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 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1653
1654 int hugepage_madvise(struct vm_area_struct *vma,
1655                      unsigned long *vm_flags, int advice)
1656 {
1657         switch (advice) {
1658         case MADV_HUGEPAGE:
1659 #ifdef CONFIG_S390
1660                 /*
1661                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1662                  * can't handle this properly after s390_enable_sie, so we simply
1663                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1664                  */
1665                 if (mm_has_pgste(vma->vm_mm))
1666                         return 0;
1667 #endif
1668                 /*
1669                  * Be somewhat over-protective like KSM for now!
1670                  */
1671                 if (*vm_flags & VM_NO_THP)
1672                         return -EINVAL;
1673                 *vm_flags &= ~VM_NOHUGEPAGE;
1674                 *vm_flags |= VM_HUGEPAGE;
1675                 /*
1676                  * If the vma become good for khugepaged to scan,
1677                  * register it here without waiting a page fault that
1678                  * may not happen any time soon.
1679                  */
1680                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1681                         return -ENOMEM;
1682                 break;
1683         case MADV_NOHUGEPAGE:
1684                 /*
1685                  * Be somewhat over-protective like KSM for now!
1686                  */
1687                 if (*vm_flags & VM_NO_THP)
1688                         return -EINVAL;
1689                 *vm_flags &= ~VM_HUGEPAGE;
1690                 *vm_flags |= VM_NOHUGEPAGE;
1691                 /*
1692                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1693                  * this vma even if we leave the mm registered in khugepaged if
1694                  * it got registered before VM_NOHUGEPAGE was set.
1695                  */
1696                 break;
1697         }
1698
1699         return 0;
1700 }
1701
1702 static int __init khugepaged_slab_init(void)
1703 {
1704         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1705                                           sizeof(struct mm_slot),
1706                                           __alignof__(struct mm_slot), 0, NULL);
1707         if (!mm_slot_cache)
1708                 return -ENOMEM;
1709
1710         return 0;
1711 }
1712
1713 static void __init khugepaged_slab_exit(void)
1714 {
1715         kmem_cache_destroy(mm_slot_cache);
1716 }
1717
1718 static inline struct mm_slot *alloc_mm_slot(void)
1719 {
1720         if (!mm_slot_cache)     /* initialization failed */
1721                 return NULL;
1722         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1723 }
1724
1725 static inline void free_mm_slot(struct mm_slot *mm_slot)
1726 {
1727         kmem_cache_free(mm_slot_cache, mm_slot);
1728 }
1729
1730 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1731 {
1732         struct mm_slot *mm_slot;
1733
1734         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1735                 if (mm == mm_slot->mm)
1736                         return mm_slot;
1737
1738         return NULL;
1739 }
1740
1741 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1742                                     struct mm_slot *mm_slot)
1743 {
1744         mm_slot->mm = mm;
1745         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1746 }
1747
1748 static inline int khugepaged_test_exit(struct mm_struct *mm)
1749 {
1750         return atomic_read(&mm->mm_users) == 0;
1751 }
1752
1753 int __khugepaged_enter(struct mm_struct *mm)
1754 {
1755         struct mm_slot *mm_slot;
1756         int wakeup;
1757
1758         mm_slot = alloc_mm_slot();
1759         if (!mm_slot)
1760                 return -ENOMEM;
1761
1762         /* __khugepaged_exit() must not run from under us */
1763         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1764         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1765                 free_mm_slot(mm_slot);
1766                 return 0;
1767         }
1768
1769         spin_lock(&khugepaged_mm_lock);
1770         insert_to_mm_slots_hash(mm, mm_slot);
1771         /*
1772          * Insert just behind the scanning cursor, to let the area settle
1773          * down a little.
1774          */
1775         wakeup = list_empty(&khugepaged_scan.mm_head);
1776         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1777         spin_unlock(&khugepaged_mm_lock);
1778
1779         atomic_inc(&mm->mm_count);
1780         if (wakeup)
1781                 wake_up_interruptible(&khugepaged_wait);
1782
1783         return 0;
1784 }
1785
1786 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1787                                unsigned long vm_flags)
1788 {
1789         unsigned long hstart, hend;
1790         if (!vma->anon_vma)
1791                 /*
1792                  * Not yet faulted in so we will register later in the
1793                  * page fault if needed.
1794                  */
1795                 return 0;
1796         if (vma->vm_ops)
1797                 /* khugepaged not yet working on file or special mappings */
1798                 return 0;
1799         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1800         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1801         hend = vma->vm_end & HPAGE_PMD_MASK;
1802         if (hstart < hend)
1803                 return khugepaged_enter(vma, vm_flags);
1804         return 0;
1805 }
1806
1807 void __khugepaged_exit(struct mm_struct *mm)
1808 {
1809         struct mm_slot *mm_slot;
1810         int free = 0;
1811
1812         spin_lock(&khugepaged_mm_lock);
1813         mm_slot = get_mm_slot(mm);
1814         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1815                 hash_del(&mm_slot->hash);
1816                 list_del(&mm_slot->mm_node);
1817                 free = 1;
1818         }
1819         spin_unlock(&khugepaged_mm_lock);
1820
1821         if (free) {
1822                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1823                 free_mm_slot(mm_slot);
1824                 mmdrop(mm);
1825         } else if (mm_slot) {
1826                 /*
1827                  * This is required to serialize against
1828                  * khugepaged_test_exit() (which is guaranteed to run
1829                  * under mmap sem read mode). Stop here (after we
1830                  * return all pagetables will be destroyed) until
1831                  * khugepaged has finished working on the pagetables
1832                  * under the mmap_sem.
1833                  */
1834                 down_write(&mm->mmap_sem);
1835                 up_write(&mm->mmap_sem);
1836         }
1837 }
1838
1839 static void release_pte_page(struct page *page)
1840 {
1841         /* 0 stands for page_is_file_cache(page) == false */
1842         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1843         unlock_page(page);
1844         putback_lru_page(page);
1845 }
1846
1847 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1848 {
1849         while (--_pte >= pte) {
1850                 pte_t pteval = *_pte;
1851                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1852                         release_pte_page(pte_page(pteval));
1853         }
1854 }
1855
1856 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1857                                         unsigned long address,
1858                                         pte_t *pte)
1859 {
1860         struct page *page = NULL;
1861         pte_t *_pte;
1862         int none_or_zero = 0, result = 0;
1863         bool referenced = false, writable = false;
1864
1865         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1866              _pte++, address += PAGE_SIZE) {
1867                 pte_t pteval = *_pte;
1868                 if (pte_none(pteval) || (pte_present(pteval) &&
1869                                 is_zero_pfn(pte_pfn(pteval)))) {
1870                         if (!userfaultfd_armed(vma) &&
1871                             ++none_or_zero <= khugepaged_max_ptes_none) {
1872                                 continue;
1873                         } else {
1874                                 result = SCAN_EXCEED_NONE_PTE;
1875                                 goto out;
1876                         }
1877                 }
1878                 if (!pte_present(pteval)) {
1879                         result = SCAN_PTE_NON_PRESENT;
1880                         goto out;
1881                 }
1882                 page = vm_normal_page(vma, address, pteval);
1883                 if (unlikely(!page)) {
1884                         result = SCAN_PAGE_NULL;
1885                         goto out;
1886                 }
1887
1888                 VM_BUG_ON_PAGE(PageCompound(page), page);
1889                 VM_BUG_ON_PAGE(!PageAnon(page), page);
1890                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1891
1892                 /*
1893                  * We can do it before isolate_lru_page because the
1894                  * page can't be freed from under us. NOTE: PG_lock
1895                  * is needed to serialize against split_huge_page
1896                  * when invoked from the VM.
