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/debugfs.h>
  27 #include <linux/migrate.h>
  28 #include <linux/hashtable.h>
  29 #include <linux/userfaultfd_k.h>
  30 #include <linux/page_idle.h>
  31
  32 #include <asm/tlb.h>
  33 #include <asm/pgalloc.h>
  34 #include "internal.h"
  35
  36 enum scan_result {
  37         SCAN_FAIL,
  38         SCAN_SUCCEED,
  39         SCAN_PMD_NULL,
  40         SCAN_EXCEED_NONE_PTE,
  41         SCAN_PTE_NON_PRESENT,
  42         SCAN_PAGE_RO,
  43         SCAN_NO_REFERENCED_PAGE,
  44         SCAN_PAGE_NULL,
  45         SCAN_SCAN_ABORT,
  46         SCAN_PAGE_COUNT,
  47         SCAN_PAGE_LRU,
  48         SCAN_PAGE_LOCK,
  49         SCAN_PAGE_ANON,
  50         SCAN_PAGE_COMPOUND,
  51         SCAN_ANY_PROCESS,
  52         SCAN_VMA_NULL,
  53         SCAN_VMA_CHECK,
  54         SCAN_ADDRESS_RANGE,
  55         SCAN_SWAP_CACHE_PAGE,
  56         SCAN_DEL_PAGE_LRU,
  57         SCAN_ALLOC_HUGE_PAGE_FAIL,
  58         SCAN_CGROUP_CHARGE_FAIL
  59 };
  60
  61 #define CREATE_TRACE_POINTS
  62 #include <trace/events/huge_memory.h>
  63
  64 /*
  65  * By default transparent hugepage support is disabled in order that avoid
  66  * to risk increase the memory footprint of applications without a guaranteed
  67  * benefit. When transparent hugepage support is enabled, is for all mappings,
  68  * and khugepaged scans all mappings.
  69  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  70  * for all hugepage allocations.
  71  */
  72 unsigned long transparent_hugepage_flags __read_mostly =
  73 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  74         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  75 #endif
  76 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  77         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  78 #endif
  79         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  80         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  81         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  82
  83 /* default scan 8*512 pte (or vmas) every 30 second */
  84 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  85 static unsigned int khugepaged_pages_collapsed;
  86 static unsigned int khugepaged_full_scans;
  87 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  88 /* during fragmentation poll the hugepage allocator once every minute */
  89 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  90 static struct task_struct *khugepaged_thread __read_mostly;
  91 static DEFINE_MUTEX(khugepaged_mutex);
  92 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  93 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  94 /*
  95  * default collapse hugepages if there is at least one pte mapped like
  96  * it would have happened if the vma was large enough during page
  97  * fault.
  98  */
  99 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
 100
 101 static int khugepaged(void *none);
 102 static int khugepaged_slab_init(void);
 103 static void khugepaged_slab_exit(void);
 104
 105 #define MM_SLOTS_HASH_BITS 10
 106 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 107
 108 static struct kmem_cache *mm_slot_cache __read_mostly;
 109
 110 /**
 111  * struct mm_slot - hash lookup from mm to mm_slot
 112  * @hash: hash collision list
 113  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 114  * @mm: the mm that this information is valid for
 115  */
 116 struct mm_slot {
 117         struct hlist_node hash;
 118         struct list_head mm_node;
 119         struct mm_struct *mm;
 120 };
 121
 122 /**
 123  * struct khugepaged_scan - cursor for scanning
 124  * @mm_head: the head of the mm list to scan
 125  * @mm_slot: the current mm_slot we are scanning
 126  * @address: the next address inside that to be scanned
 127  *
 128  * There is only the one khugepaged_scan instance of this cursor structure.
 129  */
 130 struct khugepaged_scan {
 131         struct list_head mm_head;
 132         struct mm_slot *mm_slot;
 133         unsigned long address;
 134 };
 135 static struct khugepaged_scan khugepaged_scan = {
 136         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 137 };
 138
 139 static DEFINE_SPINLOCK(split_queue_lock);
 140 static LIST_HEAD(split_queue);
 141 static unsigned long split_queue_len;
 142 static struct shrinker deferred_split_shrinker;
 143
 144 static void set_recommended_min_free_kbytes(void)
 145 {
 146         struct zone *zone;
 147         int nr_zones = 0;
 148         unsigned long recommended_min;
 149
 150         for_each_populated_zone(zone)
 151                 nr_zones++;
 152
 153         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 154         recommended_min = pageblock_nr_pages * nr_zones * 2;
 155
 156         /*
 157          * Make sure that on average at least two pageblocks are almost free
 158          * of another type, one for a migratetype to fall back to and a
 159          * second to avoid subsequent fallbacks of other types There are 3
 160          * MIGRATE_TYPES we care about.
 161          */
 162         recommended_min += pageblock_nr_pages * nr_zones *
 163                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 164
 165         /* don't ever allow to reserve more than 5% of the lowmem */
 166         recommended_min = min(recommended_min,
 167                               (unsigned long) nr_free_buffer_pages() / 20);
 168         recommended_min <<= (PAGE_SHIFT-10);
 169
 170         if (recommended_min > min_free_kbytes) {
 171                 if (user_min_free_kbytes >= 0)
 172                         pr_info("raising min_free_kbytes from %d to %lu "
 173                                 "to help transparent hugepage allocations\n",
 174                                 min_free_kbytes, recommended_min);
 175
 176                 min_free_kbytes = recommended_min;
 177         }
 178         setup_per_zone_wmarks();
 179 }
 180
 181 static int start_stop_khugepaged(void)
 182 {
 183         int err = 0;
 184         if (khugepaged_enabled()) {
 185                 if (!khugepaged_thread)
 186                         khugepaged_thread = kthread_run(khugepaged, NULL,
 187                                                         "khugepaged");
 188                 if (IS_ERR(khugepaged_thread)) {
 189                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 190                         err = PTR_ERR(khugepaged_thread);
 191                         khugepaged_thread = NULL;
 192                         goto fail;
 193                 }
 194
 195                 if (!list_empty(&khugepaged_scan.mm_head))
 196                         wake_up_interruptible(&khugepaged_wait);
 197
 198                 set_recommended_min_free_kbytes();
 199         } else if (khugepaged_thread) {
 200                 kthread_stop(khugepaged_thread);
 201                 khugepaged_thread = NULL;
 202         }
 203 fail:
 204         return err;
 205 }
 206
 207 static atomic_t huge_zero_refcount;
 208 struct page *huge_zero_page __read_mostly;
 209
 210 struct page *get_huge_zero_page(void)
 211 {
 212         struct page *zero_page;
 213 retry:
 214         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 215                 return READ_ONCE(huge_zero_page);
 216
 217         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 218                         HPAGE_PMD_ORDER);
 219         if (!zero_page) {
 220                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 221                 return NULL;
 222         }
 223         count_vm_event(THP_ZERO_PAGE_ALLOC);
 224         preempt_disable();
 225         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 226                 preempt_enable();
 227                 __free_pages(zero_page, compound_order(zero_page));
 228                 goto retry;
 229         }
 230
 231         /* We take additional reference here. It will be put back by shrinker */
 232         atomic_set(&huge_zero_refcount, 2);
 233         preempt_enable();
 234         return READ_ONCE(huge_zero_page);
 235 }
 236
 237 static void put_huge_zero_page(void)
 238 {
 239         /*
 240          * Counter should never go to zero here. Only shrinker can put
 241          * last reference.
 242          */
 243         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 244 }
 245
 246 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 247                                         struct shrink_control *sc)
 248 {
 249         /* we can free zero page only if last reference remains */
 250         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 251 }
 252
 253 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 254                                        struct shrink_control *sc)
 255 {
 256         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 257                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 258                 BUG_ON(zero_page == NULL);
 259                 __free_pages(zero_page, compound_order(zero_page));
 260                 return HPAGE_PMD_NR;
 261         }
 262
 263         return 0;
 264 }
 265
 266 static struct shrinker huge_zero_page_shrinker = {
 267         .count_objects = shrink_huge_zero_page_count,
 268         .scan_objects = shrink_huge_zero_page_scan,
 269         .seeks = DEFAULT_SEEKS,
 270 };
 271
 272 #ifdef CONFIG_SYSFS
 273
 274 static ssize_t double_flag_show(struct kobject *kobj,
 275                                 struct kobj_attribute *attr, char *buf,
 276                                 enum transparent_hugepage_flag enabled,
 277                                 enum transparent_hugepage_flag req_madv)
 278 {
 279         if (test_bit(enabled, &transparent_hugepage_flags)) {
 280                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 281                 return sprintf(buf, "[always] madvise never\n");
 282         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 283                 return sprintf(buf, "always [madvise] never\n");
 284         else
 285                 return sprintf(buf, "always madvise [never]\n");
 286 }
 287 static ssize_t double_flag_store(struct kobject *kobj,
 288                                  struct kobj_attribute *attr,
 289                                  const char *buf, size_t count,
 290                                  enum transparent_hugepage_flag enabled,
 291                                  enum transparent_hugepage_flag req_madv)
 292 {
 293         if (!memcmp("always", buf,
 294                     min(sizeof("always")-1, count))) {
 295                 set_bit(enabled, &transparent_hugepage_flags);
 296                 clear_bit(req_madv, &transparent_hugepage_flags);
 297         } else if (!memcmp("madvise", buf,
 298                            min(sizeof("madvise")-1, count))) {
 299                 clear_bit(enabled, &transparent_hugepage_flags);
 300                 set_bit(req_madv, &transparent_hugepage_flags);
 301         } else if (!memcmp("never", buf,
 302                            min(sizeof("never")-1, count))) {
 303                 clear_bit(enabled, &transparent_hugepage_flags);
 304                 clear_bit(req_madv, &transparent_hugepage_flags);
 305         } else
 306                 return -EINVAL;
 307
 308         return count;
 309 }
 310
 311 static ssize_t enabled_show(struct kobject *kobj,
 312                             struct kobj_attribute *attr, char *buf)
 313 {
 314         return double_flag_show(kobj, attr, buf,
 315                                 TRANSPARENT_HUGEPAGE_FLAG,
 316                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 317 }
 318 static ssize_t enabled_store(struct kobject *kobj,
 319                              struct kobj_attribute *attr,
 320                              const char *buf, size_t count)
 321 {
 322         ssize_t ret;
 323
 324         ret = double_flag_store(kobj, attr, buf, count,
 325                                 TRANSPARENT_HUGEPAGE_FLAG,
 326                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 327
 328         if (ret > 0) {
 329                 int err;
 330
 331                 mutex_lock(&khugepaged_mutex);
 332                 err = start_stop_khugepaged();
 333                 mutex_unlock(&khugepaged_mutex);
 334
 335                 if (err)
 336                         ret = err;
 337         }
 338
 339         return ret;
 340 }
 341 static struct kobj_attribute enabled_attr =
 342         __ATTR(enabled, 0644, enabled_show, enabled_store);
 343
 344 static ssize_t single_flag_show(struct kobject *kobj,
 345                                 struct kobj_attribute *attr, char *buf,
 346                                 enum transparent_hugepage_flag flag)
 347 {
 348         return sprintf(buf, "%d\n",
 349                        !!test_bit(flag, &transparent_hugepage_flags));
 350 }
 351
 352 static ssize_t single_flag_store(struct kobject *kobj,
 353                                  struct kobj_attribute *attr,
 354                                  const char *buf, size_t count,
 355                                  enum transparent_hugepage_flag flag)
 356 {
 357         unsigned long value;
 358         int ret;
 359
 360         ret = kstrtoul(buf, 10, &value);
 361         if (ret < 0)
 362                 return ret;
 363         if (value > 1)
 364                 return -EINVAL;
 365
 366         if (value)
 367                 set_bit(flag, &transparent_hugepage_flags);
 368         else
 369                 clear_bit(flag, &transparent_hugepage_flags);
 370
 371         return count;
 372 }
 373
 374 /*
 375  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 376  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 377  * memory just to allocate one more hugepage.
