk of CPU bugs in that area.
2313          */
2314         _pmd = pmdp_collapse_flush(vma, address, pmd);
2315         spin_unlock(pmd_ptl);
2316         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2317
2318         spin_lock(pte_ptl);
2319         isolated = __collapse_huge_page_isolate(vma, address, pte);
2320         spin_unlock(pte_ptl);
2321
2322         if (unlikely(!isolated)) {
2323                 pte_unmap(pte);
2324                 spin_lock(pmd_ptl);
2325                 BUG_ON(!pmd_none(*pmd));
2326                 /*
2327                  * We can only use set_pmd_at when establishing
2328                  * hugepmds and never for establishing regular pmds that
2329                  * points to regular pagetables. Use pmd_populate for that
2330                  */
2331                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2332                 spin_unlock(pmd_ptl);
2333                 anon_vma_unlock_write(vma->anon_vma);
2334                 result = SCAN_FAIL;
2335                 goto out;
2336         }
2337
2338         /*
2339          * All pages are isolated and locked so anon_vma rmap
2340          * can't run anymore.
2341          */
2342         anon_vma_unlock_write(vma->anon_vma);
2343
2344         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2345         pte_unmap(pte);
2346         __SetPageUptodate(new_page);
2347         pgtable = pmd_pgtable(_pmd);
2348
2349         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2350         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2351
2352         /*
2353          * spin_lock() below is not the equivalent of smp_wmb(), so
2354          * this is needed to avoid the copy_huge_page writes to become
2355          * visible after the set_pmd_at() write.
2356          */
2357         smp_wmb();
2358
2359         spin_lock(pmd_ptl);
2360         BUG_ON(!pmd_none(*pmd));
2361         page_add_new_anon_rmap(new_page, vma, address, true);
2362         mem_cgroup_commit_charge(new_page, memcg, false, true);
2363         lru_cache_add_active_or_unevictable(new_page, vma);
2364         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2365         set_pmd_at(mm, address, pmd, _pmd);
2366         update_mmu_cache_pmd(vma, address, pmd);
2367         spin_unlock(pmd_ptl);
2368
2369         *hpage = NULL;
2370
2371         khugepaged_pages_collapsed++;
2372         result = SCAN_SUCCEED;
2373 out_up_write:
2374         up_write(&mm->mmap_sem);
2375         trace_mm_collapse_huge_page(mm, isolated, result);
2376         return;
2377
2378 out_nolock:
2379         trace_mm_collapse_huge_page(mm, isolated, result);
2380         return;
2381 out:
2382         mem_cgroup_cancel_charge(new_page, memcg, true);
2383         goto out_up_write;
2384 }
2385
2386 static int khugepaged_scan_pmd(struct mm_struct *mm,
2387                                struct vm_area_struct *vma,
2388                                unsigned long address,
2389                                struct page **hpage)
2390 {
2391         pmd_t *pmd;
2392         pte_t *pte, *_pte;
2393         int ret = 0, none_or_zero = 0, result = 0;
2394         struct page *page = NULL;
2395         unsigned long _address;
2396         spinlock_t *ptl;
2397         int node = NUMA_NO_NODE;
2398         bool writable = false, referenced = false;
2399
2400         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2401
2402         pmd = mm_find_pmd(mm, address);
2403         if (!pmd) {
2404                 result = SCAN_PMD_NULL;
2405                 goto out;
2406         }
2407
2408         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2409         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2410         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2411              _pte++, _address += PAGE_SIZE) {
2412                 pte_t pteval = *_pte;
2413                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2414                         if (!userfaultfd_armed(vma) &&
2415                             ++none_or_zero <= khugepaged_max_ptes_none) {
2416                                 continue;
2417                         } else {
2418                                 result = SCAN_EXCEED_NONE_PTE;
2419                                 goto out_unmap;
2420                         }
2421                 }
2422                 if (!pte_present(pteval)) {
2423                         result = SCAN_PTE_NON_PRESENT;
2424                         goto out_unmap;
2425                 }
2426                 if (pte_write(pteval))
2427                         writable = true;
2428
2429                 page = vm_normal_page(vma, _address, pteval);
2430                 if (unlikely(!page)) {
2431                         result = SCAN_PAGE_NULL;
2432                         goto out_unmap;
2433                 }
2434
2435                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2436                 if (PageCompound(page)) {
2437                         result = SCAN_PAGE_COMPOUND;
2438                         goto out_unmap;
2439                 }
2440
2441                 /*
2442                  * Record which node the original page is from and save this
2443                  * information to khugepaged_node_load[].