1897                  */
1898                 if (!trylock_page(page)) {
1899                         result = SCAN_PAGE_LOCK;
1900                         goto out;
1901                 }
1902
1903                 /*
1904                  * cannot use mapcount: can't collapse if there's a gup pin.
1905                  * The page must only be referenced by the scanned process
1906                  * and page swap cache.
1907                  */
1908                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1909                         unlock_page(page);
1910                         result = SCAN_PAGE_COUNT;
1911                         goto out;
1912                 }
1913                 if (pte_write(pteval)) {
1914                         writable = true;
1915                 } else {
1916                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
1917                                 unlock_page(page);
1918                                 result = SCAN_SWAP_CACHE_PAGE;
1919                                 goto out;
1920                         }
1921                         /*
1922                          * Page is not in the swap cache. It can be collapsed
1923                          * into a THP.
1924                          */
1925                 }
1926
1927                 /*
1928                  * Isolate the page to avoid collapsing an hugepage
1929                  * currently in use by the VM.
1930                  */
1931                 if (isolate_lru_page(page)) {
1932                         unlock_page(page);
1933                         result = SCAN_DEL_PAGE_LRU;
1934                         goto out;
1935                 }
1936                 /* 0 stands for page_is_file_cache(page) == false */
1937                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1938                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1939                 VM_BUG_ON_PAGE(PageLRU(page), page);
1940
1941                 /* If there is no mapped pte young don't collapse the page */
1942                 if (pte_young(pteval) ||
1943                     page_is_young(page) || PageReferenced(page) ||
1944                     mmu_notifier_test_young(vma->vm_mm, address))
1945                         referenced = true;
1946         }
1947         if (likely(writable)) {
1948                 if (likely(referenced)) {
1949                         result = SCAN_SUCCEED;
1950                         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
1951                                                             referenced, writable, result);
1952                         return 1;
1953                 }
1954         } else {
1955                 result = SCAN_PAGE_RO;
1956         }
1957
1958 out:
1959         release_pte_pages(pte, _pte);
1960         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
1961                                             referenced, writable, result);
1962         return 0;
1963 }
1964
1965 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1966                                       struct vm_area_struct *vma,
1967                                       unsigned long address,
1968                                       spinlock_t *ptl)
1969 {
1970         pte_t *_pte;
1971         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1972                 pte_t pteval = *_pte;
1973                 struct page *src_page;
1974
1975                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1976                         clear_user_highpage(page, address);
1977                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1978                         if (is_zero_pfn(pte_pfn(pteval))) {
1979                                 /*
1980                                  * ptl mostly unnecessary.
1981                                  */
1982                                 spin_lock(ptl);
1983                                 /*
1984                                  * paravirt calls inside pte_clear here are
1985                                  * superfluous.
1986                                  */
1987                                 pte_clear(vma->vm_mm, address, _pte);
1988                                 spin_unlock(ptl);
1989                         }
1990                 } else {
1991                         src_page = pte_page(pteval);
1992                         copy_user_highpage(page, src_page, address, vma);
1993                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
1994                         release_pte_page(src_page);
1995                         /*
1996                          * ptl mostly unnecessary, but preempt has to
1997                          * be disabled to update the per-cpu stats
1998                          * inside page_remove_rmap().
1999                          */
2000                         spin_lock(ptl);
2001                         /*
2002                          * paravirt calls inside pte_clear here are
2003                          * superfluous.
2004                          */
2005                         pte_clear(vma->vm_mm, address, _pte);
2006                         page_remove_rmap(src_page, false);
2007                         spin_unlock(ptl);
2008                         free_page_and_swap_cache(src_page);
2009                 }
2010
2011                 address += PAGE_SIZE;
2012                 page++;
2013         }
2014 }
2015
2016 static void khugepaged_alloc_sleep(void)
2017 {
2018         DEFINE_WAIT(wait);
2019
2020         add_wait_queue(&khugepaged_wait, &wait);
2021         freezable_schedule_timeout_interruptible(
2022                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2023         remove_wait_queue(&khugepaged_wait, &wait);
2024 }
2025
2026 static int khugepaged_node_load[MAX_NUMNODES];
2027
2028 static bool khugepaged_scan_abort(int nid)
2029 {
2030         int i;
2031
2032         /*
2033          * If zone_reclaim_mode is disabled, then no extra effort is made to
2034          * allocate memory locally.
2035          */
2036         if (!zone_reclaim_mode)
2037                 return false;
2038
2039         /* If there is a count for this node already, it must be acceptable */
2040         if (khugepaged_node_load[nid])
2041                 return false;
2042
2043         for (i = 0; i < MAX_NUMNODES; i++) {
2044                 if (!khugepaged_node_load[i])
2045                         continue;
2046                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2047                         return true;
2048         }
2049         return false;
2050 }
2051
2052 #ifdef CONFIG_NUMA
2053 static int khugepaged_find_target_node(void)
2054 {
2055         static int last_khugepaged_target_node = NUMA_NO_NODE;
2056         int nid, target_node = 0, max_value = 0;
2057
2058         /* find first node with max normal pages hit */
2059         for (nid = 0; nid < MAX_NUMNODES; nid++)
2060                 if (khugepaged_node_load[nid] > max_value) {
2061                         max_value = khugepaged_node_load[nid];
2062                         target_node = nid;
2063                 }
2064
2065         /* do some balance if several nodes have the same hit record */
2066         if (target_node <= last_khugepaged_target_node)
2067                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2068                                 nid++)
2069                         if (max_value == khugepaged_node_load[nid]) {
2070                                 target_node = nid;
2071                                 break;
2072                         }
2073
2074         last_khugepaged_target_node = target_node;
2075         return target_node;
2076 }
2077
2078 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2079 {
2080         if (IS_ERR(*hpage)) {
2081                 if (!*wait)
2082                         return false;
2083
2084                 *wait = false;
2085                 *hpage = NULL;
2086                 khugepaged_alloc_sleep();
2087         } else if (*hpage) {
2088                 put_page(*hpage);
2089                 *hpage = NULL;
2090         }
2091
2092         return true;
2093 }
2094
2095 static struct page *
2096 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2097                        unsigned long address, int node)
2098 {
2099         VM_BUG_ON_PAGE(*hpage, *hpage);
2100
2101         /*
2102          * Before allocating the hugepage, release the mmap_sem read lock.