 378  */
 379 static ssize_t defrag_show(struct kobject *kobj,
 380                            struct kobj_attribute *attr, char *buf)
 381 {
 382         return double_flag_show(kobj, attr, buf,
 383                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 384                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 385 }
 386 static ssize_t defrag_store(struct kobject *kobj,
 387                             struct kobj_attribute *attr,
 388                             const char *buf, size_t count)
 389 {
 390         return double_flag_store(kobj, attr, buf, count,
 391                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 392                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 393 }
 394 static struct kobj_attribute defrag_attr =
 395         __ATTR(defrag, 0644, defrag_show, defrag_store);
 396
 397 static ssize_t use_zero_page_show(struct kobject *kobj,
 398                 struct kobj_attribute *attr, char *buf)
 399 {
 400         return single_flag_show(kobj, attr, buf,
 401                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 402 }
 403 static ssize_t use_zero_page_store(struct kobject *kobj,
 404                 struct kobj_attribute *attr, const char *buf, size_t count)
 405 {
 406         return single_flag_store(kobj, attr, buf, count,
 407                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 408 }
 409 static struct kobj_attribute use_zero_page_attr =
 410         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 411 #ifdef CONFIG_DEBUG_VM
 412 static ssize_t debug_cow_show(struct kobject *kobj,
 413                                 struct kobj_attribute *attr, char *buf)
 414 {
 415         return single_flag_show(kobj, attr, buf,
 416                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 417 }
 418 static ssize_t debug_cow_store(struct kobject *kobj,
 419                                struct kobj_attribute *attr,
 420                                const char *buf, size_t count)
 421 {
 422         return single_flag_store(kobj, attr, buf, count,
 423                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 424 }
 425 static struct kobj_attribute debug_cow_attr =
 426         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 427 #endif /* CONFIG_DEBUG_VM */
 428
 429 static struct attribute *hugepage_attr[] = {
 430         &enabled_attr.attr,
 431         &defrag_attr.attr,
 432         &use_zero_page_attr.attr,
 433 #ifdef CONFIG_DEBUG_VM
 434         &debug_cow_attr.attr,
 435 #endif
 436         NULL,
 437 };
 438
 439 static struct attribute_group hugepage_attr_group = {
 440         .attrs = hugepage_attr,
 441 };
 442
 443 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 444                                          struct kobj_attribute *attr,
 445                                          char *buf)
 446 {
 447         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 448 }
 449
 450 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 451                                           struct kobj_attribute *attr,
 452                                           const char *buf, size_t count)
 453 {
 454         unsigned long msecs;
 455         int err;
 456
 457         err = kstrtoul(buf, 10, &msecs);
 458         if (err || msecs > UINT_MAX)
 459                 return -EINVAL;
 460
 461         khugepaged_scan_sleep_millisecs = msecs;
 462         wake_up_interruptible(&khugepaged_wait);
 463
 464         return count;
 465 }
 466 static struct kobj_attribute scan_sleep_millisecs_attr =
 467         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 468                scan_sleep_millisecs_store);
 469
 470 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 471                                           struct kobj_attribute *attr,
 472                                           char *buf)
 473 {
 474         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 475 }
 476
 477 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 478                                            struct kobj_attribute *attr,
 479                                            const char *buf, size_t count)
 480 {
 481         unsigned long msecs;
 482         int err;
 483
 484         err = kstrtoul(buf, 10, &msecs);
 485         if (err || msecs > UINT_MAX)
 486                 return -EINVAL;
 487
 488         khugepaged_alloc_sleep_millisecs = msecs;
 489         wake_up_interruptible(&khugepaged_wait);
 490
 491         return count;
 492 }
 493 static struct kobj_attribute alloc_sleep_millisecs_attr =
 494         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 495                alloc_sleep_millisecs_store);
 496
 497 static ssize_t pages_to_scan_show(struct kobject *kobj,
 498                                   struct kobj_attribute *attr,
 499                                   char *buf)
 500 {
 501         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 502 }
 503 static ssize_t pages_to_scan_store(struct kobject *kobj,
 504                                    struct kobj_attribute *attr,
 505                                    const char *buf, size_t count)
 506 {
 507         int err;
 508         unsigned long pages;
 509
 510         err = kstrtoul(buf, 10, &pages);
 511         if (err || !pages || pages > UINT_MAX)
 512                 return -EINVAL;
 513
 514         khugepaged_pages_to_scan = pages;
 515
 516         return count;
 517 }
 518 static struct kobj_attribute pages_to_scan_attr =
 519         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 520                pages_to_scan_store);
 521
 522 static ssize_t pages_collapsed_show(struct kobject *kobj,
 523                                     struct kobj_attribute *attr,
 524                                     char *buf)
 525 {
 526         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 527 }
 528 static struct kobj_attribute pages_collapsed_attr =
 529         __ATTR_RO(pages_collapsed);
 530
 531 static ssize_t full_scans_show(struct kobject *kobj,
 532                                struct kobj_attribute *attr,
 533                                char *buf)
 534 {
 535         return sprintf(buf, "%u\n", khugepaged_full_scans);
 536 }
 537 static struct kobj_attribute full_scans_attr =
 538         __ATTR_RO(full_scans);
 539
 540 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 541                                       struct kobj_attribute *attr, char *buf)
 542 {
 543         return single_flag_show(kobj, attr, buf,
 544                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 545 }
 546 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 547                                        struct kobj_attribute *attr,
 548                                        const char *buf, size_t count)
 549 {
 550         return single_flag_store(kobj, attr, buf, count,
 551                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 552 }
 553 static struct kobj_attribute khugepaged_defrag_attr =
 554         __ATTR(defrag, 0644, khugepaged_defrag_show,
 555                khugepaged_defrag_store);
 556
 557 /*
 558  * max_ptes_none controls if khugepaged should collapse hugepages over
 559  * any unmapped ptes in turn potentially increasing the memory
 560  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 561  * reduce the available free memory in the system as it
 562  * runs. Increasing max_ptes_none will instead potentially reduce the
 563  * free memory in the system during the khugepaged scan.
 564  */
 565 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 566                                              struct kobj_attribute *attr,
 567                                              char *buf)
 568 {
 569         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 570 }
 571 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 572                                               struct kobj_attribute *attr,
 573                                               const char *buf, size_t count)
 574 {
 575         int err;
 576         unsigned long max_ptes_none;
 577
 578         err = kstrtoul(buf, 10, &max_ptes_none);
 579         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 580                 return -EINVAL;
 581
 582         khugepaged_max_ptes_none = max_ptes_none;
 583
 584         return count;
 585 }
 586 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 587         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 588                khugepaged_max_ptes_none_store);
 589
 590 static struct attribute *khugepaged_attr[] = {
 591         &khugepaged_defrag_attr.attr,
 592         &khugepaged_max_ptes_none_attr.attr,
 593         &pages_to_scan_attr.attr,
 594         &pages_collapsed_attr.attr,
 595         &full_scans_attr.attr,
 596         &scan_sleep_millisecs_attr.attr,
 597         &alloc_sleep_millisecs_attr.attr,
 598         NULL,
 599 };
 600
 601 static struct attribute_group khugepaged_attr_group = {
 602         .attrs = khugepaged_attr,
 603         .name = "khugepaged",
 604 };
 605
 606 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 607 {
 608         int err;
 609
 610         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 611         if (unlikely(!*hugepage_kobj)) {
 612                 pr_err("failed to create transparent hugepage kobject\n");
 613                 return -ENOMEM;
 614         }
 615
 616         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 617         if (err) {
 618                 pr_err("failed to register transparent hugepage group\n");
 619                 goto delete_obj;
 620         }
 621
 622         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 623         if (err) {
 624                 pr_err("failed to register transparent hugepage group\n");
 625                 goto remove_hp_group;
 626         }
 627
 628         return 0;
 629
 630 remove_hp_group:
 631         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 632 delete_obj:
 633         kobject_put(*hugepage_kobj);
 634         return err;
 635 }
 636
 637 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 638 {
 639         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 640         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 641         kobject_put(hugepage_kobj);
 642 }
 643 #else
 644 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 645 {
 646         return 0;
 647 }
 648
 649 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 650 {
 651 }
 652 #endif /* CONFIG_SYSFS */
 653
 654 static int __init hugepage_init(void)
 655 {
 656         int err;
 657         struct kobject *hugepage_kobj;
 658
 659         if (!has_transparent_hugepage()) {
 660                 transparent_hugepage_flags = 0;
 661                 return -EINVAL;
 662         }
 663
 664         err = hugepage_init_sysfs(&hugepage_kobj);
 665         if (err)
 666                 goto err_sysfs;
 667
 668         err = khugepaged_slab_init();
 669         if (err)
 670                 goto err_slab;
 671
 672         err = register_shrinker(&huge_zero_page_shrinker);
 673         if (err)
 674                 goto err_hzp_shrinker;
 675         err = register_shrinker(&deferred_split_shrinker);
 676         if (err)
 677                 goto err_split_shrinker;
 678
 679         /*
 680          * By default disable transparent hugepages on smaller systems,
 681          * where the extra memory used could hurt more than TLB overhead
 682          * is likely to save.  The admin can still enable it through /sys.
 683          */
 684         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 685                 transparent_hugepage_flags = 0;
 686                 return 0;
 687         }
 688
 689         err = start_stop_khugepaged();
 690         if (err)
 691                 goto err_khugepaged;
 692
 693         return 0;
 694 err_khugepaged:
 695         unregister_shrinker(&deferred_split_shrinker);
 696 err_split_shrinker:
 697         unregister_shrinker(&huge_zero_page_shrinker);
 698 err_hzp_shrinker:
 699         khugepaged_slab_exit();
 700 err_slab:
 701         hugepage_exit_sysfs(hugepage_kobj);
 702 err_sysfs:
 703         return err;
 704 }
 705 subsys_initcall(hugepage_init);
 706
 707 static int __init setup_transparent_hugepage(char *str)
 708 {
 709         int ret = 0;
 710         if (!str)
 711                 goto out;
 712         if (!strcmp(str, "always")) {
 713                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 714                         &transparent_hugepage_flags);
 715                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 716                           &transparent_hugepage_flags);
 717                 ret = 1;
 718         } else if (!strcmp(str, "madvise")) {
 719                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 720                           &transparent_hugepage_flags);
 721                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 722                         &transparent_hugepage_flags);
 723                 ret = 1;
 724         } else if (!strcmp(str, "never")) {
 725                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 726                           &transparent_hugepage_flags);
 727                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 728                           &transparent_hugepage_flags);
 729                 ret = 1;
 730         }
 731 out:
 732         if (!ret)
 733                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 734         return ret;
 735 }
 736 __setup("transparent_hugepage=", setup_transparent_hugepage);
 737
 738 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 739 {
 740         if (likely(vma->vm_flags & VM_WRITE))
 741                 pmd = pmd_mkwrite(pmd);
 742         return pmd;
 743 }
 744
 745 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 746 {
 747         pmd_t entry;
 748         entry = mk_pmd(page, prot);
 749         entry = pmd_mkhuge(entry);
 750         return entry;
 751 }
 752
 753 static inline struct list_head *page_deferred_list(struct page *page)
 754 {
 755         /*
 756          * ->lru in the tail pages is occupied by compound_head.
 757          * Let's use ->mapping + ->index in the second tail page as list_head.
 758          */
 759         return (struct list_head *)&page[2].mapping;
 760 }
 761
 762 void prep_transhuge_page(struct page *page)
 763 {
 764         /*
 765          * we use page->mapping and page->indexlru in second tail page
 766          * as list_head: assuming THP order >= 2
 767          */
 768         BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 769
 770         INIT_LIST_HEAD(page_deferred_list(page));
 771         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 772 }
 773
 774 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 775                                         struct vm_area_struct *vma,
 776                                         unsigned long address, pmd_t *pmd,
 777                                         struct page *page, gfp_t gfp,
 778                                         unsigned int flags)
 779 {
 780         struct mem_cgroup *memcg;
 781         pgtable_t pgtable;
 782         spinlock_t *ptl;
 783         unsigned long haddr = address & HPAGE_PMD_MASK;
 784
 785         VM_BUG_ON_PAGE(!PageCompound(page), page);
 786
 787         if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
 788                 put_page(page);
 789                 count_vm_event(THP_FAULT_FALLBACK);
 790                 return VM_FAULT_FALLBACK;
 791         }
 792
 793         pgtable = pte_alloc_one(mm, haddr);
 794         if (unlikely(!pgtable)) {
 795                 mem_cgroup_cancel_charge(page, memcg, true);
 796                 put_page(page);
 797                 return VM_FAULT_OOM;
 798         }
 799
 800         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 801         /*
 802          * The memory barrier inside __SetPageUptodate makes sure that
 803          * clear_huge_page writes become visible before the set_pmd_at()
 804          * write.
