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/pfn_t.h>
  25 #include <linux/mman.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32
  33 #include <asm/tlb.h>
  34 #include <asm/pgalloc.h>
  35 #include "internal.h"
  36
  37 enum scan_result {
  38         SCAN_FAIL,
  39         SCAN_SUCCEED,
  40         SCAN_PMD_NULL,
  41         SCAN_EXCEED_NONE_PTE,
  42         SCAN_PTE_NON_PRESENT,
  43         SCAN_PAGE_RO,
  44         SCAN_NO_REFERENCED_PAGE,
  45         SCAN_PAGE_NULL,
  46         SCAN_SCAN_ABORT,
  47         SCAN_PAGE_COUNT,
  48         SCAN_PAGE_LRU,
  49         SCAN_PAGE_LOCK,
  50         SCAN_PAGE_ANON,
  51         SCAN_PAGE_COMPOUND,
  52         SCAN_ANY_PROCESS,
  53         SCAN_VMA_NULL,
  54         SCAN_VMA_CHECK,
  55         SCAN_ADDRESS_RANGE,
  56         SCAN_SWAP_CACHE_PAGE,
  57         SCAN_DEL_PAGE_LRU,
  58         SCAN_ALLOC_HUGE_PAGE_FAIL,
  59         SCAN_CGROUP_CHARGE_FAIL
  60 };
  61
  62 #define CREATE_TRACE_POINTS
  63 #include <trace/events/huge_memory.h>
  64
  65 /*
  66  * By default transparent hugepage support is disabled in order that avoid
  67  * to risk increase the memory footprint of applications without a guaranteed
  68  * benefit. When transparent hugepage support is enabled, is for all mappings,
  69  * and khugepaged scans all mappings.
  70  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  71  * for all hugepage allocations.
  72  */
  73 unsigned long transparent_hugepage_flags __read_mostly =
  74 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  75         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  76 #endif
  77 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  78         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  79 #endif
  80         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  81         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  82         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  83
  84 /* default scan 8*512 pte (or vmas) every 30 second */
  85 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  86 static unsigned int khugepaged_pages_collapsed;
  87 static unsigned int khugepaged_full_scans;
  88 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  89 /* during fragmentation poll the hugepage allocator once every minute */
  90 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  91 static struct task_struct *khugepaged_thread __read_mostly;
  92 static DEFINE_MUTEX(khugepaged_mutex);
  93 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  94 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  95 /*
  96  * default collapse hugepages if there is at least one pte mapped like
  97  * it would have happened if the vma was large enough during page
  98  * fault.
  99  */
 100 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
 101
 102 static int khugepaged(void *none);
 103 static int khugepaged_slab_init(void);
 104 static void khugepaged_slab_exit(void);
 105
 106 #define MM_SLOTS_HASH_BITS 10
 107 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 108
 109 static struct kmem_cache *mm_slot_cache __read_mostly;
 110
 111 /**
 112  * struct mm_slot - hash lookup from mm to mm_slot
 113  * @hash: hash collision list
 114  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 115  * @mm: the mm that this information is valid for
 116  */
 117 struct mm_slot {
 118         struct hlist_node hash;
 119         struct list_head mm_node;
 120         struct mm_struct *mm;
 121 };
 122
 123 /**
 124  * struct khugepaged_scan - cursor for scanning
 125  * @mm_head: the head of the mm list to scan
 126  * @mm_slot: the current mm_slot we are scanning
 127  * @address: the next address inside that to be scanned
 128  *
 129  * There is only the one khugepaged_scan instance of this cursor structure.
 130  */
 131 struct khugepaged_scan {
 132         struct list_head mm_head;
 133         struct mm_slot *mm_slot;
 134         unsigned long address;
 135 };
 136 static struct khugepaged_scan khugepaged_scan = {
 137         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 138 };
 139
 140 static DEFINE_SPINLOCK(split_queue_lock);
 141 static LIST_HEAD(split_queue);
 142 static unsigned long split_queue_len;
 143 static struct shrinker deferred_split_shrinker;
 144
 145 static void set_recommended_min_free_kbytes(void)
 146 {
 147         struct zone *zone;
 148         int nr_zones = 0;
 149         unsigned long recommended_min;
 150
 151         for_each_populated_zone(zone)
 152                 nr_zones++;
 153
 154         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 155         recommended_min = pageblock_nr_pages * nr_zones * 2;
 156
 157         /*
 158          * Make sure that on average at least two pageblocks are almost free
 159          * of another type, one for a migratetype to fall back to and a
 160          * second to avoid subsequent fallbacks of other types There are 3
 161          * MIGRATE_TYPES we care about.
 162          */
 163         recommended_min += pageblock_nr_pages * nr_zones *
 164                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 165
 166         /* don't ever allow to reserve more than 5% of the lowmem */
 167         recommended_min = min(recommended_min,
 168                               (unsigned long) nr_free_buffer_pages() / 20);
 169         recommended_min <<= (PAGE_SHIFT-10);
 170
 171         if (recommended_min > min_free_kbytes) {
 172                 if (user_min_free_kbytes >= 0)
 173                         pr_info("raising min_free_kbytes from %d to %lu "
 174                                 "to help transparent hugepage allocations\n",
 175                                 min_free_kbytes, recommended_min);
 176
 177                 min_free_kbytes = recommended_min;
 178         }
 179         setup_per_zone_wmarks();
 180 }
 181
 182 static int start_stop_khugepaged(void)
 183 {
 184         int err = 0;
 185         if (khugepaged_enabled()) {
 186                 if (!khugepaged_thread)
 187                         khugepaged_thread = kthread_run(khugepaged, NULL,
 188                                                         "khugepaged");
 189                 if (IS_ERR(khugepaged_thread)) {
 190                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 191                         err = PTR_ERR(khugepaged_thread);
 192                         khugepaged_thread = NULL;
 193                         goto fail;
 194                 }
 195
 196                 if (!list_empty(&khugepaged_scan.mm_head))
 197                         wake_up_interruptible(&khugepaged_wait);
 198
 199                 set_recommended_min_free_kbytes();
 200         } else if (khugepaged_thread) {
 201                 kthread_stop(khugepaged_thread);
 202                 khugepaged_thread = NULL;
 203         }
 204 fail:
 205         return err;
 206 }
 207
 208 static atomic_t huge_zero_refcount;
 209 struct page *huge_zero_page __read_mostly;
 210
 211 struct page *get_huge_zero_page(void)
 212 {
 213         struct page *zero_page;
 214 retry:
 215         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 216                 return READ_ONCE(huge_zero_page);
 217
 218         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 219                         HPAGE_PMD_ORDER);
 220         if (!zero_page) {
 221                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 222                 return NULL;
 223         }
 224         count_vm_event(THP_ZERO_PAGE_ALLOC);
 225         preempt_disable();
 226         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 227                 preempt_enable();
 228                 __free_pages(zero_page, compound_order(zero_page));
 229                 goto retry;
 230         }
 231
 232         /* We take additional reference here. It will be put back by shrinker */
 233         atomic_set(&huge_zero_refcount, 2);
 234         preempt_enable();
 235         return READ_ONCE(huge_zero_page);
 236 }
 237
 238 static void put_huge_zero_page(void)
 239 {
 240         /*
 241          * Counter should never go to zero here. Only shrinker can put
 242          * last reference.
 243          */
 244         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 245 }
 246
 247 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 248                                         struct shrink_control *sc)
 249 {
 250         /* we can free zero page only if last reference remains */
 251         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 252 }
 253
 254 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 255                                        struct shrink_control *sc)
 256 {
 257         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 258                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 259                 BUG_ON(zero_page == NULL);
 260                 __free_pages(zero_page, compound_order(zero_page));
 261                 return HPAGE_PMD_NR;
 262         }
 263
 264         return 0;
 265 }
 266
 267 static struct shrinker huge_zero_page_shrinker = {
 268         .count_objects = shrink_huge_zero_page_count,
 269         .scan_objects = shrink_huge_zero_page_scan,
 270         .seeks = DEFAULT_SEEKS,
 271 };
 272
 273 #ifdef CONFIG_SYSFS
 274
 275 static ssize_t double_flag_show(struct kobject *kobj,
 276                                 struct kobj_attribute *attr, char *buf,
 277                                 enum transparent_hugepage_flag enabled,
 278                                 enum transparent_hugepage_flag req_madv)
 279 {
 280         if (test_bit(enabled, &transparent_hugepage_flags)) {
 281                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 282                 return sprintf(buf, "[always] madvise never\n");
 283         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 284                 return sprintf(buf, "always [madvise] never\n");
 285         else
 286                 return sprintf(buf, "always madvise [never]\n");
 287 }
 288 static ssize_t double_flag_store(struct kobject *kobj,
 289                                  struct kobj_attribute *attr,
 290                                  const char *buf, size_t count,
 291                                  enum transparent_hugepage_flag enabled,
 292                                  enum transparent_hugepage_flag req_madv)
 293 {
 294         if (!memcmp("always", buf,
 295                     min(sizeof("always")-1, count))) {
 296                 set_bit(enabled, &transparent_hugepage_flags);
 297                 clear_bit(req_madv, &transparent_hugepage_flags);
 298         } else if (!memcmp("madvise", buf,
 299                            min(sizeof("madvise")-1, count))) {
 300                 clear_bit(enabled, &transparent_hugepage_flags);
 301                 set_bit(req_madv, &transparent_hugepage_flags);
 302         } else if (!memcmp("never", buf,
 303                            min(sizeof("never")-1, count))) {
 304                 clear_bit(enabled, &transparent_hugepage_flags);
 305                 clear_bit(req_madv, &transparent_hugepage_flags);
 306         } else
 307                 return -EINVAL;
 308
 309         return count;
 310 }
 311
 312 static ssize_t enabled_show(struct kobject *kobj,
 313                             struct kobj_attribute *attr, char *buf)
 314 {
 315         return double_flag_show(kobj, attr, buf,
 316                                 TRANSPARENT_HUGEPAGE_FLAG,
 317                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 318 }
 319 static ssize_t enabled_store(struct kobject *kobj,
 320                              struct kobj_attribute *attr,
 321                              const char *buf, size_t count)
 322 {
 323         ssize_t ret;
 324
 325         ret = double_flag_store(kobj, attr, buf, count,
 326                                 TRANSPARENT_HUGEPAGE_FLAG,
 327                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 328
 329         if (ret > 0) {
 330                 int err;
 331
 332                 mutex_lock(&khugepaged_mutex);
 333                 err = start_stop_khugepaged();
 334                 mutex_unlock(&khugepaged_mutex);
 335
 336                 if (err)
 337                         ret = err;
 338         }
 339
 340         return ret;
 341 }
 342 static struct kobj_attribute enabled_attr =
 343         __ATTR(enabled, 0644, enabled_show, enabled_store);
 344
 345 static ssize_t single_flag_show(struct kobject *kobj,
 346                                 struct kobj_attribute *attr, char *buf,
 347                                 enum transparent_hugepage_flag flag)
 348 {
 349         return sprintf(buf, "%d\n",
 350                        !!test_bit(flag, &transparent_hugepage_flags));
 351 }
 352
 353 static ssize_t single_flag_store(struct kobject *kobj,
 354                                  struct kobj_attribute *attr,
 355                                  const char *buf, size_t count,
 356                                  enum transparent_hugepage_flag flag)
 357 {
 358         unsigned long value;
 359         int ret;
 360
 361         ret = kstrtoul(buf, 10, &value);
 362         if (ret < 0)
 363                 return ret;
 364         if (value > 1)
 365                 return -EINVAL;
 366
 367         if (value)
 368                 set_bit(flag, &transparent_hugepage_flags);
 369         else
 370                 clear_bit(flag, &transparent_hugepage_flags);
 371
 372         return count;
 373 }
 374
 375 /*
 376  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 377  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 378  * memory just to allocate one more hugepage.
