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/memremap.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/debugfs.h>
  29 #include <linux/migrate.h>
  30 #include <linux/hashtable.h>
  31 #include <linux/userfaultfd_k.h>
  32 #include <linux/page_idle.h>
  33
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 enum scan_result {
  39         SCAN_FAIL,
  40         SCAN_SUCCEED,
  41         SCAN_PMD_NULL,
  42         SCAN_EXCEED_NONE_PTE,
  43         SCAN_PTE_NON_PRESENT,
  44         SCAN_PAGE_RO,
  45         SCAN_NO_REFERENCED_PAGE,
  46         SCAN_PAGE_NULL,
  47         SCAN_SCAN_ABORT,
  48         SCAN_PAGE_COUNT,
  49         SCAN_PAGE_LRU,
  50         SCAN_PAGE_LOCK,
  51         SCAN_PAGE_ANON,
  52         SCAN_PAGE_COMPOUND,
  53         SCAN_ANY_PROCESS,
  54         SCAN_VMA_NULL,
  55         SCAN_VMA_CHECK,
  56         SCAN_ADDRESS_RANGE,
  57         SCAN_SWAP_CACHE_PAGE,
  58         SCAN_DEL_PAGE_LRU,
  59         SCAN_ALLOC_HUGE_PAGE_FAIL,
  60         SCAN_CGROUP_CHARGE_FAIL
  61 };
  62
  63 #define CREATE_TRACE_POINTS
  64 #include <trace/events/huge_memory.h>
  65
  66 /*
  67  * By default transparent hugepage support is disabled in order that avoid
  68  * to risk increase the memory footprint of applications without a guaranteed
  69  * benefit. When transparent hugepage support is enabled, is for all mappings,
  70  * and khugepaged scans all mappings.
  71  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  72  * for all hugepage allocations.
  73  */
  74 unsigned long transparent_hugepage_flags __read_mostly =
  75 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  76         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  77 #endif
  78 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  79         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  80 #endif
  81         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  82         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  83         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  84
  85 /* default scan 8*512 pte (or vmas) every 30 second */
  86 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  87 static unsigned int khugepaged_pages_collapsed;
  88 static unsigned int khugepaged_full_scans;
  89 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  90 /* during fragmentation poll the hugepage allocator once every minute */
  91 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  92 static struct task_struct *khugepaged_thread __read_mostly;
  93 static DEFINE_MUTEX(khugepaged_mutex);
  94 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  95 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  96 /*
  97  * default collapse hugepages if there is at least one pte mapped like
  98  * it would have happened if the vma was large enough during page
  99  * fault.
 100  */
 101 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
 102
 103 static int khugepaged(void *none);
 104 static int khugepaged_slab_init(void);
 105 static void khugepaged_slab_exit(void);
 106
 107 #define MM_SLOTS_HASH_BITS 10
 108 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 109
 110 static struct kmem_cache *mm_slot_cache __read_mostly;
 111
 112 /**
 113  * struct mm_slot - hash lookup from mm to mm_slot
 114  * @hash: hash collision list
 115  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 116  * @mm: the mm that this information is valid for
 117  */
 118 struct mm_slot {
 119         struct hlist_node hash;
 120         struct list_head mm_node;
 121         struct mm_struct *mm;
 122 };
 123
 124 /**
 125  * struct khugepaged_scan - cursor for scanning
 126  * @mm_head: the head of the mm list to scan
 127  * @mm_slot: the current mm_slot we are scanning
 128  * @address: the next address inside that to be scanned
 129  *
 130  * There is only the one khugepaged_scan instance of this cursor structure.
 131  */
 132 struct khugepaged_scan {
 133         struct list_head mm_head;
 134         struct mm_slot *mm_slot;
 135         unsigned long address;
 136 };
 137 static struct khugepaged_scan khugepaged_scan = {
 138         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 139 };
 140
 141 static struct shrinker deferred_split_shrinker;
 142
 143 static void set_recommended_min_free_kbytes(void)
 144 {
 145         struct zone *zone;
 146         int nr_zones = 0;
 147         unsigned long recommended_min;
 148
 149         for_each_populated_zone(zone)
 150                 nr_zones++;
 151
 152         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 153         recommended_min = pageblock_nr_pages * nr_zones * 2;
 154
 155         /*
 156          * Make sure that on average at least two pageblocks are almost free
 157          * of another type, one for a migratetype to fall back to and a
 158          * second to avoid subsequent fallbacks of other types There are 3
 159          * MIGRATE_TYPES we care about.
 160          */
 161         recommended_min += pageblock_nr_pages * nr_zones *
 162                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 163
 164         /* don't ever allow to reserve more than 5% of the lowmem */
 165         recommended_min = min(recommended_min,
 166                               (unsigned long) nr_free_buffer_pages() / 20);
 167         recommended_min <<= (PAGE_SHIFT-10);
 168
 169         if (recommended_min > min_free_kbytes) {
 170                 if (user_min_free_kbytes >= 0)
 171                         pr_info("raising min_free_kbytes from %d to %lu "
 172                                 "to help transparent hugepage allocations\n",
 173                                 min_free_kbytes, recommended_min);
 174
 175                 min_free_kbytes = recommended_min;
 176         }
 177         setup_per_zone_wmarks();
 178 }
 179
 180 static int start_stop_khugepaged(void)
 181 {
 182         int err = 0;
 183         if (khugepaged_enabled()) {
 184                 if (!khugepaged_thread)
 185                         khugepaged_thread = kthread_run(khugepaged, NULL,
 186                                                         "khugepaged");
 187                 if (IS_ERR(khugepaged_thread)) {
 188                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 189                         err = PTR_ERR(khugepaged_thread);
 190                         khugepaged_thread = NULL;
 191                         goto fail;
 192                 }
 193
 194                 if (!list_empty(&khugepaged_scan.mm_head))
 195                         wake_up_interruptible(&khugepaged_wait);
 196
 197                 set_recommended_min_free_kbytes();
 198         } else if (khugepaged_thread) {
 199                 kthread_stop(khugepaged_thread);
 200                 khugepaged_thread = NULL;
 201         }
 202 fail:
 203         return err;
 204 }
 205
 206 static atomic_t huge_zero_refcount;
 207 struct page *huge_zero_page __read_mostly;
 208
 209 struct page *get_huge_zero_page(void)
 210 {
 211         struct page *zero_page;
 212 retry:
 213         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 214                 return READ_ONCE(huge_zero_page);
 215
 216         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 217                         HPAGE_PMD_ORDER);
 218         if (!zero_page) {
 219                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 220                 return NULL;
 221         }
 222         count_vm_event(THP_ZERO_PAGE_ALLOC);
 223         preempt_disable();
 224         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 225                 preempt_enable();
 226                 __free_pages(zero_page, compound_order(zero_page));
 227                 goto retry;
 228         }
 229
 230         /* We take additional reference here. It will be put back by shrinker */
 231         atomic_set(&huge_zero_refcount, 2);
 232         preempt_enable();
 233         return READ_ONCE(huge_zero_page);
 234 }
 235
 236 static void put_huge_zero_page(void)
 237 {
 238         /*
 239          * Counter should never go to zero here. Only shrinker can put
 240          * last reference.
 241          */
 242         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 243 }
 244
 245 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 246                                         struct shrink_control *sc)
 247 {
 248         /* we can free zero page only if last reference remains */
 249         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 250 }
 251
 252 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 253                                        struct shrink_control *sc)
 254 {
 255         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 256                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 257                 BUG_ON(zero_page == NULL);
 258                 __free_pages(zero_page, compound_order(zero_page));
 259                 return HPAGE_PMD_NR;
 260         }
 261
 262         return 0;
 263 }
 264
 265 static struct shrinker huge_zero_page_shrinker = {
 266         .count_objects = shrink_huge_zero_page_count,
 267         .scan_objects = shrink_huge_zero_page_scan,
 268         .seeks = DEFAULT_SEEKS,
 269 };
 270
 271 #ifdef CONFIG_SYSFS
 272
 273 static ssize_t double_flag_show(struct kobject *kobj,
 274                                 struct kobj_attribute *attr, char *buf,
 275                                 enum transparent_hugepage_flag enabled,
 276                                 enum transparent_hugepage_flag req_madv)
 277 {
 278         if (test_bit(enabled, &transparent_hugepage_flags)) {
 279                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 280                 return sprintf(buf, "[always] madvise never\n");
 281         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 282                 return sprintf(buf, "always [madvise] never\n");
 283         else
 284                 return sprintf(buf, "always madvise [never]\n");
 285 }
 286 static ssize_t double_flag_store(struct kobject *kobj,
 287                                  struct kobj_attribute *attr,
 288                                  const char *buf, size_t count,
 289                                  enum transparent_hugepage_flag enabled,
 290                                  enum transparent_hugepage_flag req_madv)
 291 {
 292         if (!memcmp("always", buf,
 293                     min(sizeof("always")-1, count))) {
 294                 set_bit(enabled, &transparent_hugepage_flags);
 295                 clear_bit(req_madv, &transparent_hugepage_flags);
 296         } else if (!memcmp("madvise", buf,
 297                            min(sizeof("madvise")-1, count))) {
 298                 clear_bit(enabled, &transparent_hugepage_flags);
 299                 set_bit(req_madv, &transparent_hugepage_flags);
 300         } else if (!memcmp("never", buf,
 301                            min(sizeof("never")-1, count))) {
 302                 clear_bit(enabled, &transparent_hugepage_flags);
 303                 clear_bit(req_madv, &transparent_hugepage_flags);
 304         } else
 305                 return -EINVAL;
 306
 307         return count;
 308 }
 309
 310 static ssize_t enabled_show(struct kobject *kobj,
 311                             struct kobj_attribute *attr, char *buf)
 312 {
 313         return double_flag_show(kobj, attr, buf,
 314                                 TRANSPARENT_HUGEPAGE_FLAG,
 315                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 316 }
 317 static ssize_t enabled_store(struct kobject *kobj,
 318                              struct kobj_attribute *attr,
 319                              const char *buf, size_t count)
 320 {
 321         ssize_t ret;
 322
 323         ret = double_flag_store(kobj, attr, buf, count,
 324                                 TRANSPARENT_HUGEPAGE_FLAG,
 325                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 326
 327         if (ret > 0) {
 328                 int err;
 329
 330                 mutex_lock(&khugepaged_mutex);
 331                 err = start_stop_khugepaged();
 332                 mutex_unlock(&khugepaged_mutex);
 333
 334                 if (err)
 335                         ret = err;
 336         }
 337
 338         return ret;
 339 }
 340 static struct kobj_attribute enabled_attr =
 341         __ATTR(enabled, 0644, enabled_show, enabled_store);
 342
 343 static ssize_t single_flag_show(struct kobject *kobj,
 344                                 struct kobj_attribute *attr, char *buf,
 345                                 enum transparent_hugepage_flag flag)
 346 {
 347         return sprintf(buf, "%d\n",
 348                        !!test_bit(flag, &transparent_hugepage_flags));
 349 }
 350
 351 static ssize_t single_flag_store(struct kobject *kobj,
 352                                  struct kobj_attribute *attr,
 353                                  const char *buf, size_t count,
 354                                  enum transparent_hugepage_flag flag)
 355 {
 356         unsigned long value;
 357         int ret;
 358
 359         ret = kstrtoul(buf, 10, &value);
 360         if (ret < 0)
 361                 return ret;
 362         if (value > 1)
 363                 return -EINVAL;
 364
 365         if (value)
 366                 set_bit(flag, &transparent_hugepage_flags);
 367         else
 368                 clear_bit(flag, &transparent_hugepage_flags);
 369
 370         return count;
 371 }
 372
 373 /*
 374  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 375  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 376  * memory just to allocate one more hugepage.