 805          */
 806         __SetPageUptodate(page);
 807
 808         ptl = pmd_lock(mm, pmd);
 809         if (unlikely(!pmd_none(*pmd))) {
 810                 spin_unlock(ptl);
 811                 mem_cgroup_cancel_charge(page, memcg, true);
 812                 put_page(page);
 813                 pte_free(mm, pgtable);
 814         } else {
 815                 pmd_t entry;
 816
 817                 /* Deliver the page fault to userland */
 818                 if (userfaultfd_missing(vma)) {
 819                         int ret;
 820
 821                         spin_unlock(ptl);
 822                         mem_cgroup_cancel_charge(page, memcg, true);
 823                         put_page(page);
 824                         pte_free(mm, pgtable);
 825                         ret = handle_userfault(vma, address, flags,
 826                                                VM_UFFD_MISSING);
 827                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 828                         return ret;
 829                 }
 830
 831                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 832                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 833                 page_add_new_anon_rmap(page, vma, haddr, true);
 834                 mem_cgroup_commit_charge(page, memcg, false, true);
 835                 lru_cache_add_active_or_unevictable(page, vma);
 836                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 837                 set_pmd_at(mm, haddr, pmd, entry);
 838                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 839                 atomic_long_inc(&mm->nr_ptes);
 840                 spin_unlock(ptl);
 841                 count_vm_event(THP_FAULT_ALLOC);
 842         }
 843
 844         return 0;
 845 }
 846
 847 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 848 {
 849         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 850 }
 851
 852 /* Caller must hold page table lock. */
 853 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 854                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 855                 struct page *zero_page)
 856 {
 857         pmd_t entry;
 858         if (!pmd_none(*pmd))
 859                 return false;
 860         entry = mk_pmd(zero_page, vma->vm_page_prot);
 861         entry = pmd_mkhuge(entry);
 862         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 863         set_pmd_at(mm, haddr, pmd, entry);
 864         atomic_long_inc(&mm->nr_ptes);
 865         return true;
 866 }
 867
 868 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 869                                unsigned long address, pmd_t *pmd,
 870                                unsigned int flags)
 871 {
 872         gfp_t gfp;
 873         struct page *page;
 874         unsigned long haddr = address & HPAGE_PMD_MASK;
 875
 876         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 877                 return VM_FAULT_FALLBACK;
 878         if (unlikely(anon_vma_prepare(vma)))
 879                 return VM_FAULT_OOM;
 880         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 881                 return VM_FAULT_OOM;
 882         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 883                         transparent_hugepage_use_zero_page()) {
 884                 spinlock_t *ptl;
 885                 pgtable_t pgtable;
 886                 struct page *zero_page;
 887                 bool set;
 888                 int ret;
 889                 pgtable = pte_alloc_one(mm, haddr);
 890                 if (unlikely(!pgtable))
 891                         return VM_FAULT_OOM;
 892                 zero_page = get_huge_zero_page();
 893                 if (unlikely(!zero_page)) {
 894                         pte_free(mm, pgtable);
 895                         count_vm_event(THP_FAULT_FALLBACK);
 896                         return VM_FAULT_FALLBACK;
 897                 }
 898                 ptl = pmd_lock(mm, pmd);
 899                 ret = 0;
 900                 set = false;
 901                 if (pmd_none(*pmd)) {
 902                         if (userfaultfd_missing(vma)) {
 903                                 spin_unlock(ptl);
 904                                 ret = handle_userfault(vma, address, flags,
 905                                                        VM_UFFD_MISSING);
 906                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 907                         } else {
 908                                 set_huge_zero_page(pgtable, mm, vma,
 909                                                    haddr, pmd,
 910                                                    zero_page);
 911                                 spin_unlock(ptl);
 912                                 set = true;
 913                         }
 914                 } else
 915                         spin_unlock(ptl);
 916                 if (!set) {
 917                         pte_free(mm, pgtable);
 918                         put_huge_zero_page();
 919                 }
 920                 return ret;
 921         }
 922         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 923         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 924         if (unlikely(!page)) {
 925                 count_vm_event(THP_FAULT_FALLBACK);
 926                 return VM_FAULT_FALLBACK;
 927         }
 928         prep_transhuge_page(page);
 929         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 930                                             flags);
 931 }
 932
 933 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 934                 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
 935 {
 936         struct mm_struct *mm = vma->vm_mm;
 937         pmd_t entry;
 938         spinlock_t *ptl;
 939
 940         ptl = pmd_lock(mm, pmd);
 941         if (pmd_none(*pmd)) {
 942                 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
 943                 if (write) {
 944                         entry = pmd_mkyoung(pmd_mkdirty(entry));
 945                         entry = maybe_pmd_mkwrite(entry, vma);
 946                 }
 947                 set_pmd_at(mm, addr, pmd, entry);
 948                 update_mmu_cache_pmd(vma, addr, pmd);
 949         }
 950         spin_unlock(ptl);
 951 }
 952
 953 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 954                         pmd_t *pmd, unsigned long pfn, bool write)
 955 {
 956         pgprot_t pgprot = vma->vm_page_prot;
 957         /*
 958          * If we had pmd_special, we could avoid all these restrictions,
 959          * but we need to be consistent with PTEs and architectures that
 960          * can't support a 'special' bit.
 961          */
 962         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 963         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 964                                                 (VM_PFNMAP|VM_MIXEDMAP));
 965         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 966         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
 967
 968         if (addr < vma->vm_start || addr >= vma->vm_end)
 969                 return VM_FAULT_SIGBUS;
 970         if (track_pfn_insert(vma, &pgprot, pfn))
 971                 return VM_FAULT_SIGBUS;
 972         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 973         return VM_FAULT_NOPAGE;
 974 }
 975
 976 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 977                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 978                   struct vm_area_struct *vma)
 979 {
 980         spinlock_t *dst_ptl, *src_ptl;
 981         struct page *src_page;
 982         pmd_t pmd;
 983         pgtable_t pgtable;
 984         int ret;
 985
 986         ret = -ENOMEM;
 987         pgtable = pte_alloc_one(dst_mm, addr);
 988         if (unlikely(!pgtable))
 989                 goto out;
 990
 991         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 992         src_ptl = pmd_lockptr(src_mm, src_pmd);
 993         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 994
 995         ret = -EAGAIN;
 996         pmd = *src_pmd;
 997         if (unlikely(!pmd_trans_huge(pmd))) {
 998                 pte_free(dst_mm, pgtable);
 999                 goto out_unlock;
1000         }
1001         /*
1002          * When page table lock is held, the huge zero pmd should not be
1003          * under splitting since we don't split the page itself, only pmd to
1004          * a page table.
1005          */
1006         if (is_huge_zero_pmd(pmd)) {
1007                 struct page *zero_page;
1008                 /*
1009                  * get_huge_zero_page() will never allocate a new page here,
1010                  * since we already have a zero page to copy. It just takes a
1011                  * reference.
1012                  */
1013                 zero_page = get_huge_zero_page();
1014                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1015                                 zero_page);
1016                 ret = 0;
1017                 goto out_unlock;
1018         }
1019
1020         src_page = pmd_page(pmd);
1021         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1022         get_page(src_page);
1023         page_dup_rmap(src_page, true);
1024         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1025
1026         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1027         pmd = pmd_mkold(pmd_wrprotect(pmd));
1028         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1029         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1030         atomic_long_inc(&dst_mm->nr_ptes);
1031
1032         ret = 0;
1033 out_unlock:
1034         spin_unlock(src_ptl);
1035         spin_unlock(dst_ptl);
1036 out:
1037         return ret;
1038 }
1039
1040 void huge_pmd_set_accessed(struct mm_struct *mm,
1041                            struct vm_area_struct *vma,
1042                            unsigned long address,
1043                            pmd_t *pmd, pmd_t orig_pmd,
1044                            int dirty)
1045 {
1046         spinlock_t *ptl;
1047         pmd_t entry;
1048         unsigned long haddr;
1049
1050         ptl = pmd_lock(mm, pmd);
1051         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1052                 goto unlock;
1053
1054         entry = pmd_mkyoung(orig_pmd);
1055         haddr = address & HPAGE_PMD_MASK;
1056         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1057                 update_mmu_cache_pmd(vma, address, pmd);
1058
1059 unlock:
1060         spin_unlock(ptl);
1061 }
1062
1063 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1064                                         struct vm_area_struct *vma,
1065                                         unsigned long address,
1066                                         pmd_t *pmd, pmd_t orig_pmd,
1067                                         struct page *page,
1068                                         unsigned long haddr)
1069 {
1070         struct mem_cgroup *memcg;
1071         spinlock_t *ptl;
1072         pgtable_t pgtable;
1073         pmd_t _pmd;
1074         int ret = 0, i;
1075         struct page **pages;
1076         unsigned long mmun_start;       /* For mmu_notifiers */
1077         unsigned long mmun_end;         /* For mmu_notifiers */
1078
1079         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1080                         GFP_KERNEL);
1081         if (unlikely(!pages)) {
1082                 ret |= VM_FAULT_OOM;
1083                 goto out;
1084         }
1085
1086         for (i = 0; i < HPAGE_PMD_NR; i++) {
1087                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1088                                                __GFP_OTHER_NODE,
1089                                                vma, address, page_to_nid(page));
1090                 if (unlikely(!pages[i] ||
1091                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1092                                                    &memcg, false))) {
1093                         if (pages[i])
1094                                 put_page(pages[i]);
1095                         while (--i >= 0) {
1096                                 memcg = (void *)page_private(pages[i]);
1097                                 set_page_private(pages[i], 0);
1098                                 mem_cgroup_cancel_charge(pages[i], memcg,
1099                                                 false);
1100                                 put_page(pages[i]);
1101                         }
1102                         kfree(pages);
1103                         ret |= VM_FAULT_OOM;
1104                         goto out;
1105                 }
1106                 set_page_private(pages[i], (unsigned long)memcg);
1107         }
1108
1109         for (i = 0; i < HPAGE_PMD_NR; i++) {
1110                 copy_user_highpage(pages[i], page + i,
1111                                    haddr + PAGE_SIZE * i, vma);
1112                 __SetPageUptodate(pages[i]);
1113                 cond_resched();
1114         }
1115
1116         mmun_start = haddr;
1117         mmun_end   = haddr + HPAGE_PMD_SIZE;
1118         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1119
1120         ptl = pmd_lock(mm, pmd);
1121         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1122                 goto out_free_pages;
1123         VM_BUG_ON_PAGE(!PageHead(page), page);
1124
1125         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1126         /* leave pmd empty until pte is filled */
1127
1128         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1129         pmd_populate(mm, &_pmd, pgtable);
1130
1131         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1132                 pte_t *pte, entry;
1133                 entry = mk_pte(pages[i], vma->vm_page_prot);
1134                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1135                 memcg = (void *)page_private(pages[i]);
1136                 set_page_private(pages[i], 0);
1137                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1138                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1139                 lru_cache_add_active_or_unevictable(pages[i], vma);
1140                 pte = pte_offset_map(&_pmd, haddr);
1141                 VM_BUG_ON(!pte_none(*pte));
1142                 set_pte_at(mm, haddr, pte, entry);
1143                 pte_unmap(pte);
1144         }
1145         kfree(pages);
1146
1147         smp_wmb(); /* make pte visible before pmd */
1148         pmd_populate(mm, pmd, pgtable);
1149         page_remove_rmap(page, true);
1150         spin_unlock(ptl);
1151
1152         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1153
1154         ret |= VM_FAULT_WRITE;
1155         put_page(page);
1156
1157 out:
1158         return ret;
1159
1160 out_free_pages:
1161         spin_unlock(ptl);
1162         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1163         for (i = 0; i < HPAGE_PMD_NR; i++) {
1164                 memcg = (void *)page_private(pages[i]);
1165                 set_page_private(pages[i], 0);
1166                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1167                 put_page(pages[i]);
1168         }
1169         kfree(pages);
1170         goto out;
1171 }
1172
1173 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1174                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1175 {
1176         spinlock_t *ptl;
1177         int ret = 0;
1178         struct page *page = NULL, *new_page;
1179         struct mem_cgroup *memcg;
1180         unsigned long haddr;
1181         unsigned long mmun_start;       /* For mmu_notifiers */
1182         unsigned long mmun_end;         /* For mmu_notifiers */
1183         gfp_t huge_gfp;                 /* for allocation and charge */
1184
1185         ptl = pmd_lockptr(mm, pmd);
1186         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1187         haddr = address & HPAGE_PMD_MASK;
1188         if (is_huge_zero_pmd(orig_pmd))
1189                 goto alloc;
1190         spin_lock(ptl);
1191         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1192                 goto out_unlock;
1193
1194         page = pmd_page(orig_pmd);
1195         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1196         /*
1197          * We can only reuse the page if nobody else maps the huge page or it's
1198          * part. We can do it by checking page_mapcount() on each sub-page, but
1199          * it's expensive.
1200          * The cheaper way is to check page_count() to be equal 1: every
1201          * mapcount takes page reference reference, so this way we can
1202          * guarantee, that the PMD is the only mapping.
1203          * This can give false negative if somebody pinned the page, but that's
1204          * fine.
1205          */
1206         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1207                 pmd_t entry;
1208                 entry = pmd_mkyoung(orig_pmd);
1209                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1210                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1211                         update_mmu_cache_pmd(vma, address, pmd);
1212                 ret |= VM_FAULT_WRITE;
1213                 goto out_unlock;
1214         }
1215         get_page(page);
1216         spin_unlock(ptl);
1217 alloc:
1218         if (transparent_hugepage_enabled(vma) &&
1219             !transparent_hugepage_debug_cow()) {
1220                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1221                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1222         } else
1223                 new_page = NULL;
1224
1225         if (likely(new_page)) {
1226                 prep_transhuge_page(new_page);
1227         } else {
1228                 if (!page) {
1229                         split_huge_pmd(vma, pmd, address);
1230                         ret |= VM_FAULT_FALLBACK;
1231                 } else {
1232                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1233                                         pmd, orig_pmd, page, haddr);
1234                         if (ret & VM_FAULT_OOM) {
1235                                 split_huge_pmd(vma, pmd, address);
1236                                 ret |= VM_FAULT_FALLBACK;
1237                         }
1238                         put_page(page);
1239                 }
1240                 count_vm_event(THP_FAULT_FALLBACK);
1241                 goto out;
1242         }
1243
1244         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1245                                            true))) {
1246                 put_page(new_page);
1247                 if (page) {
1248                         split_huge_pmd(vma, pmd, address);
1249                         put_page(page);
1250                 } else
1251                         split_huge_pmd(vma, pmd, address);
1252                 ret |= VM_FAULT_FALLBACK;
1253                 count_vm_event(THP_FAULT_FALLBACK);
1254                 goto out;
1255         }
1256
1257         count_vm_event(THP_FAULT_ALLOC);
1258
1259         if (!page)
1260                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1261         else
1262                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1263         __SetPageUptodate(new_page);
1264
1265         mmun_start = haddr;
1266         mmun_end   = haddr + HPAGE_PMD_SIZE;
1267         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1268
1269         spin_lock(ptl);
1270         if (page)
1271                 put_page(page);
1272         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1273                 spin_unlock(ptl);
1274                 mem_cgroup_cancel_charge(new_page, memcg, true);
1275                 put_page(new_page);
1276                 goto out_mn;
1277         } else {
1278                 pmd_t entry;
1279                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1280                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1281                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1282                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1283                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1284                 lru_cache_add_active_or_unevictable(new_page, vma);
1285                 set_pmd_at(mm, haddr, pmd, entry);
1286                 update_mmu_cache_pmd(vma, address, pmd);
1287                 if (!page) {
1288                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1289                         put_huge_zero_page();
1290                 } else {
1291                         VM_BUG_ON_PAGE(!PageHead(page), page);
1292                         page_remove_rmap(page, true);
1293                         put_page(page);
1294                 }
1295                 ret |= VM_FAULT_WRITE;
1296         }
1297         spin_unlock(ptl);
1298 out_mn:
1299         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1300 out:
1301         return ret;
1302 out_unlock:
1303         spin_unlock(ptl);
1304         return ret;
1305 }
1306
1307 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1308                                    unsigned long addr,
1309                                    pmd_t *pmd,
1310                                    unsigned int flags)
1311 {
1312         struct mm_struct *mm = vma->vm_mm;
1313         struct page *page = NULL;
1314
1315         assert_spin_locked(pmd_lockptr(mm, pmd));
1316
1317         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1318                 goto out;
1319
1320         /* Avoid dumping huge zero page */
1321         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1322                 return ERR_PTR(-EFAULT);
1323
1324         /* Full NUMA hinting faults to serialise migration in fault paths */
1325         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1326                 goto out;
1327
1328         page = pmd_page(*pmd);
1329         VM_BUG_ON_PAGE(!PageHead(page), page);
1330         if (flags & FOLL_TOUCH) {
1331                 pmd_t _pmd;
1332                 /*
1333                  * We should set the dirty bit only for FOLL_WRITE but
1334                  * for now the dirty bit in the pmd is meaningless.