2103          * The allocation can take potentially a long time if it involves
2104          * sync compaction, and we do not need to hold the mmap_sem during
2105          * that. We will recheck the vma after taking it again in write mode.
2106          */
2107         up_read(&mm->mmap_sem);
2108
2109         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2110         if (unlikely(!*hpage)) {
2111                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2112                 *hpage = ERR_PTR(-ENOMEM);
2113                 return NULL;
2114         }
2115
2116         prep_transhuge_page(*hpage);
2117         count_vm_event(THP_COLLAPSE_ALLOC);
2118         return *hpage;
2119 }
2120 #else
2121 static int khugepaged_find_target_node(void)
2122 {
2123         return 0;
2124 }
2125
2126 static inline struct page *alloc_hugepage(int defrag)
2127 {
2128         struct page *page;
2129
2130         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2131         if (page)
2132                 prep_transhuge_page(page);
2133         return page;
2134 }
2135
2136 static struct page *khugepaged_alloc_hugepage(bool *wait)
2137 {
2138         struct page *hpage;
2139
2140         do {
2141                 hpage = alloc_hugepage(khugepaged_defrag());
2142                 if (!hpage) {
2143                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2144                         if (!*wait)
2145                                 return NULL;
2146
2147                         *wait = false;
2148                         khugepaged_alloc_sleep();
2149                 } else
2150                         count_vm_event(THP_COLLAPSE_ALLOC);
2151         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2152
2153         return hpage;
2154 }
2155
2156 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2157 {
2158         if (!*hpage)
2159                 *hpage = khugepaged_alloc_hugepage(wait);
2160
2161         if (unlikely(!*hpage))
2162                 return false;
2163
2164         return true;
2165 }
2166
2167 static struct page *
2168 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2169                        unsigned long address, int node)
2170 {
2171         up_read(&mm->mmap_sem);
2172         VM_BUG_ON(!*hpage);
2173
2174         return  *hpage;
2175 }
2176 #endif
2177
2178 static bool hugepage_vma_check(struct vm_area_struct *vma)
2179 {
2180         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2181             (vma->vm_flags & VM_NOHUGEPAGE))
2182                 return false;
2183         if (!vma->anon_vma || vma->vm_ops)
2184                 return false;
2185         if (is_vma_temporary_stack(vma))
2186                 return false;
2187         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2188         return true;
2189 }
2190
2191 static void collapse_huge_page(struct mm_struct *mm,
2192                                    unsigned long address,
2193                                    struct page **hpage,
2194                                    struct vm_area_struct *vma,
2195                                    int node)
2196 {
2197         pmd_t *pmd, _pmd;
2198         pte_t *pte;
2199         pgtable_t pgtable;
2200         struct page *new_page;
2201         spinlock_t *pmd_ptl, *pte_ptl;
2202         int isolated, result = 0;
2203         unsigned long hstart, hend;
2204         struct mem_cgroup *memcg;
2205         unsigned long mmun_start;       /* For mmu_notifiers */
2206         unsigned long mmun_end;         /* For mmu_notifiers */
2207         gfp_t gfp;
2208
2209         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2210
2211         /* Only allocate from the target node */
2212         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2213                 __GFP_THISNODE;
2214
2215         /* release the mmap_sem read lock. */
2216         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2217         if (!new_page) {
2218                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2219                 goto out_nolock;
2220         }
2221
2222         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2223                 result = SCAN_CGROUP_CHARGE_FAIL;
2224                 goto out_nolock;
2225         }
2226
2227         /*
2228          * Prevent all access to pagetables with the exception of
2229          * gup_fast later hanlded by the ptep_clear_flush and the VM
2230          * handled by the anon_vma lock + PG_lock.
2231          */
2232         down_write(&mm->mmap_sem);
2233         if (unlikely(khugepaged_test_exit(mm))) {
2234                 result = SCAN_ANY_PROCESS;
2235                 goto out;
2236         }
2237
2238         vma = find_vma(mm, address);
2239         if (!vma) {
2240                 result = SCAN_VMA_NULL;
2241                 goto out;
2242         }
2243         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2244         hend = vma->vm_end & HPAGE_PMD_MASK;
2245         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2246                 result = SCAN_ADDRESS_RANGE;
2247                 goto out;
2248         }
2249         if (!hugepage_vma_check(vma)) {
2250                 result = SCAN_VMA_CHECK;
2251                 goto out;
2252         }
2253         pmd = mm_find_pmd(mm, address);
2254         if (!pmd) {
2255                 result = SCAN_PMD_NULL;
2256                 goto out;
2257         }
2258
2259         anon_vma_lock_write(vma->anon_vma);
2260
2261         pte = pte_offset_map(pmd, address);
2262         pte_ptl = pte_lockptr(mm, pmd);
2263
2264         mmun_start = address;
2265         mmun_end   = address + HPAGE_PMD_SIZE;
2266         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2267         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2268         /*
2269          * After this gup_fast can't run anymore. This also removes
2270          * any huge TLB entry from the CPU so we won't allow
2271          * huge and small TLB entries for the same virtual address
2272          * to avoid the risk of CPU bugs in that area.
2273          */
2274         _pmd = pmdp_collapse_flush(vma, address, pmd);
2275         spin_unlock(pmd_ptl);
2276         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2277
2278         spin_lock(pte_ptl);
2279         isolated = __collapse_huge_page_isolate(vma, address, pte);
2280         spin_unlock(pte_ptl);
2281
2282         if (unlikely(!isolated)) {
2283                 pte_unmap(pte);
2284                 spin_lock(pmd_ptl);
2285                 BUG_ON(!pmd_none(*pmd));
2286                 /*
2287                  * We can only use set_pmd_at when establishing
2288                  * hugepmds and never for establishing regular pmds that
2289                  * points to regular pagetables. Use pmd_populate for that
2290                  */
2291                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2292                 spin_unlock(pmd_ptl);
2293                 anon_vma_unlock_write(vma->anon_vma);
2294                 result = SCAN_FAIL;
2295                 goto out;
2296         }
2297
2298         /*
2299          * All pages are isolated and locked so anon_vma rmap
2300          * can't run anymore.
2301          */
2302         anon_vma_unlock_write(vma->anon_vma);
2303
2304         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2305         pte_unmap(pte);
2306         __SetPageUptodate(new_page);
2307         pgtable = pmd_pgtable(_pmd);
2308
2309         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2310         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2311
2312         /*
2313          * spin_lock() below is not the equivalent of smp_wmb(), so
2314          * this is needed to avoid the copy_huge_page writes to become
2315          * visible after the set_pmd_at() write.