 377  */
 378 static ssize_t defrag_show(struct kobject *kobj,
 379                            struct kobj_attribute *attr, char *buf)
 380 {
 381         return double_flag_show(kobj, attr, buf,
 382                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 383                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 384 }
 385 static ssize_t defrag_store(struct kobject *kobj,
 386                             struct kobj_attribute *attr,
 387                             const char *buf, size_t count)
 388 {
 389         return double_flag_store(kobj, attr, buf, count,
 390                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 391                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 392 }
 393 static struct kobj_attribute defrag_attr =
 394         __ATTR(defrag, 0644, defrag_show, defrag_store);
 395
 396 static ssize_t use_zero_page_show(struct kobject *kobj,
 397                 struct kobj_attribute *attr, char *buf)
 398 {
 399         return single_flag_show(kobj, attr, buf,
 400                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 401 }
 402 static ssize_t use_zero_page_store(struct kobject *kobj,
 403                 struct kobj_attribute *attr, const char *buf, size_t count)
 404 {
 405         return single_flag_store(kobj, attr, buf, count,
 406                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 407 }
 408 static struct kobj_attribute use_zero_page_attr =
 409         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 410 #ifdef CONFIG_DEBUG_VM
 411 static ssize_t debug_cow_show(struct kobject *kobj,
 412                                 struct kobj_attribute *attr, char *buf)
 413 {
 414         return single_flag_show(kobj, attr, buf,
 415                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 416 }
 417 static ssize_t debug_cow_store(struct kobject *kobj,
 418                                struct kobj_attribute *attr,
 419                                const char *buf, size_t count)
 420 {
 421         return single_flag_store(kobj, attr, buf, count,
 422                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 423 }
 424 static struct kobj_attribute debug_cow_attr =
 425         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 426 #endif /* CONFIG_DEBUG_VM */
 427
 428 static struct attribute *hugepage_attr[] = {
 429         &enabled_attr.attr,
 430         &defrag_attr.attr,
 431         &use_zero_page_attr.attr,
 432 #ifdef CONFIG_DEBUG_VM
 433         &debug_cow_attr.attr,
 434 #endif
 435         NULL,
 436 };
 437
 438 static struct attribute_group hugepage_attr_group = {
 439         .attrs = hugepage_attr,
 440 };
 441
 442 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 443                                          struct kobj_attribute *attr,
 444                                          char *buf)
 445 {
 446         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 447 }
 448
 449 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 450                                           struct kobj_attribute *attr,
 451                                           const char *buf, size_t count)
 452 {
 453         unsigned long msecs;
 454         int err;
 455
 456         err = kstrtoul(buf, 10, &msecs);
 457         if (err || msecs > UINT_MAX)
 458                 return -EINVAL;
 459
 460         khugepaged_scan_sleep_millisecs = msecs;
 461         wake_up_interruptible(&khugepaged_wait);
 462
 463         return count;
 464 }
 465 static struct kobj_attribute scan_sleep_millisecs_attr =
 466         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 467                scan_sleep_millisecs_store);
 468
 469 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 470                                           struct kobj_attribute *attr,
 471                                           char *buf)
 472 {
 473         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 474 }
 475
 476 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 477                                            struct kobj_attribute *attr,
 478                                            const char *buf, size_t count)
 479 {
 480         unsigned long msecs;
 481         int err;
 482
 483         err = kstrtoul(buf, 10, &msecs);
 484         if (err || msecs > UINT_MAX)
 485                 return -EINVAL;
 486
 487         khugepaged_alloc_sleep_millisecs = msecs;
 488         wake_up_interruptible(&khugepaged_wait);
 489
 490         return count;
 491 }
 492 static struct kobj_attribute alloc_sleep_millisecs_attr =
 493         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 494                alloc_sleep_millisecs_store);
 495
 496 static ssize_t pages_to_scan_show(struct kobject *kobj,
 497                                   struct kobj_attribute *attr,
 498                                   char *buf)
 499 {
 500         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 501 }
 502 static ssize_t pages_to_scan_store(struct kobject *kobj,
 503                                    struct kobj_attribute *attr,
 504                                    const char *buf, size_t count)
 505 {
 506         int err;
 507         unsigned long pages;
 508
 509         err = kstrtoul(buf, 10, &pages);
 510         if (err || !pages || pages > UINT_MAX)
 511                 return -EINVAL;
 512
 513         khugepaged_pages_to_scan = pages;
 514
 515         return count;
 516 }
 517 static struct kobj_attribute pages_to_scan_attr =
 518         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 519                pages_to_scan_store);
 520
 521 static ssize_t pages_collapsed_show(struct kobject *kobj,
 522                                     struct kobj_attribute *attr,
 523                                     char *buf)
 524 {
 525         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 526 }
 527 static struct kobj_attribute pages_collapsed_attr =
 528         __ATTR_RO(pages_collapsed);
 529
 530 static ssize_t full_scans_show(struct kobject *kobj,
 531                                struct kobj_attribute *attr,
 532                                char *buf)
 533 {
 534         return sprintf(buf, "%u\n", khugepaged_full_scans);
 535 }
 536 static struct kobj_attribute full_scans_attr =
 537         __ATTR_RO(full_scans);
 538
 539 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 540                                       struct kobj_attribute *attr, char *buf)
 541 {
 542         return single_flag_show(kobj, attr, buf,
 543                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 544 }
 545 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 546                                        struct kobj_attribute *attr,
 547                                        const char *buf, size_t count)
 548 {
 549         return single_flag_store(kobj, attr, buf, count,
 550                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 551 }
 552 static struct kobj_attribute khugepaged_defrag_attr =
 553         __ATTR(defrag, 0644, khugepaged_defrag_show,
 554                khugepaged_defrag_store);
 555
 556 /*
 557  * max_ptes_none controls if khugepaged should collapse hugepages over
 558  * any unmapped ptes in turn potentially increasing the memory
 559  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 560  * reduce the available free memory in the system as it
 561  * runs. Increasing max_ptes_none will instead potentially reduce the
 562  * free memory in the system during the khugepaged scan.
 563  */
 564 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 565                                              struct kobj_attribute *attr,
 566                                              char *buf)
 567 {
 568         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 569 }
 570 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 571                                               struct kobj_attribute *attr,
 572                                               const char *buf, size_t count)
 573 {
 574         int err;
 575         unsigned long max_ptes_none;
 576
 577         err = kstrtoul(buf, 10, &max_ptes_none);
 578         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 579                 return -EINVAL;
 580
 581         khugepaged_max_ptes_none = max_ptes_none;
 582
 583         return count;
 584 }
 585 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 586         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 587                khugepaged_max_ptes_none_store);
 588
 589 static struct attribute *khugepaged_attr[] = {
 590         &khugepaged_defrag_attr.attr,
 591         &khugepaged_max_ptes_none_attr.attr,
 592         &pages_to_scan_attr.attr,
 593         &pages_collapsed_attr.attr,
 594         &full_scans_attr.attr,
 595         &scan_sleep_millisecs_attr.attr,
 596         &alloc_sleep_millisecs_attr.attr,
 597         NULL,
 598 };
 599
 600 static struct attribute_group khugepaged_attr_group = {
 601         .attrs = khugepaged_attr,
 602         .name = "khugepaged",
 603 };
 604
 605 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 606 {
 607         int err;
 608
 609         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 610         if (unlikely(!*hugepage_kobj)) {
 611                 pr_err("failed to create transparent hugepage kobject\n");
 612                 return -ENOMEM;
 613         }
 614
 615         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 616         if (err) {
 617                 pr_err("failed to register transparent hugepage group\n");
 618                 goto delete_obj;
 619         }
 620
 621         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 622         if (err) {
 623                 pr_err("failed to register transparent hugepage group\n");
 624                 goto remove_hp_group;
 625         }
 626
 627         return 0;
 628
 629 remove_hp_group:
 630         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 631 delete_obj:
 632         kobject_put(*hugepage_kobj);
 633         return err;
 634 }
 635
 636 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 637 {
 638         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 639         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 640         kobject_put(hugepage_kobj);
 641 }
 642 #else
 643 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 644 {
 645         return 0;
 646 }
 647
 648 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 649 {
 650 }
 651 #endif /* CONFIG_SYSFS */
 652
 653 static int __init hugepage_init(void)
 654 {
 655         int err;
 656         struct kobject *hugepage_kobj;
 657
 658         if (!has_transparent_hugepage()) {
 659                 transparent_hugepage_flags = 0;
 660                 return -EINVAL;
 661         }
 662
 663         err = hugepage_init_sysfs(&hugepage_kobj);
 664         if (err)
 665                 goto err_sysfs;
 666
 667         err = khugepaged_slab_init();
 668         if (err)
 669                 goto err_slab;
 670
 671         err = register_shrinker(&huge_zero_page_shrinker);
 672         if (err)
 673                 goto err_hzp_shrinker;
 674         err = register_shrinker(&deferred_split_shrinker);
 675         if (err)
 676                 goto err_split_shrinker;
 677
 678         /*
 679          * By default disable transparent hugepages on smaller systems,
 680          * where the extra memory used could hurt more than TLB overhead
 681          * is likely to save.  The admin can still enable it through /sys.
 682          */
 683         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 684                 transparent_hugepage_flags = 0;
 685                 return 0;
 686         }
 687
 688         err = start_stop_khugepaged();
 689         if (err)
 690                 goto err_khugepaged;
 691
 692         return 0;
 693 err_khugepaged:
 694         unregister_shrinker(&deferred_split_shrinker);
 695 err_split_shrinker:
 696         unregister_shrinker(&huge_zero_page_shrinker);
 697 err_hzp_shrinker:
 698         khugepaged_slab_exit();
 699 err_slab:
 700         hugepage_exit_sysfs(hugepage_kobj);
 701 err_sysfs:
 702         return err;
 703 }
 704 subsys_initcall(hugepage_init);
 705
 706 static int __init setup_transparent_hugepage(char *str)
 707 {
 708         int ret = 0;
 709         if (!str)
 710                 goto out;
 711         if (!strcmp(str, "always")) {
 712                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 713                         &transparent_hugepage_flags);
 714                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 715                           &transparent_hugepage_flags);
 716                 ret = 1;
 717         } else if (!strcmp(str, "madvise")) {
 718                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 719                           &transparent_hugepage_flags);
 720                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 721                         &transparent_hugepage_flags);
 722                 ret = 1;
 723         } else if (!strcmp(str, "never")) {
 724                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 725                           &transparent_hugepage_flags);
 726                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 727                           &transparent_hugepage_flags);
 728                 ret = 1;
 729         }
 730 out:
 731         if (!ret)
 732                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 733         return ret;
 734 }
 735 __setup("transparent_hugepage=", setup_transparent_hugepage);
 736
 737 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 738 {
 739         if (likely(vma->vm_flags & VM_WRITE))
 740                 pmd = pmd_mkwrite(pmd);
 741         return pmd;
 742 }
 743
 744 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 745 {
 746         pmd_t entry;
 747         entry = mk_pmd(page, prot);
 748         entry = pmd_mkhuge(entry);
 749         return entry;
 750 }
 751
 752 static inline struct list_head *page_deferred_list(struct page *page)
 753 {
 754         /*
 755          * ->lru in the tail pages is occupied by compound_head.
 756          * Let's use ->mapping + ->index in the second tail page as list_head.
 757          */
 758         return (struct list_head *)&page[2].mapping;
 759 }
 760
 761 void prep_transhuge_page(struct page *page)
 762 {
 763         /*
 764          * we use page->mapping and page->indexlru in second tail page
 765          * as list_head: assuming THP order >= 2
 766          */
 767         BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 768
 769         INIT_LIST_HEAD(page_deferred_list(page));
 770         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 771 }
 772
 773 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 774                                         struct vm_area_struct *vma,
 775                                         unsigned long address, pmd_t *pmd,
 776                                         struct page *page, gfp_t gfp,
 777                                         unsigned int flags)
 778 {
 779         struct mem_cgroup *memcg;
 780         pgtable_t pgtable;
 781         spinlock_t *ptl;
 782         unsigned long haddr = address & HPAGE_PMD_MASK;
 783
 784         VM_BUG_ON_PAGE(!PageCompound(page), page);
 785
 786         if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
 787                 put_page(page);
 788                 count_vm_event(THP_FAULT_FALLBACK);
 789                 return VM_FAULT_FALLBACK;
 790         }
 791
 792         pgtable = pte_alloc_one(mm, haddr);
 793         if (unlikely(!pgtable)) {
 794                 mem_cgroup_cancel_charge(page, memcg, true);
 795                 put_page(page);
 796                 return VM_FAULT_OOM;
 797         }
 798
 799         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 800         /*
 801          * The memory barrier inside __SetPageUptodate makes sure that
 802          * clear_huge_page writes become visible before the set_pmd_at()
 803          * write.