2444                  * Khupaged will allocate hugepage from the node has the max
2445                  * hit record.
2446                  */
2447                 node = page_to_nid(page);
2448                 if (khugepaged_scan_abort(node)) {
2449                         result = SCAN_SCAN_ABORT;
2450                         goto out_unmap;
2451                 }
2452                 khugepaged_node_load[node]++;
2453                 if (!PageLRU(page)) {
2454                         result = SCAN_SCAN_ABORT;
2455                         goto out_unmap;
2456                 }
2457                 if (PageLocked(page)) {
2458                         result = SCAN_PAGE_LOCK;
2459                         goto out_unmap;
2460                 }
2461                 if (!PageAnon(page)) {
2462                         result = SCAN_PAGE_ANON;
2463                         goto out_unmap;
2464                 }
2465
2466                 /*
2467                  * cannot use mapcount: can't collapse if there's a gup pin.
2468                  * The page must only be referenced by the scanned process
2469                  * and page swap cache.
2470                  */
2471                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2472                         result = SCAN_PAGE_COUNT;
2473                         goto out_unmap;
2474                 }
2475                 if (pte_young(pteval) ||
2476                     page_is_young(page) || PageReferenced(page) ||
2477                     mmu_notifier_test_young(vma->vm_mm, address))
2478                         referenced = true;
2479         }
2480         if (writable) {
2481                 if (referenced) {
2482                         result = SCAN_SUCCEED;
2483                         ret = 1;
2484                 } else {
2485                         result = SCAN_NO_REFERENCED_PAGE;
2486                 }
2487         } else {
2488                 result = SCAN_PAGE_RO;
2489         }
2490 out_unmap:
2491         pte_unmap_unlock(pte, ptl);
2492         if (ret) {
2493                 node = khugepaged_find_target_node();
2494                 /* collapse_huge_page will return with the mmap_sem released */
2495                 collapse_huge_page(mm, address, hpage, vma, node);
2496         }
2497 out:
2498         trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced,
2499                                      none_or_zero, result);
2500         return ret;
2501 }
2502
2503 static void collect_mm_slot(struct mm_slot *mm_slot)
2504 {
2505         struct mm_struct *mm = mm_slot->mm;
2506
2507         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2508
2509         if (khugepaged_test_exit(mm)) {
2510                 /* free mm_slot */
2511                 hash_del(&mm_slot->hash);
2512                 list_del(&mm_slot->mm_node);
2513
2514                 /*
2515                  * Not strictly needed because the mm exited already.
2516                  *
2517                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2518                  */
2519
2520                 /* khugepaged_mm_lock actually not necessary for the below */
2521                 free_mm_slot(mm_slot);
2522                 mmdrop(mm);
2523         }
2524 }
2525
2526 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2527                                             struct page **hpage)
2528         __releases(&khugepaged_mm_lock)
2529         __acquires(&khugepaged_mm_lock)
2530 {
2531         struct mm_slot *mm_slot;
2532         struct mm_struct *mm;
2533         struct vm_area_struct *vma;
2534         int progress = 0;
2535
2536         VM_BUG_ON(!pages);
2537         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2538
2539         if (khugepaged_scan.mm_slot)
2540                 mm_slot = khugepaged_scan.mm_slot;
2541         else {
2542                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2543                                      struct mm_slot, mm_node);
2544                 khugepaged_scan.address = 0;
2545                 khugepaged_scan.mm_slot = mm_slot;
2546         }
2547         spin_unlock(&khugepaged_mm_lock);
2548
2549         mm = mm_slot->mm;
2550         down_read(&mm->mmap_sem);
2551         if (unlikely(khugepaged_test_exit(mm)))
2552                 vma = NULL;
2553         else
2554                 vma = find_vma(mm, khugepaged_scan.address);
2555
2556         progress++;
2557         for (; vma; vma = vma->vm_next) {
2558                 unsigned long hstart, hend;
2559
2560                 cond_resched();
2561                 if (unlikely(khugepaged_test_exit(mm))) {
2562                         progress++;
2563                         break;
2564                 }
2565                 if (!hugepage_vma_check(vma)) {
2566 skip:
2567                         progress++;
2568                         continue;
2569                 }
2570                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2571                 hend = vma->vm_end & HPAGE_PMD_MASK;
2572                 if (hstart >= hend)
2573                         goto skip;
2574                 if (khugepaged_scan.address > hend)
2575                         goto skip;
2576                 if (khugepaged_scan.address < hstart)
2577                         khugepaged_scan.address = hstart;
2578                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2579
2580                 while (khugepaged_scan.address < hend) {
2581                         int ret;
2582                         cond_resched();
2583                         if (unlikely(khugepaged_test_exit(mm)))
2584                                 goto breakouterloop;
2585
2586                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2587                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2588                                   hend);
2589                         ret = khugepaged_scan_pmd(mm, vma,
2590                                                   khugepaged_scan.address,
2591                                                   hpage);
2592                         /* move to next address */
2593                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2594                         progress += HPAGE_PMD_NR;
2595                         if (ret)
2596                                 /* we released mmap_sem so break loop */
2597                                 goto breakouterloop_mmap_sem;
2598                         if (progress >= pages)
2599                                 goto breakouterloop;
2600                 }
2601         }
2602 breakouterloop:
2603         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2604 breakouterloop_mmap_sem:
2605
2606         spin_lock(&khugepaged_mm_lock);
2607         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2608         /*
2609          * Release the current mm_slot if this mm is about to die, or
2610          * if we scanned all vmas of this mm.