 379  */
 380 static ssize_t defrag_show(struct kobject *kobj,
 381                            struct kobj_attribute *attr, char *buf)
 382 {
 383         return double_flag_show(kobj, attr, buf,
 384                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 385                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 386 }
 387 static ssize_t defrag_store(struct kobject *kobj,
 388                             struct kobj_attribute *attr,
 389                             const char *buf, size_t count)
 390 {
 391         return double_flag_store(kobj, attr, buf, count,
 392                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 393                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 394 }
 395 static struct kobj_attribute defrag_attr =
 396         __ATTR(defrag, 0644, defrag_show, defrag_store);
 397
 398 static ssize_t use_zero_page_show(struct kobject *kobj,
 399                 struct kobj_attribute *attr, char *buf)
 400 {
 401         return single_flag_show(kobj, attr, buf,
 402                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 403 }
 404 static ssize_t use_zero_page_store(struct kobject *kobj,
 405                 struct kobj_attribute *attr, const char *buf, size_t count)
 406 {
 407         return single_flag_store(kobj, attr, buf, count,
 408                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 409 }
 410 static struct kobj_attribute use_zero_page_attr =
 411         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 412 #ifdef CONFIG_DEBUG_VM
 413 static ssize_t debug_cow_show(struct kobject *kobj,
 414                                 struct kobj_attribute *attr, char *buf)
 415 {
 416         return single_flag_show(kobj, attr, buf,
 417                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 418 }
 419 static ssize_t debug_cow_store(struct kobject *kobj,
 420                                struct kobj_attribute *attr,
 421                                const char *buf, size_t count)
 422 {
 423         return single_flag_store(kobj, attr, buf, count,
 424                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 425 }
 426 static struct kobj_attribute debug_cow_attr =
 427         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 428 #endif /* CONFIG_DEBUG_VM */
 429
 430 static struct attribute *hugepage_attr[] = {
 431         &enabled_attr.attr,
 432         &defrag_attr.attr,
 433         &use_zero_page_attr.attr,
 434 #ifdef CONFIG_DEBUG_VM
 435         &debug_cow_attr.attr,
 436 #endif
 437         NULL,
 438 };
 439
 440 static struct attribute_group hugepage_attr_group = {
 441         .attrs = hugepage_attr,
 442 };
 443
 444 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 445                                          struct kobj_attribute *attr,
 446                                          char *buf)
 447 {
 448         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 449 }
 450
 451 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 452                                           struct kobj_attribute *attr,
 453                                           const char *buf, size_t count)
 454 {
 455         unsigned long msecs;
 456         int err;
 457
 458         err = kstrtoul(buf, 10, &msecs);
 459         if (err || msecs > UINT_MAX)
 460                 return -EINVAL;
 461
 462         khugepaged_scan_sleep_millisecs = msecs;
 463         wake_up_interruptible(&khugepaged_wait);
 464
 465         return count;
 466 }
 467 static struct kobj_attribute scan_sleep_millisecs_attr =
 468         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 469                scan_sleep_millisecs_store);
 470
 471 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 472                                           struct kobj_attribute *attr,
 473                                           char *buf)
 474 {
 475         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 476 }
 477
 478 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 479                                            struct kobj_attribute *attr,
 480                                            const char *buf, size_t count)
 481 {
 482         unsigned long msecs;
 483         int err;
 484
 485         err = kstrtoul(buf, 10, &msecs);
 486         if (err || msecs > UINT_MAX)
 487                 return -EINVAL;
 488
 489         khugepaged_alloc_sleep_millisecs = msecs;
 490         wake_up_interruptible(&khugepaged_wait);
 491
 492         return count;
 493 }
 494 static struct kobj_attribute alloc_sleep_millisecs_attr =
 495         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 496                alloc_sleep_millisecs_store);
 497
 498 static ssize_t pages_to_scan_show(struct kobject *kobj,
 499                                   struct kobj_attribute *attr,
 500                                   char *buf)
 501 {
 502         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 503 }
 504 static ssize_t pages_to_scan_store(struct kobject *kobj,
 505                                    struct kobj_attribute *attr,
 506                                    const char *buf, size_t count)
 507 {
 508         int err;
 509         unsigned long pages;
 510
 511         err = kstrtoul(buf, 10, &pages);
 512         if (err || !pages || pages > UINT_MAX)
 513                 return -EINVAL;
 514
 515         khugepaged_pages_to_scan = pages;
 516
 517         return count;
 518 }
 519 static struct kobj_attribute pages_to_scan_attr =
 520         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 521                pages_to_scan_store);
 522
 523 static ssize_t pages_collapsed_show(struct kobject *kobj,
 524                                     struct kobj_attribute *attr,
 525                                     char *buf)
 526 {
 527         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 528 }
 529 static struct kobj_attribute pages_collapsed_attr =
 530         __ATTR_RO(pages_collapsed);
 531
 532 static ssize_t full_scans_show(struct kobject *kobj,
 533                                struct kobj_attribute *attr,
 534                                char *buf)
 535 {
 536         return sprintf(buf, "%u\n", khugepaged_full_scans);
 537 }
 538 static struct kobj_attribute full_scans_attr =
 539         __ATTR_RO(full_scans);
 540
 541 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 542                                       struct kobj_attribute *attr, char *buf)
 543 {
 544         return single_flag_show(kobj, attr, buf,
 545                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 546 }
 547 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 548                                        struct kobj_attribute *attr,
 549                                        const char *buf, size_t count)
 550 {
 551         return single_flag_store(kobj, attr, buf, count,
 552                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 553 }
 554 static struct kobj_attribute khugepaged_defrag_attr =
 555         __ATTR(defrag, 0644, khugepaged_defrag_show,
 556                khugepaged_defrag_store);
 557
 558 /*
 559  * max_ptes_none controls if khugepaged should collapse hugepages over
 560  * any unmapped ptes in turn potentially increasing the memory
 561  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 562  * reduce the available free memory in the system as it
 563  * runs. Increasing max_ptes_none will instead potentially reduce the
 564  * free memory in the system during the khugepaged scan.
 565  */
 566 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 567                                              struct kobj_attribute *attr,
 568                                              char *buf)
 569 {
 570         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 571 }
 572 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 573                                               struct kobj_attribute *attr,
 574                                               const char *buf, size_t count)
 575 {
 576         int err;
 577         unsigned long max_ptes_none;
 578
 579         err = kstrtoul(buf, 10, &max_ptes_none);
 580         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 581                 return -EINVAL;
 582
 583         khugepaged_max_ptes_none = max_ptes_none;
 584
 585         return count;
 586 }
 587 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 588         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 589                khugepaged_max_ptes_none_store);
 590
 591 static struct attribute *khugepaged_attr[] = {
 592         &khugepaged_defrag_attr.attr,
 593         &khugepaged_max_ptes_none_attr.attr,
 594         &pages_to_scan_attr.attr,
 595         &pages_collapsed_attr.attr,
 596         &full_scans_attr.attr,
 597         &scan_sleep_millisecs_attr.attr,
 598         &alloc_sleep_millisecs_attr.attr,
 599         NULL,
 600 };
 601
 602 static struct attribute_group khugepaged_attr_group = {
 603         .attrs = khugepaged_attr,
 604         .name = "khugepaged",
 605 };
 606
 607 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 608 {
 609         int err;
 610
 611         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 612         if (unlikely(!*hugepage_kobj)) {
 613                 pr_err("failed to create transparent hugepage kobject\n");
 614                 return -ENOMEM;
 615         }
 616
 617         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 618         if (err) {
 619                 pr_err("failed to register transparent hugepage group\n");
 620                 goto delete_obj;
 621         }
 622
 623         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 624         if (err) {
 625                 pr_err("failed to register transparent hugepage group\n");
 626                 goto remove_hp_group;
 627         }
 628
 629         return 0;
 630
 631 remove_hp_group:
 632         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 633 delete_obj:
 634         kobject_put(*hugepage_kobj);
 635         return err;
 636 }
 637
 638 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 639 {
 640         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 641         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 642         kobject_put(hugepage_kobj);
 643 }
 644 #else
 645 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 646 {
 647         return 0;
 648 }
 649
 650 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 651 {
 652 }
 653 #endif /* CONFIG_SYSFS */
 654
 655 static int __init hugepage_init(void)
 656 {
 657         int err;
 658         struct kobject *hugepage_kobj;
 659
 660         if (!has_transparent_hugepage()) {
 661                 transparent_hugepage_flags = 0;
 662                 return -EINVAL;
 663         }
 664
 665         err = hugepage_init_sysfs(&hugepage_kobj);
 666         if (err)
 667                 goto err_sysfs;
 668
 669         err = khugepaged_slab_init();
 670         if (err)
 671                 goto err_slab;
 672
 673         err = register_shrinker(&huge_zero_page_shrinker);
 674         if (err)
 675                 goto err_hzp_shrinker;
 676         err = register_shrinker(&deferred_split_shrinker);
 677         if (err)
 678                 goto err_split_shrinker;
 679
 680         /*
 681          * By default disable transparent hugepages on smaller systems,
 682          * where the extra memory used could hurt more than TLB overhead
 683          * is likely to save.  The admin can still enable it through /sys.
 684          */
 685         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 686                 transparent_hugepage_flags = 0;
 687                 return 0;
 688         }
 689
 690         err = start_stop_khugepaged();
 691         if (err)
 692                 goto err_khugepaged;
 693
 694         return 0;
 695 err_khugepaged:
 696         unregister_shrinker(&deferred_split_shrinker);
 697 err_split_shrinker:
 698         unregister_shrinker(&huge_zero_page_shrinker);
 699 err_hzp_shrinker:
 700         khugepaged_slab_exit();
 701 err_slab:
 702         hugepage_exit_sysfs(hugepage_kobj);
 703 err_sysfs:
 704         return err;
 705 }
 706 subsys_initcall(hugepage_init);
 707
 708 static int __init setup_transparent_hugepage(char *str)
 709 {
 710         int ret = 0;
 711         if (!str)
 712                 goto out;
 713         if (!strcmp(str, "always")) {
 714                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 715                         &transparent_hugepage_flags);
 716                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 717                           &transparent_hugepage_flags);
 718                 ret = 1;
 719         } else if (!strcmp(str, "madvise")) {
 720                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 721                           &transparent_hugepage_flags);
 722                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 723                         &transparent_hugepage_flags);
 724                 ret = 1;
 725         } else if (!strcmp(str, "never")) {
 726                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 727                           &transparent_hugepage_flags);
 728                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 729                           &transparent_hugepage_flags);
 730                 ret = 1;
 731         }
 732 out:
 733         if (!ret)
 734                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 735         return ret;
 736 }
 737 __setup("transparent_hugepage=", setup_transparent_hugepage);
 738
 739 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 740 {
 741         if (likely(vma->vm_flags & VM_WRITE))
 742                 pmd = pmd_mkwrite(pmd);
 743         return pmd;
 744 }
 745
 746 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 747 {
 748         pmd_t entry;
 749         entry = mk_pmd(page, prot);
 750         entry = pmd_mkhuge(entry);
 751         return entry;
 752 }
 753
 754 static inline struct list_head *page_deferred_list(struct page *page)
 755 {
 756         /*
 757          * ->lru in the tail pages is occupied by compound_head.
 758          * Let's use ->mapping + ->index in the second tail page as list_head.
 759          */
 760         return (struct list_head *)&page[2].mapping;
 761 }
 762
 763 void prep_transhuge_page(struct page *page)
 764 {
 765         /*
 766          * we use page->mapping and page->indexlru in second tail page
 767          * as list_head: assuming THP order >= 2
 768          */
 769         BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 770
 771         INIT_LIST_HEAD(page_deferred_list(page));
 772         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 773 }
 774
 775 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 776                                         struct vm_area_struct *vma,
 777                                         unsigned long address, pmd_t *pmd,
 778                                         struct page *page, gfp_t gfp,
 779                                         unsigned int flags)
 780 {
 781         struct mem_cgroup *memcg;
 782         pgtable_t pgtable;
 783         spinlock_t *ptl;
 784         unsigned long haddr = address & HPAGE_PMD_MASK;
 785
 786         VM_BUG_ON_PAGE(!PageCompound(page), page);
 787
 788         if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
 789                 put_page(page);
 790                 count_vm_event(THP_FAULT_FALLBACK);
 791                 return VM_FAULT_FALLBACK;
 792         }
 793
 794         pgtable = pte_alloc_one(mm, haddr);
 795         if (unlikely(!pgtable)) {
 796                 mem_cgroup_cancel_charge(page, memcg, true);
 797                 put_page(page);
 798                 return VM_FAULT_OOM;
 799         }
 800
 801         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 802         /*
 803          * The memory barrier inside __SetPageUptodate makes sure that
 804          * clear_huge_page writes become visible before the set_pmd_at()
 805          * write.
 806          */
 807         __SetPageUptodate(page);
 808
 809         ptl = pmd_lock(mm, pmd);
 810         if (unlikely(!pmd_none(*pmd))) {
 811                 spin_unlock(ptl);
 812                 mem_cgroup_cancel_charge(page, memcg, true);
 813                 put_page(page);
 814                 pte_free(mm, pgtable);
 815         } else {
 816                 pmd_t entry;
 817
 818                 /* Deliver the page fault to userland */
 819                 if (userfaultfd_missing(vma)) {
 820                         int ret;
 821
 822                         spin_unlock(ptl);
 823                         mem_cgroup_cancel_charge(page, memcg, true);
 824                         put_page(page);
 825                         pte_free(mm, pgtable);
 826                         ret = handle_userfault(vma, address, flags,
 827                                                VM_UFFD_MISSING);
 828                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 829                         return ret;
 830                 }
 831
 832                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 833                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 834                 page_add_new_anon_rmap(page, vma, haddr, true);
 835                 mem_cgroup_commit_charge(page, memcg, false, true);
 836                 lru_cache_add_active_or_unevictable(page, vma);
 837                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 838                 set_pmd_at(mm, haddr, pmd, entry);
 839                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 840                 atomic_long_inc(&mm->nr_ptes);
 841                 spin_unlock(ptl);
 842                 count_vm_event(THP_FAULT_ALLOC);
 843         }
 844
 845         return 0;
 846 }
 847
 848 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 849 {
 850         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 851 }
 852
 853 /* Caller must hold page table lock. */
 854 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 855                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 856                 struct page *zero_page)
 857 {
 858         pmd_t entry;
 859         if (!pmd_none(*pmd))
 860                 return false;
 861         entry = mk_pmd(zero_page, vma->vm_page_prot);
 862         entry = pmd_mkhuge(entry);
 863         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 864         set_pmd_at(mm, haddr, pmd, entry);
 865         atomic_long_inc(&mm->nr_ptes);
 866         return true;
 867 }
 868
 869 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 870                                unsigned long address, pmd_t *pmd,
 871                                unsigned int flags)
 872 {
 873         gfp_t gfp;
 874         struct page *page;
 875         unsigned long haddr = address & HPAGE_PMD_MASK;
 876
 877         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 878                 return VM_FAULT_FALLBACK;
 879         if (unlikely(anon_vma_prepare(vma)))
 880                 return VM_FAULT_OOM;
 881         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 882                 return VM_FAULT_OOM;
 883         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 884                         transparent_hugepage_use_zero_page()) {
 885                 spinlock_t *ptl;
 886                 pgtable_t pgtable;
 887                 struct page *zero_page;
 888                 bool set;
 889                 int ret;
 890                 pgtable = pte_alloc_one(mm, haddr);
 891                 if (unlikely(!pgtable))
 892                         return VM_FAULT_OOM;
 893                 zero_page = get_huge_zero_page();
 894                 if (unlikely(!zero_page)) {
 895                         pte_free(mm, pgtable);
 896                         count_vm_event(THP_FAULT_FALLBACK);
 897                         return VM_FAULT_FALLBACK;
 898                 }
 899                 ptl = pmd_lock(mm, pmd);
 900                 ret = 0;
 901                 set = false;
 902                 if (pmd_none(*pmd)) {
 903                         if (userfaultfd_missing(vma)) {
 904                                 spin_unlock(ptl);
 905                                 ret = handle_userfault(vma, address, flags,
 906                                                        VM_UFFD_MISSING);
 907                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 908                         } else {
 909                                 set_huge_zero_page(pgtable, mm, vma,
 910                                                    haddr, pmd,
 911                                                    zero_page);
 912                                 spin_unlock(ptl);
 913                                 set = true;
 914                         }
 915                 } else
 916                         spin_unlock(ptl);
 917                 if (!set) {
 918                         pte_free(mm, pgtable);
 919                         put_huge_zero_page();
 920                 }
 921                 return ret;
 922         }
 923         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 924         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 925         if (unlikely(!page)) {
 926                 count_vm_event(THP_FAULT_FALLBACK);
 927                 return VM_FAULT_FALLBACK;
 928         }
 929         prep_transhuge_page(page);
 930         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 931                                             flags);
 932 }
 933
 934 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 935                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 936 {
 937         struct mm_struct *mm = vma->vm_mm;
 938         pmd_t entry;
 939         spinlock_t *ptl;
 940
 941         ptl = pmd_lock(mm, pmd);
 942         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 943         if (pfn_t_devmap(pfn))
 944                 entry = pmd_mkdevmap(entry);
 945         if (write) {
 946                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 947                 entry = maybe_pmd_mkwrite(entry, vma);
 948         }
 949         set_pmd_at(mm, addr, pmd, entry);
 950         update_mmu_cache_pmd(vma, addr, pmd);
 951         spin_unlock(ptl);
 952 }
 953
 954 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 955                         pmd_t *pmd, pfn_t pfn, bool write)
 956 {
 957         pgprot_t pgprot = vma->vm_page_prot;
 958         /*
 959          * If we had pmd_special, we could avoid all these restrictions,
 960          * but we need to be consistent with PTEs and architectures that
 961          * can't support a 'special' bit.