2316          */
2317         smp_wmb();
2318
2319         spin_lock(pmd_ptl);
2320         BUG_ON(!pmd_none(*pmd));
2321         page_add_new_anon_rmap(new_page, vma, address, true);
2322         mem_cgroup_commit_charge(new_page, memcg, false, true);
2323         lru_cache_add_active_or_unevictable(new_page, vma);
2324         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2325         set_pmd_at(mm, address, pmd, _pmd);
2326         update_mmu_cache_pmd(vma, address, pmd);
2327         spin_unlock(pmd_ptl);
2328
2329         *hpage = NULL;
2330
2331         khugepaged_pages_collapsed++;
2332         result = SCAN_SUCCEED;
2333 out_up_write:
2334         up_write(&mm->mmap_sem);
2335         trace_mm_collapse_huge_page(mm, isolated, result);
2336         return;
2337
2338 out_nolock:
2339         trace_mm_collapse_huge_page(mm, isolated, result);
2340         return;
2341 out:
2342         mem_cgroup_cancel_charge(new_page, memcg, true);
2343         goto out_up_write;
2344 }
2345
2346 static int khugepaged_scan_pmd(struct mm_struct *mm,
2347                                struct vm_area_struct *vma,
2348                                unsigned long address,
2349                                struct page **hpage)
2350 {
2351         pmd_t *pmd;
2352         pte_t *pte, *_pte;
2353         int ret = 0, none_or_zero = 0, result = 0;
2354         struct page *page = NULL;
2355         unsigned long _address;
2356         spinlock_t *ptl;
2357         int node = NUMA_NO_NODE;
2358         bool writable = false, referenced = false;
2359
2360         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2361
2362         pmd = mm_find_pmd(mm, address);
2363         if (!pmd) {
2364                 result = SCAN_PMD_NULL;
2365                 goto out;
2366         }
2367
2368         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2369         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2370         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2371              _pte++, _address += PAGE_SIZE) {
2372                 pte_t pteval = *_pte;
2373                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2374                         if (!userfaultfd_armed(vma) &&
2375                             ++none_or_zero <= khugepaged_max_ptes_none) {
2376                                 continue;
2377                         } else {
2378                                 result = SCAN_EXCEED_NONE_PTE;
2379                                 goto out_unmap;
2380                         }
2381                 }
2382                 if (!pte_present(pteval)) {
2383                         result = SCAN_PTE_NON_PRESENT;
2384                         goto out_unmap;
2385                 }
2386                 if (pte_write(pteval))
2387                         writable = true;
2388
2389                 page = vm_normal_page(vma, _address, pteval);
2390                 if (unlikely(!page)) {
2391                         result = SCAN_PAGE_NULL;
2392                         goto out_unmap;
2393                 }
2394
2395                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2396                 if (PageCompound(page)) {
2397                         result = SCAN_PAGE_COMPOUND;
2398                         goto out_unmap;
2399                 }
2400
2401                 /*
2402                  * Record which node the original page is from and save this
2403                  * information to khugepaged_node_load[].
2404                  * Khupaged will allocate hugepage from the node has the max
2405                  * hit record.
2406                  */
2407                 node = page_to_nid(page);
2408                 if (khugepaged_scan_abort(node)) {
2409                         result = SCAN_SCAN_ABORT;
2410                         goto out_unmap;
2411                 }
2412                 khugepaged_node_load[node]++;
2413                 if (!PageLRU(page)) {
2414                         result = SCAN_SCAN_ABORT;
2415                         goto out_unmap;
2416                 }
2417                 if (PageLocked(page)) {
2418                         result = SCAN_PAGE_LOCK;
2419                         goto out_unmap;
2420                 }
2421                 if (!PageAnon(page)) {
2422                         result = SCAN_PAGE_ANON;
2423                         goto out_unmap;
2424                 }
2425
2426                 /*
2427                  * cannot use mapcount: can't collapse if there's a gup pin.
2428                  * The page must only be referenced by the scanned process
2429                  * and page swap cache.
2430                  */
2431                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2432                         result = SCAN_PAGE_COUNT;
2433                         goto out_unmap;
2434                 }
2435                 if (pte_young(pteval) ||
2436                     page_is_young(page) || PageReferenced(page) ||
2437                     mmu_notifier_test_young(vma->vm_mm, address))
2438                         referenced = true;
2439         }
2440         if (writable) {
2441                 if (referenced) {
2442                         result = SCAN_SUCCEED;
2443                         ret = 1;
2444                 } else {
2445                         result = SCAN_NO_REFERENCED_PAGE;
2446                 }
2447         } else {
2448                 result = SCAN_PAGE_RO;
2449         }
2450 out_unmap:
2451         pte_unmap_unlock(pte, ptl);
2452         if (ret) {
2453                 node = khugepaged_find_target_node();
2454                 /* collapse_huge_page will return with the mmap_sem released */
2455                 collapse_huge_page(mm, address, hpage, vma, node);
2456         }
2457 out:
2458         trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced,
2459                                      none_or_zero, result);
2460         return ret;
2461 }
2462
2463 static void collect_mm_slot(struct mm_slot *mm_slot)
2464 {
2465         struct mm_struct *mm = mm_slot->mm;
2466
2467         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2468
2469         if (khugepaged_test_exit(mm)) {
2470                 /* free mm_slot */
2471                 hash_del(&mm_slot->hash);
2472                 list_del(&mm_slot->mm_node);
2473
2474                 /*
2475                  * Not strictly needed because the mm exited already.
2476                  *
2477                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2478                  */
2479
2480                 /* khugepaged_mm_lock actually not necessary for the below */
2481                 free_mm_slot(mm_slot);
2482                 mmdrop(mm);
2483         }
2484 }
2485
2486 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2487                                             struct page **hpage)
2488         __releases(&khugepaged_mm_lock)
2489         __acquires(&khugepaged_mm_lock)
2490 {
2491         struct mm_slot *mm_slot;
2492         struct mm_struct *mm;
2493         struct vm_area_struct *vma;
2494         int progress = 0;
2495
2496         VM_BUG_ON(!pages);
2497         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2498
2499         if (khugepaged_scan.mm_slot)
2500                 mm_slot = khugepaged_scan.mm_slot;
2501         else {
2502                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2503                                      struct mm_slot, mm_node);
2504                 khugepaged_scan.address = 0;
2505                 khugepaged_scan.mm_slot = mm_slot;
2506         }
2507         spin_unlock(&khugepaged_mm_lock);
2508
2509         mm = mm_slot->mm;
2510         down_read(&mm->mmap_sem);
2511         if (unlikely(khugepaged_test_exit(mm)))
2512                 vma = NULL;
2513         else
2514                 vma = find_vma(mm, khugepaged_scan.address);
2515
2516         progress++;
2517         for (; vma; vma = vma->vm_next) {
2518                 unsigned long hstart, hend;
2519
2520                 cond_resched();
2521                 if (unlikely(khugepaged_test_exit(mm))) {
2522                         progress++;
2523                         break;
2524                 }
2525                 if (!hugepage_vma_check(vma)) {
2526 skip:
2527                         progress++;
2528                         continue;
2529                 }
2530                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2531                 hend = vma->vm_end & HPAGE_PMD_MASK;
2532                 if (hstart >= hend)
2533                         goto skip;
2534                 if (khugepaged_scan.address > hend)
2535                         goto skip;
2536                 if (khugepaged_scan.address < hstart)
2537                         khugepaged_scan.address = hstart;
2538                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2539
2540                 while (khugepaged_scan.address < hend) {
2541                         int ret;
2542                         cond_resched();
2543                         if (unlikely(khugepaged_test_exit(mm)))
2544                                 goto breakouterloop;
2545
2546                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2547                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2548                                   hend);
2549                         ret = khugepaged_scan_pmd(mm, vma,
2550                                                   khugepaged_scan.address,
2551                                                   hpage);
2552                         /* move to next address */
2553                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2554                         progress += HPAGE_PMD_NR;
2555                         if (ret)
2556                                 /* we released mmap_sem so break loop */
2557                                 goto breakouterloop_mmap_sem;
2558                         if (progress >= pages)
2559                                 goto breakouterloop;
2560                 }
2561         }
2562 breakouterloop:
2563         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2564 breakouterloop_mmap_sem:
2565
2566         spin_lock(&khugepaged_mm_lock);
2567         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2568         /*
2569          * Release the current mm_slot if this mm is about to die, or
2570          * if we scanned all vmas of this mm.