1335                  * And if the dirty bit will become meaningful and
1336                  * we'll only set it with FOLL_WRITE, an atomic
1337                  * set_bit will be required on the pmd to set the
1338                  * young bit, instead of the current set_pmd_at.
1339                  */
1340                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1341                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1342                                           pmd, _pmd,  1))
1343                         update_mmu_cache_pmd(vma, addr, pmd);
1344         }
1345         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1346                 /*
1347                  * We don't mlock() pte-mapped THPs. This way we can avoid
1348                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1349                  *
1350                  * In most cases the pmd is the only mapping of the page as we
1351                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1352                  * writable private mappings in populate_vma_page_range().
1353                  *
1354                  * The only scenario when we have the page shared here is if we
1355                  * mlocking read-only mapping shared over fork(). We skip
1356                  * mlocking such pages.
1357                  */
1358                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1359                                 page->mapping && trylock_page(page)) {
1360                         lru_add_drain();
1361                         if (page->mapping)
1362                                 mlock_vma_page(page);
1363                         unlock_page(page);
1364                 }
1365         }
1366         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1367         VM_BUG_ON_PAGE(!PageCompound(page), page);
1368         if (flags & FOLL_GET)
1369                 get_page(page);
1370
1371 out:
1372         return page;
1373 }
1374
1375 /* NUMA hinting page fault entry point for trans huge pmds */
1376 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1377                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1378 {
1379         spinlock_t *ptl;
1380         struct anon_vma *anon_vma = NULL;
1381         struct page *page;
1382         unsigned long haddr = addr & HPAGE_PMD_MASK;
1383         int page_nid = -1, this_nid = numa_node_id();
1384         int target_nid, last_cpupid = -1;
1385         bool page_locked;
1386         bool migrated = false;
1387         bool was_writable;
1388         int flags = 0;
1389
1390         /* A PROT_NONE fault should not end up here */
1391         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1392
1393         ptl = pmd_lock(mm, pmdp);
1394         if (unlikely(!pmd_same(pmd, *pmdp)))
1395                 goto out_unlock;
1396
1397         /*
1398          * If there are potential migrations, wait for completion and retry
1399          * without disrupting NUMA hinting information. Do not relock and
1400          * check_same as the page may no longer be mapped.
1401          */
1402         if (unlikely(pmd_trans_migrating(*pmdp))) {
1403                 page = pmd_page(*pmdp);
1404                 spin_unlock(ptl);
1405                 wait_on_page_locked(page);
1406                 goto out;
1407         }
1408
1409         page = pmd_page(pmd);
1410         BUG_ON(is_huge_zero_page(page));
1411         page_nid = page_to_nid(page);
1412         last_cpupid = page_cpupid_last(page);
1413         count_vm_numa_event(NUMA_HINT_FAULTS);
1414         if (page_nid == this_nid) {
1415                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1416                 flags |= TNF_FAULT_LOCAL;
1417         }
1418
1419         /* See similar comment in do_numa_page for explanation */
1420         if (!(vma->vm_flags & VM_WRITE))
1421                 flags |= TNF_NO_GROUP;
1422
1423         /*
1424          * Acquire the page lock to serialise THP migrations but avoid dropping
1425          * page_table_lock if at all possible
1426          */
1427         page_locked = trylock_page(page);
1428         target_nid = mpol_misplaced(page, vma, haddr);
1429         if (target_nid == -1) {
1430                 /* If the page was locked, there are no parallel migrations */
1431                 if (page_locked)
1432                         goto clear_pmdnuma;
1433         }
1434
1435         /* Migration could have started since the pmd_trans_migrating check */
1436         if (!page_locked) {
1437                 spin_unlock(ptl);
1438                 wait_on_page_locked(page);
1439                 page_nid = -1;
1440                 goto out;
1441         }
1442
1443         /*
1444          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1445          * to serialises splits
1446          */
1447         get_page(page);
1448         spin_unlock(ptl);
1449         anon_vma = page_lock_anon_vma_read(page);
1450
1451         /* Confirm the PMD did not change while page_table_lock was released */
1452         spin_lock(ptl);
1453         if (unlikely(!pmd_same(pmd, *pmdp))) {
1454                 unlock_page(page);
1455                 put_page(page);
1456                 page_nid = -1;
1457                 goto out_unlock;
1458         }
1459
1460         /* Bail if we fail to protect against THP splits for any reason */
1461         if (unlikely(!anon_vma)) {
1462                 put_page(page);
1463                 page_nid = -1;
1464                 goto clear_pmdnuma;
1465         }
1466
1467         /*
1468          * Migrate the THP to the requested node, returns with page unlocked
1469          * and access rights restored.
1470          */
1471         spin_unlock(ptl);
1472         migrated = migrate_misplaced_transhuge_page(mm, vma,
1473                                 pmdp, pmd, addr, page, target_nid);
1474         if (migrated) {
1475                 flags |= TNF_MIGRATED;
1476                 page_nid = target_nid;
1477         } else
1478                 flags |= TNF_MIGRATE_FAIL;
1479
1480         goto out;
1481 clear_pmdnuma:
1482         BUG_ON(!PageLocked(page));
1483         was_writable = pmd_write(pmd);
1484         pmd = pmd_modify(pmd, vma->vm_page_prot);
1485         pmd = pmd_mkyoung(pmd);
1486         if (was_writable)
1487                 pmd = pmd_mkwrite(pmd);
1488         set_pmd_at(mm, haddr, pmdp, pmd);
1489         update_mmu_cache_pmd(vma, addr, pmdp);
1490         unlock_page(page);
1491 out_unlock:
1492         spin_unlock(ptl);
1493
1494 out:
1495         if (anon_vma)
1496                 page_unlock_anon_vma_read(anon_vma);
1497
1498         if (page_nid != -1)
1499                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1500
1501         return 0;
1502 }
1503
1504 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1505                  pmd_t *pmd, unsigned long addr)
1506 {
1507         pmd_t orig_pmd;
1508         spinlock_t *ptl;
1509
1510         if (!__pmd_trans_huge_lock(pmd, vma, &ptl))
1511                 return 0;
1512         /*
1513          * For architectures like ppc64 we look at deposited pgtable
1514          * when calling pmdp_huge_get_and_clear. So do the
1515          * pgtable_trans_huge_withdraw after finishing pmdp related
1516          * operations.
1517          */
1518         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1519                         tlb->fullmm);
1520         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1521         if (vma_is_dax(vma)) {
1522                 spin_unlock(ptl);
1523                 if (is_huge_zero_pmd(orig_pmd))
1524                         put_huge_zero_page();
1525         } else if (is_huge_zero_pmd(orig_pmd)) {
1526                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1527                 atomic_long_dec(&tlb->mm->nr_ptes);
1528                 spin_unlock(ptl);
1529                 put_huge_zero_page();
1530         } else {
1531                 struct page *page = pmd_page(orig_pmd);
1532                 page_remove_rmap(page, true);
1533                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1534                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1535                 VM_BUG_ON_PAGE(!PageHead(page), page);
1536                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1537                 atomic_long_dec(&tlb->mm->nr_ptes);
1538                 spin_unlock(ptl);
1539                 tlb_remove_page(tlb, page);
1540         }
1541         return 1;
1542 }
1543
1544 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1545                   unsigned long old_addr,
1546                   unsigned long new_addr, unsigned long old_end,
1547                   pmd_t *old_pmd, pmd_t *new_pmd)
1548 {
1549         spinlock_t *old_ptl, *new_ptl;
1550         pmd_t pmd;
1551
1552         struct mm_struct *mm = vma->vm_mm;
1553
1554         if ((old_addr & ~HPAGE_PMD_MASK) ||
1555             (new_addr & ~HPAGE_PMD_MASK) ||
1556             old_end - old_addr < HPAGE_PMD_SIZE ||
1557             (new_vma->vm_flags & VM_NOHUGEPAGE))
1558                 return false;
1559
1560         /*
1561          * The destination pmd shouldn't be established, free_pgtables()
1562          * should have release it.
1563          */
1564         if (WARN_ON(!pmd_none(*new_pmd))) {
1565                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1566                 return false;
1567         }
1568
1569         /*
1570          * We don't have to worry about the ordering of src and dst
1571          * ptlocks because exclusive mmap_sem prevents deadlock.
1572          */
1573         if (__pmd_trans_huge_lock(old_pmd, vma, &old_ptl)) {
1574                 new_ptl = pmd_lockptr(mm, new_pmd);
1575                 if (new_ptl != old_ptl)
1576                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1577                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1578                 VM_BUG_ON(!pmd_none(*new_pmd));
1579
1580                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1581                         pgtable_t pgtable;
1582                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1583                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1584                 }
1585                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1586                 if (new_ptl != old_ptl)
1587                         spin_unlock(new_ptl);
1588                 spin_unlock(old_ptl);
1589                 return true;
1590         }
1591         return false;
1592 }
1593
1594 /*
1595  * Returns
1596  *  - 0 if PMD could not be locked
1597  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1598  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1599  */
1600 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1601                 unsigned long addr, pgprot_t newprot, int prot_numa)
1602 {
1603         struct mm_struct *mm = vma->vm_mm;
1604         spinlock_t *ptl;
1605         int ret = 0;
1606
1607         if (__pmd_trans_huge_lock(pmd, vma, &ptl)) {
1608                 pmd_t entry;
1609                 bool preserve_write = prot_numa && pmd_write(*pmd);
1610                 ret = 1;
1611
1612                 /*
1613                  * Avoid trapping faults against the zero page. The read-only
1614                  * data is likely to be read-cached on the local CPU and
1615                  * local/remote hits to the zero page are not interesting.
1616                  */
1617                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1618                         spin_unlock(ptl);
1619                         return ret;
1620                 }
1621
1622                 if (!prot_numa || !pmd_protnone(*pmd)) {
1623                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1624                         entry = pmd_modify(entry, newprot);
1625                         if (preserve_write)
1626                                 entry = pmd_mkwrite(entry);
1627                         ret = HPAGE_PMD_NR;
1628                         set_pmd_at(mm, addr, pmd, entry);
1629                         BUG_ON(!preserve_write && pmd_write(entry));
1630                 }
1631                 spin_unlock(ptl);
1632         }
1633
1634         return ret;
1635 }
1636
1637 /*
1638  * Returns true if a given pmd maps a thp, false otherwise.
1639  *
1640  * Note that if it returns true, this routine returns without unlocking page
1641  * table lock. So callers must unlock it.
1642  */
1643 bool __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1644                 spinlock_t **ptl)
1645 {
1646         *ptl = pmd_lock(vma->vm_mm, pmd);
1647         if (likely(pmd_trans_huge(*pmd)))
1648                 return true;
1649         spin_unlock(*ptl);
1650         return false;
1651 }
1652
1653 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1654
1655 int hugepage_madvise(struct vm_area_struct *vma,
1656                      unsigned long *vm_flags, int advice)
1657 {
1658         switch (advice) {
1659         case MADV_HUGEPAGE:
1660 #ifdef CONFIG_S390
1661                 /*
1662                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1663                  * can't handle this properly after s390_enable_sie, so we simply
1664                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1665                  */
1666                 if (mm_has_pgste(vma->vm_mm))
1667                         return 0;
1668 #endif
1669                 /*
1670                  * Be somewhat over-protective like KSM for now!
1671                  */
1672                 if (*vm_flags & VM_NO_THP)
1673                         return -EINVAL;
1674                 *vm_flags &= ~VM_NOHUGEPAGE;
1675                 *vm_flags |= VM_HUGEPAGE;
1676                 /*
1677                  * If the vma become good for khugepaged to scan,
1678                  * register it here without waiting a page fault that
1679                  * may not happen any time soon.
1680                  */
1681                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1682                         return -ENOMEM;
1683                 break;
1684         case MADV_NOHUGEPAGE:
1685                 /*
1686                  * Be somewhat over-protective like KSM for now!
1687                  */
1688                 if (*vm_flags & VM_NO_THP)
1689                         return -EINVAL;
1690                 *vm_flags &= ~VM_HUGEPAGE;
1691                 *vm_flags |= VM_NOHUGEPAGE;
1692                 /*
1693                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1694                  * this vma even if we leave the mm registered in khugepaged if
1695                  * it got registered before VM_NOHUGEPAGE was set.