 804          */
 805         __SetPageUptodate(page);
 806
 807         ptl = pmd_lock(mm, pmd);
 808         if (unlikely(!pmd_none(*pmd))) {
 809                 spin_unlock(ptl);
 810                 mem_cgroup_cancel_charge(page, memcg, true);
 811                 put_page(page);
 812                 pte_free(mm, pgtable);
 813         } else {
 814                 pmd_t entry;
 815
 816                 /* Deliver the page fault to userland */
 817                 if (userfaultfd_missing(vma)) {
 818                         int ret;
 819
 820                         spin_unlock(ptl);
 821                         mem_cgroup_cancel_charge(page, memcg, true);
 822                         put_page(page);
 823                         pte_free(mm, pgtable);
 824                         ret = handle_userfault(vma, address, flags,
 825                                                VM_UFFD_MISSING);
 826                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 827                         return ret;
 828                 }
 829
 830                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 831                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 832                 page_add_new_anon_rmap(page, vma, haddr, true);
 833                 mem_cgroup_commit_charge(page, memcg, false, true);
 834                 lru_cache_add_active_or_unevictable(page, vma);
 835                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 836                 set_pmd_at(mm, haddr, pmd, entry);
 837                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 838                 atomic_long_inc(&mm->nr_ptes);
 839                 spin_unlock(ptl);
 840                 count_vm_event(THP_FAULT_ALLOC);
 841         }
 842
 843         return 0;
 844 }
 845
 846 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 847 {
 848         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 849 }
 850
 851 /* Caller must hold page table lock. */
 852 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 853                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 854                 struct page *zero_page)
 855 {
 856         pmd_t entry;
 857         if (!pmd_none(*pmd))
 858                 return false;
 859         entry = mk_pmd(zero_page, vma->vm_page_prot);
 860         entry = pmd_mkhuge(entry);
 861         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 862         set_pmd_at(mm, haddr, pmd, entry);
 863         atomic_long_inc(&mm->nr_ptes);
 864         return true;
 865 }
 866
 867 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 868                                unsigned long address, pmd_t *pmd,
 869                                unsigned int flags)
 870 {
 871         gfp_t gfp;
 872         struct page *page;
 873         unsigned long haddr = address & HPAGE_PMD_MASK;
 874
 875         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 876                 return VM_FAULT_FALLBACK;
 877         if (unlikely(anon_vma_prepare(vma)))
 878                 return VM_FAULT_OOM;
 879         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 880                 return VM_FAULT_OOM;
 881         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 882                         transparent_hugepage_use_zero_page()) {
 883                 spinlock_t *ptl;
 884                 pgtable_t pgtable;
 885                 struct page *zero_page;
 886                 bool set;
 887                 int ret;
 888                 pgtable = pte_alloc_one(mm, haddr);
 889                 if (unlikely(!pgtable))
 890                         return VM_FAULT_OOM;
 891                 zero_page = get_huge_zero_page();
 892                 if (unlikely(!zero_page)) {
 893                         pte_free(mm, pgtable);
 894                         count_vm_event(THP_FAULT_FALLBACK);
 895                         return VM_FAULT_FALLBACK;
 896                 }
 897                 ptl = pmd_lock(mm, pmd);
 898                 ret = 0;
 899                 set = false;
 900                 if (pmd_none(*pmd)) {
 901                         if (userfaultfd_missing(vma)) {
 902                                 spin_unlock(ptl);
 903                                 ret = handle_userfault(vma, address, flags,
 904                                                        VM_UFFD_MISSING);
 905                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 906                         } else {
 907                                 set_huge_zero_page(pgtable, mm, vma,
 908                                                    haddr, pmd,
 909                                                    zero_page);
 910                                 spin_unlock(ptl);
 911                                 set = true;
 912                         }
 913                 } else
 914                         spin_unlock(ptl);
 915                 if (!set) {
 916                         pte_free(mm, pgtable);
 917                         put_huge_zero_page();
 918                 }
 919                 return ret;
 920         }
 921         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 922         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 923         if (unlikely(!page)) {
 924                 count_vm_event(THP_FAULT_FALLBACK);
 925                 return VM_FAULT_FALLBACK;
 926         }
 927         prep_transhuge_page(page);
 928         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 929                                             flags);
 930 }
 931
 932 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 933                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 934 {
 935         struct mm_struct *mm = vma->vm_mm;
 936         pmd_t entry;
 937         spinlock_t *ptl;
 938
 939         ptl = pmd_lock(mm, pmd);
 940         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 941         if (pfn_t_devmap(pfn))
 942                 entry = pmd_mkdevmap(entry);
 943         if (write) {
 944                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 945                 entry = maybe_pmd_mkwrite(entry, vma);
 946         }
 947         set_pmd_at(mm, addr, pmd, entry);
 948         update_mmu_cache_pmd(vma, addr, pmd);
 949         spin_unlock(ptl);
 950 }
 951
 952 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 953                         pmd_t *pmd, pfn_t pfn, bool write)
 954 {
 955         pgprot_t pgprot = vma->vm_page_prot;
 956         /*
 957          * If we had pmd_special, we could avoid all these restrictions,
 958          * but we need to be consistent with PTEs and architectures that
 959          * can't support a 'special' bit.
 960          */
 961         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 962         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 963                                                 (VM_PFNMAP|VM_MIXEDMAP));
 964         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 965         BUG_ON(!pfn_t_devmap(pfn));
 966
 967         if (addr < vma->vm_start || addr >= vma->vm_end)
 968                 return VM_FAULT_SIGBUS;
 969         if (track_pfn_insert(vma, &pgprot, pfn))
 970                 return VM_FAULT_SIGBUS;
 971         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 972         return VM_FAULT_NOPAGE;
 973 }
 974
 975 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 976                 pmd_t *pmd)
 977 {
 978         pmd_t _pmd;
 979
 980         /*
 981          * We should set the dirty bit only for FOLL_WRITE but for now
 982          * the dirty bit in the pmd is meaningless.  And if the dirty
 983          * bit will become meaningful and we'll only set it with
 984          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 985          * set the young bit, instead of the current set_pmd_at.
 986          */
 987         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 988         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 989                                 pmd, _pmd,  1))
 990                 update_mmu_cache_pmd(vma, addr, pmd);
 991 }
 992
 993 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 994                 pmd_t *pmd, int flags)
 995 {
 996         unsigned long pfn = pmd_pfn(*pmd);
 997         struct mm_struct *mm = vma->vm_mm;
 998         struct dev_pagemap *pgmap;
 999         struct page *page;
1000
1001         assert_spin_locked(pmd_lockptr(mm, pmd));
1002
1003         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1004                 return NULL;
1005
1006         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1007                 /* pass */;
1008         else
1009                 return NULL;
1010
1011         if (flags & FOLL_TOUCH)
1012                 touch_pmd(vma, addr, pmd);
1013
1014         /*
1015          * device mapped pages can only be returned if the
1016          * caller will manage the page reference count.
1017          */
1018         if (!(flags & FOLL_GET))
1019                 return ERR_PTR(-EEXIST);
1020
1021         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1022         pgmap = get_dev_pagemap(pfn, NULL);
1023         if (!pgmap)
1024                 return ERR_PTR(-EFAULT);
1025         page = pfn_to_page(pfn);
1026         get_page(page);
1027         put_dev_pagemap(pgmap);
1028
1029         return page;
1030 }
1031
1032 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1033                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1034                   struct vm_area_struct *vma)
1035 {
1036         spinlock_t *dst_ptl, *src_ptl;
1037         struct page *src_page;
1038         pmd_t pmd;
1039         pgtable_t pgtable;
1040         int ret;
1041
1042         ret = -ENOMEM;
1043         pgtable = pte_alloc_one(dst_mm, addr);
1044         if (unlikely(!pgtable))
1045                 goto out;
1046
1047         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1048         src_ptl = pmd_lockptr(src_mm, src_pmd);
1049         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1050
1051         ret = -EAGAIN;
1052         pmd = *src_pmd;
1053         if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1054                 pte_free(dst_mm, pgtable);
1055                 goto out_unlock;
1056         }
1057         /*
1058          * When page table lock is held, the huge zero pmd should not be
1059          * under splitting since we don't split the page itself, only pmd to
1060          * a page table.
1061          */
1062         if (is_huge_zero_pmd(pmd)) {
1063                 struct page *zero_page;
1064                 /*
1065                  * get_huge_zero_page() will never allocate a new page here,
1066                  * since we already have a zero page to copy. It just takes a
1067                  * reference.
1068                  */
1069                 zero_page = get_huge_zero_page();
1070                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1071                                 zero_page);
1072                 ret = 0;
1073                 goto out_unlock;
1074         }
1075
1076         if (pmd_trans_huge(pmd)) {
1077                 /* thp accounting separate from pmd_devmap accounting */
1078                 src_page = pmd_page(pmd);
1079                 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1080                 get_page(src_page);
1081                 page_dup_rmap(src_page, true);
1082                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1083                 atomic_long_inc(&dst_mm->nr_ptes);
1084                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1085         }
1086
1087         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1088         pmd = pmd_mkold(pmd_wrprotect(pmd));
1089         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1090
1091         ret = 0;
1092 out_unlock:
1093         spin_unlock(src_ptl);
1094         spin_unlock(dst_ptl);
1095 out:
1096         return ret;
1097 }
1098
1099 void huge_pmd_set_accessed(struct mm_struct *mm,
1100                            struct vm_area_struct *vma,
1101                            unsigned long address,
1102                            pmd_t *pmd, pmd_t orig_pmd,
1103                            int dirty)
1104 {
1105         spinlock_t *ptl;
1106         pmd_t entry;
1107         unsigned long haddr;
1108
1109         ptl = pmd_lock(mm, pmd);
1110         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1111                 goto unlock;
1112
1113         entry = pmd_mkyoung(orig_pmd);
1114         haddr = address & HPAGE_PMD_MASK;
1115         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1116                 update_mmu_cache_pmd(vma, address, pmd);
1117
1118 unlock:
1119         spin_unlock(ptl);
1120 }
1121
1122 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1123                                         struct vm_area_struct *vma,
1124                                         unsigned long address,
1125                                         pmd_t *pmd, pmd_t orig_pmd,
1126                                         struct page *page,
1127                                         unsigned long haddr)
1128 {
1129         struct mem_cgroup *memcg;
1130         spinlock_t *ptl;
1131         pgtable_t pgtable;
1132         pmd_t _pmd;
1133         int ret = 0, i;
1134         struct page **pages;
1135         unsigned long mmun_start;       /* For mmu_notifiers */
1136         unsigned long mmun_end;         /* For mmu_notifiers */
1137
1138         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1139                         GFP_KERNEL);
1140         if (unlikely(!pages)) {
1141                 ret |= VM_FAULT_OOM;
1142                 goto out;
1143         }
1144
1145         for (i = 0; i < HPAGE_PMD_NR; i++) {
1146                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1147                                                __GFP_OTHER_NODE,
1148                                                vma, address, page_to_nid(page));
1149                 if (unlikely(!pages[i] ||
1150                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1151                                                    &memcg, false))) {
1152                         if (pages[i])
1153                                 put_page(pages[i]);
1154                         while (--i >= 0) {
1155                                 memcg = (void *)page_private(pages[i]);
1156                                 set_page_private(pages[i], 0);
1157                                 mem_cgroup_cancel_charge(pages[i], memcg,
1158                                                 false);
1159                                 put_page(pages[i]);
1160                         }
1161                         kfree(pages);
1162                         ret |= VM_FAULT_OOM;
1163                         goto out;
1164                 }
1165                 set_page_private(pages[i], (unsigned long)memcg);
1166         }
1167
1168         for (i = 0; i < HPAGE_PMD_NR; i++) {
1169                 copy_user_highpage(pages[i], page + i,
1170                                    haddr + PAGE_SIZE * i, vma);
1171                 __SetPageUptodate(pages[i]);
1172                 cond_resched();
1173         }
1174
1175         mmun_start = haddr;
1176         mmun_end   = haddr + HPAGE_PMD_SIZE;
1177         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1178
1179         ptl = pmd_lock(mm, pmd);
1180         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1181                 goto out_free_pages;
1182         VM_BUG_ON_PAGE(!PageHead(page), page);
1183
1184         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1185         /* leave pmd empty until pte is filled */
1186
1187         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1188         pmd_populate(mm, &_pmd, pgtable);
1189
1190         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1191                 pte_t *pte, entry;
1192                 entry = mk_pte(pages[i], vma->vm_page_prot);
1193                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1194                 memcg = (void *)page_private(pages[i]);
1195                 set_page_private(pages[i], 0);
1196                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1197                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1198                 lru_cache_add_active_or_unevictable(pages[i], vma);
1199                 pte = pte_offset_map(&_pmd, haddr);
1200                 VM_BUG_ON(!pte_none(*pte));
1201                 set_pte_at(mm, haddr, pte, entry);
1202                 pte_unmap(pte);
1203         }
1204         kfree(pages);
1205
1206         smp_wmb(); /* make pte visible before pmd */
1207         pmd_populate(mm, pmd, pgtable);
1208         page_remove_rmap(page, true);
1209         spin_unlock(ptl);
1210
1211         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1212
1213         ret |= VM_FAULT_WRITE;
1214         put_page(page);
1215
1216 out:
1217         return ret;
1218
1219 out_free_pages:
1220         spin_unlock(ptl);
1221         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1222         for (i = 0; i < HPAGE_PMD_NR; i++) {
1223                 memcg = (void *)page_private(pages[i]);
1224                 set_page_private(pages[i], 0);
1225                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1226                 put_page(pages[i]);
1227         }
1228         kfree(pages);
1229         goto out;
1230 }
1231
1232 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1233                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1234 {
1235         spinlock_t *ptl;
1236         int ret = 0;
1237         struct page *page = NULL, *new_page;
1238         struct mem_cgroup *memcg;
1239         unsigned long haddr;
1240         unsigned long mmun_start;       /* For mmu_notifiers */
1241         unsigned long mmun_end;         /* For mmu_notifiers */
1242         gfp_t huge_gfp;                 /* for allocation and charge */
1243
1244         ptl = pmd_lockptr(mm, pmd);
1245         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1246         haddr = address & HPAGE_PMD_MASK;
1247         if (is_huge_zero_pmd(orig_pmd))
1248                 goto alloc;
1249         spin_lock(ptl);
1250         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1251                 goto out_unlock;
1252
1253         page = pmd_page(orig_pmd);
1254         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1255         /*
1256          * We can only reuse the page if nobody else maps the huge page or it's
1257          * part. We can do it by checking page_mapcount() on each sub-page, but
1258          * it's expensive.
1259          * The cheaper way is to check page_count() to be equal 1: every
1260          * mapcount takes page reference reference, so this way we can
1261          * guarantee, that the PMD is the only mapping.
1262          * This can give false negative if somebody pinned the page, but that's
1263          * fine.