2571          */
2572         if (khugepaged_test_exit(mm) || !vma) {
2573                 /*
2574                  * Make sure that if mm_users is reaching zero while
2575                  * khugepaged runs here, khugepaged_exit will find
2576                  * mm_slot not pointing to the exiting mm.
2577                  */
2578                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2579                         khugepaged_scan.mm_slot = list_entry(
2580                                 mm_slot->mm_node.next,
2581                                 struct mm_slot, mm_node);
2582                         khugepaged_scan.address = 0;
2583                 } else {
2584                         khugepaged_scan.mm_slot = NULL;
2585                         khugepaged_full_scans++;
2586                 }
2587
2588                 collect_mm_slot(mm_slot);
2589         }
2590
2591         return progress;
2592 }
2593
2594 static int khugepaged_has_work(void)
2595 {
2596         return !list_empty(&khugepaged_scan.mm_head) &&
2597                 khugepaged_enabled();
2598 }
2599
2600 static int khugepaged_wait_event(void)
2601 {
2602         return !list_empty(&khugepaged_scan.mm_head) ||
2603                 kthread_should_stop();
2604 }
2605
2606 static void khugepaged_do_scan(void)
2607 {
2608         struct page *hpage = NULL;
2609         unsigned int progress = 0, pass_through_head = 0;
2610         unsigned int pages = khugepaged_pages_to_scan;
2611         bool wait = true;
2612
2613         barrier(); /* write khugepaged_pages_to_scan to local stack */
2614
2615         while (progress < pages) {
2616                 if (!khugepaged_prealloc_page(&hpage, &wait))
2617                         break;
2618
2619                 cond_resched();
2620
2621                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2622                         break;
2623
2624                 spin_lock(&khugepaged_mm_lock);
2625                 if (!khugepaged_scan.mm_slot)
2626                         pass_through_head++;
2627                 if (khugepaged_has_work() &&
2628                     pass_through_head < 2)
2629                         progress += khugepaged_scan_mm_slot(pages - progress,
2630                                                             &hpage);
2631                 else
2632                         progress = pages;
2633                 spin_unlock(&khugepaged_mm_lock);
2634         }
2635
2636         if (!IS_ERR_OR_NULL(hpage))
2637                 put_page(hpage);
2638 }
2639
2640 static void khugepaged_wait_work(void)
2641 {
2642         if (khugepaged_has_work()) {
2643                 if (!khugepaged_scan_sleep_millisecs)
2644                         return;
2645
2646                 wait_event_freezable_timeout(khugepaged_wait,
2647                                              kthread_should_stop(),
2648                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2649                 return;
2650         }
2651
2652         if (khugepaged_enabled())
2653                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2654 }
2655
2656 static int khugepaged(void *none)
2657 {
2658         struct mm_slot *mm_slot;
2659
2660         set_freezable();
2661         set_user_nice(current, MAX_NICE);
2662
2663         while (!kthread_should_stop()) {
2664                 khugepaged_do_scan();
2665                 khugepaged_wait_work();
2666         }
2667
2668         spin_lock(&khugepaged_mm_lock);
2669         mm_slot = khugepaged_scan.mm_slot;
2670         khugepaged_scan.mm_slot = NULL;
2671         if (mm_slot)
2672                 collect_mm_slot(mm_slot);
2673         spin_unlock(&khugepaged_mm_lock);
2674         return 0;
2675 }
2676
2677 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2678                 unsigned long haddr, pmd_t *pmd)
2679 {
2680         struct mm_struct *mm = vma->vm_mm;
2681         pgtable_t pgtable;
2682         pmd_t _pmd;
2683         int i;
2684
2685         /* leave pmd empty until pte is filled */
2686         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2687
2688         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2689         pmd_populate(mm, &_pmd, pgtable);
2690
2691         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2692                 pte_t *pte, entry;
2693                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2694                 entry = pte_mkspecial(entry);
2695                 pte = pte_offset_map(&_pmd, haddr);
2696                 VM_BUG_ON(!pte_none(*pte));
2697                 set_pte_at(mm, haddr, pte, entry);
2698                 pte_unmap(pte);
2699         }
2700         smp_wmb(); /* make pte visible before pmd */
2701         pmd_populate(mm, pmd, pgtable);
2702         put_huge_zero_page();
2703 }
2704
2705 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2706                 unsigned long haddr, bool freeze)
2707 {
2708         struct mm_struct *mm = vma->vm_mm;
2709         struct page *page;
2710         pgtable_t pgtable;
2711         pmd_t _pmd;
2712         bool young, write;
2713         int i;
2714
2715         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2716         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2717         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2718         VM_BUG_ON(!pmd_trans_huge(*pmd));
2719
2720         count_vm_event(THP_SPLIT_PMD);
2721
2722         if (vma_is_dax(vma)) {
2723                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2724                 if (is_huge_zero_pmd(_pmd))
2725                         put_huge_zero_page();
2726                 return;
2727         } else if (is_huge_zero_pmd(*pmd)) {
2728                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2729         }
2730
2731         page = pmd_page(*pmd);
2732         VM_BUG_ON_PAGE(!page_count(page), page);
2733         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2734         write = pmd_write(*pmd);
2735         young = pmd_young(*pmd);
2736
2737         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2738         pmd_populate(mm, &_pmd, pgtable);
2739
2740         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2741                 pte_t entry, *pte;
2742                 /*
2743                  * Note that NUMA hinting access restrictions are not
2744                  * transferred to avoid any possibility of altering
2745                  * permissions across VMAs.