1696                  */
1697                 break;
1698         }
1699
1700         return 0;
1701 }
1702
1703 static int __init khugepaged_slab_init(void)
1704 {
1705         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1706                                           sizeof(struct mm_slot),
1707                                           __alignof__(struct mm_slot), 0, NULL);
1708         if (!mm_slot_cache)
1709                 return -ENOMEM;
1710
1711         return 0;
1712 }
1713
1714 static void __init khugepaged_slab_exit(void)
1715 {
1716         kmem_cache_destroy(mm_slot_cache);
1717 }
1718
1719 static inline struct mm_slot *alloc_mm_slot(void)
1720 {
1721         if (!mm_slot_cache)     /* initialization failed */
1722                 return NULL;
1723         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1724 }
1725
1726 static inline void free_mm_slot(struct mm_slot *mm_slot)
1727 {
1728         kmem_cache_free(mm_slot_cache, mm_slot);
1729 }
1730
1731 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1732 {
1733         struct mm_slot *mm_slot;
1734
1735         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1736                 if (mm == mm_slot->mm)
1737                         return mm_slot;
1738
1739         return NULL;
1740 }
1741
1742 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1743                                     struct mm_slot *mm_slot)
1744 {
1745         mm_slot->mm = mm;
1746         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1747 }
1748
1749 static inline int khugepaged_test_exit(struct mm_struct *mm)
1750 {
1751         return atomic_read(&mm->mm_users) == 0;
1752 }
1753
1754 int __khugepaged_enter(struct mm_struct *mm)
1755 {
1756         struct mm_slot *mm_slot;
1757         int wakeup;
1758
1759         mm_slot = alloc_mm_slot();
1760         if (!mm_slot)
1761                 return -ENOMEM;
1762
1763         /* __khugepaged_exit() must not run from under us */
1764         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1765         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1766                 free_mm_slot(mm_slot);
1767                 return 0;
1768         }
1769
1770         spin_lock(&khugepaged_mm_lock);
1771         insert_to_mm_slots_hash(mm, mm_slot);
1772         /*
1773          * Insert just behind the scanning cursor, to let the area settle
1774          * down a little.
1775          */
1776         wakeup = list_empty(&khugepaged_scan.mm_head);
1777         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1778         spin_unlock(&khugepaged_mm_lock);
1779
1780         atomic_inc(&mm->mm_count);
1781         if (wakeup)
1782                 wake_up_interruptible(&khugepaged_wait);
1783
1784         return 0;
1785 }
1786
1787 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1788                                unsigned long vm_flags)
1789 {
1790         unsigned long hstart, hend;
1791         if (!vma->anon_vma)
1792                 /*
1793                  * Not yet faulted in so we will register later in the
1794                  * page fault if needed.
1795                  */
1796                 return 0;
1797         if (vma->vm_ops)
1798                 /* khugepaged not yet working on file or special mappings */
1799                 return 0;
1800         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1801         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1802         hend = vma->vm_end & HPAGE_PMD_MASK;
1803         if (hstart < hend)
1804                 return khugepaged_enter(vma, vm_flags);
1805         return 0;
1806 }
1807
1808 void __khugepaged_exit(struct mm_struct *mm)
1809 {
1810         struct mm_slot *mm_slot;
1811         int free = 0;
1812
1813         spin_lock(&khugepaged_mm_lock);
1814         mm_slot = get_mm_slot(mm);
1815         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1816                 hash_del(&mm_slot->hash);
1817                 list_del(&mm_slot->mm_node);
1818                 free = 1;
1819         }
1820         spin_unlock(&khugepaged_mm_lock);
1821
1822         if (free) {
1823                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1824                 free_mm_slot(mm_slot);
1825                 mmdrop(mm);
1826         } else if (mm_slot) {
1827                 /*
1828                  * This is required to serialize against
1829                  * khugepaged_test_exit() (which is guaranteed to run
1830                  * under mmap sem read mode). Stop here (after we
1831                  * return all pagetables will be destroyed) until
1832                  * khugepaged has finished working on the pagetables
1833                  * under the mmap_sem.
1834                  */
1835                 down_write(&mm->mmap_sem);
1836                 up_write(&mm->mmap_sem);
1837         }
1838 }
1839
1840 static void release_pte_page(struct page *page)
1841 {
1842         /* 0 stands for page_is_file_cache(page) == false */
1843         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1844         unlock_page(page);
1845         putback_lru_page(page);
1846 }
1847
1848 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1849 {
1850         while (--_pte >= pte) {
1851                 pte_t pteval = *_pte;
1852                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1853                         release_pte_page(pte_page(pteval));
1854         }
1855 }
1856
1857 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1858                                         unsigned long address,
1859                                         pte_t *pte)
1860 {
1861         struct page *page = NULL;
1862         pte_t *_pte;
1863         int none_or_zero = 0, result = 0;
1864         bool referenced = false, writable = false;
1865
1866         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1867              _pte++, address += PAGE_SIZE) {
1868                 pte_t pteval = *_pte;
1869                 if (pte_none(pteval) || (pte_present(pteval) &&
1870                                 is_zero_pfn(pte_pfn(pteval)))) {
1871                         if (!userfaultfd_armed(vma) &&
1872                             ++none_or_zero <= khugepaged_max_ptes_none) {
1873                                 continue;
1874                         } else {
1875                                 result = SCAN_EXCEED_NONE_PTE;
1876                                 goto out;
1877                         }
1878                 }
1879                 if (!pte_present(pteval)) {
1880                         result = SCAN_PTE_NON_PRESENT;
1881                         goto out;
1882                 }
1883                 page = vm_normal_page(vma, address, pteval);
1884                 if (unlikely(!page)) {
1885                         result = SCAN_PAGE_NULL;
1886                         goto out;
1887                 }
1888
1889                 VM_BUG_ON_PAGE(PageCompound(page), page);
1890                 VM_BUG_ON_PAGE(!PageAnon(page), page);
1891                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1892
1893                 /*
1894                  * We can do it before isolate_lru_page because the
1895                  * page can't be freed from under us. NOTE: PG_lock
1896                  * is needed to serialize against split_huge_page
1897                  * when invoked from the VM.
1898                  */
1899                 if (!trylock_page(page)) {
1900                         result = SCAN_PAGE_LOCK;
1901                         goto out;
1902                 }
1903
1904                 /*
1905                  * cannot use mapcount: can't collapse if there's a gup pin.
1906                  * The page must only be referenced by the scanned process
1907                  * and page swap cache.
1908                  */
1909                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1910                         unlock_page(page);
1911                         result = SCAN_PAGE_COUNT;
1912                         goto out;
1913                 }
1914                 if (pte_write(pteval)) {
1915                         writable = true;
1916                 } else {
1917                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
1918                                 unlock_page(page);
1919                                 result = SCAN_SWAP_CACHE_PAGE;
1920                                 goto out;
1921                         }
1922                         /*
1923                          * Page is not in the swap cache. It can be collapsed
1924                          * into a THP.
1925                          */
1926                 }
1927
1928                 /*
1929                  * Isolate the page to avoid collapsing an hugepage
1930                  * currently in use by the VM.
1931                  */
1932                 if (isolate_lru_page(page)) {
1933                         unlock_page(page);
1934                         result = SCAN_DEL_PAGE_LRU;
1935                         goto out;
1936                 }
1937                 /* 0 stands for page_is_file_cache(page) == false */
1938                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1939                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1940                 VM_BUG_ON_PAGE(PageLRU(page), page);
1941
1942                 /* If there is no mapped pte young don't collapse the page */
1943                 if (pte_young(pteval) ||
1944                     page_is_young(page) || PageReferenced(page) ||
1945                     mmu_notifier_test_young(vma->vm_mm, address))
1946                         referenced = true;
1947         }
1948         if (likely(writable)) {
1949                 if (likely(referenced)) {
1950                         result = SCAN_SUCCEED;
1951                         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
1952                                                             referenced, writable, result);
1953                         return 1;
1954                 }
1955         } else {
1956                 result = SCAN_PAGE_RO;
1957         }
1958
1959 out:
1960         release_pte_pages(pte, _pte);
1961         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
1962                                             referenced, writable, result);
1963         return 0;
1964 }
1965
1966 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1967                                       struct vm_area_struct *vma,
1968                                       unsigned long address,
1969                                       spinlock_t *ptl)
1970 {
1971         pte_t *_pte;
1972         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1973                 pte_t pteval = *_pte;
1974                 struct page *src_page;
1975
1976                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1977                         clear_user_highpage(page, address);
1978                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1979                         if (is_zero_pfn(pte_pfn(pteval))) {
1980                                 /*
1981                                  * ptl mostly unnecessary.
1982                                  */
1983                                 spin_lock(ptl);
1984                                 /*
1985                                  * paravirt calls inside pte_clear here are
1986                                  * superfluous.
1987                                  */
1988                                 pte_clear(vma->vm_mm, address, _pte);
1989                                 spin_unlock(ptl);
1990                         }
1991                 } else {
1992                         src_page = pte_page(pteval);
1993                         copy_user_highpage(page, src_page, address, vma);
1994                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
1995                         release_pte_page(src_page);
1996                         /*
1997                          * ptl mostly unnecessary, but preempt has to
1998                          * be disabled to update the per-cpu stats
1999                          * inside page_remove_rmap().
2000                          */
2001                         spin_lock(ptl);
2002                         /*
2003                          * paravirt calls inside pte_clear here are
2004                          * superfluous.
2005                          */
2006                         pte_clear(vma->vm_mm, address, _pte);
2007                         page_remove_rmap(src_page, false);
2008                         spin_unlock(ptl);
2009                         free_page_and_swap_cache(src_page);
2010                 }
2011
2012                 address += PAGE_SIZE;
2013                 page++;
2014         }
2015 }
2016
2017 static void khugepaged_alloc_sleep(void)
2018 {
2019         DEFINE_WAIT(wait);
2020
2021         add_wait_queue(&khugepaged_wait, &wait);
2022         freezable_schedule_timeout_interruptible(
2023                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2024         remove_wait_queue(&khugepaged_wait, &wait);
2025 }
2026
2027 static int khugepaged_node_load[MAX_NUMNODES];
2028
2029 static bool khugepaged_scan_abort(int nid)
2030 {
2031         int i;
2032
2033         /*
2034          * If zone_reclaim_mode is disabled, then no extra effort is made to
2035          * allocate memory locally.
2036          */
2037         if (!zone_reclaim_mode)
2038                 return false;
2039
2040         /* If there is a count for this node already, it must be acceptable */
2041         if (khugepaged_node_load[nid])
2042                 return false;
2043
2044         for (i = 0; i < MAX_NUMNODES; i++) {
2045                 if (!khugepaged_node_load[i])
2046                         continue;
2047                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2048                         return true;
2049         }
2050         return false;
2051 }
2052
2053 #ifdef CONFIG_NUMA
2054 static int khugepaged_find_target_node(void)
2055 {
2056         static int last_khugepaged_target_node = NUMA_NO_NODE;
2057         int nid, target_node = 0, max_value = 0;
2058
2059         /* find first node with max normal pages hit */
2060         for (nid = 0; nid < MAX_NUMNODES; nid++)
2061                 if (khugepaged_node_load[nid] > max_value) {
2062                         max_value = khugepaged_node_load[nid];
2063                         target_node = nid;
2064                 }
2065
2066         /* do some balance if several nodes have the same hit record */
2067         if (target_node <= last_khugepaged_target_node)
2068                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2069                                 nid++)
2070                         if (max_value == khugepaged_node_load[nid]) {
2071                                 target_node = nid;
2072                                 break;
2073                         }
2074
2075         last_khugepaged_target_node = target_node;
2076         return target_node;
2077 }
2078
2079 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2080 {
2081         if (IS_ERR(*hpage)) {
2082                 if (!*wait)
2083                         return false;
2084
2085                 *wait = false;
2086                 *hpage = NULL;
2087                 khugepaged_alloc_sleep();
2088         } else if (*hpage) {
2089                 put_page(*hpage);
2090                 *hpage = NULL;
2091         }
2092
2093         return true;
2094 }
2095
2096 static struct page *
2097 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2098                        unsigned long address, int node)
2099 {
2100         VM_BUG_ON_PAGE(*hpage, *hpage);
2101
2102         /*
2103          * Before allocating the hugepage, release the mmap_sem read lock.
2104          * The allocation can take potentially a long time if it involves
2105          * sync compaction, and we do not need to hold the mmap_sem during
2106          * that. We will recheck the vma after taking it again in write mode.