1264          */
1265         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1266                 pmd_t entry;
1267                 entry = pmd_mkyoung(orig_pmd);
1268                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1269                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1270                         update_mmu_cache_pmd(vma, address, pmd);
1271                 ret |= VM_FAULT_WRITE;
1272                 goto out_unlock;
1273         }
1274         get_page(page);
1275         spin_unlock(ptl);
1276 alloc:
1277         if (transparent_hugepage_enabled(vma) &&
1278             !transparent_hugepage_debug_cow()) {
1279                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1280                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1281         } else
1282                 new_page = NULL;
1283
1284         if (likely(new_page)) {
1285                 prep_transhuge_page(new_page);
1286         } else {
1287                 if (!page) {
1288                         split_huge_pmd(vma, pmd, address);
1289                         ret |= VM_FAULT_FALLBACK;
1290                 } else {
1291                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1292                                         pmd, orig_pmd, page, haddr);
1293                         if (ret & VM_FAULT_OOM) {
1294                                 split_huge_pmd(vma, pmd, address);
1295                                 ret |= VM_FAULT_FALLBACK;
1296                         }
1297                         put_page(page);
1298                 }
1299                 count_vm_event(THP_FAULT_FALLBACK);
1300                 goto out;
1301         }
1302
1303         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1304                                            true))) {
1305                 put_page(new_page);
1306                 if (page) {
1307                         split_huge_pmd(vma, pmd, address);
1308                         put_page(page);
1309                 } else
1310                         split_huge_pmd(vma, pmd, address);
1311                 ret |= VM_FAULT_FALLBACK;
1312                 count_vm_event(THP_FAULT_FALLBACK);
1313                 goto out;
1314         }
1315
1316         count_vm_event(THP_FAULT_ALLOC);
1317
1318         if (!page)
1319                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1320         else
1321                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1322         __SetPageUptodate(new_page);
1323
1324         mmun_start = haddr;
1325         mmun_end   = haddr + HPAGE_PMD_SIZE;
1326         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1327
1328         spin_lock(ptl);
1329         if (page)
1330                 put_page(page);
1331         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1332                 spin_unlock(ptl);
1333                 mem_cgroup_cancel_charge(new_page, memcg, true);
1334                 put_page(new_page);
1335                 goto out_mn;
1336         } else {
1337                 pmd_t entry;
1338                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1339                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1340                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1341                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1342                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1343                 lru_cache_add_active_or_unevictable(new_page, vma);
1344                 set_pmd_at(mm, haddr, pmd, entry);
1345                 update_mmu_cache_pmd(vma, address, pmd);
1346                 if (!page) {
1347                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1348                         put_huge_zero_page();
1349                 } else {
1350                         VM_BUG_ON_PAGE(!PageHead(page), page);
1351                         page_remove_rmap(page, true);
1352                         put_page(page);
1353                 }
1354                 ret |= VM_FAULT_WRITE;
1355         }
1356         spin_unlock(ptl);
1357 out_mn:
1358         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1359 out:
1360         return ret;
1361 out_unlock:
1362         spin_unlock(ptl);
1363         return ret;
1364 }
1365
1366 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1367                                    unsigned long addr,
1368                                    pmd_t *pmd,
1369                                    unsigned int flags)
1370 {
1371         struct mm_struct *mm = vma->vm_mm;
1372         struct page *page = NULL;
1373
1374         assert_spin_locked(pmd_lockptr(mm, pmd));
1375
1376         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1377                 goto out;
1378
1379         /* Avoid dumping huge zero page */
1380         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1381                 return ERR_PTR(-EFAULT);
1382
1383         /* Full NUMA hinting faults to serialise migration in fault paths */
1384         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1385                 goto out;
1386
1387         page = pmd_page(*pmd);
1388         VM_BUG_ON_PAGE(!PageHead(page), page);
1389         if (flags & FOLL_TOUCH)
1390                 touch_pmd(vma, addr, pmd);
1391         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1392                 /*
1393                  * We don't mlock() pte-mapped THPs. This way we can avoid
1394                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1395                  *
1396                  * In most cases the pmd is the only mapping of the page as we
1397                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1398                  * writable private mappings in populate_vma_page_range().
1399                  *
1400                  * The only scenario when we have the page shared here is if we
1401                  * mlocking read-only mapping shared over fork(). We skip
1402                  * mlocking such pages.
1403                  */
1404                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1405                                 page->mapping && trylock_page(page)) {
1406                         lru_add_drain();
1407                         if (page->mapping)
1408                                 mlock_vma_page(page);
1409                         unlock_page(page);
1410                 }
1411         }
1412         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1413         VM_BUG_ON_PAGE(!PageCompound(page), page);
1414         if (flags & FOLL_GET)
1415                 get_page(page);
1416
1417 out:
1418         return page;
1419 }
1420
1421 /* NUMA hinting page fault entry point for trans huge pmds */
1422 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1423                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1424 {
1425         spinlock_t *ptl;
1426         struct anon_vma *anon_vma = NULL;
1427         struct page *page;
1428         unsigned long haddr = addr & HPAGE_PMD_MASK;
1429         int page_nid = -1, this_nid = numa_node_id();
1430         int target_nid, last_cpupid = -1;
1431         bool page_locked;
1432         bool migrated = false;
1433         bool was_writable;
1434         int flags = 0;
1435
1436         /* A PROT_NONE fault should not end up here */
1437         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1438
1439         ptl = pmd_lock(mm, pmdp);
1440         if (unlikely(!pmd_same(pmd, *pmdp)))
1441                 goto out_unlock;
1442
1443         /*
1444          * If there are potential migrations, wait for completion and retry
1445          * without disrupting NUMA hinting information. Do not relock and
1446          * check_same as the page may no longer be mapped.
1447          */
1448         if (unlikely(pmd_trans_migrating(*pmdp))) {
1449                 page = pmd_page(*pmdp);
1450                 spin_unlock(ptl);
1451                 wait_on_page_locked(page);
1452                 goto out;
1453         }
1454
1455         page = pmd_page(pmd);
1456         BUG_ON(is_huge_zero_page(page));
1457         page_nid = page_to_nid(page);
1458         last_cpupid = page_cpupid_last(page);
1459         count_vm_numa_event(NUMA_HINT_FAULTS);
1460         if (page_nid == this_nid) {
1461                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1462                 flags |= TNF_FAULT_LOCAL;
1463         }
1464
1465         /* See similar comment in do_numa_page for explanation */
1466         if (!(vma->vm_flags & VM_WRITE))
1467                 flags |= TNF_NO_GROUP;
1468
1469         /*
1470          * Acquire the page lock to serialise THP migrations but avoid dropping
1471          * page_table_lock if at all possible
1472          */
1473         page_locked = trylock_page(page);
1474         target_nid = mpol_misplaced(page, vma, haddr);
1475         if (target_nid == -1) {
1476                 /* If the page was locked, there are no parallel migrations */
1477                 if (page_locked)
1478                         goto clear_pmdnuma;
1479         }
1480
1481         /* Migration could have started since the pmd_trans_migrating check */
1482         if (!page_locked) {
1483                 spin_unlock(ptl);
1484                 wait_on_page_locked(page);
1485                 page_nid = -1;
1486                 goto out;
1487         }
1488
1489         /*
1490          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1491          * to serialises splits
1492          */
1493         get_page(page);
1494         spin_unlock(ptl);
1495         anon_vma = page_lock_anon_vma_read(page);
1496
1497         /* Confirm the PMD did not change while page_table_lock was released */
1498         spin_lock(ptl);
1499         if (unlikely(!pmd_same(pmd, *pmdp))) {
1500                 unlock_page(page);
1501                 put_page(page);
1502                 page_nid = -1;
1503                 goto out_unlock;
1504         }
1505
1506         /* Bail if we fail to protect against THP splits for any reason */
1507         if (unlikely(!anon_vma)) {
1508                 put_page(page);
1509                 page_nid = -1;
1510                 goto clear_pmdnuma;
1511         }
1512
1513         /*
1514          * Migrate the THP to the requested node, returns with page unlocked
1515          * and access rights restored.
1516          */
1517         spin_unlock(ptl);
1518         migrated = migrate_misplaced_transhuge_page(mm, vma,
1519                                 pmdp, pmd, addr, page, target_nid);
1520         if (migrated) {
1521                 flags |= TNF_MIGRATED;
1522                 page_nid = target_nid;
1523         } else
1524                 flags |= TNF_MIGRATE_FAIL;
1525
1526         goto out;
1527 clear_pmdnuma:
1528         BUG_ON(!PageLocked(page));
1529         was_writable = pmd_write(pmd);
1530         pmd = pmd_modify(pmd, vma->vm_page_prot);
1531         pmd = pmd_mkyoung(pmd);
1532         if (was_writable)
1533                 pmd = pmd_mkwrite(pmd);
1534         set_pmd_at(mm, haddr, pmdp, pmd);
1535         update_mmu_cache_pmd(vma, addr, pmdp);
1536         unlock_page(page);
1537 out_unlock:
1538         spin_unlock(ptl);
1539
1540 out:
1541         if (anon_vma)
1542                 page_unlock_anon_vma_read(anon_vma);
1543
1544         if (page_nid != -1)
1545                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1546
1547         return 0;
1548 }
1549
1550 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1551                 pmd_t *pmd, unsigned long addr, unsigned long next)
1552
1553 {
1554         spinlock_t *ptl;
1555         pmd_t orig_pmd;
1556         struct page *page;
1557         struct mm_struct *mm = tlb->mm;
1558         int ret = 0;
1559
1560         ptl = pmd_trans_huge_lock(pmd, vma);
1561         if (!ptl)
1562                 goto out_unlocked;
1563
1564         orig_pmd = *pmd;
1565         if (is_huge_zero_pmd(orig_pmd)) {
1566                 ret = 1;
1567                 goto out;
1568         }
1569
1570         page = pmd_page(orig_pmd);
1571         /*
1572          * If other processes are mapping this page, we couldn't discard
1573          * the page unless they all do MADV_FREE so let's skip the page.
1574          */
1575         if (page_mapcount(page) != 1)
1576                 goto out;
1577
1578         if (!trylock_page(page))
1579                 goto out;
1580
1581         /*
1582          * If user want to discard part-pages of THP, split it so MADV_FREE
1583          * will deactivate only them.
1584          */
1585         if (next - addr != HPAGE_PMD_SIZE) {
1586                 get_page(page);
1587                 spin_unlock(ptl);
1588                 if (split_huge_page(page)) {
1589                         put_page(page);
1590                         unlock_page(page);
1591                         goto out_unlocked;
1592                 }
1593                 put_page(page);
1594                 unlock_page(page);
1595                 ret = 1;
1596                 goto out_unlocked;
1597         }
1598
1599         if (PageDirty(page))
1600                 ClearPageDirty(page);
1601         unlock_page(page);
1602
1603         if (PageActive(page))
1604                 deactivate_page(page);
1605
1606         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1607                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1608                         tlb->fullmm);
1609                 orig_pmd = pmd_mkold(orig_pmd);
1610                 orig_pmd = pmd_mkclean(orig_pmd);
1611
1612                 set_pmd_at(mm, addr, pmd, orig_pmd);
1613                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1614         }
1615         ret = 1;
1616 out:
1617         spin_unlock(ptl);
1618 out_unlocked:
1619         return ret;
1620 }
1621
1622 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1623                  pmd_t *pmd, unsigned long addr)
1624 {
1625         pmd_t orig_pmd;
1626         spinlock_t *ptl;
1627
1628         ptl = __pmd_trans_huge_lock(pmd, vma);
1629         if (!ptl)
1630                 return 0;
1631         /*
1632          * For architectures like ppc64 we look at deposited pgtable
1633          * when calling pmdp_huge_get_and_clear. So do the
1634          * pgtable_trans_huge_withdraw after finishing pmdp related
1635          * operations.
1636          */
1637         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1638                         tlb->fullmm);
1639         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1640         if (vma_is_dax(vma)) {
1641                 spin_unlock(ptl);
1642                 if (is_huge_zero_pmd(orig_pmd))
1643                         put_huge_zero_page();
1644         } else if (is_huge_zero_pmd(orig_pmd)) {
1645                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1646                 atomic_long_dec(&tlb->mm->nr_ptes);
1647                 spin_unlock(ptl);
1648                 put_huge_zero_page();
1649         } else {
1650                 struct page *page = pmd_page(orig_pmd);
1651                 page_remove_rmap(page, true);
1652                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1653                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1654                 VM_BUG_ON_PAGE(!PageHead(page), page);
1655                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1656                 atomic_long_dec(&tlb->mm->nr_ptes);
1657                 spin_unlock(ptl);
1658                 tlb_remove_page(tlb, page);
1659         }
1660         return 1;
1661 }
1662
1663 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1664                   unsigned long old_addr,
1665                   unsigned long new_addr, unsigned long old_end,
1666                   pmd_t *old_pmd, pmd_t *new_pmd)
1667 {
1668         spinlock_t *old_ptl, *new_ptl;
1669         pmd_t pmd;
1670
1671         struct mm_struct *mm = vma->vm_mm;
1672
1673         if ((old_addr & ~HPAGE_PMD_MASK) ||
1674             (new_addr & ~HPAGE_PMD_MASK) ||
1675             old_end - old_addr < HPAGE_PMD_SIZE ||
1676             (new_vma->vm_flags & VM_NOHUGEPAGE))
1677                 return false;
1678
1679         /*
1680          * The destination pmd shouldn't be established, free_pgtables()
1681          * should have release it.
1682          */
1683         if (WARN_ON(!pmd_none(*new_pmd))) {
1684                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1685                 return false;
1686         }
1687
1688         /*
1689          * We don't have to worry about the ordering of src and dst
1690          * ptlocks because exclusive mmap_sem prevents deadlock.
1691          */
1692         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1693         if (old_ptl) {
1694                 new_ptl = pmd_lockptr(mm, new_pmd);
1695                 if (new_ptl != old_ptl)
1696                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1697                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1698                 VM_BUG_ON(!pmd_none(*new_pmd));
1699
1700                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1701                         pgtable_t pgtable;
1702                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1703                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1704                 }
1705                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1706                 if (new_ptl != old_ptl)
1707                         spin_unlock(new_ptl);
1708                 spin_unlock(old_ptl);
1709                 return true;
1710         }
1711         return false;
1712 }
1713
1714 /*
1715  * Returns
1716  *  - 0 if PMD could not be locked
1717  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1718  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1719  */
1720 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1721                 unsigned long addr, pgprot_t newprot, int prot_numa)
1722 {
1723         struct mm_struct *mm = vma->vm_mm;
1724         spinlock_t *ptl;
1725         int ret = 0;
1726
1727         ptl = __pmd_trans_huge_lock(pmd, vma);
1728         if (ptl) {
1729                 pmd_t entry;
1730                 bool preserve_write = prot_numa && pmd_write(*pmd);
1731                 ret = 1;
1732
1733                 /*
1734                  * Avoid trapping faults against the zero page. The read-only
1735                  * data is likely to be read-cached on the local CPU and
1736                  * local/remote hits to the zero page are not interesting.