 962          */
 963         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 964         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 965                                                 (VM_PFNMAP|VM_MIXEDMAP));
 966         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 967         BUG_ON(!pfn_t_devmap(pfn));
 968
 969         if (addr < vma->vm_start || addr >= vma->vm_end)
 970                 return VM_FAULT_SIGBUS;
 971         if (track_pfn_insert(vma, &pgprot, pfn))
 972                 return VM_FAULT_SIGBUS;
 973         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 974         return VM_FAULT_NOPAGE;
 975 }
 976
 977 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 978                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 979                   struct vm_area_struct *vma)
 980 {
 981         spinlock_t *dst_ptl, *src_ptl;
 982         struct page *src_page;
 983         pmd_t pmd;
 984         pgtable_t pgtable;
 985         int ret;
 986
 987         ret = -ENOMEM;
 988         pgtable = pte_alloc_one(dst_mm, addr);
 989         if (unlikely(!pgtable))
 990                 goto out;
 991
 992         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 993         src_ptl = pmd_lockptr(src_mm, src_pmd);
 994         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 995
 996         ret = -EAGAIN;
 997         pmd = *src_pmd;
 998         if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
 999                 pte_free(dst_mm, pgtable);
1000                 goto out_unlock;
1001         }
1002         /*
1003          * When page table lock is held, the huge zero pmd should not be
1004          * under splitting since we don't split the page itself, only pmd to
1005          * a page table.
1006          */
1007         if (is_huge_zero_pmd(pmd)) {
1008                 struct page *zero_page;
1009                 /*
1010                  * get_huge_zero_page() will never allocate a new page here,
1011                  * since we already have a zero page to copy. It just takes a
1012                  * reference.
1013                  */
1014                 zero_page = get_huge_zero_page();
1015                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1016                                 zero_page);
1017                 ret = 0;
1018                 goto out_unlock;
1019         }
1020
1021         if (pmd_trans_huge(pmd)) {
1022                 /* thp accounting separate from pmd_devmap accounting */
1023                 src_page = pmd_page(pmd);
1024                 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1025                 get_page(src_page);
1026                 page_dup_rmap(src_page, true);
1027                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1028                 atomic_long_inc(&dst_mm->nr_ptes);
1029                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1030         }
1031
1032         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1033         pmd = pmd_mkold(pmd_wrprotect(pmd));
1034         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1035
1036         ret = 0;
1037 out_unlock:
1038         spin_unlock(src_ptl);
1039         spin_unlock(dst_ptl);
1040 out:
1041         return ret;
1042 }
1043
1044 void huge_pmd_set_accessed(struct mm_struct *mm,
1045                            struct vm_area_struct *vma,
1046                            unsigned long address,
1047                            pmd_t *pmd, pmd_t orig_pmd,
1048                            int dirty)
1049 {
1050         spinlock_t *ptl;
1051         pmd_t entry;
1052         unsigned long haddr;
1053
1054         ptl = pmd_lock(mm, pmd);
1055         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1056                 goto unlock;
1057
1058         entry = pmd_mkyoung(orig_pmd);
1059         haddr = address & HPAGE_PMD_MASK;
1060         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1061                 update_mmu_cache_pmd(vma, address, pmd);
1062
1063 unlock:
1064         spin_unlock(ptl);
1065 }
1066
1067 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1068                                         struct vm_area_struct *vma,
1069                                         unsigned long address,
1070                                         pmd_t *pmd, pmd_t orig_pmd,
1071                                         struct page *page,
1072                                         unsigned long haddr)
1073 {
1074         struct mem_cgroup *memcg;
1075         spinlock_t *ptl;
1076         pgtable_t pgtable;
1077         pmd_t _pmd;
1078         int ret = 0, i;
1079         struct page **pages;
1080         unsigned long mmun_start;       /* For mmu_notifiers */
1081         unsigned long mmun_end;         /* For mmu_notifiers */
1082
1083         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1084                         GFP_KERNEL);
1085         if (unlikely(!pages)) {
1086                 ret |= VM_FAULT_OOM;
1087                 goto out;
1088         }
1089
1090         for (i = 0; i < HPAGE_PMD_NR; i++) {
1091                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1092                                                __GFP_OTHER_NODE,
1093                                                vma, address, page_to_nid(page));
1094                 if (unlikely(!pages[i] ||
1095                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1096                                                    &memcg, false))) {
1097                         if (pages[i])
1098                                 put_page(pages[i]);
1099                         while (--i >= 0) {
1100                                 memcg = (void *)page_private(pages[i]);
1101                                 set_page_private(pages[i], 0);
1102                                 mem_cgroup_cancel_charge(pages[i], memcg,
1103                                                 false);
1104                                 put_page(pages[i]);
1105                         }
1106                         kfree(pages);
1107                         ret |= VM_FAULT_OOM;
1108                         goto out;
1109                 }
1110                 set_page_private(pages[i], (unsigned long)memcg);
1111         }
1112
1113         for (i = 0; i < HPAGE_PMD_NR; i++) {
1114                 copy_user_highpage(pages[i], page + i,
1115                                    haddr + PAGE_SIZE * i, vma);
1116                 __SetPageUptodate(pages[i]);
1117                 cond_resched();
1118         }
1119
1120         mmun_start = haddr;
1121         mmun_end   = haddr + HPAGE_PMD_SIZE;
1122         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1123
1124         ptl = pmd_lock(mm, pmd);
1125         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1126                 goto out_free_pages;
1127         VM_BUG_ON_PAGE(!PageHead(page), page);
1128
1129         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1130         /* leave pmd empty until pte is filled */
1131
1132         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1133         pmd_populate(mm, &_pmd, pgtable);
1134
1135         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1136                 pte_t *pte, entry;
1137                 entry = mk_pte(pages[i], vma->vm_page_prot);
1138                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1139                 memcg = (void *)page_private(pages[i]);
1140                 set_page_private(pages[i], 0);
1141                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1142                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1143                 lru_cache_add_active_or_unevictable(pages[i], vma);
1144                 pte = pte_offset_map(&_pmd, haddr);
1145                 VM_BUG_ON(!pte_none(*pte));
1146                 set_pte_at(mm, haddr, pte, entry);
1147                 pte_unmap(pte);
1148         }
1149         kfree(pages);
1150
1151         smp_wmb(); /* make pte visible before pmd */
1152         pmd_populate(mm, pmd, pgtable);
1153         page_remove_rmap(page, true);
1154         spin_unlock(ptl);
1155
1156         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1157
1158         ret |= VM_FAULT_WRITE;
1159         put_page(page);
1160
1161 out:
1162         return ret;
1163
1164 out_free_pages:
1165         spin_unlock(ptl);
1166         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1167         for (i = 0; i < HPAGE_PMD_NR; i++) {
1168                 memcg = (void *)page_private(pages[i]);
1169                 set_page_private(pages[i], 0);
1170                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1171                 put_page(pages[i]);
1172         }
1173         kfree(pages);
1174         goto out;
1175 }
1176
1177 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1178                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1179 {
1180         spinlock_t *ptl;
1181         int ret = 0;
1182         struct page *page = NULL, *new_page;
1183         struct mem_cgroup *memcg;
1184         unsigned long haddr;
1185         unsigned long mmun_start;       /* For mmu_notifiers */
1186         unsigned long mmun_end;         /* For mmu_notifiers */
1187         gfp_t huge_gfp;                 /* for allocation and charge */
1188
1189         ptl = pmd_lockptr(mm, pmd);
1190         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1191         haddr = address & HPAGE_PMD_MASK;
1192         if (is_huge_zero_pmd(orig_pmd))
1193                 goto alloc;
1194         spin_lock(ptl);
1195         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1196                 goto out_unlock;
1197
1198         page = pmd_page(orig_pmd);
1199         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1200         /*
1201          * We can only reuse the page if nobody else maps the huge page or it's
1202          * part. We can do it by checking page_mapcount() on each sub-page, but
1203          * it's expensive.
1204          * The cheaper way is to check page_count() to be equal 1: every
1205          * mapcount takes page reference reference, so this way we can
1206          * guarantee, that the PMD is the only mapping.
1207          * This can give false negative if somebody pinned the page, but that's
1208          * fine.
1209          */
1210         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1211                 pmd_t entry;
1212                 entry = pmd_mkyoung(orig_pmd);
1213                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1214                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1215                         update_mmu_cache_pmd(vma, address, pmd);
1216                 ret |= VM_FAULT_WRITE;
1217                 goto out_unlock;
1218         }
1219         get_page(page);
1220         spin_unlock(ptl);
1221 alloc:
1222         if (transparent_hugepage_enabled(vma) &&
1223             !transparent_hugepage_debug_cow()) {
1224                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1225                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1226         } else
1227                 new_page = NULL;
1228
1229         if (likely(new_page)) {
1230                 prep_transhuge_page(new_page);
1231         } else {
1232                 if (!page) {
1233                         split_huge_pmd(vma, pmd, address);
1234                         ret |= VM_FAULT_FALLBACK;
1235                 } else {
1236                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1237                                         pmd, orig_pmd, page, haddr);
1238                         if (ret & VM_FAULT_OOM) {
1239                                 split_huge_pmd(vma, pmd, address);
1240                                 ret |= VM_FAULT_FALLBACK;
1241                         }
1242                         put_page(page);
1243                 }
1244                 count_vm_event(THP_FAULT_FALLBACK);
1245                 goto out;
1246         }
1247
1248         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1249                                            true))) {
1250                 put_page(new_page);
1251                 if (page) {
1252                         split_huge_pmd(vma, pmd, address);
1253                         put_page(page);
1254                 } else
1255                         split_huge_pmd(vma, pmd, address);
1256                 ret |= VM_FAULT_FALLBACK;
1257                 count_vm_event(THP_FAULT_FALLBACK);
1258                 goto out;
1259         }
1260
1261         count_vm_event(THP_FAULT_ALLOC);
1262
1263         if (!page)
1264                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1265         else
1266                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1267         __SetPageUptodate(new_page);
1268
1269         mmun_start = haddr;
1270         mmun_end   = haddr + HPAGE_PMD_SIZE;
1271         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1272
1273         spin_lock(ptl);
1274         if (page)
1275                 put_page(page);
1276         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1277                 spin_unlock(ptl);
1278                 mem_cgroup_cancel_charge(new_page, memcg, true);
1279                 put_page(new_page);
1280                 goto out_mn;
1281         } else {
1282                 pmd_t entry;
1283                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1284                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1285                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1286                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1287                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1288                 lru_cache_add_active_or_unevictable(new_page, vma);
1289                 set_pmd_at(mm, haddr, pmd, entry);
1290                 update_mmu_cache_pmd(vma, address, pmd);
1291                 if (!page) {
1292                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1293                         put_huge_zero_page();
1294                 } else {
1295                         VM_BUG_ON_PAGE(!PageHead(page), page);
1296                         page_remove_rmap(page, true);
1297                         put_page(page);
1298                 }
1299                 ret |= VM_FAULT_WRITE;
1300         }
1301         spin_unlock(ptl);
1302 out_mn:
1303         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1304 out:
1305         return ret;
1306 out_unlock:
1307         spin_unlock(ptl);
1308         return ret;
1309 }
1310
1311 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1312                                    unsigned long addr,
1313                                    pmd_t *pmd,
1314                                    unsigned int flags)
1315 {
1316         struct mm_struct *mm = vma->vm_mm;
1317         struct page *page = NULL;
1318
1319         assert_spin_locked(pmd_lockptr(mm, pmd));
1320
1321         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1322                 goto out;
1323
1324         /* Avoid dumping huge zero page */
1325         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1326                 return ERR_PTR(-EFAULT);
1327
1328         /* Full NUMA hinting faults to serialise migration in fault paths */
1329         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1330                 goto out;
1331
1332         page = pmd_page(*pmd);
1333         VM_BUG_ON_PAGE(!PageHead(page), page);
1334         if (flags & FOLL_TOUCH) {
1335                 pmd_t _pmd;
1336                 /*
1337                  * We should set the dirty bit only for FOLL_WRITE but
1338                  * for now the dirty bit in the pmd is meaningless.
1339                  * And if the dirty bit will become meaningful and
1340                  * we'll only set it with FOLL_WRITE, an atomic
1341                  * set_bit will be required on the pmd to set the
1342                  * young bit, instead of the current set_pmd_at.