2107          */
2108         up_read(&mm->mmap_sem);
2109
2110         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2111         if (unlikely(!*hpage)) {
2112                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2113                 *hpage = ERR_PTR(-ENOMEM);
2114                 return NULL;
2115         }
2116
2117         prep_transhuge_page(*hpage);
2118         count_vm_event(THP_COLLAPSE_ALLOC);
2119         return *hpage;
2120 }
2121 #else
2122 static int khugepaged_find_target_node(void)
2123 {
2124         return 0;
2125 }
2126
2127 static inline struct page *alloc_hugepage(int defrag)
2128 {
2129         struct page *page;
2130
2131         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2132         if (page)
2133                 prep_transhuge_page(page);
2134         return page;
2135 }
2136
2137 static struct page *khugepaged_alloc_hugepage(bool *wait)
2138 {
2139         struct page *hpage;
2140
2141         do {
2142                 hpage = alloc_hugepage(khugepaged_defrag());
2143                 if (!hpage) {
2144                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2145                         if (!*wait)
2146                                 return NULL;
2147
2148                         *wait = false;
2149                         khugepaged_alloc_sleep();
2150                 } else
2151                         count_vm_event(THP_COLLAPSE_ALLOC);
2152         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2153
2154         return hpage;
2155 }
2156
2157 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2158 {
2159         if (!*hpage)
2160                 *hpage = khugepaged_alloc_hugepage(wait);
2161
2162         if (unlikely(!*hpage))
2163                 return false;
2164
2165         return true;
2166 }
2167
2168 static struct page *
2169 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2170                        unsigned long address, int node)
2171 {
2172         up_read(&mm->mmap_sem);
2173         VM_BUG_ON(!*hpage);
2174
2175         return  *hpage;
2176 }
2177 #endif
2178
2179 static bool hugepage_vma_check(struct vm_area_struct *vma)
2180 {
2181         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2182             (vma->vm_flags & VM_NOHUGEPAGE))
2183                 return false;
2184         if (!vma->anon_vma || vma->vm_ops)
2185                 return false;
2186         if (is_vma_temporary_stack(vma))
2187                 return false;
2188         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2189         return true;
2190 }
2191
2192 static void collapse_huge_page(struct mm_struct *mm,
2193                                    unsigned long address,
2194                                    struct page **hpage,
2195                                    struct vm_area_struct *vma,
2196                                    int node)
2197 {
2198         pmd_t *pmd, _pmd;
2199         pte_t *pte;
2200         pgtable_t pgtable;
2201         struct page *new_page;
2202         spinlock_t *pmd_ptl, *pte_ptl;
2203         int isolated, result = 0;
2204         unsigned long hstart, hend;
2205         struct mem_cgroup *memcg;
2206         unsigned long mmun_start;       /* For mmu_notifiers */
2207         unsigned long mmun_end;         /* For mmu_notifiers */
2208         gfp_t gfp;
2209
2210         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2211
2212         /* Only allocate from the target node */
2213         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2214                 __GFP_THISNODE;
2215
2216         /* release the mmap_sem read lock. */
2217         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2218         if (!new_page) {
2219                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2220                 goto out_nolock;
2221         }
2222
2223         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2224                 result = SCAN_CGROUP_CHARGE_FAIL;
2225                 goto out_nolock;
2226         }
2227
2228         /*
2229          * Prevent all access to pagetables with the exception of
2230          * gup_fast later hanlded by the ptep_clear_flush and the VM
2231          * handled by the anon_vma lock + PG_lock.
2232          */
2233         down_write(&mm->mmap_sem);
2234         if (unlikely(khugepaged_test_exit(mm))) {
2235                 result = SCAN_ANY_PROCESS;
2236                 goto out;
2237         }
2238
2239         vma = find_vma(mm, address);
2240         if (!vma) {
2241                 result = SCAN_VMA_NULL;
2242                 goto out;
2243         }
2244         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2245         hend = vma->vm_end & HPAGE_PMD_MASK;
2246         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2247                 result = SCAN_ADDRESS_RANGE;
2248                 goto out;
2249         }
2250         if (!hugepage_vma_check(vma)) {
2251                 result = SCAN_VMA_CHECK;
2252                 goto out;
2253         }
2254         pmd = mm_find_pmd(mm, address);
2255         if (!pmd) {
2256                 result = SCAN_PMD_NULL;
2257                 goto out;
2258         }
2259
2260         anon_vma_lock_write(vma->anon_vma);
2261
2262         pte = pte_offset_map(pmd, address);
2263         pte_ptl = pte_lockptr(mm, pmd);
2264
2265         mmun_start = address;
2266         mmun_end   = address + HPAGE_PMD_SIZE;
2267         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2268         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2269         /*
2270          * After this gup_fast can't run anymore. This also removes
2271          * any huge TLB entry from the CPU so we won't allow
2272          * huge and small TLB entries for the same virtual address
2273          * to avoid the risk of CPU bugs in that area.
2274          */
2275         _pmd = pmdp_collapse_flush(vma, address, pmd);
2276         spin_unlock(pmd_ptl);
2277         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2278
2279         spin_lock(pte_ptl);
2280         isolated = __collapse_huge_page_isolate(vma, address, pte);
2281         spin_unlock(pte_ptl);
2282
2283         if (unlikely(!isolated)) {
2284                 pte_unmap(pte);
2285                 spin_lock(pmd_ptl);
2286                 BUG_ON(!pmd_none(*pmd));
2287                 /*
2288                  * We can only use set_pmd_at when establishing
2289                  * hugepmds and never for establishing regular pmds that
2290                  * points to regular pagetables. Use pmd_populate for that
2291                  */
2292                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2293                 spin_unlock(pmd_ptl);
2294                 anon_vma_unlock_write(vma->anon_vma);
2295                 result = SCAN_FAIL;
2296                 goto out;
2297         }
2298
2299         /*
2300          * All pages are isolated and locked so anon_vma rmap
2301          * can't run anymore.
2302          */
2303         anon_vma_unlock_write(vma->anon_vma);
2304
2305         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2306         pte_unmap(pte);
2307         __SetPageUptodate(new_page);
2308         pgtable = pmd_pgtable(_pmd);
2309
2310         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2311         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2312
2313         /*
2314          * spin_lock() below is not the equivalent of smp_wmb(), so
2315          * this is needed to avoid the copy_huge_page writes to become
2316          * visible after the set_pmd_at() write.
2317          */
2318         smp_wmb();
2319
2320         spin_lock(pmd_ptl);
2321         BUG_ON(!pmd_none(*pmd));
2322         page_add_new_anon_rmap(new_page, vma, address, true);
2323         mem_cgroup_commit_charge(new_page, memcg, false, true);
2324         lru_cache_add_active_or_unevictable(new_page, vma);
2325         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2326         set_pmd_at(mm, address, pmd, _pmd);
2327         update_mmu_cache_pmd(vma, address, pmd);
2328         spin_unlock(pmd_ptl);
2329
2330         *hpage = NULL;
2331
2332         khugepaged_pages_collapsed++;
2333         result = SCAN_SUCCEED;
2334 out_up_write:
2335         up_write(&mm->mmap_sem);
2336         trace_mm_collapse_huge_page(mm, isolated, result);
2337         return;
2338
2339 out_nolock:
2340         trace_mm_collapse_huge_page(mm, isolated, result);
2341         return;
2342 out:
2343         mem_cgroup_cancel_charge(new_page, memcg, true);
2344         goto out_up_write;
2345 }
2346
2347 static int khugepaged_scan_pmd(struct mm_struct *mm,
2348                                struct vm_area_struct *vma,
2349                                unsigned long address,
2350                                struct page **hpage)
2351 {
2352         pmd_t *pmd;
2353         pte_t *pte, *_pte;
2354         int ret = 0, none_or_zero = 0, result = 0;
2355         struct page *page = NULL;
2356         unsigned long _address;
2357         spinlock_t *ptl;
2358         int node = NUMA_NO_NODE;
2359         bool writable = false, referenced = false;
2360
2361         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2362
2363         pmd = mm_find_pmd(mm, address);
2364         if (!pmd) {
2365                 result = SCAN_PMD_NULL;
2366                 goto out;
2367         }
2368
2369         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2370         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2371         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2372              _pte++, _address += PAGE_SIZE) {
2373                 pte_t pteval = *_pte;
2374                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2375                         if (!userfaultfd_armed(vma) &&
2376                             ++none_or_zero <= khugepaged_max_ptes_none) {
2377                                 continue;
2378                         } else {
2379                                 result = SCAN_EXCEED_NONE_PTE;
2380                                 goto out_unmap;
2381                         }
2382                 }
2383                 if (!pte_present(pteval)) {
2384                         result = SCAN_PTE_NON_PRESENT;
2385                         goto out_unmap;
2386                 }
2387                 if (pte_write(pteval))
2388                         writable = true;
2389
2390                 page = vm_normal_page(vma, _address, pteval);
2391                 if (unlikely(!page)) {
2392                         result = SCAN_PAGE_NULL;
2393                         goto out_unmap;
2394                 }
2395
2396                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2397                 if (PageCompound(page)) {
2398                         result = SCAN_PAGE_COMPOUND;
2399                         goto out_unmap;
2400                 }
2401
2402                 /*
2403                  * Record which node the original page is from and save this
2404                  * information to khugepaged_node_load[].
2405                  * Khupaged will allocate hugepage from the node has the max
2406                  * hit record.
2407                  */
2408                 node = page_to_nid(page);
2409                 if (khugepaged_scan_abort(node)) {
2410                         result = SCAN_SCAN_ABORT;
2411                         goto out_unmap;
2412                 }
2413                 khugepaged_node_load[node]++;
2414                 if (!PageLRU(page)) {
2415                         result = SCAN_SCAN_ABORT;
2416                         goto out_unmap;
2417                 }
2418                 if (PageLocked(page)) {
2419                         result = SCAN_PAGE_LOCK;
2420                         goto out_unmap;
2421                 }
2422                 if (!PageAnon(page)) {
2423                         result = SCAN_PAGE_ANON;
2424                         goto out_unmap;
2425                 }
2426
2427                 /*
2428                  * cannot use mapcount: can't collapse if there's a gup pin.
2429                  * The page must only be referenced by the scanned process
2430                  * and page swap cache.
2431                  */
2432                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2433                         result = SCAN_PAGE_COUNT;
2434                         goto out_unmap;
2435                 }
2436                 if (pte_young(pteval) ||
2437                     page_is_young(page) || PageReferenced(page) ||
2438                     mmu_notifier_test_young(vma->vm_mm, address))
2439                         referenced = true;
2440         }
2441         if (writable) {
2442                 if (referenced) {
2443                         result = SCAN_SUCCEED;
2444                         ret = 1;
2445                 } else {
2446                         result = SCAN_NO_REFERENCED_PAGE;
2447                 }
2448         } else {
2449                 result = SCAN_PAGE_RO;
2450         }
2451 out_unmap:
2452         pte_unmap_unlock(pte, ptl);
2453         if (ret) {
2454                 node = khugepaged_find_target_node();
2455                 /* collapse_huge_page will return with the mmap_sem released */
2456                 collapse_huge_page(mm, address, hpage, vma, node);
2457         }
2458 out:
2459         trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced,
2460                                      none_or_zero, result);
2461         return ret;
2462 }
2463
2464 static void collect_mm_slot(struct mm_slot *mm_slot)
2465 {
2466         struct mm_struct *mm = mm_slot->mm;
2467
2468         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2469
2470         if (khugepaged_test_exit(mm)) {
2471                 /* free mm_slot */
2472                 hash_del(&mm_slot->hash);
2473                 list_del(&mm_slot->mm_node);
2474
2475                 /*
2476                  * Not strictly needed because the mm exited already.
2477                  *
2478                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2479                  */
2480
2481                 /* khugepaged_mm_lock actually not necessary for the below */
2482                 free_mm_slot(mm_slot);
2483                 mmdrop(mm);
2484         }
2485 }
2486
2487 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2488                                             struct page **hpage)
2489         __releases(&khugepaged_mm_lock)
2490         __acquires(&khugepaged_mm_lock)
2491 {
2492         struct mm_slot *mm_slot;
2493         struct mm_struct *mm;
2494         struct vm_area_struct *vma;
2495         int progress = 0;
2496
2497         VM_BUG_ON(!pages);
2498         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2499
2500         if (khugepaged_scan.mm_slot)
2501                 mm_slot = khugepaged_scan.mm_slot;
2502         else {
2503                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2504                                      struct mm_slot, mm_node);
2505                 khugepaged_scan.address = 0;
2506                 khugepaged_scan.mm_slot = mm_slot;
2507         }
2508         spin_unlock(&khugepaged_mm_lock);
2509
2510         mm = mm_slot->mm;
2511         down_read(&mm->mmap_sem);
2512         if (unlikely(khugepaged_test_exit(mm)))
2513                 vma = NULL;
2514         else
2515                 vma = find_vma(mm, khugepaged_scan.address);
2516
2517         progress++;
2518         for (; vma; vma = vma->vm_next) {
2519                 unsigned long hstart, hend;
2520
2521                 cond_resched();
2522                 if (unlikely(khugepaged_test_exit(mm))) {
2523                         progress++;
2524                         break;
2525                 }
2526                 if (!hugepage_vma_check(vma)) {
2527 skip:
2528                         progress++;
2529                         continue;
2530                 }
2531                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2532                 hend = vma->vm_end & HPAGE_PMD_MASK;
2533                 if (hstart >= hend)
2534                         goto skip;
2535                 if (khugepaged_scan.address > hend)
2536                         goto skip;
2537                 if (khugepaged_scan.address < hstart)
2538                         khugepaged_scan.address = hstart;
2539                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2540
2541                 while (khugepaged_scan.address < hend) {
2542                         int ret;
2543                         cond_resched();
2544                         if (unlikely(khugepaged_test_exit(mm)))
2545                                 goto breakouterloop;
2546
2547                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2548                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2549                                   hend);
2550                         ret = khugepaged_scan_pmd(mm, vma,
2551                                                   khugepaged_scan.address,
2552                                                   hpage);
2553                         /* move to next address */
2554                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2555                         progress += HPAGE_PMD_NR;
2556                         if (ret)
2557                                 /* we released mmap_sem so break loop */
2558                                 goto breakouterloop_mmap_sem;
2559                         if (progress >= pages)
2560                                 goto breakouterloop;
2561                 }
2562         }
2563 breakouterloop:
2564         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2565 breakouterloop_mmap_sem:
2566
2567         spin_lock(&khugepaged_mm_lock);
2568         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2569         /*
2570          * Release the current mm_slot if this mm is about to die, or
2571          * if we scanned all vmas of this mm.
2572          */
2573         if (khugepaged_test_exit(mm) || !vma) {
2574                 /*
2575                  * Make sure that if mm_users is reaching zero while
2576                  * khugepaged runs here, khugepaged_exit will find
2577                  * mm_slot not pointing to the exiting mm.