2746                  */
2747                 if (freeze) {
2748                         swp_entry_t swp_entry;
2749                         swp_entry = make_migration_entry(page + i, write);
2750                         entry = swp_entry_to_pte(swp_entry);
2751                 } else {
2752                         entry = mk_pte(page + i, vma->vm_page_prot);
2753                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2754                         if (!write)
2755                                 entry = pte_wrprotect(entry);
2756                         if (!young)
2757                                 entry = pte_mkold(entry);
2758                 }
2759                 pte = pte_offset_map(&_pmd, haddr);
2760                 BUG_ON(!pte_none(*pte));
2761                 set_pte_at(mm, haddr, pte, entry);
2762                 atomic_inc(&page[i]._mapcount);
2763                 pte_unmap(pte);
2764         }
2765
2766         /*
2767          * Set PG_double_map before dropping compound_mapcount to avoid
2768          * false-negative page_mapped().
2769          */
2770         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2771                 for (i = 0; i < HPAGE_PMD_NR; i++)
2772                         atomic_inc(&page[i]._mapcount);
2773         }
2774
2775         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2776                 /* Last compound_mapcount is gone. */
2777                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2778                 if (TestClearPageDoubleMap(page)) {
2779                         /* No need in mapcount reference anymore */
2780                         for (i = 0; i < HPAGE_PMD_NR; i++)
2781                                 atomic_dec(&page[i]._mapcount);
2782                 }
2783         }
2784
2785         smp_wmb(); /* make pte visible before pmd */
2786         /*
2787          * Up to this point the pmd is present and huge and userland has the
2788          * whole access to the hugepage during the split (which happens in
2789          * place). If we overwrite the pmd with the not-huge version pointing
2790          * to the pte here (which of course we could if all CPUs were bug
2791          * free), userland could trigger a small page size TLB miss on the
2792          * small sized TLB while the hugepage TLB entry is still established in
2793          * the huge TLB. Some CPU doesn't like that.
2794          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2795          * 383 on page 93. Intel should be safe but is also warns that it's
2796          * only safe if the permission and cache attributes of the two entries
2797          * loaded in the two TLB is identical (which should be the case here).
2798          * But it is generally safer to never allow small and huge TLB entries
2799          * for the same virtual address to be loaded simultaneously. So instead
2800          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2801          * current pmd notpresent (atomically because here the pmd_trans_huge
2802          * and pmd_trans_splitting must remain set at all times on the pmd
2803          * until the split is complete for this pmd), then we flush the SMP TLB
2804          * and finally we write the non-huge version of the pmd entry with
2805          * pmd_populate.
2806          */
2807         pmdp_invalidate(vma, haddr, pmd);
2808         pmd_populate(mm, pmd, pgtable);
2809
2810         if (freeze) {
2811                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2812                         page_remove_rmap(page + i, false);
2813                         put_page(page + i);
2814                 }
2815         }
2816 }
2817
2818 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2819                 unsigned long address)
2820 {
2821         spinlock_t *ptl;
2822         struct mm_struct *mm = vma->vm_mm;
2823         struct page *page = NULL;
2824         unsigned long haddr = address & HPAGE_PMD_MASK;
2825
2826         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2827         ptl = pmd_lock(mm, pmd);
2828         if (unlikely(!pmd_trans_huge(*pmd)))
2829                 goto out;
2830         page = pmd_page(*pmd);
2831         __split_huge_pmd_locked(vma, pmd, haddr, false);
2832         if (PageMlocked(page))
2833                 get_page(page);
2834         else
2835                 page = NULL;
2836 out:
2837         spin_unlock(ptl);
2838         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2839         if (page) {
2840                 lock_page(page);
2841                 munlock_vma_page(page);
2842                 unlock_page(page);
2843                 put_page(page);
2844         }
2845 }
2846
2847 static void split_huge_pmd_address(struct vm_area_struct *vma,
2848                                     unsigned long address)
2849 {
2850         pgd_t *pgd;
2851         pud_t *pud;
2852         pmd_t *pmd;
2853
2854         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2855
2856         pgd = pgd_offset(vma->vm_mm, address);
2857         if (!pgd_present(*pgd))
2858                 return;
2859
2860         pud = pud_offset(pgd, address);
2861         if (!pud_present(*pud))
2862                 return;
2863
2864         pmd = pmd_offset(pud, address);
2865         if (!pmd_present(*pmd) || !pmd_trans_huge(*pmd))
2866                 return;
2867         /*
2868          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2869          * materialize from under us.
2870          */
2871         split_huge_pmd(vma, pmd, address);
2872 }
2873
2874 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2875                              unsigned long start,
2876                              unsigned long end,
2877                              long adjust_next)
2878 {
2879         /*
2880          * If the new start address isn't hpage aligned and it could
2881          * previously contain an hugepage: check if we need to split
2882          * an huge pmd.
2883          */
2884         if (start & ~HPAGE_PMD_MASK &&
2885             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2886             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2887                 split_huge_pmd_address(vma, start);
2888
2889         /*
2890          * If the new end address isn't hpage aligned and it could
2891          * previously contain an hugepage: check if we need to split
2892          * an huge pmd.
2893          */
2894         if (end & ~HPAGE_PMD_MASK &&
2895             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2896             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2897                 split_huge_pmd_address(vma, end);
2898
2899         /*
2900          * If we're also updating the vma->vm_next->vm_start, if the new
2901          * vm_next->vm_start isn't page aligned and it could previously
2902          * contain an hugepage: check if we need to split an huge pmd.