1343                  */
1344                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1345                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1346                                           pmd, _pmd,  1))
1347                         update_mmu_cache_pmd(vma, addr, pmd);
1348         }
1349         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1350                 /*
1351                  * We don't mlock() pte-mapped THPs. This way we can avoid
1352                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1353                  *
1354                  * In most cases the pmd is the only mapping of the page as we
1355                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1356                  * writable private mappings in populate_vma_page_range().
1357                  *
1358                  * The only scenario when we have the page shared here is if we
1359                  * mlocking read-only mapping shared over fork(). We skip
1360                  * mlocking such pages.
1361                  */
1362                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1363                                 page->mapping && trylock_page(page)) {
1364                         lru_add_drain();
1365                         if (page->mapping)
1366                                 mlock_vma_page(page);
1367                         unlock_page(page);
1368                 }
1369         }
1370         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1371         VM_BUG_ON_PAGE(!PageCompound(page), page);
1372         if (flags & FOLL_GET)
1373                 get_page(page);
1374
1375 out:
1376         return page;
1377 }
1378
1379 /* NUMA hinting page fault entry point for trans huge pmds */
1380 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1381                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1382 {
1383         spinlock_t *ptl;
1384         struct anon_vma *anon_vma = NULL;
1385         struct page *page;
1386         unsigned long haddr = addr & HPAGE_PMD_MASK;
1387         int page_nid = -1, this_nid = numa_node_id();
1388         int target_nid, last_cpupid = -1;
1389         bool page_locked;
1390         bool migrated = false;
1391         bool was_writable;
1392         int flags = 0;
1393
1394         /* A PROT_NONE fault should not end up here */
1395         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1396
1397         ptl = pmd_lock(mm, pmdp);
1398         if (unlikely(!pmd_same(pmd, *pmdp)))
1399                 goto out_unlock;
1400
1401         /*
1402          * If there are potential migrations, wait for completion and retry
1403          * without disrupting NUMA hinting information. Do not relock and
1404          * check_same as the page may no longer be mapped.
1405          */
1406         if (unlikely(pmd_trans_migrating(*pmdp))) {
1407                 page = pmd_page(*pmdp);
1408                 spin_unlock(ptl);
1409                 wait_on_page_locked(page);
1410                 goto out;
1411         }
1412
1413         page = pmd_page(pmd);
1414         BUG_ON(is_huge_zero_page(page));
1415         page_nid = page_to_nid(page);
1416         last_cpupid = page_cpupid_last(page);
1417         count_vm_numa_event(NUMA_HINT_FAULTS);
1418         if (page_nid == this_nid) {
1419                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1420                 flags |= TNF_FAULT_LOCAL;
1421         }
1422
1423         /* See similar comment in do_numa_page for explanation */
1424         if (!(vma->vm_flags & VM_WRITE))
1425                 flags |= TNF_NO_GROUP;
1426
1427         /*
1428          * Acquire the page lock to serialise THP migrations but avoid dropping
1429          * page_table_lock if at all possible
1430          */
1431         page_locked = trylock_page(page);
1432         target_nid = mpol_misplaced(page, vma, haddr);
1433         if (target_nid == -1) {
1434                 /* If the page was locked, there are no parallel migrations */
1435                 if (page_locked)
1436                         goto clear_pmdnuma;
1437         }
1438
1439         /* Migration could have started since the pmd_trans_migrating check */
1440         if (!page_locked) {
1441                 spin_unlock(ptl);
1442                 wait_on_page_locked(page);
1443                 page_nid = -1;
1444                 goto out;
1445         }
1446
1447         /*
1448          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1449          * to serialises splits
1450          */
1451         get_page(page);
1452         spin_unlock(ptl);
1453         anon_vma = page_lock_anon_vma_read(page);
1454
1455         /* Confirm the PMD did not change while page_table_lock was released */
1456         spin_lock(ptl);
1457         if (unlikely(!pmd_same(pmd, *pmdp))) {
1458                 unlock_page(page);
1459                 put_page(page);
1460                 page_nid = -1;
1461                 goto out_unlock;
1462         }
1463
1464         /* Bail if we fail to protect against THP splits for any reason */
1465         if (unlikely(!anon_vma)) {
1466                 put_page(page);
1467                 page_nid = -1;
1468                 goto clear_pmdnuma;
1469         }
1470
1471         /*
1472          * Migrate the THP to the requested node, returns with page unlocked
1473          * and access rights restored.
1474          */
1475         spin_unlock(ptl);
1476         migrated = migrate_misplaced_transhuge_page(mm, vma,
1477                                 pmdp, pmd, addr, page, target_nid);
1478         if (migrated) {
1479                 flags |= TNF_MIGRATED;
1480                 page_nid = target_nid;
1481         } else
1482                 flags |= TNF_MIGRATE_FAIL;
1483
1484         goto out;
1485 clear_pmdnuma:
1486         BUG_ON(!PageLocked(page));
1487         was_writable = pmd_write(pmd);
1488         pmd = pmd_modify(pmd, vma->vm_page_prot);
1489         pmd = pmd_mkyoung(pmd);
1490         if (was_writable)
1491                 pmd = pmd_mkwrite(pmd);
1492         set_pmd_at(mm, haddr, pmdp, pmd);
1493         update_mmu_cache_pmd(vma, addr, pmdp);
1494         unlock_page(page);
1495 out_unlock:
1496         spin_unlock(ptl);
1497
1498 out:
1499         if (anon_vma)
1500                 page_unlock_anon_vma_read(anon_vma);
1501
1502         if (page_nid != -1)
1503                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1504
1505         return 0;
1506 }
1507
1508 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1509                 pmd_t *pmd, unsigned long addr, unsigned long next)
1510
1511 {
1512         spinlock_t *ptl;
1513         pmd_t orig_pmd;
1514         struct page *page;
1515         struct mm_struct *mm = tlb->mm;
1516         int ret = 0;
1517
1518         if (!pmd_trans_huge_lock(pmd, vma, &ptl))
1519                 goto out;
1520
1521         orig_pmd = *pmd;
1522         if (is_huge_zero_pmd(orig_pmd)) {
1523                 ret = 1;
1524                 goto out;
1525         }
1526
1527         page = pmd_page(orig_pmd);
1528         /*
1529          * If other processes are mapping this page, we couldn't discard
1530          * the page unless they all do MADV_FREE so let's skip the page.
1531          */
1532         if (page_mapcount(page) != 1)
1533                 goto out;
1534
1535         if (!trylock_page(page))
1536                 goto out;
1537
1538         /*
1539          * If user want to discard part-pages of THP, split it so MADV_FREE
1540          * will deactivate only them.
1541          */
1542         if (next - addr != HPAGE_PMD_SIZE) {
1543                 get_page(page);
1544                 spin_unlock(ptl);
1545                 if (split_huge_page(page)) {
1546                         put_page(page);
1547                         unlock_page(page);
1548                         goto out_unlocked;
1549                 }
1550                 put_page(page);
1551                 unlock_page(page);
1552                 ret = 1;
1553                 goto out_unlocked;
1554         }
1555
1556         if (PageDirty(page))
1557                 ClearPageDirty(page);
1558         unlock_page(page);
1559
1560         if (PageActive(page))
1561                 deactivate_page(page);
1562
1563         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1564                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1565                         tlb->fullmm);
1566                 orig_pmd = pmd_mkold(orig_pmd);
1567                 orig_pmd = pmd_mkclean(orig_pmd);
1568
1569                 set_pmd_at(mm, addr, pmd, orig_pmd);
1570                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1571         }
1572         ret = 1;
1573 out:
1574         spin_unlock(ptl);
1575 out_unlocked:
1576         return ret;
1577 }
1578
1579 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1580                  pmd_t *pmd, unsigned long addr)
1581 {
1582         pmd_t orig_pmd;
1583         spinlock_t *ptl;
1584
1585         if (!__pmd_trans_huge_lock(pmd, vma, &ptl))
1586                 return 0;
1587         /*
1588          * For architectures like ppc64 we look at deposited pgtable
1589          * when calling pmdp_huge_get_and_clear. So do the
1590          * pgtable_trans_huge_withdraw after finishing pmdp related
1591          * operations.
1592          */
1593         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1594                         tlb->fullmm);
1595         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1596         if (vma_is_dax(vma)) {
1597                 spin_unlock(ptl);
1598                 if (is_huge_zero_pmd(orig_pmd))
1599                         put_huge_zero_page();
1600         } else if (is_huge_zero_pmd(orig_pmd)) {
1601                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1602                 atomic_long_dec(&tlb->mm->nr_ptes);
1603                 spin_unlock(ptl);
1604                 put_huge_zero_page();
1605         } else {
1606                 struct page *page = pmd_page(orig_pmd);
1607                 page_remove_rmap(page, true);
1608                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1609                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1610                 VM_BUG_ON_PAGE(!PageHead(page), page);
1611                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1612                 atomic_long_dec(&tlb->mm->nr_ptes);
1613                 spin_unlock(ptl);
1614                 tlb_remove_page(tlb, page);
1615         }
1616         return 1;
1617 }
1618
1619 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1620                   unsigned long old_addr,
1621                   unsigned long new_addr, unsigned long old_end,
1622                   pmd_t *old_pmd, pmd_t *new_pmd)
1623 {
1624         spinlock_t *old_ptl, *new_ptl;
1625         pmd_t pmd;
1626
1627         struct mm_struct *mm = vma->vm_mm;
1628
1629         if ((old_addr & ~HPAGE_PMD_MASK) ||
1630             (new_addr & ~HPAGE_PMD_MASK) ||
1631             old_end - old_addr < HPAGE_PMD_SIZE ||
1632             (new_vma->vm_flags & VM_NOHUGEPAGE))
1633                 return false;
1634
1635         /*
1636          * The destination pmd shouldn't be established, free_pgtables()
1637          * should have release it.
1638          */
1639         if (WARN_ON(!pmd_none(*new_pmd))) {
1640                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1641                 return false;
1642         }
1643
1644         /*
1645          * We don't have to worry about the ordering of src and dst
1646          * ptlocks because exclusive mmap_sem prevents deadlock.
1647          */
1648         if (__pmd_trans_huge_lock(old_pmd, vma, &old_ptl)) {
1649                 new_ptl = pmd_lockptr(mm, new_pmd);
1650                 if (new_ptl != old_ptl)
1651                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1652                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1653                 VM_BUG_ON(!pmd_none(*new_pmd));
1654
1655                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1656                         pgtable_t pgtable;
1657                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1658                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1659                 }
1660                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1661                 if (new_ptl != old_ptl)
1662                         spin_unlock(new_ptl);
1663                 spin_unlock(old_ptl);
1664                 return true;
1665         }
1666         return false;
1667 }
1668
1669 /*
1670  * Returns
1671  *  - 0 if PMD could not be locked
1672  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1673  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1674  */
1675 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1676                 unsigned long addr, pgprot_t newprot, int prot_numa)
1677 {
1678         struct mm_struct *mm = vma->vm_mm;
1679         spinlock_t *ptl;
1680         int ret = 0;
1681
1682         if (__pmd_trans_huge_lock(pmd, vma, &ptl)) {
1683                 pmd_t entry;
1684                 bool preserve_write = prot_numa && pmd_write(*pmd);
1685                 ret = 1;
1686
1687                 /*
1688                  * Avoid trapping faults against the zero page. The read-only
1689                  * data is likely to be read-cached on the local CPU and
1690                  * local/remote hits to the zero page are not interesting.
1691                  */
1692                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1693                         spin_unlock(ptl);
1694                         return ret;
1695                 }
1696
1697                 if (!prot_numa || !pmd_protnone(*pmd)) {
1698                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1699                         entry = pmd_modify(entry, newprot);
1700                         if (preserve_write)
1701                                 entry = pmd_mkwrite(entry);
1702                         ret = HPAGE_PMD_NR;
1703                         set_pmd_at(mm, addr, pmd, entry);
1704                         BUG_ON(!preserve_write && pmd_write(entry));
1705                 }
1706                 spin_unlock(ptl);
1707         }
1708
1709         return ret;
1710 }
1711
1712 /*
1713  * Returns true if a given pmd maps a thp, false otherwise.
1714  *
1715  * Note that if it returns true, this routine returns without unlocking page
1716  * table lock. So callers must unlock it.
1717  */
1718 bool __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1719                 spinlock_t **ptl)
1720 {
1721         *ptl = pmd_lock(vma->vm_mm, pmd);
1722         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1723                 return true;
1724         spin_unlock(*ptl);
1725         return false;
1726 }
1727
1728 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1729
1730 int hugepage_madvise(struct vm_area_struct *vma,
1731                      unsigned long *vm_flags, int advice)
1732 {
1733         switch (advice) {
1734         case MADV_HUGEPAGE:
1735 #ifdef CONFIG_S390
1736                 /*
1737                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1738                  * can't handle this properly after s390_enable_sie, so we simply
1739                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1740                  */
1741                 if (mm_has_pgste(vma->vm_mm))
1742                         return 0;
1743 #endif
1744                 /*
1745                  * Be somewhat over-protective like KSM for now!
1746                  */
1747                 if (*vm_flags & VM_NO_THP)
1748                         return -EINVAL;
1749                 *vm_flags &= ~VM_NOHUGEPAGE;
1750                 *vm_flags |= VM_HUGEPAGE;
1751                 /*
1752                  * If the vma become good for khugepaged to scan,
1753                  * register it here without waiting a page fault that
1754                  * may not happen any time soon.
1755                  */
1756                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1757                         return -ENOMEM;
1758                 break;
1759         case MADV_NOHUGEPAGE:
1760                 /*
1761                  * Be somewhat over-protective like KSM for now!
1762                  */
1763                 if (*vm_flags & VM_NO_THP)
1764                         return -EINVAL;
1765                 *vm_flags &= ~VM_HUGEPAGE;
1766                 *vm_flags |= VM_NOHUGEPAGE;
1767                 /*
1768                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1769                  * this vma even if we leave the mm registered in khugepaged if
1770                  * it got registered before VM_NOHUGEPAGE was set.