2578                  */
2579                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2580                         khugepaged_scan.mm_slot = list_entry(
2581                                 mm_slot->mm_node.next,
2582                                 struct mm_slot, mm_node);
2583                         khugepaged_scan.address = 0;
2584                 } else {
2585                         khugepaged_scan.mm_slot = NULL;
2586                         khugepaged_full_scans++;
2587                 }
2588
2589                 collect_mm_slot(mm_slot);
2590         }
2591
2592         return progress;
2593 }
2594
2595 static int khugepaged_has_work(void)
2596 {
2597         return !list_empty(&khugepaged_scan.mm_head) &&
2598                 khugepaged_enabled();
2599 }
2600
2601 static int khugepaged_wait_event(void)
2602 {
2603         return !list_empty(&khugepaged_scan.mm_head) ||
2604                 kthread_should_stop();
2605 }
2606
2607 static void khugepaged_do_scan(void)
2608 {
2609         struct page *hpage = NULL;
2610         unsigned int progress = 0, pass_through_head = 0;
2611         unsigned int pages = khugepaged_pages_to_scan;
2612         bool wait = true;
2613
2614         barrier(); /* write khugepaged_pages_to_scan to local stack */
2615
2616         while (progress < pages) {
2617                 if (!khugepaged_prealloc_page(&hpage, &wait))
2618                         break;
2619
2620                 cond_resched();
2621
2622                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2623                         break;
2624
2625                 spin_lock(&khugepaged_mm_lock);
2626                 if (!khugepaged_scan.mm_slot)
2627                         pass_through_head++;
2628                 if (khugepaged_has_work() &&
2629                     pass_through_head < 2)
2630                         progress += khugepaged_scan_mm_slot(pages - progress,
2631                                                             &hpage);
2632                 else
2633                         progress = pages;
2634                 spin_unlock(&khugepaged_mm_lock);
2635         }
2636
2637         if (!IS_ERR_OR_NULL(hpage))
2638                 put_page(hpage);
2639 }
2640
2641 static void khugepaged_wait_work(void)
2642 {
2643         if (khugepaged_has_work()) {
2644                 if (!khugepaged_scan_sleep_millisecs)
2645                         return;
2646
2647                 wait_event_freezable_timeout(khugepaged_wait,
2648                                              kthread_should_stop(),
2649                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2650                 return;
2651         }
2652
2653         if (khugepaged_enabled())
2654                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2655 }
2656
2657 static int khugepaged(void *none)
2658 {
2659         struct mm_slot *mm_slot;
2660
2661         set_freezable();
2662         set_user_nice(current, MAX_NICE);
2663
2664         while (!kthread_should_stop()) {
2665                 khugepaged_do_scan();
2666                 khugepaged_wait_work();
2667         }
2668
2669         spin_lock(&khugepaged_mm_lock);
2670         mm_slot = khugepaged_scan.mm_slot;
2671         khugepaged_scan.mm_slot = NULL;
2672         if (mm_slot)
2673                 collect_mm_slot(mm_slot);
2674         spin_unlock(&khugepaged_mm_lock);
2675         return 0;
2676 }
2677
2678 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2679                 unsigned long haddr, pmd_t *pmd)
2680 {
2681         struct mm_struct *mm = vma->vm_mm;
2682         pgtable_t pgtable;
2683         pmd_t _pmd;
2684         int i;
2685
2686         /* leave pmd empty until pte is filled */
2687         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2688
2689         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2690         pmd_populate(mm, &_pmd, pgtable);
2691
2692         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2693                 pte_t *pte, entry;
2694                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2695                 entry = pte_mkspecial(entry);
2696                 pte = pte_offset_map(&_pmd, haddr);
2697                 VM_BUG_ON(!pte_none(*pte));
2698                 set_pte_at(mm, haddr, pte, entry);
2699                 pte_unmap(pte);
2700         }
2701         smp_wmb(); /* make pte visible before pmd */
2702         pmd_populate(mm, pmd, pgtable);
2703         put_huge_zero_page();
2704 }
2705
2706 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2707                 unsigned long haddr, bool freeze)
2708 {
2709         struct mm_struct *mm = vma->vm_mm;
2710         struct page *page;
2711         pgtable_t pgtable;
2712         pmd_t _pmd;
2713         bool young, write;
2714         int i;
2715
2716         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2717         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2718         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2719         VM_BUG_ON(!pmd_trans_huge(*pmd));
2720
2721         count_vm_event(THP_SPLIT_PMD);
2722
2723         if (vma_is_dax(vma)) {
2724                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2725                 if (is_huge_zero_pmd(_pmd))
2726                         put_huge_zero_page();
2727                 return;
2728         } else if (is_huge_zero_pmd(*pmd)) {
2729                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2730         }
2731
2732         page = pmd_page(*pmd);
2733         VM_BUG_ON_PAGE(!page_count(page), page);
2734         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2735         write = pmd_write(*pmd);
2736         young = pmd_young(*pmd);
2737
2738         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2739         pmd_populate(mm, &_pmd, pgtable);
2740
2741         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2742                 pte_t entry, *pte;
2743                 /*
2744                  * Note that NUMA hinting access restrictions are not
2745                  * transferred to avoid any possibility of altering
2746                  * permissions across VMAs.
2747                  */
2748                 if (freeze) {
2749                         swp_entry_t swp_entry;
2750                         swp_entry = make_migration_entry(page + i, write);
2751                         entry = swp_entry_to_pte(swp_entry);
2752                 } else {
2753                         entry = mk_pte(page + i, vma->vm_page_prot);
2754                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2755                         if (!write)
2756                                 entry = pte_wrprotect(entry);
2757                         if (!young)
2758                                 entry = pte_mkold(entry);
2759                 }
2760                 pte = pte_offset_map(&_pmd, haddr);
2761                 BUG_ON(!pte_none(*pte));
2762                 set_pte_at(mm, haddr, pte, entry);
2763                 atomic_inc(&page[i]._mapcount);
2764                 pte_unmap(pte);
2765         }
2766
2767         /*
2768          * Set PG_double_map before dropping compound_mapcount to avoid
2769          * false-negative page_mapped().
2770          */
2771         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2772                 for (i = 0; i < HPAGE_PMD_NR; i++)
2773                         atomic_inc(&page[i]._mapcount);
2774         }
2775
2776         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2777                 /* Last compound_mapcount is gone. */
2778                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2779                 if (TestClearPageDoubleMap(page)) {
2780                         /* No need in mapcount reference anymore */
2781                         for (i = 0; i < HPAGE_PMD_NR; i++)
2782                                 atomic_dec(&page[i]._mapcount);
2783                 }
2784         }
2785
2786         smp_wmb(); /* make pte visible before pmd */
2787         /*
2788          * Up to this point the pmd is present and huge and userland has the
2789          * whole access to the hugepage during the split (which happens in
2790          * place). If we overwrite the pmd with the not-huge version pointing
2791          * to the pte here (which of course we could if all CPUs were bug
2792          * free), userland could trigger a small page size TLB miss on the
2793          * small sized TLB while the hugepage TLB entry is still established in
2794          * the huge TLB. Some CPU doesn't like that.
2795          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2796          * 383 on page 93. Intel should be safe but is also warns that it's
2797          * only safe if the permission and cache attributes of the two entries
2798          * loaded in the two TLB is identical (which should be the case here).
2799          * But it is generally safer to never allow small and huge TLB entries
2800          * for the same virtual address to be loaded simultaneously. So instead
2801          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2802          * current pmd notpresent (atomically because here the pmd_trans_huge
2803          * and pmd_trans_splitting must remain set at all times on the pmd
2804          * until the split is complete for this pmd), then we flush the SMP TLB
2805          * and finally we write the non-huge version of the pmd entry with
2806          * pmd_populate.
2807          */
2808         pmdp_invalidate(vma, haddr, pmd);
2809         pmd_populate(mm, pmd, pgtable);
2810
2811         if (freeze) {
2812                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2813                         page_remove_rmap(page + i, false);
2814                         put_page(page + i);
2815                 }
2816         }
2817 }
2818
2819 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2820                 unsigned long address)
2821 {
2822         spinlock_t *ptl;
2823         struct mm_struct *mm = vma->vm_mm;
2824         struct page *page = NULL;
2825         unsigned long haddr = address & HPAGE_PMD_MASK;
2826
2827         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2828         ptl = pmd_lock(mm, pmd);
2829         if (unlikely(!pmd_trans_huge(*pmd)))
2830                 goto out;
2831         page = pmd_page(*pmd);
2832         __split_huge_pmd_locked(vma, pmd, haddr, false);
2833         if (PageMlocked(page))
2834                 get_page(page);
2835         else
2836                 page = NULL;
2837 out:
2838         spin_unlock(ptl);
2839         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2840         if (page) {
2841                 lock_page(page);
2842                 munlock_vma_page(page);
2843                 unlock_page(page);
2844                 put_page(page);
2845         }
2846 }
2847
2848 static void split_huge_pmd_address(struct vm_area_struct *vma,
2849                                     unsigned long address)
2850 {
2851         pgd_t *pgd;
2852         pud_t *pud;
2853         pmd_t *pmd;
2854
2855         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2856
2857         pgd = pgd_offset(vma->vm_mm, address);
2858         if (!pgd_present(*pgd))
2859                 return;
2860
2861         pud = pud_offset(pgd, address);
2862         if (!pud_present(*pud))
2863                 return;
2864
2865         pmd = pmd_offset(pud, address);
2866         if (!pmd_present(*pmd) || !pmd_trans_huge(*pmd))
2867                 return;
2868         /*
2869          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2870          * materialize from under us.
2871          */
2872         split_huge_pmd(vma, pmd, address);
2873 }
2874
2875 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2876                              unsigned long start,
2877                              unsigned long end,
2878                              long adjust_next)
2879 {
2880         /*
2881          * If the new start address isn't hpage aligned and it could
2882          * previously contain an hugepage: check if we need to split
2883          * an huge pmd.
2884          */
2885         if (start & ~HPAGE_PMD_MASK &&
2886             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2887             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2888                 split_huge_pmd_address(vma, start);
2889
2890         /*
2891          * If the new end address isn't hpage aligned and it could
2892          * previously contain an hugepage: check if we need to split
2893          * an huge pmd.
2894          */
2895         if (end & ~HPAGE_PMD_MASK &&
2896             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2897             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2898                 split_huge_pmd_address(vma, end);
2899
2900         /*
2901          * If we're also updating the vma->vm_next->vm_start, if the new
2902          * vm_next->vm_start isn't page aligned and it could previously
2903          * contain an hugepage: check if we need to split an huge pmd.
2904          */
2905         if (adjust_next > 0) {
2906                 struct vm_area_struct *next = vma->vm_next;
2907                 unsigned long nstart = next->vm_start;
2908                 nstart += adjust_next << PAGE_SHIFT;
2909                 if (nstart & ~HPAGE_PMD_MASK &&
2910                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2911                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2912                         split_huge_pmd_address(next, nstart);
2913         }
2914 }
2915
2916 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
2917                 unsigned long address)
2918 {
2919         spinlock_t *ptl;
2920         pgd_t *pgd;
2921         pud_t *pud;
2922         pmd_t *pmd;
2923         pte_t *pte;
2924         int i, nr = HPAGE_PMD_NR;
2925
2926         /* Skip pages which doesn't belong to the VMA */
2927         if (address < vma->vm_start) {
2928                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
2929                 page += off;
2930                 nr -= off;
2931                 address = vma->vm_start;
2932         }
2933
2934         pgd = pgd_offset(vma->vm_mm, address);
2935         if (!pgd_present(*pgd))
2936                 return;
2937         pud = pud_offset(pgd, address);
2938         if (!pud_present(*pud))
2939                 return;
2940         pmd = pmd_offset(pud, address);
2941         ptl = pmd_lock(vma->vm_mm, pmd);
2942         if (!pmd_present(*pmd)) {
2943                 spin_unlock(ptl);
2944                 return;
2945         }
2946         if (pmd_trans_huge(*pmd)) {
2947                 if (page == pmd_page(*pmd))
2948                         __split_huge_pmd_locked(vma, pmd, address, true);
2949                 spin_unlock(ptl);
2950                 return;
2951         }
2952         spin_unlock(ptl);
2953
2954         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
2955         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
2956                 pte_t entry, swp_pte;
2957                 swp_entry_t swp_entry;
2958
2959                 if (!pte_present(pte[i]))
2960                         continue;
2961                 if (page_to_pfn(page) != pte_pfn(pte[i]))
2962                         continue;
2963                 flush_cache_page(vma, address, page_to_pfn(page));
2964                 entry = ptep_clear_flush(vma, address, pte + i);
2965                 swp_entry = make_migration_entry(page, pte_write(entry));
2966                 swp_pte = swp_entry_to_pte(swp_entry);
2967                 if (pte_soft_dirty(entry))
2968                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
2969                 set_pte_at(vma->vm_mm, address, pte + i, swp_pte);
2970                 page_remove_rmap(page, false);
2971                 put_page(page);
2972         }
2973         pte_unmap_unlock(pte, ptl);
2974 }
2975
2976 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
2977 {
2978         struct anon_vma_chain *avc;
2979         pgoff_t pgoff = page_to_pgoff(page);
2980
2981         VM_BUG_ON_PAGE(!PageHead(page), page);
2982
2983         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
2984                         pgoff + HPAGE_PMD_NR - 1) {
2985                 unsigned long haddr;
2986
2987                 haddr = __vma_address(page, avc->vma) & HPAGE_PMD_MASK;
2988                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
2989                                 haddr, haddr + HPAGE_PMD_SIZE);
2990                 freeze_page_vma(avc->vma, page, haddr);
2991                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
2992                                 haddr, haddr + HPAGE_PMD_SIZE);
2993         }
2994 }
2995
2996 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
2997                 unsigned long address)
2998 {
2999         spinlock_t *ptl;
3000         pmd_t *pmd;
3001         pte_t *pte, entry;
3002         swp_entry_t swp_entry;
3003         int i, nr = HPAGE_PMD_NR;
3004
3005         /* Skip pages which doesn't belong to the VMA */
3006         if (address < vma->vm_start) {
3007                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3008                 page += off;
3009                 nr -= off;
3010                 address = vma->vm_start;
3011         }
3012
3013         pmd = mm_find_pmd(vma->vm_mm, address);
3014         if (!pmd)
3015                 return;
3016         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3017         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
3018                 if (!is_swap_pte(pte[i]))
3019                         continue;
3020
3021                 swp_entry = pte_to_swp_entry(pte[i]);
3022                 if (!is_migration_entry(swp_entry))
3023                         continue;
3024                 if (migration_entry_to_page(swp_entry) != page)
3025                         continue;
3026
3027                 get_page(page);
3028                 page_add_anon_rmap(page, vma, address, false);
3029
3030                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3031                 entry = pte_mkdirty(entry);
3032                 if (is_write_migration_entry(swp_entry))
3033                         entry = maybe_mkwrite(entry, vma);
3034
3035                 flush_dcache_page(page);
3036                 set_pte_at(vma->vm_mm, address, pte + i, entry);
3037
3038                 /* No need to invalidate - it was non-present before */
3039                 update_mmu_cache(vma, address, pte + i);
3040         }
3041         pte_unmap_unlock(pte, ptl);
3042 }
3043
3044 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3045 {
3046         struct anon_vma_chain *avc;
3047         pgoff_t pgoff = page_to_pgoff(page);
3048
3049         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3050                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3051                 unsigned long address = __vma_address(page, avc->vma);
3052
3053                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3054                                 address, address + HPAGE_PMD_SIZE);
3055                 unfreeze_page_vma(avc->vma, page, address);
3056                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3057                                 address, address + HPAGE_PMD_SIZE);
3058         }
3059 }
3060
3061 static int __split_huge_page_tail(struct page *head, int tail,
3062                 struct lruvec *lruvec, struct list_head *list)
3063 {
3064         int mapcount;
3065         struct page *page_tail = head + tail;
3066
3067         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3068         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3069
3070         /*
3071          * tail_page->_count is zero and not changing from under us. But
3072          * get_page_unless_zero() may be running from under us on the
3073          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3074          * would then run atomic_set() concurrently with
3075          * get_page_unless_zero(), and atomic_set() is implemented in C not
3076          * using locked ops. spin_unlock on x86 sometime uses locked ops
3077          * because of PPro errata 66, 92, so unless somebody can guarantee
3078          * atomic_set() here would be safe on all archs (and not only on x86),
3079          * it's safer to use atomic_add().