2611          */
2612         if (khugepaged_test_exit(mm) || !vma) {
2613                 /*
2614                  * Make sure that if mm_users is reaching zero while
2615                  * khugepaged runs here, khugepaged_exit will find
2616                  * mm_slot not pointing to the exiting mm.
2617                  */
2618                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2619                         khugepaged_scan.mm_slot = list_entry(
2620                                 mm_slot->mm_node.next,
2621                                 struct mm_slot, mm_node);
2622                         khugepaged_scan.address = 0;
2623                 } else {
2624                         khugepaged_scan.mm_slot = NULL;
2625                         khugepaged_full_scans++;
2626                 }
2627
2628                 collect_mm_slot(mm_slot);
2629         }
2630
2631         return progress;
2632 }
2633
2634 static int khugepaged_has_work(void)
2635 {
2636         return !list_empty(&khugepaged_scan.mm_head) &&
2637                 khugepaged_enabled();
2638 }
2639
2640 static int khugepaged_wait_event(void)
2641 {
2642         return !list_empty(&khugepaged_scan.mm_head) ||
2643                 kthread_should_stop();
2644 }
2645
2646 static void khugepaged_do_scan(void)
2647 {
2648         struct page *hpage = NULL;
2649         unsigned int progress = 0, pass_through_head = 0;
2650         unsigned int pages = khugepaged_pages_to_scan;
2651         bool wait = true;
2652
2653         barrier(); /* write khugepaged_pages_to_scan to local stack */
2654
2655         while (progress < pages) {
2656                 if (!khugepaged_prealloc_page(&hpage, &wait))
2657                         break;
2658
2659                 cond_resched();
2660
2661                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2662                         break;
2663
2664                 spin_lock(&khugepaged_mm_lock);
2665                 if (!khugepaged_scan.mm_slot)
2666                         pass_through_head++;
2667                 if (khugepaged_has_work() &&
2668                     pass_through_head < 2)
2669                         progress += khugepaged_scan_mm_slot(pages - progress,
2670                                                             &hpage);
2671                 else
2672                         progress = pages;
2673                 spin_unlock(&khugepaged_mm_lock);
2674         }
2675
2676         if (!IS_ERR_OR_NULL(hpage))
2677                 put_page(hpage);
2678 }
2679
2680 static void khugepaged_wait_work(void)
2681 {
2682         if (khugepaged_has_work()) {
2683                 if (!khugepaged_scan_sleep_millisecs)
2684                         return;
2685
2686                 wait_event_freezable_timeout(khugepaged_wait,
2687                                              kthread_should_stop(),
2688                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2689                 return;
2690         }
2691
2692         if (khugepaged_enabled())
2693                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2694 }
2695
2696 static int khugepaged(void *none)
2697 {
2698         struct mm_slot *mm_slot;
2699
2700         set_freezable();
2701         set_user_nice(current, MAX_NICE);
2702
2703         while (!kthread_should_stop()) {
2704                 khugepaged_do_scan();
2705                 khugepaged_wait_work();
2706         }
2707
2708         spin_lock(&khugepaged_mm_lock);
2709         mm_slot = khugepaged_scan.mm_slot;
2710         khugepaged_scan.mm_slot = NULL;
2711         if (mm_slot)
2712                 collect_mm_slot(mm_slot);
2713         spin_unlock(&khugepaged_mm_lock);
2714         return 0;
2715 }
2716
2717 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2718                 unsigned long haddr, pmd_t *pmd)
2719 {
2720         struct mm_struct *mm = vma->vm_mm;
2721         pgtable_t pgtable;
2722         pmd_t _pmd;
2723         int i;
2724
2725         /* leave pmd empty until pte is filled */
2726         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2727
2728         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2729         pmd_populate(mm, &_pmd, pgtable);
2730
2731         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2732                 pte_t *pte, entry;
2733                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2734                 entry = pte_mkspecial(entry);
2735                 pte = pte_offset_map(&_pmd, haddr);
2736                 VM_BUG_ON(!pte_none(*pte));
2737                 set_pte_at(mm, haddr, pte, entry);
2738                 pte_unmap(pte);
2739         }
2740         smp_wmb(); /* make pte visible before pmd */
2741         pmd_populate(mm, pmd, pgtable);
2742         put_huge_zero_page();
2743 }
2744
2745 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2746                 unsigned long haddr, bool freeze)