1771                  */
1772                 break;
1773         }
1774
1775         return 0;
1776 }
1777
1778 static int __init khugepaged_slab_init(void)
1779 {
1780         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1781                                           sizeof(struct mm_slot),
1782                                           __alignof__(struct mm_slot), 0, NULL);
1783         if (!mm_slot_cache)
1784                 return -ENOMEM;
1785
1786         return 0;
1787 }
1788
1789 static void __init khugepaged_slab_exit(void)
1790 {
1791         kmem_cache_destroy(mm_slot_cache);
1792 }
1793
1794 static inline struct mm_slot *alloc_mm_slot(void)
1795 {
1796         if (!mm_slot_cache)     /* initialization failed */
1797                 return NULL;
1798         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1799 }
1800
1801 static inline void free_mm_slot(struct mm_slot *mm_slot)
1802 {
1803         kmem_cache_free(mm_slot_cache, mm_slot);
1804 }
1805
1806 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1807 {
1808         struct mm_slot *mm_slot;
1809
1810         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1811                 if (mm == mm_slot->mm)
1812                         return mm_slot;
1813
1814         return NULL;
1815 }
1816
1817 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1818                                     struct mm_slot *mm_slot)
1819 {
1820         mm_slot->mm = mm;
1821         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1822 }
1823
1824 static inline int khugepaged_test_exit(struct mm_struct *mm)
1825 {
1826         return atomic_read(&mm->mm_users) == 0;
1827 }
1828
1829 int __khugepaged_enter(struct mm_struct *mm)
1830 {
1831         struct mm_slot *mm_slot;
1832         int wakeup;
1833
1834         mm_slot = alloc_mm_slot();
1835         if (!mm_slot)
1836                 return -ENOMEM;
1837
1838         /* __khugepaged_exit() must not run from under us */
1839         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1840         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1841                 free_mm_slot(mm_slot);
1842                 return 0;
1843         }
1844
1845         spin_lock(&khugepaged_mm_lock);
1846         insert_to_mm_slots_hash(mm, mm_slot);
1847         /*
1848          * Insert just behind the scanning cursor, to let the area settle
1849          * down a little.
1850          */
1851         wakeup = list_empty(&khugepaged_scan.mm_head);
1852         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1853         spin_unlock(&khugepaged_mm_lock);
1854
1855         atomic_inc(&mm->mm_count);
1856         if (wakeup)
1857                 wake_up_interruptible(&khugepaged_wait);
1858
1859         return 0;
1860 }
1861
1862 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1863                                unsigned long vm_flags)
1864 {
1865         unsigned long hstart, hend;
1866         if (!vma->anon_vma)
1867                 /*
1868                  * Not yet faulted in so we will register later in the
1869                  * page fault if needed.
1870                  */
1871                 return 0;
1872         if (vma->vm_ops)
1873                 /* khugepaged not yet working on file or special mappings */
1874                 return 0;
1875         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1876         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1877         hend = vma->vm_end & HPAGE_PMD_MASK;
1878         if (hstart < hend)
1879                 return khugepaged_enter(vma, vm_flags);
1880         return 0;
1881 }
1882
1883 void __khugepaged_exit(struct mm_struct *mm)
1884 {
1885         struct mm_slot *mm_slot;
1886         int free = 0;
1887
1888         spin_lock(&khugepaged_mm_lock);
1889         mm_slot = get_mm_slot(mm);
1890         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1891                 hash_del(&mm_slot->hash);
1892                 list_del(&mm_slot->mm_node);
1893                 free = 1;
1894         }
1895         spin_unlock(&khugepaged_mm_lock);
1896
1897         if (free) {
1898                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1899                 free_mm_slot(mm_slot);
1900                 mmdrop(mm);
1901         } else if (mm_slot) {
1902                 /*
1903                  * This is required to serialize against
1904                  * khugepaged_test_exit() (which is guaranteed to run
1905                  * under mmap sem read mode). Stop here (after we
1906                  * return all pagetables will be destroyed) until
1907                  * khugepaged has finished working on the pagetables
1908                  * under the mmap_sem.
1909                  */
1910                 down_write(&mm->mmap_sem);
1911                 up_write(&mm->mmap_sem);
1912         }
1913 }
1914
1915 static void release_pte_page(struct page *page)
1916 {
1917         /* 0 stands for page_is_file_cache(page) == false */
1918         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1919         unlock_page(page);
1920         putback_lru_page(page);
1921 }
1922
1923 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1924 {
1925         while (--_pte >= pte) {
1926                 pte_t pteval = *_pte;
1927                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1928                         release_pte_page(pte_page(pteval));
1929         }
1930 }
1931
1932 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1933                                         unsigned long address,
1934                                         pte_t *pte)
1935 {
1936         struct page *page = NULL;
1937         pte_t *_pte;
1938         int none_or_zero = 0, result = 0;
1939         bool referenced = false, writable = false;
1940
1941         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1942              _pte++, address += PAGE_SIZE) {
1943                 pte_t pteval = *_pte;
1944                 if (pte_none(pteval) || (pte_present(pteval) &&
1945                                 is_zero_pfn(pte_pfn(pteval)))) {
1946                         if (!userfaultfd_armed(vma) &&
1947                             ++none_or_zero <= khugepaged_max_ptes_none) {
1948                                 continue;
1949                         } else {
1950                                 result = SCAN_EXCEED_NONE_PTE;
1951                                 goto out;
1952                         }
1953                 }
1954                 if (!pte_present(pteval)) {
1955                         result = SCAN_PTE_NON_PRESENT;
1956                         goto out;
1957                 }
1958                 page = vm_normal_page(vma, address, pteval);
1959                 if (unlikely(!page)) {
1960                         result = SCAN_PAGE_NULL;
1961                         goto out;
1962                 }
1963
1964                 VM_BUG_ON_PAGE(PageCompound(page), page);
1965                 VM_BUG_ON_PAGE(!PageAnon(page), page);
1966                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1967
1968                 /*
1969                  * We can do it before isolate_lru_page because the
1970                  * page can't be freed from under us. NOTE: PG_lock
1971                  * is needed to serialize against split_huge_page
1972                  * when invoked from the VM.
1973                  */
1974                 if (!trylock_page(page)) {
1975                         result = SCAN_PAGE_LOCK;
1976                         goto out;
1977                 }
1978
1979                 /*
1980                  * cannot use mapcount: can't collapse if there's a gup pin.
1981                  * The page must only be referenced by the scanned process
1982                  * and page swap cache.
1983                  */
1984                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1985                         unlock_page(page);
1986                         result = SCAN_PAGE_COUNT;
1987                         goto out;
1988                 }
1989                 if (pte_write(pteval)) {
1990                         writable = true;
1991                 } else {
1992                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
1993                                 unlock_page(page);
1994                                 result = SCAN_SWAP_CACHE_PAGE;
1995                                 goto out;
1996                         }
1997                         /*
1998                          * Page is not in the swap cache. It can be collapsed
1999                          * into a THP.
2000                          */
2001                 }
2002
2003                 /*
2004                  * Isolate the page to avoid collapsing an hugepage
2005                  * currently in use by the VM.
2006                  */
2007                 if (isolate_lru_page(page)) {
2008                         unlock_page(page);
2009                         result = SCAN_DEL_PAGE_LRU;
2010                         goto out;
2011                 }
2012                 /* 0 stands for page_is_file_cache(page) == false */
2013                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2014                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2015                 VM_BUG_ON_PAGE(PageLRU(page), page);
2016
2017                 /* If there is no mapped pte young don't collapse the page */
2018                 if (pte_young(pteval) ||
2019                     page_is_young(page) || PageReferenced(page) ||
2020                     mmu_notifier_test_young(vma->vm_mm, address))
2021                         referenced = true;
2022         }
2023         if (likely(writable)) {
2024                 if (likely(referenced)) {
2025                         result = SCAN_SUCCEED;
2026                         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
2027                                                             referenced, writable, result);
2028                         return 1;
2029                 }
2030         } else {
2031                 result = SCAN_PAGE_RO;
2032         }
2033
2034 out:
2035         release_pte_pages(pte, _pte);
2036         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
2037                                             referenced, writable, result);
2038         return 0;
2039 }
2040
2041 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2042                                       struct vm_area_struct *vma,
2043                                       unsigned long address,
2044                                       spinlock_t *ptl)
2045 {
2046         pte_t *_pte;
2047         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2048                 pte_t pteval = *_pte;
2049                 struct page *src_page;
2050
2051                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2052                         clear_user_highpage(page, address);
2053                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2054                         if (is_zero_pfn(pte_pfn(pteval))) {
2055                                 /*
2056                                  * ptl mostly unnecessary.
2057                                  */
2058                                 spin_lock(ptl);
2059                                 /*
2060                                  * paravirt calls inside pte_clear here are
2061                                  * superfluous.
2062                                  */
2063                                 pte_clear(vma->vm_mm, address, _pte);
2064                                 spin_unlock(ptl);
2065                         }
2066                 } else {
2067                         src_page = pte_page(pteval);
2068                         copy_user_highpage(page, src_page, address, vma);
2069                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2070                         release_pte_page(src_page);
2071                         /*
2072                          * ptl mostly unnecessary, but preempt has to
2073                          * be disabled to update the per-cpu stats
2074                          * inside page_remove_rmap().
2075                          */
2076                         spin_lock(ptl);
2077                         /*
2078                          * paravirt calls inside pte_clear here are
2079                          * superfluous.
2080                          */
2081                         pte_clear(vma->vm_mm, address, _pte);
2082                         page_remove_rmap(src_page, false);
2083                         spin_unlock(ptl);
2084                         free_page_and_swap_cache(src_page);
2085                 }
2086
2087                 address += PAGE_SIZE;
2088                 page++;
2089         }
2090 }
2091
2092 static void khugepaged_alloc_sleep(void)
2093 {
2094         DEFINE_WAIT(wait);
2095
2096         add_wait_queue(&khugepaged_wait, &wait);
2097         freezable_schedule_timeout_interruptible(
2098                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2099         remove_wait_queue(&khugepaged_wait, &wait);
2100 }
2101
2102 static int khugepaged_node_load[MAX_NUMNODES];
2103
2104 static bool khugepaged_scan_abort(int nid)
2105 {
2106         int i;
2107
2108         /*
2109          * If zone_reclaim_mode is disabled, then no extra effort is made to
2110          * allocate memory locally.
2111          */
2112         if (!zone_reclaim_mode)
2113                 return false;
2114
2115         /* If there is a count for this node already, it must be acceptable */
2116         if (khugepaged_node_load[nid])
2117                 return false;
2118
2119         for (i = 0; i < MAX_NUMNODES; i++) {
2120                 if (!khugepaged_node_load[i])
2121                         continue;
2122                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2123                         return true;
2124         }
2125         return false;
2126 }
2127
2128 #ifdef CONFIG_NUMA
2129 static int khugepaged_find_target_node(void)
2130 {
2131         static int last_khugepaged_target_node = NUMA_NO_NODE;
2132         int nid, target_node = 0, max_value = 0;
2133
2134         /* find first node with max normal pages hit */
2135         for (nid = 0; nid < MAX_NUMNODES; nid++)
2136                 if (khugepaged_node_load[nid] > max_value) {
2137                         max_value = khugepaged_node_load[nid];
2138                         target_node = nid;
2139                 }
2140
2141         /* do some balance if several nodes have the same hit record */
2142         if (target_node <= last_khugepaged_target_node)
2143                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2144                                 nid++)
2145                         if (max_value == khugepaged_node_load[nid]) {
2146                                 target_node = nid;
2147                                 break;
2148                         }
2149
2150         last_khugepaged_target_node = target_node;
2151         return target_node;
2152 }
2153
2154 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2155 {
2156         if (IS_ERR(*hpage)) {
2157                 if (!*wait)
2158                         return false;
2159
2160                 *wait = false;
2161                 *hpage = NULL;
2162                 khugepaged_alloc_sleep();
2163         } else if (*hpage) {
2164                 put_page(*hpage);
2165                 *hpage = NULL;
2166         }
2167
2168         return true;
2169 }
2170
2171 static struct page *
2172 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2173                        unsigned long address, int node)
2174 {
2175         VM_BUG_ON_PAGE(*hpage, *hpage);
2176
2177         /*
2178          * Before allocating the hugepage, release the mmap_sem read lock.
2179          * The allocation can take potentially a long time if it involves
2180          * sync compaction, and we do not need to hold the mmap_sem during
2181          * that. We will recheck the vma after taking it again in write mode.