1737                  */
1738                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1739                         spin_unlock(ptl);
1740                         return ret;
1741                 }
1742
1743                 if (!prot_numa || !pmd_protnone(*pmd)) {
1744                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1745                         entry = pmd_modify(entry, newprot);
1746                         if (preserve_write)
1747                                 entry = pmd_mkwrite(entry);
1748                         ret = HPAGE_PMD_NR;
1749                         set_pmd_at(mm, addr, pmd, entry);
1750                         BUG_ON(!preserve_write && pmd_write(entry));
1751                 }
1752                 spin_unlock(ptl);
1753         }
1754
1755         return ret;
1756 }
1757
1758 /*
1759  * Returns true if a given pmd maps a thp, false otherwise.
1760  *
1761  * Note that if it returns true, this routine returns without unlocking page
1762  * table lock. So callers must unlock it.
1763  */
1764 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1765 {
1766         spinlock_t *ptl;
1767         ptl = pmd_lock(vma->vm_mm, pmd);
1768         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1769                 return ptl;
1770         spin_unlock(ptl);
1771         return NULL;
1772 }
1773
1774 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1775
1776 int hugepage_madvise(struct vm_area_struct *vma,
1777                      unsigned long *vm_flags, int advice)
1778 {
1779         switch (advice) {
1780         case MADV_HUGEPAGE:
1781 #ifdef CONFIG_S390
1782                 /*
1783                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1784                  * can't handle this properly after s390_enable_sie, so we simply
1785                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1786                  */
1787                 if (mm_has_pgste(vma->vm_mm))
1788                         return 0;
1789 #endif
1790                 /*
1791                  * Be somewhat over-protective like KSM for now!
1792                  */
1793                 if (*vm_flags & VM_NO_THP)
1794                         return -EINVAL;
1795                 *vm_flags &= ~VM_NOHUGEPAGE;
1796                 *vm_flags |= VM_HUGEPAGE;
1797                 /*
1798                  * If the vma become good for khugepaged to scan,
1799                  * register it here without waiting a page fault that
1800                  * may not happen any time soon.
1801                  */
1802                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1803                         return -ENOMEM;
1804                 break;
1805         case MADV_NOHUGEPAGE:
1806                 /*
1807                  * Be somewhat over-protective like KSM for now!
1808                  */
1809                 if (*vm_flags & VM_NO_THP)
1810                         return -EINVAL;
1811                 *vm_flags &= ~VM_HUGEPAGE;
1812                 *vm_flags |= VM_NOHUGEPAGE;
1813                 /*
1814                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1815                  * this vma even if we leave the mm registered in khugepaged if
1816                  * it got registered before VM_NOHUGEPAGE was set.
1817                  */
1818                 break;
1819         }
1820
1821         return 0;
1822 }
1823
1824 static int __init khugepaged_slab_init(void)
1825 {
1826         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1827                                           sizeof(struct mm_slot),
1828                                           __alignof__(struct mm_slot), 0, NULL);
1829         if (!mm_slot_cache)
1830                 return -ENOMEM;
1831
1832         return 0;
1833 }
1834
1835 static void __init khugepaged_slab_exit(void)
1836 {
1837         kmem_cache_destroy(mm_slot_cache);
1838 }
1839
1840 static inline struct mm_slot *alloc_mm_slot(void)
1841 {
1842         if (!mm_slot_cache)     /* initialization failed */
1843                 return NULL;
1844         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1845 }
1846
1847 static inline void free_mm_slot(struct mm_slot *mm_slot)
1848 {
1849         kmem_cache_free(mm_slot_cache, mm_slot);
1850 }
1851
1852 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1853 {
1854         struct mm_slot *mm_slot;
1855
1856         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1857                 if (mm == mm_slot->mm)
1858                         return mm_slot;
1859
1860         return NULL;
1861 }
1862
1863 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1864                                     struct mm_slot *mm_slot)
1865 {
1866         mm_slot->mm = mm;
1867         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1868 }
1869
1870 static inline int khugepaged_test_exit(struct mm_struct *mm)
1871 {
1872         return atomic_read(&mm->mm_users) == 0;
1873 }
1874
1875 int __khugepaged_enter(struct mm_struct *mm)
1876 {
1877         struct mm_slot *mm_slot;
1878         int wakeup;
1879
1880         mm_slot = alloc_mm_slot();
1881         if (!mm_slot)
1882                 return -ENOMEM;
1883
1884         /* __khugepaged_exit() must not run from under us */
1885         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1886         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1887                 free_mm_slot(mm_slot);
1888                 return 0;
1889         }
1890
1891         spin_lock(&khugepaged_mm_lock);
1892         insert_to_mm_slots_hash(mm, mm_slot);
1893         /*
1894          * Insert just behind the scanning cursor, to let the area settle
1895          * down a little.
1896          */
1897         wakeup = list_empty(&khugepaged_scan.mm_head);
1898         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1899         spin_unlock(&khugepaged_mm_lock);
1900
1901         atomic_inc(&mm->mm_count);
1902         if (wakeup)
1903                 wake_up_interruptible(&khugepaged_wait);
1904
1905         return 0;
1906 }
1907
1908 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1909                                unsigned long vm_flags)
1910 {
1911         unsigned long hstart, hend;
1912         if (!vma->anon_vma)
1913                 /*
1914                  * Not yet faulted in so we will register later in the
1915                  * page fault if needed.
1916                  */
1917                 return 0;
1918         if (vma->vm_ops)
1919                 /* khugepaged not yet working on file or special mappings */
1920                 return 0;
1921         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1922         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1923         hend = vma->vm_end & HPAGE_PMD_MASK;
1924         if (hstart < hend)
1925                 return khugepaged_enter(vma, vm_flags);
1926         return 0;
1927 }
1928
1929 void __khugepaged_exit(struct mm_struct *mm)
1930 {
1931         struct mm_slot *mm_slot;
1932         int free = 0;
1933
1934         spin_lock(&khugepaged_mm_lock);
1935         mm_slot = get_mm_slot(mm);
1936         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1937                 hash_del(&mm_slot->hash);
1938                 list_del(&mm_slot->mm_node);
1939                 free = 1;
1940         }
1941         spin_unlock(&khugepaged_mm_lock);
1942
1943         if (free) {
1944                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1945                 free_mm_slot(mm_slot);
1946                 mmdrop(mm);
1947         } else if (mm_slot) {
1948                 /*
1949                  * This is required to serialize against
1950                  * khugepaged_test_exit() (which is guaranteed to run
1951                  * under mmap sem read mode). Stop here (after we
1952                  * return all pagetables will be destroyed) until
1953                  * khugepaged has finished working on the pagetables
1954                  * under the mmap_sem.
1955                  */
1956                 down_write(&mm->mmap_sem);
1957                 up_write(&mm->mmap_sem);
1958         }
1959 }
1960
1961 static void release_pte_page(struct page *page)
1962 {
1963         /* 0 stands for page_is_file_cache(page) == false */
1964         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1965         unlock_page(page);
1966         putback_lru_page(page);
1967 }
1968
1969 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1970 {
1971         while (--_pte >= pte) {
1972                 pte_t pteval = *_pte;
1973                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1974                         release_pte_page(pte_page(pteval));
1975         }
1976 }
1977
1978 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1979                                         unsigned long address,
1980                                         pte_t *pte)
1981 {
1982         struct page *page = NULL;
1983         pte_t *_pte;
1984         int none_or_zero = 0, result = 0;
1985         bool referenced = false, writable = false;
1986
1987         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1988              _pte++, address += PAGE_SIZE) {
1989                 pte_t pteval = *_pte;
1990                 if (pte_none(pteval) || (pte_present(pteval) &&
1991                                 is_zero_pfn(pte_pfn(pteval)))) {
1992                         if (!userfaultfd_armed(vma) &&
1993                             ++none_or_zero <= khugepaged_max_ptes_none) {
1994                                 continue;
1995                         } else {
1996                                 result = SCAN_EXCEED_NONE_PTE;
1997                                 goto out;
1998                         }
1999                 }
2000                 if (!pte_present(pteval)) {
2001                         result = SCAN_PTE_NON_PRESENT;
2002                         goto out;
2003                 }
2004                 page = vm_normal_page(vma, address, pteval);
2005                 if (unlikely(!page)) {
2006                         result = SCAN_PAGE_NULL;
2007                         goto out;
2008                 }
2009
2010                 VM_BUG_ON_PAGE(PageCompound(page), page);
2011                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2012                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2013
2014                 /*
2015                  * We can do it before isolate_lru_page because the
2016                  * page can't be freed from under us. NOTE: PG_lock
2017                  * is needed to serialize against split_huge_page
2018                  * when invoked from the VM.
2019                  */
2020                 if (!trylock_page(page)) {
2021                         result = SCAN_PAGE_LOCK;
2022                         goto out;
2023                 }
2024
2025                 /*
2026                  * cannot use mapcount: can't collapse if there's a gup pin.
2027                  * The page must only be referenced by the scanned process
2028                  * and page swap cache.
2029                  */
2030                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2031                         unlock_page(page);
2032                         result = SCAN_PAGE_COUNT;
2033                         goto out;
2034                 }
2035                 if (pte_write(pteval)) {
2036                         writable = true;
2037                 } else {
2038                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2039                                 unlock_page(page);
2040                                 result = SCAN_SWAP_CACHE_PAGE;
2041                                 goto out;
2042                         }
2043                         /*
2044                          * Page is not in the swap cache. It can be collapsed
2045                          * into a THP.
2046                          */
2047                 }
2048
2049                 /*
2050                  * Isolate the page to avoid collapsing an hugepage
2051                  * currently in use by the VM.
2052                  */
2053                 if (isolate_lru_page(page)) {
2054                         unlock_page(page);
2055                         result = SCAN_DEL_PAGE_LRU;
2056                         goto out;
2057                 }
2058                 /* 0 stands for page_is_file_cache(page) == false */
2059                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2060                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2061                 VM_BUG_ON_PAGE(PageLRU(page), page);
2062
2063                 /* If there is no mapped pte young don't collapse the page */
2064                 if (pte_young(pteval) ||
2065                     page_is_young(page) || PageReferenced(page) ||
2066                     mmu_notifier_test_young(vma->vm_mm, address))
2067                         referenced = true;
2068         }
2069         if (likely(writable)) {
2070                 if (likely(referenced)) {
2071                         result = SCAN_SUCCEED;
2072                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2073                                                             referenced, writable, result);
2074                         return 1;
2075                 }
2076         } else {
2077                 result = SCAN_PAGE_RO;
2078         }
2079
2080 out:
2081         release_pte_pages(pte, _pte);
2082         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2083                                             referenced, writable, result);
2084         return 0;
2085 }
2086
2087 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2088                                       struct vm_area_struct *vma,
2089                                       unsigned long address,
2090                                       spinlock_t *ptl)
2091 {
2092         pte_t *_pte;
2093         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2094                 pte_t pteval = *_pte;
2095                 struct page *src_page;
2096
2097                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2098                         clear_user_highpage(page, address);
2099                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2100                         if (is_zero_pfn(pte_pfn(pteval))) {
2101                                 /*
2102                                  * ptl mostly unnecessary.
2103                                  */
2104                                 spin_lock(ptl);
2105                                 /*
2106                                  * paravirt calls inside pte_clear here are
2107                                  * superfluous.
2108                                  */
2109                                 pte_clear(vma->vm_mm, address, _pte);
2110                                 spin_unlock(ptl);
2111                         }
2112                 } else {
2113                         src_page = pte_page(pteval);
2114                         copy_user_highpage(page, src_page, address, vma);
2115                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2116                         release_pte_page(src_page);
2117                         /*
2118                          * ptl mostly unnecessary, but preempt has to
2119                          * be disabled to update the per-cpu stats
2120                          * inside page_remove_rmap().
2121                          */
2122                         spin_lock(ptl);
2123                         /*
2124                          * paravirt calls inside pte_clear here are
2125                          * superfluous.
2126                          */
2127                         pte_clear(vma->vm_mm, address, _pte);
2128                         page_remove_rmap(src_page, false);
2129                         spin_unlock(ptl);
2130                         free_page_and_swap_cache(src_page);
2131                 }
2132
2133                 address += PAGE_SIZE;
2134                 page++;
2135         }
2136 }
2137
2138 static void khugepaged_alloc_sleep(void)
2139 {
2140         DEFINE_WAIT(wait);
2141
2142         add_wait_queue(&khugepaged_wait, &wait);
2143         freezable_schedule_timeout_interruptible(
2144                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2145         remove_wait_queue(&khugepaged_wait, &wait);
2146 }
2147
2148 static int khugepaged_node_load[MAX_NUMNODES];
2149
2150 static bool khugepaged_scan_abort(int nid)
2151 {
2152         int i;
2153
2154         /*
2155          * If zone_reclaim_mode is disabled, then no extra effort is made to
2156          * allocate memory locally.