2747 {
2748         struct mm_struct *mm = vma->vm_mm;
2749         struct page *page;
2750         pgtable_t pgtable;
2751         pmd_t _pmd;
2752         bool young, write;
2753         int i;
2754
2755         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2756         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2757         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2758         VM_BUG_ON(!pmd_trans_huge(*pmd));
2759
2760         count_vm_event(THP_SPLIT_PMD);
2761
2762         if (vma_is_dax(vma)) {
2763                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2764                 if (is_huge_zero_pmd(_pmd))
2765                         put_huge_zero_page();
2766                 return;
2767         } else if (is_huge_zero_pmd(*pmd)) {
2768                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2769         }
2770
2771         page = pmd_page(*pmd);
2772         VM_BUG_ON_PAGE(!page_count(page), page);
2773         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2774         write = pmd_write(*pmd);
2775         young = pmd_young(*pmd);
2776
2777         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2778         pmd_populate(mm, &_pmd, pgtable);
2779
2780         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2781                 pte_t entry, *pte;
2782                 /*
2783                  * Note that NUMA hinting access restrictions are not
2784                  * transferred to avoid any possibility of altering
2785                  * permissions across VMAs.
2786                  */
2787                 if (freeze) {
2788                         swp_entry_t swp_entry;
2789                         swp_entry = make_migration_entry(page + i, write);
2790                         entry = swp_entry_to_pte(swp_entry);
2791                 } else {
2792                         entry = mk_pte(page + i, vma->vm_page_prot);
2793                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2794                         if (!write)
2795                                 entry = pte_wrprotect(entry);
2796                         if (!young)
2797                                 entry = pte_mkold(entry);
2798                 }
2799                 pte = pte_offset_map(&_pmd, haddr);
2800                 BUG_ON(!pte_none(*pte));
2801                 set_pte_at(mm, haddr, pte, entry);
2802                 atomic_inc(&page[i]._mapcount);
2803                 pte_unmap(pte);
2804         }
2805
2806         /*
2807          * Set PG_double_map before dropping compound_mapcount to avoid
2808          * false-negative page_mapped().
2809          */
2810         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2811                 for (i = 0; i < HPAGE_PMD_NR; i++)
2812                         atomic_inc(&page[i]._mapcount);
2813         }
2814
2815         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2816                 /* Last compound_mapcount is gone. */
2817                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2818                 if (TestClearPageDoubleMap(page)) {
2819                         /* No need in mapcount reference anymore */
2820                         for (i = 0; i < HPAGE_PMD_NR; i++)
2821                                 atomic_dec(&page[i]._mapcount);
2822                 }
2823         }
2824
2825         smp_wmb(); /* make pte visible before pmd */
2826         /*
2827          * Up to this point the pmd is present and huge and userland has the
2828          * whole access to the hugepage during the split (which happens in
2829          * place). If we overwrite the pmd with the not-huge version pointing
2830          * to the pte here (which of course we could if all CPUs were bug
2831          * free), userland could trigger a small page size TLB miss on the
2832          * small sized TLB while the hugepage TLB entry is still established in
2833          * the huge TLB. Some CPU doesn't like that.
2834          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2835          * 383 on page 93. Intel should be safe but is also warns that it's
2836          * only safe if the permission and cache attributes of the two entries
2837          * loaded in the two TLB is identical (which should be the case here).
2838          * But it is generally safer to never allow small and huge TLB entries
2839          * for the same virtual address to be loaded simultaneously. So instead
2840          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2841          * current pmd notpresent (atomically because here the pmd_trans_huge
2842          * and pmd_trans_splitting must remain set at all times on the pmd
2843          * until the split is complete for this pmd), then we flush the SMP TLB
2844          * and finally we write the non-huge version of the pmd entry with
2845          * pmd_populate.