2182          */
2183         up_read(&mm->mmap_sem);
2184
2185         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2186         if (unlikely(!*hpage)) {
2187                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2188                 *hpage = ERR_PTR(-ENOMEM);
2189                 return NULL;
2190         }
2191
2192         prep_transhuge_page(*hpage);
2193         count_vm_event(THP_COLLAPSE_ALLOC);
2194         return *hpage;
2195 }
2196 #else
2197 static int khugepaged_find_target_node(void)
2198 {
2199         return 0;
2200 }
2201
2202 static inline struct page *alloc_hugepage(int defrag)
2203 {
2204         struct page *page;
2205
2206         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2207         if (page)
2208                 prep_transhuge_page(page);
2209         return page;
2210 }
2211
2212 static struct page *khugepaged_alloc_hugepage(bool *wait)
2213 {
2214         struct page *hpage;
2215
2216         do {
2217                 hpage = alloc_hugepage(khugepaged_defrag());
2218                 if (!hpage) {
2219                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2220                         if (!*wait)
2221                                 return NULL;
2222
2223                         *wait = false;
2224                         khugepaged_alloc_sleep();
2225                 } else
2226                         count_vm_event(THP_COLLAPSE_ALLOC);
2227         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2228
2229         return hpage;
2230 }
2231
2232 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2233 {
2234         if (!*hpage)
2235                 *hpage = khugepaged_alloc_hugepage(wait);
2236
2237         if (unlikely(!*hpage))
2238                 return false;
2239
2240         return true;
2241 }
2242
2243 static struct page *
2244 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2245                        unsigned long address, int node)
2246 {
2247         up_read(&mm->mmap_sem);
2248         VM_BUG_ON(!*hpage);
2249
2250         return  *hpage;
2251 }
2252 #endif
2253
2254 static bool hugepage_vma_check(struct vm_area_struct *vma)
2255 {
2256         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2257             (vma->vm_flags & VM_NOHUGEPAGE))
2258                 return false;
2259         if (!vma->anon_vma || vma->vm_ops)
2260                 return false;
2261         if (is_vma_temporary_stack(vma))
2262                 return false;
2263         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2264         return true;
2265 }
2266
2267 static void collapse_huge_page(struct mm_struct *mm,
2268                                    unsigned long address,
2269                                    struct page **hpage,
2270                                    struct vm_area_struct *vma,
2271                                    int node)
2272 {
2273         pmd_t *pmd, _pmd;
2274         pte_t *pte;
2275         pgtable_t pgtable;
2276         struct page *new_page;
2277         spinlock_t *pmd_ptl, *pte_ptl;
2278         int isolated, result = 0;
2279         unsigned long hstart, hend;
2280         struct mem_cgroup *memcg;
2281         unsigned long mmun_start;       /* For mmu_notifiers */
2282         unsigned long mmun_end;         /* For mmu_notifiers */
2283         gfp_t gfp;
2284
2285         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2286
2287         /* Only allocate from the target node */
2288         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2289                 __GFP_THISNODE;
2290
2291         /* release the mmap_sem read lock. */
2292         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2293         if (!new_page) {
2294                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2295                 goto out_nolock;
2296         }
2297
2298         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2299                 result = SCAN_CGROUP_CHARGE_FAIL;
2300                 goto out_nolock;
2301         }
2302
2303         /*
2304          * Prevent all access to pagetables with the exception of
2305          * gup_fast later hanlded by the ptep_clear_flush and the VM
2306          * handled by the anon_vma lock + PG_lock.
2307          */
2308         down_write(&mm->mmap_sem);
2309         if (unlikely(khugepaged_test_exit(mm))) {
2310                 result = SCAN_ANY_PROCESS;
2311                 goto out;
2312         }
2313
2314         vma = find_vma(mm, address);
2315         if (!vma) {
2316                 result = SCAN_VMA_NULL;
2317                 goto out;
2318         }
2319         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2320         hend = vma->vm_end & HPAGE_PMD_MASK;
2321         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2322                 result = SCAN_ADDRESS_RANGE;
2323                 goto out;
2324         }
2325         if (!hugepage_vma_check(vma)) {
2326                 result = SCAN_VMA_CHECK;
2327                 goto out;
2328         }
2329         pmd = mm_find_pmd(mm, address);
2330         if (!pmd) {
2331                 result = SCAN_PMD_NULL;
2332                 goto out;
2333         }
2334
2335         anon_vma_lock_write(vma->anon_vma);
2336
2337         pte = pte_offset_map(pmd, address);
2338         pte_ptl = pte_lockptr(mm, pmd);
2339
2340         mmun_start = address;
2341         mmun_end   = address + HPAGE_PMD_SIZE;
2342         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2343         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2344         /*
2345          * After this gup_fast can't run anymore. This also removes
2346          * any huge TLB entry from the CPU so we won't allow
2347          * huge and small TLB entries for the same virtual address
2348          * to avoid the risk of CPU bugs in that area.
2349          */
2350         _pmd = pmdp_collapse_flush(vma, address, pmd);
2351         spin_unlock(pmd_ptl);
2352         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2353
2354         spin_lock(pte_ptl);
2355         isolated = __collapse_huge_page_isolate(vma, address, pte);
2356         spin_unlock(pte_ptl);
2357
2358         if (unlikely(!isolated)) {
2359                 pte_unmap(pte);
2360                 spin_lock(pmd_ptl);
2361                 BUG_ON(!pmd_none(*pmd));
2362                 /*
2363                  * We can only use set_pmd_at when establishing
2364                  * hugepmds and never for establishing regular pmds that
2365                  * points to regular pagetables. Use pmd_populate for that
2366                  */
2367                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2368                 spin_unlock(pmd_ptl);
2369                 anon_vma_unlock_write(vma->anon_vma);
2370                 result = SCAN_FAIL;
2371                 goto out;
2372         }
2373
2374         /*
2375          * All pages are isolated and locked so anon_vma rmap
2376          * can't run anymore.
2377          */
2378         anon_vma_unlock_write(vma->anon_vma);
2379
2380         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2381         pte_unmap(pte);
2382         __SetPageUptodate(new_page);
2383         pgtable = pmd_pgtable(_pmd);
2384
2385         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2386         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2387
2388         /*
2389          * spin_lock() below is not the equivalent of smp_wmb(), so
2390          * this is needed to avoid the copy_huge_page writes to become
2391          * visible after the set_pmd_at() write.
2392          */
2393         smp_wmb();
2394
2395         spin_lock(pmd_ptl);
2396         BUG_ON(!pmd_none(*pmd));
2397         page_add_new_anon_rmap(new_page, vma, address, true);
2398         mem_cgroup_commit_charge(new_page, memcg, false, true);
2399         lru_cache_add_active_or_unevictable(new_page, vma);
2400         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2401         set_pmd_at(mm, address, pmd, _pmd);
2402         update_mmu_cache_pmd(vma, address, pmd);
2403         spin_unlock(pmd_ptl);
2404
2405         *hpage = NULL;
2406
2407         khugepaged_pages_collapsed++;
2408         result = SCAN_SUCCEED;
2409 out_up_write:
2410         up_write(&mm->mmap_sem);
2411         trace_mm_collapse_huge_page(mm, isolated, result);
2412         return;
2413
2414 out_nolock:
2415         trace_mm_collapse_huge_page(mm, isolated, result);
2416         return;
2417 out:
2418         mem_cgroup_cancel_charge(new_page, memcg, true);
2419         goto out_up_write;
2420 }
2421
2422 static int khugepaged_scan_pmd(struct mm_struct *mm,
2423                                struct vm_area_struct *vma,
2424                                unsigned long address,
2425                                struct page **hpage)
2426 {
2427         pmd_t *pmd;
2428         pte_t *pte, *_pte;
2429         int ret = 0, none_or_zero = 0, result = 0;
2430         struct page *page = NULL;
2431         unsigned long _address;
2432         spinlock_t *ptl;
2433         int node = NUMA_NO_NODE;
2434         bool writable = false, referenced = false;
2435
2436         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2437
2438         pmd = mm_find_pmd(mm, address);
2439         if (!pmd) {
2440                 result = SCAN_PMD_NULL;
2441                 goto out;
2442         }
2443
2444         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2445         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2446         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2447              _pte++, _address += PAGE_SIZE) {
2448                 pte_t pteval = *_pte;
2449                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2450                         if (!userfaultfd_armed(vma) &&
2451                             ++none_or_zero <= khugepaged_max_ptes_none) {
2452                                 continue;
2453                         } else {
2454                                 result = SCAN_EXCEED_NONE_PTE;
2455                                 goto out_unmap;
2456                         }
2457                 }
2458                 if (!pte_present(pteval)) {
2459                         result = SCAN_PTE_NON_PRESENT;
2460                         goto out_unmap;
2461                 }
2462                 if (pte_write(pteval))
2463                         writable = true;
2464
2465                 page = vm_normal_page(vma, _address, pteval);
2466                 if (unlikely(!page)) {
2467                         result = SCAN_PAGE_NULL;
2468                         goto out_unmap;
2469                 }
2470
2471                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2472                 if (PageCompound(page)) {
2473                         result = SCAN_PAGE_COMPOUND;
2474                         goto out_unmap;
2475                 }
2476
2477                 /*
2478                  * Record which node the original page is from and save this
2479                  * information to khugepaged_node_load[].
2480                  * Khupaged will allocate hugepage from the node has the max
2481                  * hit record.
2482                  */
2483                 node = page_to_nid(page);
2484                 if (khugepaged_scan_abort(node)) {
2485                         result = SCAN_SCAN_ABORT;
2486                         goto out_unmap;
2487                 }
2488                 khugepaged_node_load[node]++;
2489                 if (!PageLRU(page)) {
2490                         result = SCAN_SCAN_ABORT;
2491                         goto out_unmap;
2492                 }
2493                 if (PageLocked(page)) {
2494                         result = SCAN_PAGE_LOCK;
2495                         goto out_unmap;
2496                 }
2497                 if (!PageAnon(page)) {
2498                         result = SCAN_PAGE_ANON;
2499                         goto out_unmap;
2500                 }
2501
2502                 /*
2503                  * cannot use mapcount: can't collapse if there's a gup pin.
2504                  * The page must only be referenced by the scanned process
2505                  * and page swap cache.
2506                  */
2507                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2508                         result = SCAN_PAGE_COUNT;
2509                         goto out_unmap;
2510                 }
2511                 if (pte_young(pteval) ||
2512                     page_is_young(page) || PageReferenced(page) ||
2513                     mmu_notifier_test_young(vma->vm_mm, address))
2514                         referenced = true;
2515         }
2516         if (writable) {
2517                 if (referenced) {
2518                         result = SCAN_SUCCEED;
2519                         ret = 1;
2520                 } else {
2521                         result = SCAN_NO_REFERENCED_PAGE;
2522                 }
2523         } else {
2524                 result = SCAN_PAGE_RO;
2525         }
2526 out_unmap:
2527         pte_unmap_unlock(pte, ptl);
2528         if (ret) {
2529                 node = khugepaged_find_target_node();
2530                 /* collapse_huge_page will return with the mmap_sem released */
2531                 collapse_huge_page(mm, address, hpage, vma, node);
2532         }
2533 out:
2534         trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced,
2535                                      none_or_zero, result);
2536         return ret;
2537 }
2538
2539 static void collect_mm_slot(struct mm_slot *mm_slot)
2540 {
2541         struct mm_struct *mm = mm_slot->mm;
2542
2543         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2544
2545         if (khugepaged_test_exit(mm)) {
2546                 /* free mm_slot */
2547                 hash_del(&mm_slot->hash);
2548                 list_del(&mm_slot->mm_node);
2549
2550                 /*
2551                  * Not strictly needed because the mm exited already.
2552                  *
2553                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2554                  */
2555
2556                 /* khugepaged_mm_lock actually not necessary for the below */
2557                 free_mm_slot(mm_slot);
2558                 mmdrop(mm);
2559         }
2560 }
2561
2562 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2563                                             struct page **hpage)
2564         __releases(&khugepaged_mm_lock)
2565         __acquires(&khugepaged_mm_lock)
2566 {
2567         struct mm_slot *mm_slot;
2568         struct mm_struct *mm;
2569         struct vm_area_struct *vma;
2570         int progress = 0;
2571
2572         VM_BUG_ON(!pages);
2573         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2574
2575         if (khugepaged_scan.mm_slot)
2576                 mm_slot = khugepaged_scan.mm_slot;
2577         else {
2578                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2579                                      struct mm_slot, mm_node);
2580                 khugepaged_scan.address = 0;
2581                 khugepaged_scan.mm_slot = mm_slot;
2582         }
2583         spin_unlock(&khugepaged_mm_lock);
2584
2585         mm = mm_slot->mm;
2586         down_read(&mm->mmap_sem);
2587         if (unlikely(khugepaged_test_exit(mm)))
2588                 vma = NULL;
2589         else
2590                 vma = find_vma(mm, khugepaged_scan.address);
2591
2592         progress++;
2593         for (; vma; vma = vma->vm_next) {
2594                 unsigned long hstart, hend;
2595
2596                 cond_resched();
2597                 if (unlikely(khugepaged_test_exit(mm))) {
2598                         progress++;
2599                         break;
2600                 }
2601                 if (!hugepage_vma_check(vma)) {
2602 skip:
2603                         progress++;
2604                         continue;
2605                 }
2606                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2607                 hend = vma->vm_end & HPAGE_PMD_MASK;
2608                 if (hstart >= hend)
2609                         goto skip;
2610                 if (khugepaged_scan.address > hend)
2611                         goto skip;
2612                 if (khugepaged_scan.address < hstart)
2613                         khugepaged_scan.address = hstart;
2614                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2615
2616                 while (khugepaged_scan.address < hend) {
2617                         int ret;
2618                         cond_resched();
2619                         if (unlikely(khugepaged_test_exit(mm)))
2620                                 goto breakouterloop;
2621
2622                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2623                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2624                                   hend);
2625                         ret = khugepaged_scan_pmd(mm, vma,
2626                                                   khugepaged_scan.address,
2627                                                   hpage);
2628                         /* move to next address */
2629                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2630                         progress += HPAGE_PMD_NR;
2631                         if (ret)
2632                                 /* we released mmap_sem so break loop */
2633                                 goto breakouterloop_mmap_sem;
2634                         if (progress >= pages)
2635                                 goto breakouterloop;
2636                 }
2637         }
2638 breakouterloop:
2639         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2640 breakouterloop_mmap_sem:
2641
2642         spin_lock(&khugepaged_mm_lock);
2643         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2644         /*
2645          * Release the current mm_slot if this mm is about to die, or
2646          * if we scanned all vmas of this mm.
2647          */
2648         if (khugepaged_test_exit(mm) || !vma) {
2649                 /*
2650                  * Make sure that if mm_users is reaching zero while
2651                  * khugepaged runs here, khugepaged_exit will find
2652                  * mm_slot not pointing to the exiting mm.