2157          */
2158         if (!zone_reclaim_mode)
2159                 return false;
2160
2161         /* If there is a count for this node already, it must be acceptable */
2162         if (khugepaged_node_load[nid])
2163                 return false;
2164
2165         for (i = 0; i < MAX_NUMNODES; i++) {
2166                 if (!khugepaged_node_load[i])
2167                         continue;
2168                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2169                         return true;
2170         }
2171         return false;
2172 }
2173
2174 #ifdef CONFIG_NUMA
2175 static int khugepaged_find_target_node(void)
2176 {
2177         static int last_khugepaged_target_node = NUMA_NO_NODE;
2178         int nid, target_node = 0, max_value = 0;
2179
2180         /* find first node with max normal pages hit */
2181         for (nid = 0; nid < MAX_NUMNODES; nid++)
2182                 if (khugepaged_node_load[nid] > max_value) {
2183                         max_value = khugepaged_node_load[nid];
2184                         target_node = nid;
2185                 }
2186
2187         /* do some balance if several nodes have the same hit record */
2188         if (target_node <= last_khugepaged_target_node)
2189                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2190                                 nid++)
2191                         if (max_value == khugepaged_node_load[nid]) {
2192                                 target_node = nid;
2193                                 break;
2194                         }
2195
2196         last_khugepaged_target_node = target_node;
2197         return target_node;
2198 }
2199
2200 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2201 {
2202         if (IS_ERR(*hpage)) {
2203                 if (!*wait)
2204                         return false;
2205
2206                 *wait = false;
2207                 *hpage = NULL;
2208                 khugepaged_alloc_sleep();
2209         } else if (*hpage) {
2210                 put_page(*hpage);
2211                 *hpage = NULL;
2212         }
2213
2214         return true;
2215 }
2216
2217 static struct page *
2218 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2219                        unsigned long address, int node)
2220 {
2221         VM_BUG_ON_PAGE(*hpage, *hpage);
2222
2223         /*
2224          * Before allocating the hugepage, release the mmap_sem read lock.
2225          * The allocation can take potentially a long time if it involves
2226          * sync compaction, and we do not need to hold the mmap_sem during
2227          * that. We will recheck the vma after taking it again in write mode.
2228          */
2229         up_read(&mm->mmap_sem);
2230
2231         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2232         if (unlikely(!*hpage)) {
2233                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2234                 *hpage = ERR_PTR(-ENOMEM);
2235                 return NULL;
2236         }
2237
2238         prep_transhuge_page(*hpage);
2239         count_vm_event(THP_COLLAPSE_ALLOC);
2240         return *hpage;
2241 }
2242 #else
2243 static int khugepaged_find_target_node(void)
2244 {
2245         return 0;
2246 }
2247
2248 static inline struct page *alloc_hugepage(int defrag)
2249 {
2250         struct page *page;
2251
2252         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2253         if (page)
2254                 prep_transhuge_page(page);
2255         return page;
2256 }
2257
2258 static struct page *khugepaged_alloc_hugepage(bool *wait)
2259 {
2260         struct page *hpage;
2261
2262         do {
2263                 hpage = alloc_hugepage(khugepaged_defrag());
2264                 if (!hpage) {
2265                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2266                         if (!*wait)
2267                                 return NULL;
2268
2269                         *wait = false;
2270                         khugepaged_alloc_sleep();
2271                 } else
2272                         count_vm_event(THP_COLLAPSE_ALLOC);
2273         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2274
2275         return hpage;
2276 }
2277
2278 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2279 {
2280         if (!*hpage)
2281                 *hpage = khugepaged_alloc_hugepage(wait);
2282
2283         if (unlikely(!*hpage))
2284                 return false;
2285
2286         return true;
2287 }
2288
2289 static struct page *
2290 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2291                        unsigned long address, int node)
2292 {
2293         up_read(&mm->mmap_sem);
2294         VM_BUG_ON(!*hpage);
2295
2296         return  *hpage;
2297 }
2298 #endif
2299
2300 static bool hugepage_vma_check(struct vm_area_struct *vma)
2301 {
2302         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2303             (vma->vm_flags & VM_NOHUGEPAGE))
2304                 return false;
2305         if (!vma->anon_vma || vma->vm_ops)
2306                 return false;
2307         if (is_vma_temporary_stack(vma))
2308                 return false;
2309         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2310         return true;
2311 }
2312
2313 static void collapse_huge_page(struct mm_struct *mm,
2314                                    unsigned long address,
2315                                    struct page **hpage,
2316                                    struct vm_area_struct *vma,
2317                                    int node)
2318 {
2319         pmd_t *pmd, _pmd;
2320         pte_t *pte;
2321         pgtable_t pgtable;
2322         struct page *new_page;
2323         spinlock_t *pmd_ptl, *pte_ptl;
2324         int isolated = 0, result = 0;
2325         unsigned long hstart, hend;
2326         struct mem_cgroup *memcg;
2327         unsigned long mmun_start;       /* For mmu_notifiers */
2328         unsigned long mmun_end;         /* For mmu_notifiers */
2329         gfp_t gfp;
2330
2331         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2332
2333         /* Only allocate from the target node */
2334         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2335                 __GFP_THISNODE;
2336
2337         /* release the mmap_sem read lock. */
2338         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2339         if (!new_page) {
2340                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2341                 goto out_nolock;
2342         }
2343
2344         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2345                 result = SCAN_CGROUP_CHARGE_FAIL;
2346                 goto out_nolock;
2347         }
2348
2349         /*
2350          * Prevent all access to pagetables with the exception of
2351          * gup_fast later hanlded by the ptep_clear_flush and the VM
2352          * handled by the anon_vma lock + PG_lock.
2353          */
2354         down_write(&mm->mmap_sem);
2355         if (unlikely(khugepaged_test_exit(mm))) {
2356                 result = SCAN_ANY_PROCESS;
2357                 goto out;
2358         }
2359
2360         vma = find_vma(mm, address);
2361         if (!vma) {
2362                 result = SCAN_VMA_NULL;
2363                 goto out;
2364         }
2365         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2366         hend = vma->vm_end & HPAGE_PMD_MASK;
2367         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2368                 result = SCAN_ADDRESS_RANGE;
2369                 goto out;
2370         }
2371         if (!hugepage_vma_check(vma)) {
2372                 result = SCAN_VMA_CHECK;
2373                 goto out;
2374         }
2375         pmd = mm_find_pmd(mm, address);
2376         if (!pmd) {
2377                 result = SCAN_PMD_NULL;
2378                 goto out;
2379         }
2380
2381         anon_vma_lock_write(vma->anon_vma);
2382
2383         pte = pte_offset_map(pmd, address);
2384         pte_ptl = pte_lockptr(mm, pmd);
2385
2386         mmun_start = address;
2387         mmun_end   = address + HPAGE_PMD_SIZE;
2388         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2389         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2390         /*
2391          * After this gup_fast can't run anymore. This also removes
2392          * any huge TLB entry from the CPU so we won't allow
2393          * huge and small TLB entries for the same virtual address
2394          * to avoid the risk of CPU bugs in that area.
2395          */
2396         _pmd = pmdp_collapse_flush(vma, address, pmd);
2397         spin_unlock(pmd_ptl);
2398         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2399
2400         spin_lock(pte_ptl);
2401         isolated = __collapse_huge_page_isolate(vma, address, pte);
2402         spin_unlock(pte_ptl);
2403
2404         if (unlikely(!isolated)) {
2405                 pte_unmap(pte);
2406                 spin_lock(pmd_ptl);
2407                 BUG_ON(!pmd_none(*pmd));
2408                 /*
2409                  * We can only use set_pmd_at when establishing
2410                  * hugepmds and never for establishing regular pmds that
2411                  * points to regular pagetables. Use pmd_populate for that
2412                  */
2413                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2414                 spin_unlock(pmd_ptl);
2415                 anon_vma_unlock_write(vma->anon_vma);
2416                 result = SCAN_FAIL;
2417                 goto out;
2418         }
2419
2420         /*
2421          * All pages are isolated and locked so anon_vma rmap
2422          * can't run anymore.
2423          */
2424         anon_vma_unlock_write(vma->anon_vma);
2425
2426         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2427         pte_unmap(pte);
2428         __SetPageUptodate(new_page);
2429         pgtable = pmd_pgtable(_pmd);
2430
2431         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2432         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2433
2434         /*
2435          * spin_lock() below is not the equivalent of smp_wmb(), so
2436          * this is needed to avoid the copy_huge_page writes to become
2437          * visible after the set_pmd_at() write.
2438          */
2439         smp_wmb();
2440
2441         spin_lock(pmd_ptl);
2442         BUG_ON(!pmd_none(*pmd));
2443         page_add_new_anon_rmap(new_page, vma, address, true);
2444         mem_cgroup_commit_charge(new_page, memcg, false, true);
2445         lru_cache_add_active_or_unevictable(new_page, vma);
2446         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2447         set_pmd_at(mm, address, pmd, _pmd);
2448         update_mmu_cache_pmd(vma, address, pmd);
2449         spin_unlock(pmd_ptl);
2450
2451         *hpage = NULL;
2452
2453         khugepaged_pages_collapsed++;
2454         result = SCAN_SUCCEED;
2455 out_up_write:
2456         up_write(&mm->mmap_sem);
2457         trace_mm_collapse_huge_page(mm, isolated, result);
2458         return;
2459
2460 out_nolock:
2461         trace_mm_collapse_huge_page(mm, isolated, result);
2462         return;
2463 out:
2464         mem_cgroup_cancel_charge(new_page, memcg, true);
2465         goto out_up_write;
2466 }
2467
2468 static int khugepaged_scan_pmd(struct mm_struct *mm,
2469                                struct vm_area_struct *vma,
2470                                unsigned long address,
2471                                struct page **hpage)
2472 {
2473         pmd_t *pmd;
2474         pte_t *pte, *_pte;
2475         int ret = 0, none_or_zero = 0, result = 0;
2476         struct page *page = NULL;
2477         unsigned long _address;
2478         spinlock_t *ptl;
2479         int node = NUMA_NO_NODE;
2480         bool writable = false, referenced = false;
2481
2482         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2483
2484         pmd = mm_find_pmd(mm, address);
2485         if (!pmd) {
2486                 result = SCAN_PMD_NULL;
2487                 goto out;
2488         }
2489
2490         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2491         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2492         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2493              _pte++, _address += PAGE_SIZE) {
2494                 pte_t pteval = *_pte;
2495                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2496                         if (!userfaultfd_armed(vma) &&
2497                             ++none_or_zero <= khugepaged_max_ptes_none) {
2498                                 continue;
2499                         } else {
2500                                 result = SCAN_EXCEED_NONE_PTE;
2501                                 goto out_unmap;
2502                         }
2503                 }
2504                 if (!pte_present(pteval)) {
2505                         result = SCAN_PTE_NON_PRESENT;
2506                         goto out_unmap;
2507                 }
2508                 if (pte_write(pteval))
2509                         writable = true;
2510
2511                 page = vm_normal_page(vma, _address, pteval);
2512                 if (unlikely(!page)) {
2513                         result = SCAN_PAGE_NULL;
2514                         goto out_unmap;
2515                 }
2516
2517                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2518                 if (PageCompound(page)) {
2519                         result = SCAN_PAGE_COMPOUND;
2520                         goto out_unmap;
2521                 }
2522
2523                 /*
2524                  * Record which node the original page is from and save this
2525                  * information to khugepaged_node_load[].
2526                  * Khupaged will allocate hugepage from the node has the max
2527                  * hit record.
2528                  */
2529                 node = page_to_nid(page);
2530                 if (khugepaged_scan_abort(node)) {
2531                         result = SCAN_SCAN_ABORT;
2532                         goto out_unmap;
2533                 }
2534                 khugepaged_node_load[node]++;
2535                 if (!PageLRU(page)) {
2536                         result = SCAN_SCAN_ABORT;
2537                         goto out_unmap;
2538                 }
2539                 if (PageLocked(page)) {
2540                         result = SCAN_PAGE_LOCK;
2541                         goto out_unmap;
2542                 }
2543                 if (!PageAnon(page)) {
2544                         result = SCAN_PAGE_ANON;
2545                         goto out_unmap;
2546                 }
2547
2548                 /*
2549                  * cannot use mapcount: can't collapse if there's a gup pin.
2550                  * The page must only be referenced by the scanned process
2551                  * and page swap cache.
2552                  */
2553                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2554                         result = SCAN_PAGE_COUNT;
2555                         goto out_unmap;
2556                 }
2557                 if (pte_young(pteval) ||
2558                     page_is_young(page) || PageReferenced(page) ||
2559                     mmu_notifier_test_young(vma->vm_mm, address))
2560                         referenced = true;
2561         }
2562         if (writable) {
2563                 if (referenced) {
2564                         result = SCAN_SUCCEED;
2565                         ret = 1;
2566                 } else {
2567                         result = SCAN_NO_REFERENCED_PAGE;
2568                 }
2569         } else {
2570                 result = SCAN_PAGE_RO;
2571         }
2572 out_unmap:
2573         pte_unmap_unlock(pte, ptl);
2574         if (ret) {
2575                 node = khugepaged_find_target_node();
2576                 /* collapse_huge_page will return with the mmap_sem released */
2577                 collapse_huge_page(mm, address, hpage, vma, node);
2578         }
2579 out:
2580         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2581                                      none_or_zero, result);
2582         return ret;
2583 }
2584
2585 static void collect_mm_slot(struct mm_slot *mm_slot)
2586 {
2587         struct mm_struct *mm = mm_slot->mm;
2588
2589         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2590
2591         if (khugepaged_test_exit(mm)) {
2592                 /* free mm_slot */
2593                 hash_del(&mm_slot->hash);
2594                 list_del(&mm_slot->mm_node);
2595
2596                 /*
2597                  * Not strictly needed because the mm exited already.