2846          */
2847         pmdp_invalidate(vma, haddr, pmd);
2848         pmd_populate(mm, pmd, pgtable);
2849
2850         if (freeze) {
2851                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2852                         page_remove_rmap(page + i, false);
2853                         put_page(page + i);
2854                 }
2855         }
2856 }
2857
2858 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2859                 unsigned long address)
2860 {
2861         spinlock_t *ptl;
2862         struct mm_struct *mm = vma->vm_mm;
2863         struct page *page = NULL;
2864         unsigned long haddr = address & HPAGE_PMD_MASK;
2865
2866         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2867         ptl = pmd_lock(mm, pmd);
2868         if (unlikely(!pmd_trans_huge(*pmd)))
2869                 goto out;
2870         page = pmd_page(*pmd);
2871         __split_huge_pmd_locked(vma, pmd, haddr, false);
2872         if (PageMlocked(page))
2873                 get_page(page);
2874         else
2875                 page = NULL;
2876 out:
2877         spin_unlock(ptl);
2878         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2879         if (page) {
2880                 lock_page(page);
2881                 munlock_vma_page(page);
2882                 unlock_page(page);
2883                 put_page(page);
2884         }
2885 }
2886
2887 static void split_huge_pmd_address(struct vm_area_struct *vma,
2888                                     unsigned long address)
2889 {
2890         pgd_t *pgd;
2891         pud_t *pud;
2892         pmd_t *pmd;
2893
2894         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2895
2896         pgd = pgd_offset(vma->vm_mm, address);
2897         if (!pgd_present(*pgd))
2898                 return;
2899
2900         pud = pud_offset(pgd, address);
2901         if (!pud_present(*pud))
2902                 return;
2903
2904         pmd = pmd_offset(pud, address);
2905         if (!pmd_present(*pmd) || !pmd_trans_huge(*pmd))
2906                 return;
2907         /*
2908          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2909          * materialize from under us.
2910          */
2911         split_huge_pmd(vma, pmd, address);
2912 }
2913
2914 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2915                              unsigned long start,
2916                              unsigned long end,
2917                              long adjust_next)
2918 {
2919         /*
2920          * If the new start address isn't hpage aligned and it could
2921          * previously contain an hugepage: check if we need to split
2922          * an huge pmd.
2923          */
2924         if (start & ~HPAGE_PMD_MASK &&
2925             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2926             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2927                 split_huge_pmd_address(vma, start);
2928
2929         /*
2930          * If the new end address isn't hpage aligned and it could
2931          * previously contain an hugepage: check if we need to split
2932          * an huge pmd.
2933          */
2934         if (end & ~HPAGE_PMD_MASK &&
2935             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2936             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2937                 split_huge_pmd_address(vma, end);
2938
2939         /*
2940          * If we're also updating the vma->vm_next->vm_start, if the new
2941          * vm_next->vm_start isn't page aligned and it could previously
2942          * contain an hugepage: check if we need to split an huge pmd.
2943          */
2944         if (adjust_next > 0) {
2945                 struct vm_area_struct *next = vma->vm_next;
2946                 unsigned long nstart = next->vm_start;
2947                 nstart += adjust_next << PAGE_SHIFT;
2948                 if (nstart & ~HPAGE_PMD_MASK &&
2949                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2950                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2951                         split_huge_pmd_address(next, nstart);
2952         }
2953 }
2954
2955 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
2956                 unsigned long address)
2957 {
2958         spinlock_t *ptl;
2959         pgd_t *pgd;
2960         pud_t *pud;
2961         pmd_t *pmd;
2962         pte_t *pte;
2963         int i, nr = HPAGE_PMD_NR;
2964
2965         /* Skip pages which doesn't belong to the VMA */
2966         if (address < vma->vm_start) {
2967                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
2968                 page += off;
2969                 nr -= off;
2970                 address = vma->vm_start;
2971         }
2972
2973         pgd = pgd_offset(vma->vm_mm, address);
2974         if (!pgd_present(*pgd))
2975                 return;
2976         pud = pud_offset(pgd, address);
2977         if (!pud_present(*pud))
2978                 return;
2979         pmd = pmd_offset(pud, address);
2980         ptl = pmd_lock(vma->vm_mm, pmd);
2981         if (!pmd_present(*pmd)) {
2982                 spin_unlock(ptl);
2983                 return;
2984         }
2985         if (pmd_trans_huge(*pmd)) {
2986                 if (page == pmd_page(*pmd))
2987                         __split_huge_pmd_locked(vma, pmd, address, true);
2988                 spin_unlock(ptl);
2989                 return;
2990         }
2991         spin_unlock(ptl);
2992
2993         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
2994         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
2995                 pte_t entry, swp_pte;
2996                 swp_entry_t swp_entry;