2653                  */
2654                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2655                         khugepaged_scan.mm_slot = list_entry(
2656                                 mm_slot->mm_node.next,
2657                                 struct mm_slot, mm_node);
2658                         khugepaged_scan.address = 0;
2659                 } else {
2660                         khugepaged_scan.mm_slot = NULL;
2661                         khugepaged_full_scans++;
2662                 }
2663
2664                 collect_mm_slot(mm_slot);
2665         }
2666
2667         return progress;
2668 }
2669
2670 static int khugepaged_has_work(void)
2671 {
2672         return !list_empty(&khugepaged_scan.mm_head) &&
2673                 khugepaged_enabled();
2674 }
2675
2676 static int khugepaged_wait_event(void)
2677 {
2678         return !list_empty(&khugepaged_scan.mm_head) ||
2679                 kthread_should_stop();
2680 }
2681
2682 static void khugepaged_do_scan(void)
2683 {
2684         struct page *hpage = NULL;
2685         unsigned int progress = 0, pass_through_head = 0;
2686         unsigned int pages = khugepaged_pages_to_scan;
2687         bool wait = true;
2688
2689         barrier(); /* write khugepaged_pages_to_scan to local stack */
2690
2691         while (progress < pages) {
2692                 if (!khugepaged_prealloc_page(&hpage, &wait))
2693                         break;
2694
2695                 cond_resched();
2696
2697                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2698                         break;
2699
2700                 spin_lock(&khugepaged_mm_lock);
2701                 if (!khugepaged_scan.mm_slot)
2702                         pass_through_head++;
2703                 if (khugepaged_has_work() &&
2704                     pass_through_head < 2)
2705                         progress += khugepaged_scan_mm_slot(pages - progress,
2706                                                             &hpage);
2707                 else
2708                         progress = pages;
2709                 spin_unlock(&khugepaged_mm_lock);
2710         }
2711
2712         if (!IS_ERR_OR_NULL(hpage))
2713                 put_page(hpage);
2714 }
2715
2716 static void khugepaged_wait_work(void)
2717 {
2718         if (khugepaged_has_work()) {
2719                 if (!khugepaged_scan_sleep_millisecs)
2720                         return;
2721
2722                 wait_event_freezable_timeout(khugepaged_wait,
2723                                              kthread_should_stop(),
2724                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2725                 return;
2726         }
2727
2728         if (khugepaged_enabled())
2729                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2730 }
2731
2732 static int khugepaged(void *none)
2733 {
2734         struct mm_slot *mm_slot;
2735
2736         set_freezable();
2737         set_user_nice(current, MAX_NICE);
2738
2739         while (!kthread_should_stop()) {
2740                 khugepaged_do_scan();
2741                 khugepaged_wait_work();
2742         }
2743
2744         spin_lock(&khugepaged_mm_lock);
2745         mm_slot = khugepaged_scan.mm_slot;
2746         khugepaged_scan.mm_slot = NULL;
2747         if (mm_slot)
2748                 collect_mm_slot(mm_slot);
2749         spin_unlock(&khugepaged_mm_lock);
2750         return 0;
2751 }
2752
2753 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2754                 unsigned long haddr, pmd_t *pmd)
2755 {
2756         struct mm_struct *mm = vma->vm_mm;
2757         pgtable_t pgtable;
2758         pmd_t _pmd;
2759         int i;
2760
2761         /* leave pmd empty until pte is filled */
2762         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2763
2764         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2765         pmd_populate(mm, &_pmd, pgtable);
2766
2767         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2768                 pte_t *pte, entry;
2769                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2770                 entry = pte_mkspecial(entry);
2771                 pte = pte_offset_map(&_pmd, haddr);
2772                 VM_BUG_ON(!pte_none(*pte));
2773                 set_pte_at(mm, haddr, pte, entry);
2774                 pte_unmap(pte);
2775         }
2776         smp_wmb(); /* make pte visible before pmd */
2777         pmd_populate(mm, pmd, pgtable);
2778         put_huge_zero_page();
2779 }
2780
2781 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2782                 unsigned long haddr, bool freeze)
2783 {
2784         struct mm_struct *mm = vma->vm_mm;
2785         struct page *page;
2786         pgtable_t pgtable;
2787         pmd_t _pmd;
2788         bool young, write, dirty;
2789         int i;
2790
2791         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2792         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2793         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2794         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2795
2796         count_vm_event(THP_SPLIT_PMD);
2797
2798         if (vma_is_dax(vma)) {
2799                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2800                 if (is_huge_zero_pmd(_pmd))
2801                         put_huge_zero_page();
2802                 return;
2803         } else if (is_huge_zero_pmd(*pmd)) {
2804                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2805         }
2806
2807         page = pmd_page(*pmd);
2808         VM_BUG_ON_PAGE(!page_count(page), page);
2809         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2810         write = pmd_write(*pmd);
2811         young = pmd_young(*pmd);
2812         dirty = pmd_dirty(*pmd);
2813
2814         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2815         pmd_populate(mm, &_pmd, pgtable);
2816
2817         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2818                 pte_t entry, *pte;
2819                 /*
2820                  * Note that NUMA hinting access restrictions are not
2821                  * transferred to avoid any possibility of altering
2822                  * permissions across VMAs.
2823                  */
2824                 if (freeze) {
2825                         swp_entry_t swp_entry;
2826                         swp_entry = make_migration_entry(page + i, write);
2827                         entry = swp_entry_to_pte(swp_entry);
2828                 } else {
2829                         entry = mk_pte(page + i, vma->vm_page_prot);
2830                         entry = maybe_mkwrite(entry, vma);
2831                         if (!write)
2832                                 entry = pte_wrprotect(entry);
2833                         if (!young)
2834                                 entry = pte_mkold(entry);
2835                 }
2836                 if (dirty)
2837                         SetPageDirty(page + i);
2838                 pte = pte_offset_map(&_pmd, haddr);
2839                 BUG_ON(!pte_none(*pte));
2840                 set_pte_at(mm, haddr, pte, entry);
2841                 atomic_inc(&page[i]._mapcount);
2842                 pte_unmap(pte);
2843         }
2844
2845         /*
2846          * Set PG_double_map before dropping compound_mapcount to avoid
2847          * false-negative page_mapped().
2848          */
2849         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2850                 for (i = 0; i < HPAGE_PMD_NR; i++)
2851                         atomic_inc(&page[i]._mapcount);
2852         }
2853
2854         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2855                 /* Last compound_mapcount is gone. */
2856                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2857                 if (TestClearPageDoubleMap(page)) {
2858                         /* No need in mapcount reference anymore */
2859                         for (i = 0; i < HPAGE_PMD_NR; i++)
2860                                 atomic_dec(&page[i]._mapcount);
2861                 }
2862         }
2863
2864         smp_wmb(); /* make pte visible before pmd */
2865         /*
2866          * Up to this point the pmd is present and huge and userland has the
2867          * whole access to the hugepage during the split (which happens in
2868          * place). If we overwrite the pmd with the not-huge version pointing
2869          * to the pte here (which of course we could if all CPUs were bug
2870          * free), userland could trigger a small page size TLB miss on the
2871          * small sized TLB while the hugepage TLB entry is still established in
2872          * the huge TLB. Some CPU doesn't like that.
2873          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2874          * 383 on page 93. Intel should be safe but is also warns that it's
2875          * only safe if the permission and cache attributes of the two entries
2876          * loaded in the two TLB is identical (which should be the case here).
2877          * But it is generally safer to never allow small and huge TLB entries
2878          * for the same virtual address to be loaded simultaneously. So instead
2879          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2880          * current pmd notpresent (atomically because here the pmd_trans_huge
2881          * and pmd_trans_splitting must remain set at all times on the pmd
2882          * until the split is complete for this pmd), then we flush the SMP TLB
2883          * and finally we write the non-huge version of the pmd entry with
2884          * pmd_populate.
2885          */
2886         pmdp_invalidate(vma, haddr, pmd);
2887         pmd_populate(mm, pmd, pgtable);
2888
2889         if (freeze) {
2890                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2891                         page_remove_rmap(page + i, false);
2892                         put_page(page + i);
2893                 }
2894         }
2895 }
2896
2897 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2898                 unsigned long address)
2899 {
2900         spinlock_t *ptl;
2901         struct mm_struct *mm = vma->vm_mm;
2902         struct page *page = NULL;
2903         unsigned long haddr = address & HPAGE_PMD_MASK;
2904
2905         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2906         ptl = pmd_lock(mm, pmd);
2907         if (pmd_trans_huge(*pmd)) {
2908                 page = pmd_page(*pmd);
2909                 if (PageMlocked(page))
2910                         get_page(page);
2911                 else
2912                         page = NULL;
2913         } else if (!pmd_devmap(*pmd))
2914                 goto out;
2915         __split_huge_pmd_locked(vma, pmd, haddr, false);
2916 out:
2917         spin_unlock(ptl);
2918         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2919         if (page) {
2920                 lock_page(page);
2921                 munlock_vma_page(page);
2922                 unlock_page(page);
2923                 put_page(page);
2924         }
2925 }
2926
2927 static void split_huge_pmd_address(struct vm_area_struct *vma,
2928                                     unsigned long address)
2929 {
2930         pgd_t *pgd;
2931         pud_t *pud;
2932         pmd_t *pmd;
2933
2934         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2935
2936         pgd = pgd_offset(vma->vm_mm, address);
2937         if (!pgd_present(*pgd))
2938                 return;
2939
2940         pud = pud_offset(pgd, address);
2941         if (!pud_present(*pud))
2942                 return;
2943
2944         pmd = pmd_offset(pud, address);
2945         if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
2946                 return;
2947         /*
2948          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2949          * materialize from under us.
2950          */
2951         split_huge_pmd(vma, pmd, address);
2952 }
2953
2954 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2955                              unsigned long start,
2956                              unsigned long end,
2957                              long adjust_next)
2958 {
2959         /*
2960          * If the new start address isn't hpage aligned and it could
2961          * previously contain an hugepage: check if we need to split
2962          * an huge pmd.
2963          */
2964         if (start & ~HPAGE_PMD_MASK &&
2965             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2966             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2967                 split_huge_pmd_address(vma, start);
2968
2969         /*
2970          * If the new end address isn't hpage aligned and it could
2971          * previously contain an hugepage: check if we need to split
2972          * an huge pmd.
2973          */
2974         if (end & ~HPAGE_PMD_MASK &&
2975             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2976             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2977                 split_huge_pmd_address(vma, end);
2978
2979         /*
2980          * If we're also updating the vma->vm_next->vm_start, if the new
2981          * vm_next->vm_start isn't page aligned and it could previously
2982          * contain an hugepage: check if we need to split an huge pmd.
2983          */
2984         if (adjust_next > 0) {
2985                 struct vm_area_struct *next = vma->vm_next;
2986                 unsigned long nstart = next->vm_start;
2987                 nstart += adjust_next << PAGE_SHIFT;
2988                 if (nstart & ~HPAGE_PMD_MASK &&
2989                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2990                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2991                         split_huge_pmd_address(next, nstart);
2992         }
2993 }
2994
2995 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
2996                 unsigned long address)
2997 {
2998         unsigned long haddr = address & HPAGE_PMD_MASK;
2999         spinlock_t *ptl;
3000         pgd_t *pgd;
3001         pud_t *pud;
3002         pmd_t *pmd;
3003         pte_t *pte;
3004         int i, nr = HPAGE_PMD_NR;
3005
3006         /* Skip pages which doesn't belong to the VMA */
3007         if (address < vma->vm_start) {
3008                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3009                 page += off;
3010                 nr -= off;
3011                 address = vma->vm_start;
3012         }
3013
3014         pgd = pgd_offset(vma->vm_mm, address);
3015         if (!pgd_present(*pgd))
3016                 return;
3017         pud = pud_offset(pgd, address);
3018         if (!pud_present(*pud))
3019                 return;
3020         pmd = pmd_offset(pud, address);
3021         ptl = pmd_lock(vma->vm_mm, pmd);
3022         if (!pmd_present(*pmd)) {
3023                 spin_unlock(ptl);
3024                 return;
3025         }
3026         if (pmd_trans_huge(*pmd)) {
3027                 if (page == pmd_page(*pmd))
3028                         __split_huge_pmd_locked(vma, pmd, haddr, true);
3029                 spin_unlock(ptl);
3030                 return;
3031         }
3032         spin_unlock(ptl);
3033
3034         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3035         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3036                 pte_t entry, swp_pte;
3037                 swp_entry_t swp_entry;
3038
3039                 /*
3040                  * We've just crossed page table boundary: need to map next one.
3041                  * It can happen if THP was mremaped to non PMD-aligned address.
3042                  */
3043                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3044                         pte_unmap_unlock(pte - 1, ptl);
3045                         pmd = mm_find_pmd(vma->vm_mm, address);
3046                         if (!pmd)
3047                                 return;
3048                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3049                                         address, &ptl);
3050                 }
3051
3052                 if (!pte_present(*pte))
3053                         continue;
3054                 if (page_to_pfn(page) != pte_pfn(*pte))
3055                         continue;
3056                 flush_cache_page(vma, address, page_to_pfn(page));
3057                 entry = ptep_clear_flush(vma, address, pte);
3058                 if (pte_dirty(entry))
3059                         SetPageDirty(page);
3060                 swp_entry = make_migration_entry(page, pte_write(entry));
3061                 swp_pte = swp_entry_to_pte(swp_entry);
3062                 if (pte_soft_dirty(entry))
3063                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
3064                 set_pte_at(vma->vm_mm, address, pte, swp_pte);
3065                 page_remove_rmap(page, false);
3066                 put_page(page);
3067         }
3068         pte_unmap_unlock(pte - 1, ptl);
3069 }
3070
3071 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
3072 {
3073         struct anon_vma_chain *avc;
3074         pgoff_t pgoff = page_to_pgoff(page);
3075
3076         VM_BUG_ON_PAGE(!PageHead(page), page);
3077
3078         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
3079                         pgoff + HPAGE_PMD_NR - 1) {
3080                 unsigned long address = __vma_address(page, avc->vma);
3081
3082                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3083                                 address, address + HPAGE_PMD_SIZE);
3084                 freeze_page_vma(avc->vma, page, address);
3085                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3086                                 address, address + HPAGE_PMD_SIZE);
3087         }
3088 }
3089
3090 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
3091                 unsigned long address)
3092 {
3093         spinlock_t *ptl;
3094         pmd_t *pmd;
3095         pte_t *pte, entry;
3096         swp_entry_t swp_entry;
3097         unsigned long haddr = address & HPAGE_PMD_MASK;
3098         int i, nr = HPAGE_PMD_NR;
3099
3100         /* Skip pages which doesn't belong to the VMA */
3101         if (address < vma->vm_start) {
3102                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3103                 page += off;
3104                 nr -= off;
3105                 address = vma->vm_start;
3106         }
3107
3108         pmd = mm_find_pmd(vma->vm_mm, address);
3109         if (!pmd)
3110                 return;
3111
3112         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3113         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3114                 /*
3115                  * We've just crossed page table boundary: need to map next one.