2903          */
2904         if (adjust_next > 0) {
2905                 struct vm_area_struct *next = vma->vm_next;
2906                 unsigned long nstart = next->vm_start;
2907                 nstart += adjust_next << PAGE_SHIFT;
2908                 if (nstart & ~HPAGE_PMD_MASK &&
2909                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2910                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2911                         split_huge_pmd_address(next, nstart);
2912         }
2913 }
2914
2915 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
2916                 unsigned long address)
2917 {
2918         spinlock_t *ptl;
2919         pgd_t *pgd;
2920         pud_t *pud;
2921         pmd_t *pmd;
2922         pte_t *pte;
2923         int i, nr = HPAGE_PMD_NR;
2924
2925         /* Skip pages which doesn't belong to the VMA */
2926         if (address < vma->vm_start) {
2927                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
2928                 page += off;
2929                 nr -= off;
2930                 address = vma->vm_start;
2931         }
2932
2933         pgd = pgd_offset(vma->vm_mm, address);
2934         if (!pgd_present(*pgd))
2935                 return;
2936         pud = pud_offset(pgd, address);
2937         if (!pud_present(*pud))
2938                 return;
2939         pmd = pmd_offset(pud, address);
2940         ptl = pmd_lock(vma->vm_mm, pmd);
2941         if (!pmd_present(*pmd)) {
2942                 spin_unlock(ptl);
2943                 return;
2944         }
2945         if (pmd_trans_huge(*pmd)) {
2946                 if (page == pmd_page(*pmd))
2947                         __split_huge_pmd_locked(vma, pmd, address, true);
2948                 spin_unlock(ptl);
2949                 return;
2950         }
2951         spin_unlock(ptl);
2952
2953         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
2954         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
2955                 pte_t entry, swp_pte;
2956                 swp_entry_t swp_entry;
2957
2958                 if (!pte_present(pte[i]))
2959                         continue;
2960                 if (page_to_pfn(page) != pte_pfn(pte[i]))
2961                         continue;
2962                 flush_cache_page(vma, address, page_to_pfn(page));
2963                 entry = ptep_clear_flush(vma, address, pte + i);
2964                 swp_entry = make_migration_entry(page, pte_write(entry));
2965                 swp_pte = swp_entry_to_pte(swp_entry);
2966                 if (pte_soft_dirty(entry))
2967                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
2968                 set_pte_at(vma->vm_mm, address, pte + i, swp_pte);
2969                 page_remove_rmap(page, false);
2970                 put_page(page);
2971         }
2972         pte_unmap_unlock(pte, ptl);
2973 }
2974
2975 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
2976 {
2977         struct anon_vma_chain *avc;
2978         pgoff_t pgoff = page_to_pgoff(page);
2979
2980         VM_BUG_ON_PAGE(!PageHead(page), page);
2981
2982         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
2983                         pgoff + HPAGE_PMD_NR - 1) {
2984                 unsigned long haddr;
2985
2986                 haddr = __vma_address(page, avc->vma) & HPAGE_PMD_MASK;
2987                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
2988                                 haddr, haddr + HPAGE_PMD_SIZE);
2989                 freeze_page_vma(avc->vma, page, haddr);
2990                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
2991                                 haddr, haddr + HPAGE_PMD_SIZE);
2992         }
2993 }
2994
2995 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
2996                 unsigned long address)
2997 {
2998         spinlock_t *ptl;
2999         pmd_t *pmd;
3000         pte_t *pte, entry;
3001         swp_entry_t swp_entry;
3002         int i, nr = HPAGE_PMD_NR;
3003
3004         /* Skip pages which doesn't belong to the VMA */
3005         if (address < vma->vm_start) {
3006                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3007                 page += off;
3008                 nr -= off;
3009                 address = vma->vm_start;
3010         }
3011
3012         pmd = mm_find_pmd(vma->vm_mm, address);
3013         if (!pmd)
3014                 return;
3015         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3016         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
3017                 if (!is_swap_pte(pte[i]))
3018                         continue;
3019
3020                 swp_entry = pte_to_swp_entry(pte[i]);
3021                 if (!is_migration_entry(swp_entry))
3022                         continue;
3023                 if (migration_entry_to_page(swp_entry) != page)
3024                         continue;
3025
3026                 get_page(page);
3027                 page_add_anon_rmap(page, vma, address, false);
3028
3029                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3030                 entry = pte_mkdirty(entry);
3031                 if (is_write_migration_entry(swp_entry))
3032                         entry = maybe_mkwrite(entry, vma);
3033
3034                 flush_dcache_page(page);
3035                 set_pte_at(vma->vm_mm, address, pte + i, entry);
3036
3037                 /* No need to invalidate - it was non-present before */
3038                 update_mmu_cache(vma, address, pte + i);
3039         }
3040         pte_unmap_unlock(pte, ptl);
3041 }
3042
3043 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3044 {
3045         struct anon_vma_chain *avc;
3046         pgoff_t pgoff = page_to_pgoff(page);
3047
3048         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3049                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3050                 unsigned long address = __vma_address(page, avc->vma);
3051
3052                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3053                                 address, address + HPAGE_PMD_SIZE);
3054                 unfreeze_page_vma(avc->vma, page, address);
3055                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3056                                 address, address + HPAGE_PMD_SIZE);
3057         }
3058 }
3059
3060 static int __split_huge_page_tail(struct page *head, int tail,
3061                 struct lruvec *lruvec, struct list_head *list)
3062 {
3063         int mapcount;
3064         struct page *page_tail = head + tail;
3065
3066         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3067         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3068
3069         /*
3070          * tail_page->_count is zero and not changing from under us. But
3071          * get_page_unless_zero() may be running from under us on the
3072          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3073          * would then run atomic_set() concurrently with
3074          * get_page_unless_zero(), and atomic_set() is implemented in C not
3075          * using locked ops. spin_unlock on x86 sometime uses locked ops
3076          * because of PPro errata 66, 92, so unless somebody can guarantee
3077          * atomic_set() here would be safe on all archs (and not only on x86),
3078          * it's safer to use atomic_add().
3079          */
3080         atomic_add(mapcount + 1, &page_tail->_count);
3081
3082
3083         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3084         page_tail->flags |= (head->flags &
3085                         ((1L << PG_referenced) |
3086                          (1L << PG_swapbacked) |
3087                          (1L << PG_mlocked) |
3088                          (1L << PG_uptodate) |
3089                          (1L << PG_active) |
3090                          (1L << PG_locked) |
3091                          (1L << PG_unevictable)));
3092         page_tail->flags |= (1L << PG_dirty);
3093
3094         /*
3095          * After clearing PageTail the gup refcount can be released.
3096          * Page flags also must be visible before we make the page non-compound.
3097          */
3098         smp_wmb();
3099
3100         clear_compound_head(page_tail);
3101
3102         if (page_is_young(head))
3103                 set_page_young(page_tail);
3104         if (page_is_idle(head))
3105                 set_page_idle(page_tail);
3106
3107         /* ->mapping in first tail page is compound_mapcount */
3108         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3109                         page_tail);
3110         page_tail->mapping = head->mapping;
3111
3112         page_tail->index = head->index + tail;
3113         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3114         lru_add_page_tail(head, page_tail, lruvec, list);
3115
3116         return mapcount;
3117 }
3118
3119 static void __split_huge_page(struct page *page, struct list_head *list)
3120 {
3121         struct page *head = compound_head(page);
3122         struct zone *zone = page_zone(head);
3123         struct lruvec *lruvec;
3124         int i, tail_mapcount;
3125
3126         /* prevent PageLRU to go away from under us, and freeze lru stats */
3127         spin_lock_irq(&zone->lru_lock);
3128         lruvec = mem_cgroup_page_lruvec(head, zone);
3129
3130         /* complete memcg works before add pages to LRU */
3131         mem_cgroup_split_huge_fixup(head);
3132
3133         tail_mapcount = 0;
3134         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3135                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3136         atomic_sub(tail_mapcount, &head->_count);
3137
3138         ClearPageCompound(head);
3139         spin_unlock_irq(&zone->lru_lock);
3140
3141         unfreeze_page(page_anon_vma(head), head);
3142
3143         for (i = 0; i < HPAGE_PMD_NR; i++) {
3144                 struct page *subpage = head + i;
3145                 if (subpage == page)
3146                         continue;
3147                 unlock_page(subpage);
3148
3149                 /*
3150                  * Subpages may be freed if there wasn't any mapping
3151                  * like if add_to_swap() is running on a lru page that
3152                  * had its mapping zapped. And freeing these pages
3153                  * requires taking the lru_lock so we do the put_page
3154                  * of the tail pages after the split is complete.