2997
2998                 if (!pte_present(pte[i]))
2999                         continue;
3000                 if (page_to_pfn(page) != pte_pfn(pte[i]))
3001                         continue;
3002                 flush_cache_page(vma, address, page_to_pfn(page));
3003                 entry = ptep_clear_flush(vma, address, pte + i);
3004                 swp_entry = make_migration_entry(page, pte_write(entry));
3005                 swp_pte = swp_entry_to_pte(swp_entry);
3006                 if (pte_soft_dirty(entry))
3007                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
3008                 set_pte_at(vma->vm_mm, address, pte + i, swp_pte);
3009                 page_remove_rmap(page, false);
3010                 put_page(page);
3011         }
3012         pte_unmap_unlock(pte, ptl);
3013 }
3014
3015 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
3016 {
3017         struct anon_vma_chain *avc;
3018         pgoff_t pgoff = page_to_pgoff(page);
3019
3020         VM_BUG_ON_PAGE(!PageHead(page), page);
3021
3022         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
3023                         pgoff + HPAGE_PMD_NR - 1) {
3024                 unsigned long haddr;
3025
3026                 haddr = __vma_address(page, avc->vma) & HPAGE_PMD_MASK;
3027                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3028                                 haddr, haddr + HPAGE_PMD_SIZE);
3029                 freeze_page_vma(avc->vma, page, haddr);
3030                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3031                                 haddr, haddr + HPAGE_PMD_SIZE);
3032         }
3033 }
3034
3035 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
3036                 unsigned long address)
3037 {
3038         spinlock_t *ptl;
3039         pmd_t *pmd;
3040         pte_t *pte, entry;
3041         swp_entry_t swp_entry;
3042         int i, nr = HPAGE_PMD_NR;
3043
3044         /* Skip pages which doesn't belong to the VMA */
3045         if (address < vma->vm_start) {
3046                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3047                 page += off;
3048                 nr -= off;
3049                 address = vma->vm_start;
3050         }
3051
3052         pmd = mm_find_pmd(vma->vm_mm, address);
3053         if (!pmd)
3054                 return;
3055         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3056         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
3057                 if (!is_swap_pte(pte[i]))
3058                         continue;
3059
3060                 swp_entry = pte_to_swp_entry(pte[i]);
3061                 if (!is_migration_entry(swp_entry))
3062                         continue;
3063                 if (migration_entry_to_page(swp_entry) != page)
3064                         continue;
3065
3066                 get_page(page);
3067                 page_add_anon_rmap(page, vma, address, false);
3068
3069                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3070                 entry = pte_mkdirty(entry);
3071                 if (is_write_migration_entry(swp_entry))
3072                         entry = maybe_mkwrite(entry, vma);
3073
3074                 flush_dcache_page(page);
3075                 set_pte_at(vma->vm_mm, address, pte + i, entry);
3076
3077                 /* No need to invalidate - it was non-present before */
3078                 update_mmu_cache(vma, address, pte + i);
3079         }
3080         pte_unmap_unlock(pte, ptl);
3081 }
3082
3083 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3084 {
3085         struct anon_vma_chain *avc;
3086         pgoff_t pgoff = page_to_pgoff(page);
3087
3088         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3089                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3090                 unsigned long address = __vma_address(page, avc->vma);
3091
3092                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3093                                 address, address + HPAGE_PMD_SIZE);
3094                 unfreeze_page_vma(avc->vma, page, address);
3095                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3096                                 address, address + HPAGE_PMD_SIZE);
3097         }
3098 }
3099
3100 static int total_mapcount(struct page *page)
3101 {
3102         int i, ret;
3103
3104         ret = compound_mapcount(page);
3105         for (i = 0; i < HPAGE_PMD_NR; i++)
3106                 ret += atomic_read(&page[i]._mapcount) + 1;
3107
3108         if (PageDoubleMap(page))
3109                 ret -= HPAGE_PMD_NR;
3110
3111         return ret;
3112 }
3113
3114 static int __split_huge_page_tail(struct page *head, int tail,
3115                 struct lruvec *lruvec, struct list_head *list)
3116 {
3117         int mapcount;
3118         struct page *page_tail = head + tail;
3119
3120         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3121         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3122
3123         /*
3124          * tail_page->_count is zero and not changing from under us. But
3125          * get_page_unless_zero() may be running from under us on the
3126          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3127          * would then run atomic_set() concurrently with
3128          * get_page_unless_zero(), and atomic_set() is implemented in C not
3129          * using locked ops. spin_unlock on x86 sometime uses locked ops
3130          * because of PPro errata 66, 92, so unless somebody can guarantee
3131          * atomic_set() here would be safe on all archs (and not only on x86),
3132          * it's safer to use atomic_add().