3116                  * It can happen if THP was mremaped to non-PMD aligned address.
3117                  */
3118                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3119                         pte_unmap_unlock(pte - 1, ptl);
3120                         pmd = mm_find_pmd(vma->vm_mm, address);
3121                         if (!pmd)
3122                                 return;
3123                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3124                                         address, &ptl);
3125                 }
3126
3127                 if (!is_swap_pte(*pte))
3128                         continue;
3129
3130                 swp_entry = pte_to_swp_entry(*pte);
3131                 if (!is_migration_entry(swp_entry))
3132                         continue;
3133                 if (migration_entry_to_page(swp_entry) != page)
3134                         continue;
3135
3136                 get_page(page);
3137                 page_add_anon_rmap(page, vma, address, false);
3138
3139                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3140                 if (PageDirty(page))
3141                         entry = pte_mkdirty(entry);
3142                 if (is_write_migration_entry(swp_entry))
3143                         entry = maybe_mkwrite(entry, vma);
3144
3145                 flush_dcache_page(page);
3146                 set_pte_at(vma->vm_mm, address, pte, entry);
3147
3148                 /* No need to invalidate - it was non-present before */
3149                 update_mmu_cache(vma, address, pte);
3150         }
3151         pte_unmap_unlock(pte - 1, ptl);
3152 }
3153
3154 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3155 {
3156         struct anon_vma_chain *avc;
3157         pgoff_t pgoff = page_to_pgoff(page);
3158
3159         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3160                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3161                 unsigned long address = __vma_address(page, avc->vma);
3162
3163                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3164                                 address, address + HPAGE_PMD_SIZE);
3165                 unfreeze_page_vma(avc->vma, page, address);
3166                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3167                                 address, address + HPAGE_PMD_SIZE);
3168         }
3169 }
3170
3171 static int __split_huge_page_tail(struct page *head, int tail,
3172                 struct lruvec *lruvec, struct list_head *list)
3173 {
3174         int mapcount;
3175         struct page *page_tail = head + tail;
3176
3177         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3178         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3179
3180         /*
3181          * tail_page->_count is zero and not changing from under us. But
3182          * get_page_unless_zero() may be running from under us on the
3183          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3184          * would then run atomic_set() concurrently with
3185          * get_page_unless_zero(), and atomic_set() is implemented in C not
3186          * using locked ops. spin_unlock on x86 sometime uses locked ops
3187          * because of PPro errata 66, 92, so unless somebody can guarantee
3188          * atomic_set() here would be safe on all archs (and not only on x86),
3189          * it's safer to use atomic_add().
3190          */
3191         atomic_add(mapcount + 1, &page_tail->_count);
3192
3193
3194         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3195         page_tail->flags |= (head->flags &
3196                         ((1L << PG_referenced) |
3197                          (1L << PG_swapbacked) |
3198                          (1L << PG_mlocked) |
3199                          (1L << PG_uptodate) |
3200                          (1L << PG_active) |
3201                          (1L << PG_locked) |
3202                          (1L << PG_unevictable) |
3203                          (1L << PG_dirty)));
3204
3205         /*
3206          * After clearing PageTail the gup refcount can be released.
3207          * Page flags also must be visible before we make the page non-compound.
3208          */
3209         smp_wmb();
3210
3211         clear_compound_head(page_tail);
3212
3213         if (page_is_young(head))
3214                 set_page_young(page_tail);
3215         if (page_is_idle(head))
3216                 set_page_idle(page_tail);
3217
3218         /* ->mapping in first tail page is compound_mapcount */
3219         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3220                         page_tail);
3221         page_tail->mapping = head->mapping;
3222
3223         page_tail->index = head->index + tail;
3224         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3225         lru_add_page_tail(head, page_tail, lruvec, list);
3226
3227         return mapcount;
3228 }
3229
3230 static void __split_huge_page(struct page *page, struct list_head *list)
3231 {
3232         struct page *head = compound_head(page);
3233         struct zone *zone = page_zone(head);
3234         struct lruvec *lruvec;
3235         int i, tail_mapcount;
3236
3237         /* prevent PageLRU to go away from under us, and freeze lru stats */
3238         spin_lock_irq(&zone->lru_lock);
3239         lruvec = mem_cgroup_page_lruvec(head, zone);
3240
3241         /* complete memcg works before add pages to LRU */
3242         mem_cgroup_split_huge_fixup(head);
3243
3244         tail_mapcount = 0;
3245         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3246                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3247         atomic_sub(tail_mapcount, &head->_count);
3248
3249         ClearPageCompound(head);
3250         spin_unlock_irq(&zone->lru_lock);
3251
3252         unfreeze_page(page_anon_vma(head), head);
3253
3254         for (i = 0; i < HPAGE_PMD_NR; i++) {
3255                 struct page *subpage = head + i;
3256                 if (subpage == page)
3257                         continue;
3258                 unlock_page(subpage);
3259
3260                 /*
3261                  * Subpages may be freed if there wasn't any mapping
3262                  * like if add_to_swap() is running on a lru page that
3263                  * had its mapping zapped. And freeing these pages
3264                  * requires taking the lru_lock so we do the put_page
3265                  * of the tail pages after the split is complete.
3266                  */
3267                 put_page(subpage);
3268         }
3269 }
3270
3271 int total_mapcount(struct page *page)
3272 {
3273         int i, ret;
3274
3275         VM_BUG_ON_PAGE(PageTail(page), page);
3276
3277         if (likely(!PageCompound(page)))
3278                 return atomic_read(&page->_mapcount) + 1;
3279
3280         ret = compound_mapcount(page);
3281         if (PageHuge(page))
3282                 return ret;
3283         for (i = 0; i < HPAGE_PMD_NR; i++)
3284                 ret += atomic_read(&page[i]._mapcount) + 1;
3285         if (PageDoubleMap(page))
3286                 ret -= HPAGE_PMD_NR;
3287         return ret;
3288 }
3289
3290 /*
3291  * This function splits huge page into normal pages. @page can point to any
3292  * subpage of huge page to split. Split doesn't change the position of @page.
3293  *
3294  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3295  * The huge page must be locked.
3296  *
3297  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3298  *
3299  * Both head page and tail pages will inherit mapping, flags, and so on from
3300  * the hugepage.
3301  *
3302  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3303  * they are not mapped.
3304  *
3305  * Returns 0 if the hugepage is split successfully.
3306  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3307  * us.
3308  */
3309 int split_huge_page_to_list(struct page *page, struct list_head *list)
3310 {
3311         struct page *head = compound_head(page);
3312         struct anon_vma *anon_vma;
3313         int count, mapcount, ret;
3314         bool mlocked;
3315
3316         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3317         VM_BUG_ON_PAGE(!PageAnon(page), page);
3318         VM_BUG_ON_PAGE(!PageLocked(page), page);
3319         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3320         VM_BUG_ON_PAGE(!PageCompound(page), page);
3321
3322         /*
3323          * The caller does not necessarily hold an mmap_sem that would prevent
3324          * the anon_vma disappearing so we first we take a reference to it
3325          * and then lock the anon_vma for write. This is similar to
3326          * page_lock_anon_vma_read except the write lock is taken to serialise
3327          * against parallel split or collapse operations.
3328          */
3329         anon_vma = page_get_anon_vma(head);
3330         if (!anon_vma) {
3331                 ret = -EBUSY;
3332                 goto out;
3333         }
3334         anon_vma_lock_write(anon_vma);
3335
3336         /*
3337          * Racy check if we can split the page, before freeze_page() will
3338          * split PMDs
3339          */
3340         if (total_mapcount(head) != page_count(head) - 1) {
3341                 ret = -EBUSY;
3342                 goto out_unlock;
3343         }
3344
3345         mlocked = PageMlocked(page);
3346         freeze_page(anon_vma, head);
3347         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3348
3349         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3350         if (mlocked)
3351                 lru_add_drain();
3352
3353         /* Prevent deferred_split_scan() touching ->_count */
3354         spin_lock(&split_queue_lock);
3355         count = page_count(head);
3356         mapcount = total_mapcount(head);
3357         if (!mapcount && count == 1) {
3358                 if (!list_empty(page_deferred_list(head))) {
3359                         split_queue_len--;
3360                         list_del(page_deferred_list(head));
3361                 }
3362                 spin_unlock(&split_queue_lock);
3363                 __split_huge_page(page, list);
3364                 ret = 0;
3365         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3366                 spin_unlock(&split_queue_lock);
3367                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3368                                 mapcount, count);
3369                 if (PageTail(page))
3370                         dump_page(head, NULL);
3371                 dump_page(page, "total_mapcount(head) > 0");
3372                 BUG();
3373         } else {
3374                 spin_unlock(&split_queue_lock);
3375                 unfreeze_page(anon_vma, head);
3376                 ret = -EBUSY;
3377         }
3378
3379 out_unlock:
3380         anon_vma_unlock_write(anon_vma);
3381         put_anon_vma(anon_vma);
3382 out:
3383         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3384         return ret;
3385 }
3386
3387 void free_transhuge_page(struct page *page)
3388 {
3389         unsigned long flags;
3390
3391         spin_lock_irqsave(&split_queue_lock, flags);
3392         if (!list_empty(page_deferred_list(page))) {
3393                 split_queue_len--;
3394                 list_del(page_deferred_list(page));
3395         }
3396         spin_unlock_irqrestore(&split_queue_lock, flags);
3397         free_compound_page(page);
3398 }
3399
3400 void deferred_split_huge_page(struct page *page)
3401 {
3402         unsigned long flags;
3403
3404         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3405
3406         spin_lock_irqsave(&split_queue_lock, flags);
3407         if (list_empty(page_deferred_list(page))) {
3408                 list_add_tail(page_deferred_list(page), &split_queue);
3409                 split_queue_len++;
3410         }
3411         spin_unlock_irqrestore(&split_queue_lock, flags);
3412 }
3413
3414 static unsigned long deferred_split_count(struct shrinker *shrink,
3415                 struct shrink_control *sc)
3416 {
3417         /*
3418          * Split a page from split_queue will free up at least one page,
3419          * at most HPAGE_PMD_NR - 1. We don't track exact number.
3420          * Let's use HPAGE_PMD_NR / 2 as ballpark.
3421          */
3422         return ACCESS_ONCE(split_queue_len) * HPAGE_PMD_NR / 2;
3423 }
3424
3425 static unsigned long deferred_split_scan(struct shrinker *shrink,
3426                 struct shrink_control *sc)
3427 {
3428         unsigned long flags;
3429         LIST_HEAD(list), *pos, *next;
3430         struct page *page;
3431         int split = 0;
3432
3433         spin_lock_irqsave(&split_queue_lock, flags);
3434         list_splice_init(&split_queue, &list);
3435
3436         /* Take pin on all head pages to avoid freeing them under us */
3437         list_for_each_safe(pos, next, &list) {
3438                 page = list_entry((void *)pos, struct page, mapping);
3439                 page = compound_head(page);
3440                 /* race with put_compound_page() */
3441                 if (!get_page_unless_zero(page)) {
3442                         list_del_init(page_deferred_list(page));
3443                         split_queue_len--;
3444                 }
3445         }
3446         spin_unlock_irqrestore(&split_queue_lock, flags);
3447
3448         list_for_each_safe(pos, next, &list) {
3449                 page = list_entry((void *)pos, struct page, mapping);
3450                 lock_page(page);
3451                 /* split_huge_page() removes page from list on success */
3452                 if (!split_huge_page(page))
3453                         split++;
3454                 unlock_page(page);
3455                 put_page(page);
3456         }
3457
3458         spin_lock_irqsave(&split_queue_lock, flags);
3459         list_splice_tail(&list, &split_queue);
3460         spin_unlock_irqrestore(&split_queue_lock, flags);
3461
3462         return split * HPAGE_PMD_NR / 2;
3463 }
3464
3465 static struct shrinker deferred_split_shrinker = {
3466         .count_objects = deferred_split_count,
3467         .scan_objects = deferred_split_scan,
3468         .seeks = DEFAULT_SEEKS,
3469 };
3470
3471 #ifdef CONFIG_DEBUG_FS
3472 static int split_huge_pages_set(void *data, u64 val)
3473 {
3474         struct zone *zone;
3475         struct page *page;
3476         unsigned long pfn, max_zone_pfn;
3477         unsigned long total = 0, split = 0;
3478
3479         if (val != 1)
3480                 return -EINVAL;
3481
3482         for_each_populated_zone(zone) {
3483                 max_zone_pfn = zone_end_pfn(zone);
3484                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3485                         if (!pfn_valid(pfn))
3486                                 continue;
3487
3488                         page = pfn_to_page(pfn);
3489                         if (!get_page_unless_zero(page))
3490                                 continue;
3491
3492                         if (zone != page_zone(page))
3493                                 goto next;
3494
3495                         if (!PageHead(page) || !PageAnon(page) ||
3496                                         PageHuge(page))
3497                                 goto next;
3498
3499                         total++;
3500                         lock_page(page);
3501                         if (!split_huge_page(page))
3502                                 split++;
3503                         unlock_page(page);
3504 next:
3505                         put_page(page);
3506                 }
3507         }
3508
3509         pr_info("%lu of %lu THP split", split, total);
3510
3511         return 0;
3512 }
3513 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3514                 "%llu\n");
3515
3516 static int __init split_huge_pages_debugfs(void)
3517 {
3518         void *ret;
3519
3520         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3521                         &split_huge_pages_fops);
3522         if (!ret)
3523                 pr_warn("Failed to create split_huge_pages in debugfs");
3524         return 0;
3525 }
3526 late_initcall(split_huge_pages_debugfs);
3527 #endif