2598                  *
2599                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2600                  */
2601
2602                 /* khugepaged_mm_lock actually not necessary for the below */
2603                 free_mm_slot(mm_slot);
2604                 mmdrop(mm);
2605         }
2606 }
2607
2608 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2609                                             struct page **hpage)
2610         __releases(&khugepaged_mm_lock)
2611         __acquires(&khugepaged_mm_lock)
2612 {
2613         struct mm_slot *mm_slot;
2614         struct mm_struct *mm;
2615         struct vm_area_struct *vma;
2616         int progress = 0;
2617
2618         VM_BUG_ON(!pages);
2619         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2620
2621         if (khugepaged_scan.mm_slot)
2622                 mm_slot = khugepaged_scan.mm_slot;
2623         else {
2624                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2625                                      struct mm_slot, mm_node);
2626                 khugepaged_scan.address = 0;
2627                 khugepaged_scan.mm_slot = mm_slot;
2628         }
2629         spin_unlock(&khugepaged_mm_lock);
2630
2631         mm = mm_slot->mm;
2632         down_read(&mm->mmap_sem);
2633         if (unlikely(khugepaged_test_exit(mm)))
2634                 vma = NULL;
2635         else
2636                 vma = find_vma(mm, khugepaged_scan.address);
2637
2638         progress++;
2639         for (; vma; vma = vma->vm_next) {
2640                 unsigned long hstart, hend;
2641
2642                 cond_resched();
2643                 if (unlikely(khugepaged_test_exit(mm))) {
2644                         progress++;
2645                         break;
2646                 }
2647                 if (!hugepage_vma_check(vma)) {
2648 skip:
2649                         progress++;
2650                         continue;
2651                 }
2652                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2653                 hend = vma->vm_end & HPAGE_PMD_MASK;
2654                 if (hstart >= hend)
2655                         goto skip;
2656                 if (khugepaged_scan.address > hend)
2657                         goto skip;
2658                 if (khugepaged_scan.address < hstart)
2659                         khugepaged_scan.address = hstart;
2660                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2661
2662                 while (khugepaged_scan.address < hend) {
2663                         int ret;
2664                         cond_resched();
2665                         if (unlikely(khugepaged_test_exit(mm)))
2666                                 goto breakouterloop;
2667
2668                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2669                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2670                                   hend);
2671                         ret = khugepaged_scan_pmd(mm, vma,
2672                                                   khugepaged_scan.address,
2673                                                   hpage);
2674                         /* move to next address */
2675                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2676                         progress += HPAGE_PMD_NR;
2677                         if (ret)
2678                                 /* we released mmap_sem so break loop */
2679                                 goto breakouterloop_mmap_sem;
2680                         if (progress >= pages)
2681                                 goto breakouterloop;
2682                 }
2683         }
2684 breakouterloop:
2685         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2686 breakouterloop_mmap_sem:
2687
2688         spin_lock(&khugepaged_mm_lock);
2689         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2690         /*
2691          * Release the current mm_slot if this mm is about to die, or
2692          * if we scanned all vmas of this mm.
2693          */
2694         if (khugepaged_test_exit(mm) || !vma) {
2695                 /*
2696                  * Make sure that if mm_users is reaching zero while
2697                  * khugepaged runs here, khugepaged_exit will find
2698                  * mm_slot not pointing to the exiting mm.
2699                  */
2700                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2701                         khugepaged_scan.mm_slot = list_entry(
2702                                 mm_slot->mm_node.next,
2703                                 struct mm_slot, mm_node);
2704                         khugepaged_scan.address = 0;
2705                 } else {
2706                         khugepaged_scan.mm_slot = NULL;
2707                         khugepaged_full_scans++;
2708                 }
2709
2710                 collect_mm_slot(mm_slot);
2711         }
2712
2713         return progress;
2714 }
2715
2716 static int khugepaged_has_work(void)
2717 {
2718         return !list_empty(&khugepaged_scan.mm_head) &&
2719                 khugepaged_enabled();
2720 }
2721
2722 static int khugepaged_wait_event(void)
2723 {
2724         return !list_empty(&khugepaged_scan.mm_head) ||
2725                 kthread_should_stop();
2726 }
2727
2728 static void khugepaged_do_scan(void)
2729 {
2730         struct page *hpage = NULL;
2731         unsigned int progress = 0, pass_through_head = 0;
2732         unsigned int pages = khugepaged_pages_to_scan;
2733         bool wait = true;
2734
2735         barrier(); /* write khugepaged_pages_to_scan to local stack */
2736
2737         while (progress < pages) {
2738                 if (!khugepaged_prealloc_page(&hpage, &wait))
2739                         break;
2740
2741                 cond_resched();
2742
2743                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2744                         break;
2745
2746                 spin_lock(&khugepaged_mm_lock);
2747                 if (!khugepaged_scan.mm_slot)
2748                         pass_through_head++;
2749                 if (khugepaged_has_work() &&
2750                     pass_through_head < 2)
2751                         progress += khugepaged_scan_mm_slot(pages - progress,
2752                                                             &hpage);
2753                 else
2754                         progress = pages;
2755                 spin_unlock(&khugepaged_mm_lock);
2756         }
2757
2758         if (!IS_ERR_OR_NULL(hpage))
2759                 put_page(hpage);
2760 }
2761
2762 static void khugepaged_wait_work(void)
2763 {
2764         if (khugepaged_has_work()) {
2765                 if (!khugepaged_scan_sleep_millisecs)
2766                         return;
2767
2768                 wait_event_freezable_timeout(khugepaged_wait,
2769                                              kthread_should_stop(),
2770                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2771                 return;
2772         }
2773
2774         if (khugepaged_enabled())
2775                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2776 }
2777
2778 static int khugepaged(void *none)
2779 {
2780         struct mm_slot *mm_slot;
2781
2782         set_freezable();
2783         set_user_nice(current, MAX_NICE);
2784
2785         while (!kthread_should_stop()) {
2786                 khugepaged_do_scan();
2787                 khugepaged_wait_work();
2788         }
2789
2790         spin_lock(&khugepaged_mm_lock);
2791         mm_slot = khugepaged_scan.mm_slot;
2792         khugepaged_scan.mm_slot = NULL;
2793         if (mm_slot)
2794                 collect_mm_slot(mm_slot);
2795         spin_unlock(&khugepaged_mm_lock);
2796         return 0;
2797 }
2798
2799 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2800                 unsigned long haddr, pmd_t *pmd)
2801 {
2802         struct mm_struct *mm = vma->vm_mm;
2803         pgtable_t pgtable;
2804         pmd_t _pmd;
2805         int i;
2806
2807         /* leave pmd empty until pte is filled */
2808         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2809
2810         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2811         pmd_populate(mm, &_pmd, pgtable);
2812
2813         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2814                 pte_t *pte, entry;
2815                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2816                 entry = pte_mkspecial(entry);
2817                 pte = pte_offset_map(&_pmd, haddr);
2818                 VM_BUG_ON(!pte_none(*pte));
2819                 set_pte_at(mm, haddr, pte, entry);
2820                 pte_unmap(pte);
2821         }
2822         smp_wmb(); /* make pte visible before pmd */
2823         pmd_populate(mm, pmd, pgtable);
2824         put_huge_zero_page();
2825 }
2826
2827 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2828                 unsigned long haddr, bool freeze)
2829 {
2830         struct mm_struct *mm = vma->vm_mm;
2831         struct page *page;
2832         pgtable_t pgtable;
2833         pmd_t _pmd;
2834         bool young, write, dirty;
2835         int i;
2836
2837         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2838         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2839         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2840         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2841
2842         count_vm_event(THP_SPLIT_PMD);
2843
2844         if (vma_is_dax(vma)) {
2845                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2846                 if (is_huge_zero_pmd(_pmd))
2847                         put_huge_zero_page();
2848                 return;
2849         } else if (is_huge_zero_pmd(*pmd)) {
2850                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2851         }
2852
2853         page = pmd_page(*pmd);
2854         VM_BUG_ON_PAGE(!page_count(page), page);
2855         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2856         write = pmd_write(*pmd);
2857         young = pmd_young(*pmd);
2858         dirty = pmd_dirty(*pmd);
2859
2860         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2861         pmd_populate(mm, &_pmd, pgtable);
2862
2863         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2864                 pte_t entry, *pte;
2865                 /*
2866                  * Note that NUMA hinting access restrictions are not
2867                  * transferred to avoid any possibility of altering
2868                  * permissions across VMAs.
2869                  */
2870                 if (freeze) {
2871                         swp_entry_t swp_entry;
2872                         swp_entry = make_migration_entry(page + i, write);
2873                         entry = swp_entry_to_pte(swp_entry);
2874                 } else {
2875                         entry = mk_pte(page + i, vma->vm_page_prot);
2876                         entry = maybe_mkwrite(entry, vma);
2877                         if (!write)
2878                                 entry = pte_wrprotect(entry);
2879                         if (!young)
2880                                 entry = pte_mkold(entry);
2881                 }
2882                 if (dirty)
2883                         SetPageDirty(page + i);
2884                 pte = pte_offset_map(&_pmd, haddr);
2885                 BUG_ON(!pte_none(*pte));
2886                 set_pte_at(mm, haddr, pte, entry);
2887                 atomic_inc(&page[i]._mapcount);
2888                 pte_unmap(pte);
2889         }
2890
2891         /*
2892          * Set PG_double_map before dropping compound_mapcount to avoid
2893          * false-negative page_mapped().
2894          */
2895         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2896                 for (i = 0; i < HPAGE_PMD_NR; i++)
2897                         atomic_inc(&page[i]._mapcount);
2898         }
2899
2900         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2901                 /* Last compound_mapcount is gone. */
2902                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2903                 if (TestClearPageDoubleMap(page)) {
2904                         /* No need in mapcount reference anymore */
2905                         for (i = 0; i < HPAGE_PMD_NR; i++)
2906                                 atomic_dec(&page[i]._mapcount);
2907                 }
2908         }
2909
2910         smp_wmb(); /* make pte visible before pmd */
2911         /*
2912          * Up to this point the pmd is present and huge and userland has the
2913          * whole access to the hugepage during the split (which happens in
2914          * place). If we overwrite the pmd with the not-huge version pointing
2915          * to the pte here (which of course we could if all CPUs were bug
2916          * free), userland could trigger a small page size TLB miss on the
2917          * small sized TLB while the hugepage TLB entry is still established in
2918          * the huge TLB. Some CPU doesn't like that.
2919          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2920          * 383 on page 93. Intel should be safe but is also warns that it's
2921          * only safe if the permission and cache attributes of the two entries
2922          * loaded in the two TLB is identical (which should be the case here).
2923          * But it is generally safer to never allow small and huge TLB entries
2924          * for the same virtual address to be loaded simultaneously. So instead
2925          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2926          * current pmd notpresent (atomically because here the pmd_trans_huge
2927          * and pmd_trans_splitting must remain set at all times on the pmd
2928          * until the split is complete for this pmd), then we flush the SMP TLB
2929          * and finally we write the non-huge version of the pmd entry with
2930          * pmd_populate.
2931          */
2932         pmdp_invalidate(vma, haddr, pmd);
2933         pmd_populate(mm, pmd, pgtable);
2934
2935         if (freeze) {
2936                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2937                         page_remove_rmap(page + i, false);
2938                         put_page(page + i);
2939                 }
2940         }
2941 }
2942
2943 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2944                 unsigned long address)
2945 {
2946         spinlock_t *ptl;
2947         struct mm_struct *mm = vma->vm_mm;
2948         struct page *page = NULL;
2949         unsigned long haddr = address & HPAGE_PMD_MASK;
2950
2951         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2952         ptl = pmd_lock(mm, pmd);
2953         if (pmd_trans_huge(*pmd)) {
2954                 page = pmd_page(*pmd);
2955                 if (PageMlocked(page))
2956                         get_page(page);
2957                 else
2958                         page = NULL;
2959         } else if (!pmd_devmap(*pmd))
2960                 goto out;
2961         __split_huge_pmd_locked(vma, pmd, haddr, false);
2962 out:
2963         spin_unlock(ptl);
2964         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2965         if (page) {
2966                 lock_page(page);
2967                 munlock_vma_page(page);
2968                 unlock_page(page);
2969                 put_page(page);
2970         }
2971 }
2972
2973 static void split_huge_pmd_address(struct vm_area_struct *vma,
2974                                     unsigned long address)
2975 {
2976         pgd_t *pgd;
2977         pud_t *pud;
2978         pmd_t *pmd;
2979
2980         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2981
2982         pgd = pgd_offset(vma->vm_mm, address);
2983         if (!pgd_present(*pgd))
2984                 return;
2985
2986         pud = pud_offset(pgd, address);
2987         if (!pud_present(*pud))
2988                 return;
2989
2990         pmd = pmd_offset(pud, address);
2991         if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
2992                 return;
2993         /*
2994          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2995          * materialize from under us.
2996          */
2997         split_huge_pmd(vma, pmd, address);
2998 }
2999
3000 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3001                              unsigned long start,
3002                              unsigned long end,
3003                              long adjust_next)
3004 {
3005         /*
3006          * If the new start address isn't hpage aligned and it could
3007          * previously contain an hugepage: check if we need to split
3008          * an huge pmd.