3133          */
3134         atomic_add(mapcount + 1, &page_tail->_count);
3135
3136
3137         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3138         page_tail->flags |= (head->flags &
3139                         ((1L << PG_referenced) |
3140                          (1L << PG_swapbacked) |
3141                          (1L << PG_mlocked) |
3142                          (1L << PG_uptodate) |
3143                          (1L << PG_active) |
3144                          (1L << PG_locked) |
3145                          (1L << PG_unevictable)));
3146         page_tail->flags |= (1L << PG_dirty);
3147
3148         /*
3149          * After clearing PageTail the gup refcount can be released.
3150          * Page flags also must be visible before we make the page non-compound.
3151          */
3152         smp_wmb();
3153
3154         clear_compound_head(page_tail);
3155
3156         if (page_is_young(head))
3157                 set_page_young(page_tail);
3158         if (page_is_idle(head))
3159                 set_page_idle(page_tail);
3160
3161         /* ->mapping in first tail page is compound_mapcount */
3162         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3163                         page_tail);
3164         page_tail->mapping = head->mapping;
3165
3166         page_tail->index = head->index + tail;
3167         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3168         lru_add_page_tail(head, page_tail, lruvec, list);
3169
3170         return mapcount;
3171 }
3172
3173 static void __split_huge_page(struct page *page, struct list_head *list)
3174 {
3175         struct page *head = compound_head(page);
3176         struct zone *zone = page_zone(head);
3177         struct lruvec *lruvec;
3178         int i, tail_mapcount;
3179
3180         /* prevent PageLRU to go away from under us, and freeze lru stats */
3181         spin_lock_irq(&zone->lru_lock);
3182         lruvec = mem_cgroup_page_lruvec(head, zone);
3183
3184         /* complete memcg works before add pages to LRU */
3185         mem_cgroup_split_huge_fixup(head);
3186
3187         tail_mapcount = 0;
3188         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3189                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3190         atomic_sub(tail_mapcount, &head->_count);
3191
3192         ClearPageCompound(head);
3193         spin_unlock_irq(&zone->lru_lock);
3194
3195         unfreeze_page(page_anon_vma(head), head);
3196
3197         for (i = 0; i < HPAGE_PMD_NR; i++) {
3198                 struct page *subpage = head + i;
3199                 if (subpage == page)
3200                         continue;
3201                 unlock_page(subpage);
3202
3203                 /*
3204                  * Subpages may be freed if there wasn't any mapping
3205                  * like if add_to_swap() is running on a lru page that
3206                  * had its mapping zapped. And freeing these pages
3207                  * requires taking the lru_lock so we do the put_page
3208                  * of the tail pages after the split is complete.
3209                  */
3210                 put_page(subpage);
3211         }
3212 }
3213
3214 /*
3215  * This function splits huge page into normal pages. @page can point to any
3216  * subpage of huge page to split. Split doesn't change the position of @page.
3217  *
3218  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3219  * The huge page must be locked.
3220  *
3221  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3222  *
3223  * Both head page and tail pages will inherit mapping, flags, and so on from
3224  * the hugepage.
3225  *
3226  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3227  * they are not mapped.
3228  *
3229  * Returns 0 if the hugepage is split successfully.
3230  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3231  * us.
3232  */
3233 int split_huge_page_to_list(struct page *page, struct list_head *list)
3234 {
3235         struct page *head = compound_head(page);
3236         struct anon_vma *anon_vma;
3237         int count, mapcount, ret;
3238
3239         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3240         VM_BUG_ON_PAGE(!PageAnon(page), page);
3241         VM_BUG_ON_PAGE(!PageLocked(page), page);
3242         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3243         VM_BUG_ON_PAGE(!PageCompound(page), page);
3244
3245         /*
3246          * The caller does not necessarily hold an mmap_sem that would prevent
3247          * the anon_vma disappearing so we first we take a reference to it
3248          * and then lock the anon_vma for write. This is similar to
3249          * page_lock_anon_vma_read except the write lock is taken to serialise
3250          * against parallel split or collapse operations.