3155                  */
3156                 put_page(subpage);
3157         }
3158 }
3159
3160 int total_mapcount(struct page *page)
3161 {
3162         int i, ret;
3163
3164         VM_BUG_ON_PAGE(PageTail(page), page);
3165
3166         if (likely(!PageCompound(page)))
3167                 return atomic_read(&page->_mapcount) + 1;
3168
3169         ret = compound_mapcount(page);
3170         if (PageHuge(page))
3171                 return ret;
3172         for (i = 0; i < HPAGE_PMD_NR; i++)
3173                 ret += atomic_read(&page[i]._mapcount) + 1;
3174         if (PageDoubleMap(page))
3175                 ret -= HPAGE_PMD_NR;
3176         return ret;
3177 }
3178
3179 /*
3180  * This function splits huge page into normal pages. @page can point to any
3181  * subpage of huge page to split. Split doesn't change the position of @page.
3182  *
3183  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3184  * The huge page must be locked.
3185  *
3186  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3187  *
3188  * Both head page and tail pages will inherit mapping, flags, and so on from
3189  * the hugepage.
3190  *
3191  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3192  * they are not mapped.
3193  *
3194  * Returns 0 if the hugepage is split successfully.
3195  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3196  * us.
3197  */
3198 int split_huge_page_to_list(struct page *page, struct list_head *list)
3199 {
3200         struct page *head = compound_head(page);
3201         struct anon_vma *anon_vma;
3202         int count, mapcount, ret;
3203
3204         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3205         VM_BUG_ON_PAGE(!PageAnon(page), page);
3206         VM_BUG_ON_PAGE(!PageLocked(page), page);
3207         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3208         VM_BUG_ON_PAGE(!PageCompound(page), page);
3209
3210         /*
3211          * The caller does not necessarily hold an mmap_sem that would prevent
3212          * the anon_vma disappearing so we first we take a reference to it
3213          * and then lock the anon_vma for write. This is similar to
3214          * page_lock_anon_vma_read except the write lock is taken to serialise
3215          * against parallel split or collapse operations.
3216          */
3217         anon_vma = page_get_anon_vma(head);
3218         if (!anon_vma) {
3219                 ret = -EBUSY;
3220                 goto out;
3221         }
3222         anon_vma_lock_write(anon_vma);
3223
3224         /*
3225          * Racy check if we can split the page, before freeze_page() will
3226          * split PMDs
3227          */
3228         if (total_mapcount(head) != page_count(head) - 1) {
3229                 ret = -EBUSY;
3230                 goto out_unlock;
3231         }
3232
3233         freeze_page(anon_vma, head);
3234         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3235
3236         /* Prevent deferred_split_scan() touching ->_count */
3237         spin_lock(&split_queue_lock);
3238         count = page_count(head);
3239         mapcount = total_mapcount(head);
3240         if (mapcount == count - 1) {
3241                 if (!list_empty(page_deferred_list(head))) {
3242                         split_queue_len--;
3243                         list_del(page_deferred_list(head));
3244                 }
3245                 spin_unlock(&split_queue_lock);
3246                 __split_huge_page(page, list);
3247                 ret = 0;
3248         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount > count - 1) {
3249                 spin_unlock(&split_queue_lock);
3250                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3251                                 mapcount, count);
3252                 if (PageTail(page))
3253                         dump_page(head, NULL);
3254                 dump_page(page, "total_mapcount(head) > page_count(head) - 1");
3255                 BUG();
3256         } else {
3257                 spin_unlock(&split_queue_lock);
3258                 unfreeze_page(anon_vma, head);
3259                 ret = -EBUSY;
3260         }
3261
3262 out_unlock:
3263         anon_vma_unlock_write(anon_vma);
3264         put_anon_vma(anon_vma);
3265 out:
3266         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3267         return ret;
3268 }
3269
3270 void free_transhuge_page(struct page *page)
3271 {
3272         unsigned long flags;
3273
3274         spin_lock_irqsave(&split_queue_lock, flags);
3275         if (!list_empty(page_deferred_list(page))) {
3276                 split_queue_len--;
3277                 list_del(page_deferred_list(page));
3278         }
3279         spin_unlock_irqrestore(&split_queue_lock, flags);
3280         free_compound_page(page);
3281 }
3282
3283 void deferred_split_huge_page(struct page *page)
3284 {
3285         unsigned long flags;
3286
3287         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3288
3289         spin_lock_irqsave(&split_queue_lock, flags);
3290         if (list_empty(page_deferred_list(page))) {
3291                 list_add_tail(page_deferred_list(page), &split_queue);
3292                 split_queue_len++;
3293         }
3294         spin_unlock_irqrestore(&split_queue_lock, flags);
3295 }
3296
3297 static unsigned long deferred_split_count(struct shrinker *shrink,
3298                 struct shrink_control *sc)
3299 {
3300         /*
3301          * Split a page from split_queue will free up at least one page,
3302          * at most HPAGE_PMD_NR - 1. We don't track exact number.
3303          * Let's use HPAGE_PMD_NR / 2 as ballpark.
3304          */
3305         return ACCESS_ONCE(split_queue_len) * HPAGE_PMD_NR / 2;
3306 }
3307
3308 static unsigned long deferred_split_scan(struct shrinker *shrink,
3309                 struct shrink_control *sc)
3310 {
3311         unsigned long flags;
3312         LIST_HEAD(list), *pos, *next;
3313         struct page *page;
3314         int split = 0;
3315
3316         spin_lock_irqsave(&split_queue_lock, flags);
3317         list_splice_init(&split_queue, &list);
3318
3319         /* Take pin on all head pages to avoid freeing them under us */
3320         list_for_each_safe(pos, next, &list) {
3321                 page = list_entry((void *)pos, struct page, mapping);
3322                 page = compound_head(page);
3323                 /* race with put_compound_page() */
3324                 if (!get_page_unless_zero(page)) {
3325                         list_del_init(page_deferred_list(page));
3326                         split_queue_len--;
3327                 }
3328         }
3329         spin_unlock_irqrestore(&split_queue_lock, flags);
3330
3331         list_for_each_safe(pos, next, &list) {
3332                 page = list_entry((void *)pos, struct page, mapping);
3333                 lock_page(page);
3334                 /* split_huge_page() removes page from list on success */
3335                 if (!split_huge_page(page))
3336                         split++;
3337                 unlock_page(page);
3338                 put_page(page);
3339         }
3340
3341         spin_lock_irqsave(&split_queue_lock, flags);
3342         list_splice_tail(&list, &split_queue);
3343         spin_unlock_irqrestore(&split_queue_lock, flags);
3344
3345         return split * HPAGE_PMD_NR / 2;
3346 }
3347
3348 static struct shrinker deferred_split_shrinker = {
3349         .count_objects = deferred_split_count,
3350         .scan_objects = deferred_split_scan,
3351         .seeks = DEFAULT_SEEKS,
3352 };