3080          */
3081         atomic_add(mapcount + 1, &page_tail->_count);
3082
3083
3084         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3085         page_tail->flags |= (head->flags &
3086                         ((1L << PG_referenced) |
3087                          (1L << PG_swapbacked) |
3088                          (1L << PG_mlocked) |
3089                          (1L << PG_uptodate) |
3090                          (1L << PG_active) |
3091                          (1L << PG_locked) |
3092                          (1L << PG_unevictable)));
3093         page_tail->flags |= (1L << PG_dirty);
3094
3095         /*
3096          * After clearing PageTail the gup refcount can be released.
3097          * Page flags also must be visible before we make the page non-compound.
3098          */
3099         smp_wmb();
3100
3101         clear_compound_head(page_tail);
3102
3103         if (page_is_young(head))
3104                 set_page_young(page_tail);
3105         if (page_is_idle(head))
3106                 set_page_idle(page_tail);
3107
3108         /* ->mapping in first tail page is compound_mapcount */
3109         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3110                         page_tail);
3111         page_tail->mapping = head->mapping;
3112
3113         page_tail->index = head->index + tail;
3114         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3115         lru_add_page_tail(head, page_tail, lruvec, list);
3116
3117         return mapcount;
3118 }
3119
3120 static void __split_huge_page(struct page *page, struct list_head *list)
3121 {
3122         struct page *head = compound_head(page);
3123         struct zone *zone = page_zone(head);
3124         struct lruvec *lruvec;
3125         int i, tail_mapcount;
3126
3127         /* prevent PageLRU to go away from under us, and freeze lru stats */
3128         spin_lock_irq(&zone->lru_lock);
3129         lruvec = mem_cgroup_page_lruvec(head, zone);
3130
3131         /* complete memcg works before add pages to LRU */
3132         mem_cgroup_split_huge_fixup(head);
3133
3134         tail_mapcount = 0;
3135         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3136                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3137         atomic_sub(tail_mapcount, &head->_count);
3138
3139         ClearPageCompound(head);
3140         spin_unlock_irq(&zone->lru_lock);
3141
3142         unfreeze_page(page_anon_vma(head), head);
3143
3144         for (i = 0; i < HPAGE_PMD_NR; i++) {
3145                 struct page *subpage = head + i;
3146                 if (subpage == page)
3147                         continue;
3148                 unlock_page(subpage);
3149
3150                 /*
3151                  * Subpages may be freed if there wasn't any mapping
3152                  * like if add_to_swap() is running on a lru page that
3153                  * had its mapping zapped. And freeing these pages
3154                  * requires taking the lru_lock so we do the put_page
3155                  * of the tail pages after the split is complete.
3156                  */
3157                 put_page(subpage);
3158         }
3159 }
3160
3161 int total_mapcount(struct page *page)
3162 {
3163         int i, ret;
3164
3165         VM_BUG_ON_PAGE(PageTail(page), page);
3166
3167         if (likely(!PageCompound(page)))
3168                 return atomic_read(&page->_mapcount) + 1;
3169
3170         ret = compound_mapcount(page);
3171         if (PageHuge(page))
3172                 return ret;
3173         for (i = 0; i < HPAGE_PMD_NR; i++)
3174                 ret += atomic_read(&page[i]._mapcount) + 1;
3175         if (PageDoubleMap(page))
3176                 ret -= HPAGE_PMD_NR;
3177         return ret;
3178 }
3179
3180 /*
3181  * This function splits huge page into normal pages. @page can point to any
3182  * subpage of huge page to split. Split doesn't change the position of @page.
3183  *
3184  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3185  * The huge page must be locked.
3186  *
3187  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3188  *
3189  * Both head page and tail pages will inherit mapping, flags, and so on from
3190  * the hugepage.
3191  *
3192  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3193  * they are not mapped.
3194  *
3195  * Returns 0 if the hugepage is split successfully.
3196  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3197  * us.
3198  */
3199 int split_huge_page_to_list(struct page *page, struct list_head *list)
3200 {
3201         struct page *head = compound_head(page);
3202         struct anon_vma *anon_vma;
3203         int count, mapcount, ret;
3204
3205         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3206         VM_BUG_ON_PAGE(!PageAnon(page), page);
3207         VM_BUG_ON_PAGE(!PageLocked(page), page);
3208         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3209         VM_BUG_ON_PAGE(!PageCompound(page), page);
3210
3211         /*
3212          * The caller does not necessarily hold an mmap_sem that would prevent
3213          * the anon_vma disappearing so we first we take a reference to it
3214          * and then lock the anon_vma for write. This is similar to
3215          * page_lock_anon_vma_read except the write lock is taken to serialise
3216          * against parallel split or collapse operations.
3217          */
3218         anon_vma = page_get_anon_vma(head);
3219         if (!anon_vma) {
3220                 ret = -EBUSY;
3221                 goto out;
3222         }
3223         anon_vma_lock_write(anon_vma);
3224
3225         /*
3226          * Racy check if we can split the page, before freeze_page() will
3227          * split PMDs
3228          */
3229         if (total_mapcount(head) != page_count(head) - 1) {
3230                 ret = -EBUSY;
3231                 goto out_unlock;
3232         }
3233
3234         freeze_page(anon_vma, head);
3235         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3236
3237         /* Prevent deferred_split_scan() touching ->_count */
3238         spin_lock(&split_queue_lock);
3239         count = page_count(head);
3240         mapcount = total_mapcount(head);
3241         if (mapcount == count - 1) {
3242                 if (!list_empty(page_deferred_list(head))) {
3243                         split_queue_len--;
3244                         list_del(page_deferred_list(head));
3245                 }
3246                 spin_unlock(&split_queue_lock);
3247                 __split_huge_page(page, list);
3248                 ret = 0;
3249         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount > count - 1) {
3250                 spin_unlock(&split_queue_lock);
3251                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3252                                 mapcount, count);
3253                 if (PageTail(page))
3254                         dump_page(head, NULL);
3255                 dump_page(page, "total_mapcount(head) > page_count(head) - 1");
3256                 BUG();
3257         } else {
3258                 spin_unlock(&split_queue_lock);
3259                 unfreeze_page(anon_vma, head);
3260                 ret = -EBUSY;
3261         }
3262
3263 out_unlock:
3264         anon_vma_unlock_write(anon_vma);
3265         put_anon_vma(anon_vma);
3266 out:
3267         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3268         return ret;
3269 }
3270
3271 void free_transhuge_page(struct page *page)
3272 {
3273         unsigned long flags;
3274
3275         spin_lock_irqsave(&split_queue_lock, flags);
3276         if (!list_empty(page_deferred_list(page))) {
3277                 split_queue_len--;
3278                 list_del(page_deferred_list(page));
3279         }
3280         spin_unlock_irqrestore(&split_queue_lock, flags);
3281         free_compound_page(page);
3282 }
3283
3284 void deferred_split_huge_page(struct page *page)
3285 {
3286         unsigned long flags;
3287
3288         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3289
3290         spin_lock_irqsave(&split_queue_lock, flags);
3291         if (list_empty(page_deferred_list(page))) {
3292                 list_add_tail(page_deferred_list(page), &split_queue);
3293                 split_queue_len++;
3294         }
3295         spin_unlock_irqrestore(&split_queue_lock, flags);
3296 }
3297
3298 static unsigned long deferred_split_count(struct shrinker *shrink,
3299                 struct shrink_control *sc)
3300 {
3301         /*
3302          * Split a page from split_queue will free up at least one page,
3303          * at most HPAGE_PMD_NR - 1. We don't track exact number.
3304          * Let's use HPAGE_PMD_NR / 2 as ballpark.
3305          */
3306         return ACCESS_ONCE(split_queue_len) * HPAGE_PMD_NR / 2;
3307 }
3308
3309 static unsigned long deferred_split_scan(struct shrinker *shrink,
3310                 struct shrink_control *sc)
3311 {
3312         unsigned long flags;
3313         LIST_HEAD(list), *pos, *next;
3314         struct page *page;
3315         int split = 0;
3316
3317         spin_lock_irqsave(&split_queue_lock, flags);
3318         list_splice_init(&split_queue, &list);
3319
3320         /* Take pin on all head pages to avoid freeing them under us */
3321         list_for_each_safe(pos, next, &list) {
3322                 page = list_entry((void *)pos, struct page, mapping);
3323                 page = compound_head(page);
3324                 /* race with put_compound_page() */
3325                 if (!get_page_unless_zero(page)) {
3326                         list_del_init(page_deferred_list(page));
3327                         split_queue_len--;
3328                 }
3329         }
3330         spin_unlock_irqrestore(&split_queue_lock, flags);
3331
3332         list_for_each_safe(pos, next, &list) {
3333                 page = list_entry((void *)pos, struct page, mapping);
3334                 lock_page(page);
3335                 /* split_huge_page() removes page from list on success */
3336                 if (!split_huge_page(page))
3337                         split++;
3338                 unlock_page(page);
3339                 put_page(page);
3340         }
3341
3342         spin_lock_irqsave(&split_queue_lock, flags);
3343         list_splice_tail(&list, &split_queue);
3344         spin_unlock_irqrestore(&split_queue_lock, flags);
3345
3346         return split * HPAGE_PMD_NR / 2;
3347 }
3348
3349 static struct shrinker deferred_split_shrinker = {
3350         .count_objects = deferred_split_count,
3351         .scan_objects = deferred_split_scan,
3352         .seeks = DEFAULT_SEEKS,
3353 };
3354
3355 #ifdef CONFIG_DEBUG_FS
3356 static int split_huge_pages_set(void *data, u64 val)
3357 {
3358         struct zone *zone;
3359         struct page *page;
3360         unsigned long pfn, max_zone_pfn;
3361         unsigned long total = 0, split = 0;
3362
3363         if (val != 1)
3364                 return -EINVAL;
3365
3366         for_each_populated_zone(zone) {
3367                 max_zone_pfn = zone_end_pfn(zone);
3368                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3369                         if (!pfn_valid(pfn))
3370                                 continue;
3371
3372                         page = pfn_to_page(pfn);
3373                         if (!get_page_unless_zero(page))
3374                                 continue;
3375
3376                         if (zone != page_zone(page))
3377                                 goto next;
3378
3379                         if (!PageHead(page) || !PageAnon(page) ||
3380                                         PageHuge(page))
3381                                 goto next;
3382
3383                         total++;
3384                         lock_page(page);
3385                         if (!split_huge_page(page))
3386                                 split++;
3387                         unlock_page(page);
3388 next:
3389                         put_page(page);
3390                 }
3391         }
3392
3393         pr_info("%lu of %lu THP split", split, total);
3394
3395         return 0;
3396 }
3397 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3398                 "%llu\n");
3399
3400 static int __init split_huge_pages_debugfs(void)
3401 {
3402         void *ret;
3403
3404         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3405                         &split_huge_pages_fops);
3406         if (!ret)
3407                 pr_warn("Failed to create split_huge_pages in debugfs");
3408         return 0;
3409 }
3410 late_initcall(split_huge_pages_debugfs);
3411 #endif