3251          */
3252         anon_vma = page_get_anon_vma(head);
3253         if (!anon_vma) {
3254                 ret = -EBUSY;
3255                 goto out;
3256         }
3257         anon_vma_lock_write(anon_vma);
3258
3259         /*
3260          * Racy check if we can split the page, before freeze_page() will
3261          * split PMDs
3262          */
3263         if (total_mapcount(head) != page_count(head) - 1) {
3264                 ret = -EBUSY;
3265                 goto out_unlock;
3266         }
3267
3268         freeze_page(anon_vma, head);
3269         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3270
3271         /* Prevent deferred_split_scan() touching ->_count */
3272         spin_lock(&split_queue_lock);
3273         count = page_count(head);
3274         mapcount = total_mapcount(head);
3275         if (mapcount == count - 1) {
3276                 if (!list_empty(page_deferred_list(head))) {
3277                         split_queue_len--;
3278                         list_del(page_deferred_list(head));
3279                 }
3280                 spin_unlock(&split_queue_lock);
3281                 __split_huge_page(page, list);
3282                 ret = 0;
3283         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount > count - 1) {
3284                 spin_unlock(&split_queue_lock);
3285                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3286                                 mapcount, count);
3287                 if (PageTail(page))
3288                         dump_page(head, NULL);
3289                 dump_page(page, "total_mapcount(head) > page_count(head) - 1");
3290                 BUG();
3291         } else {
3292                 spin_unlock(&split_queue_lock);
3293                 unfreeze_page(anon_vma, head);
3294                 ret = -EBUSY;
3295         }
3296
3297 out_unlock:
3298         anon_vma_unlock_write(anon_vma);
3299         put_anon_vma(anon_vma);
3300 out:
3301         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3302         return ret;
3303 }
3304
3305 void free_transhuge_page(struct page *page)
3306 {
3307         unsigned long flags;
3308
3309         spin_lock_irqsave(&split_queue_lock, flags);
3310         if (!list_empty(page_deferred_list(page))) {
3311                 split_queue_len--;
3312                 list_del(page_deferred_list(page));
3313         }
3314         spin_unlock_irqrestore(&split_queue_lock, flags);
3315         free_compound_page(page);
3316 }
3317
3318 void deferred_split_huge_page(struct page *page)
3319 {
3320         unsigned long flags;
3321
3322         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3323
3324         spin_lock_irqsave(&split_queue_lock, flags);
3325         if (list_empty(page_deferred_list(page))) {
3326                 list_add_tail(page_deferred_list(page), &split_queue);
3327                 split_queue_len++;
3328         }
3329         spin_unlock_irqrestore(&split_queue_lock, flags);
3330 }
3331
3332 static unsigned long deferred_split_count(struct shrinker *shrink,
3333                 struct shrink_control *sc)
3334 {
3335         /*
3336          * Split a page from split_queue will free up at least one page,
3337          * at most HPAGE_PMD_NR - 1. We don't track exact number.
3338          * Let's use HPAGE_PMD_NR / 2 as ballpark.
3339          */
3340         return ACCESS_ONCE(split_queue_len) * HPAGE_PMD_NR / 2;
3341 }
3342
3343 static unsigned long deferred_split_scan(struct shrinker *shrink,
3344                 struct shrink_control *sc)
3345 {
3346         unsigned long flags;
3347         LIST_HEAD(list), *pos, *next;
3348         struct page *page;
3349         int split = 0;
3350
3351         spin_lock_irqsave(&split_queue_lock, flags);
3352         list_splice_init(&split_queue, &list);
3353
3354         /* Take pin on all head pages to avoid freeing them under us */
3355         list_for_each_safe(pos, next, &list) {
3356                 page = list_entry((void *)pos, struct page, mapping);
3357                 page = compound_head(page);
3358                 /* race with put_compound_page() */
3359                 if (!get_page_unless_zero(page)) {
3360                         list_del_init(page_deferred_list(page));
3361                         split_queue_len--;
3362                 }
3363         }
3364         spin_unlock_irqrestore(&split_queue_lock, flags);
3365
3366         list_for_each_safe(pos, next, &list) {
3367                 page = list_entry((void *)pos, struct page, mapping);
3368                 lock_page(page);
3369                 /* split_huge_page() removes page from list on success */
3370                 if (!split_huge_page(page))
3371                         split++;
3372                 unlock_page(page);
3373                 put_page(page);
3374         }
3375
3376         spin_lock_irqsave(&split_queue_lock, flags);
3377         list_splice_tail(&list, &split_queue);
3378         spin_unlock_irqrestore(&split_queue_lock, flags);
3379
3380         return split * HPAGE_PMD_NR / 2;
3381 }
3382
3383 static struct shrinker deferred_split_shrinker = {
3384         .count_objects = deferred_split_count,
3385         .scan_objects = deferred_split_scan,
3386         .seeks = DEFAULT_SEEKS,
3387 };