3009          */
3010         if (start & ~HPAGE_PMD_MASK &&
3011             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3012             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3013                 split_huge_pmd_address(vma, start);
3014
3015         /*
3016          * If the new end address isn't hpage aligned and it could
3017          * previously contain an hugepage: check if we need to split
3018          * an huge pmd.
3019          */
3020         if (end & ~HPAGE_PMD_MASK &&
3021             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3022             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3023                 split_huge_pmd_address(vma, end);
3024
3025         /*
3026          * If we're also updating the vma->vm_next->vm_start, if the new
3027          * vm_next->vm_start isn't page aligned and it could previously
3028          * contain an hugepage: check if we need to split an huge pmd.
3029          */
3030         if (adjust_next > 0) {
3031                 struct vm_area_struct *next = vma->vm_next;
3032                 unsigned long nstart = next->vm_start;
3033                 nstart += adjust_next << PAGE_SHIFT;
3034                 if (nstart & ~HPAGE_PMD_MASK &&
3035                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3036                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3037                         split_huge_pmd_address(next, nstart);
3038         }
3039 }
3040
3041 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
3042                 unsigned long address)
3043 {
3044         unsigned long haddr = address & HPAGE_PMD_MASK;
3045         spinlock_t *ptl;
3046         pgd_t *pgd;
3047         pud_t *pud;
3048         pmd_t *pmd;
3049         pte_t *pte;
3050         int i, nr = HPAGE_PMD_NR;
3051
3052         /* Skip pages which doesn't belong to the VMA */
3053         if (address < vma->vm_start) {
3054                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3055                 page += off;
3056                 nr -= off;
3057                 address = vma->vm_start;
3058         }
3059
3060         pgd = pgd_offset(vma->vm_mm, address);
3061         if (!pgd_present(*pgd))
3062                 return;
3063         pud = pud_offset(pgd, address);
3064         if (!pud_present(*pud))
3065                 return;
3066         pmd = pmd_offset(pud, address);
3067         ptl = pmd_lock(vma->vm_mm, pmd);
3068         if (!pmd_present(*pmd)) {
3069                 spin_unlock(ptl);
3070                 return;
3071         }
3072         if (pmd_trans_huge(*pmd)) {
3073                 if (page == pmd_page(*pmd))
3074                         __split_huge_pmd_locked(vma, pmd, haddr, true);
3075                 spin_unlock(ptl);
3076                 return;
3077         }
3078         spin_unlock(ptl);
3079
3080         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3081         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3082                 pte_t entry, swp_pte;
3083                 swp_entry_t swp_entry;
3084
3085                 /*
3086                  * We've just crossed page table boundary: need to map next one.
3087                  * It can happen if THP was mremaped to non PMD-aligned address.
3088                  */
3089                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3090                         pte_unmap_unlock(pte - 1, ptl);
3091                         pmd = mm_find_pmd(vma->vm_mm, address);
3092                         if (!pmd)
3093                                 return;
3094                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3095                                         address, &ptl);
3096                 }
3097
3098                 if (!pte_present(*pte))
3099                         continue;
3100                 if (page_to_pfn(page) != pte_pfn(*pte))
3101                         continue;
3102                 flush_cache_page(vma, address, page_to_pfn(page));
3103                 entry = ptep_clear_flush(vma, address, pte);
3104                 if (pte_dirty(entry))
3105                         SetPageDirty(page);
3106                 swp_entry = make_migration_entry(page, pte_write(entry));
3107                 swp_pte = swp_entry_to_pte(swp_entry);
3108                 if (pte_soft_dirty(entry))
3109                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
3110                 set_pte_at(vma->vm_mm, address, pte, swp_pte);
3111                 page_remove_rmap(page, false);
3112                 put_page(page);
3113         }
3114         pte_unmap_unlock(pte - 1, ptl);
3115 }
3116
3117 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
3118 {
3119         struct anon_vma_chain *avc;
3120         pgoff_t pgoff = page_to_pgoff(page);
3121
3122         VM_BUG_ON_PAGE(!PageHead(page), page);
3123
3124         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
3125                         pgoff + HPAGE_PMD_NR - 1) {
3126                 unsigned long address = __vma_address(page, avc->vma);
3127
3128                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3129                                 address, address + HPAGE_PMD_SIZE);
3130                 freeze_page_vma(avc->vma, page, address);
3131                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3132                                 address, address + HPAGE_PMD_SIZE);
3133         }
3134 }
3135
3136 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
3137                 unsigned long address)
3138 {
3139         spinlock_t *ptl;
3140         pmd_t *pmd;
3141         pte_t *pte, entry;
3142         swp_entry_t swp_entry;
3143         unsigned long haddr = address & HPAGE_PMD_MASK;
3144         int i, nr = HPAGE_PMD_NR;
3145
3146         /* Skip pages which doesn't belong to the VMA */
3147         if (address < vma->vm_start) {
3148                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3149                 page += off;
3150                 nr -= off;
3151                 address = vma->vm_start;
3152         }
3153
3154         pmd = mm_find_pmd(vma->vm_mm, address);
3155         if (!pmd)
3156                 return;
3157
3158         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3159         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3160                 /*
3161                  * We've just crossed page table boundary: need to map next one.
3162                  * It can happen if THP was mremaped to non-PMD aligned address.
3163                  */
3164                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3165                         pte_unmap_unlock(pte - 1, ptl);
3166                         pmd = mm_find_pmd(vma->vm_mm, address);
3167                         if (!pmd)
3168                                 return;
3169                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3170                                         address, &ptl);
3171                 }
3172
3173                 if (!is_swap_pte(*pte))
3174                         continue;
3175
3176                 swp_entry = pte_to_swp_entry(*pte);
3177                 if (!is_migration_entry(swp_entry))
3178                         continue;
3179                 if (migration_entry_to_page(swp_entry) != page)
3180                         continue;
3181
3182                 get_page(page);
3183                 page_add_anon_rmap(page, vma, address, false);
3184
3185                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3186                 if (PageDirty(page))
3187                         entry = pte_mkdirty(entry);
3188                 if (is_write_migration_entry(swp_entry))
3189                         entry = maybe_mkwrite(entry, vma);
3190
3191                 flush_dcache_page(page);
3192                 set_pte_at(vma->vm_mm, address, pte, entry);
3193
3194                 /* No need to invalidate - it was non-present before */
3195                 update_mmu_cache(vma, address, pte);
3196         }
3197         pte_unmap_unlock(pte - 1, ptl);
3198 }
3199
3200 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3201 {
3202         struct anon_vma_chain *avc;
3203         pgoff_t pgoff = page_to_pgoff(page);
3204
3205         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3206                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3207                 unsigned long address = __vma_address(page, avc->vma);
3208
3209                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3210                                 address, address + HPAGE_PMD_SIZE);
3211                 unfreeze_page_vma(avc->vma, page, address);
3212                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3213                                 address, address + HPAGE_PMD_SIZE);
3214         }
3215 }
3216
3217 static int __split_huge_page_tail(struct page *head, int tail,
3218                 struct lruvec *lruvec, struct list_head *list)
3219 {
3220         int mapcount;
3221         struct page *page_tail = head + tail;
3222
3223         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3224         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3225
3226         /*
3227          * tail_page->_count is zero and not changing from under us. But
3228          * get_page_unless_zero() may be running from under us on the
3229          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3230          * would then run atomic_set() concurrently with
3231          * get_page_unless_zero(), and atomic_set() is implemented in C not
3232          * using locked ops. spin_unlock on x86 sometime uses locked ops
3233          * because of PPro errata 66, 92, so unless somebody can guarantee
3234          * atomic_set() here would be safe on all archs (and not only on x86),
3235          * it's safer to use atomic_add().
3236          */
3237         atomic_add(mapcount + 1, &page_tail->_count);
3238
3239
3240         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3241         page_tail->flags |= (head->flags &
3242                         ((1L << PG_referenced) |
3243                          (1L << PG_swapbacked) |
3244                          (1L << PG_mlocked) |
3245                          (1L << PG_uptodate) |
3246                          (1L << PG_active) |
3247                          (1L << PG_locked) |
3248                          (1L << PG_unevictable) |
3249                          (1L << PG_dirty)));
3250
3251         /*
3252          * After clearing PageTail the gup refcount can be released.
3253          * Page flags also must be visible before we make the page non-compound.
3254          */
3255         smp_wmb();
3256
3257         clear_compound_head(page_tail);
3258
3259         if (page_is_young(head))
3260                 set_page_young(page_tail);
3261         if (page_is_idle(head))
3262                 set_page_idle(page_tail);
3263
3264         /* ->mapping in first tail page is compound_mapcount */
3265         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3266                         page_tail);
3267         page_tail->mapping = head->mapping;
3268
3269         page_tail->index = head->index + tail;
3270         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3271         lru_add_page_tail(head, page_tail, lruvec, list);
3272
3273         return mapcount;
3274 }
3275
3276 static void __split_huge_page(struct page *page, struct list_head *list)
3277 {
3278         struct page *head = compound_head(page);
3279         struct zone *zone = page_zone(head);
3280         struct lruvec *lruvec;
3281         int i, tail_mapcount;
3282
3283         /* prevent PageLRU to go away from under us, and freeze lru stats */
3284         spin_lock_irq(&zone->lru_lock);
3285         lruvec = mem_cgroup_page_lruvec(head, zone);
3286
3287         /* complete memcg works before add pages to LRU */
3288         mem_cgroup_split_huge_fixup(head);
3289
3290         tail_mapcount = 0;
3291         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3292                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3293         atomic_sub(tail_mapcount, &head->_count);
3294
3295         ClearPageCompound(head);
3296         spin_unlock_irq(&zone->lru_lock);
3297
3298         unfreeze_page(page_anon_vma(head), head);
3299
3300         for (i = 0; i < HPAGE_PMD_NR; i++) {
3301                 struct page *subpage = head + i;
3302                 if (subpage == page)
3303                         continue;
3304                 unlock_page(subpage);
3305
3306                 /*
3307                  * Subpages may be freed if there wasn't any mapping
3308                  * like if add_to_swap() is running on a lru page that
3309                  * had its mapping zapped. And freeing these pages
3310                  * requires taking the lru_lock so we do the put_page
3311                  * of the tail pages after the split is complete.
3312                  */
3313                 put_page(subpage);
3314         }
3315 }
3316
3317 int total_mapcount(struct page *page)
3318 {
3319         int i, ret;
3320
3321         VM_BUG_ON_PAGE(PageTail(page), page);
3322
3323         if (likely(!PageCompound(page)))
3324                 return atomic_read(&page->_mapcount) + 1;
3325
3326         ret = compound_mapcount(page);
3327         if (PageHuge(page))
3328                 return ret;
3329         for (i = 0; i < HPAGE_PMD_NR; i++)
3330                 ret += atomic_read(&page[i]._mapcount) + 1;
3331         if (PageDoubleMap(page))
3332                 ret -= HPAGE_PMD_NR;
3333         return ret;
3334 }
3335
3336 /*
3337  * This function splits huge page into normal pages. @page can point to any
3338  * subpage of huge page to split. Split doesn't change the position of @page.
3339  *
3340  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3341  * The huge page must be locked.
3342  *
3343  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3344  *
3345  * Both head page and tail pages will inherit mapping, flags, and so on from
3346  * the hugepage.
3347  *
3348  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3349  * they are not mapped.
3350  *
3351  * Returns 0 if the hugepage is split successfully.
3352  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3353  * us.
3354  */
3355 int split_huge_page_to_list(struct page *page, struct list_head *list)
3356 {
3357         struct page *head = compound_head(page);
3358         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3359         struct anon_vma *anon_vma;
3360         int count, mapcount, ret;
3361         bool mlocked;
3362         unsigned long flags;
3363
3364         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3365         VM_BUG_ON_PAGE(!PageAnon(page), page);
3366         VM_BUG_ON_PAGE(!PageLocked(page), page);
3367         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3368         VM_BUG_ON_PAGE(!PageCompound(page), page);
3369
3370         /*
3371          * The caller does not necessarily hold an mmap_sem that would prevent
3372          * the anon_vma disappearing so we first we take a reference to it
3373          * and then lock the anon_vma for write. This is similar to
3374          * page_lock_anon_vma_read except the write lock is taken to serialise
3375          * against parallel split or collapse operations.
3376          */
3377         anon_vma = page_get_anon_vma(head);
3378         if (!anon_vma) {
3379                 ret = -EBUSY;
3380                 goto out;
3381         }
3382         anon_vma_lock_write(anon_vma);
3383
3384         /*
3385          * Racy check if we can split the page, before freeze_page() will
3386          * split PMDs
3387          */
3388         if (total_mapcount(head) != page_count(head) - 1) {
3389                 ret = -EBUSY;
3390                 goto out_unlock;
3391         }
3392
3393         mlocked = PageMlocked(page);
3394         freeze_page(anon_vma, head);
3395         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3396
3397         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3398         if (mlocked)
3399                 lru_add_drain();
3400
3401         /* Prevent deferred_split_scan() touching ->_count */
3402         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3403         count = page_count(head);
3404         mapcount = total_mapcount(head);
3405         if (!mapcount && count == 1) {
3406                 if (!list_empty(page_deferred_list(head))) {
3407                         pgdata->split_queue_len--;
3408                         list_del(page_deferred_list(head));
3409                 }
3410                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3411                 __split_huge_page(page, list);
3412                 ret = 0;
3413         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3414                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3415                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3416                                 mapcount, count);
3417                 if (PageTail(page))
3418                         dump_page(head, NULL);
3419                 dump_page(page, "total_mapcount(head) > 0");
3420                 BUG();
3421         } else {
3422                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3423                 unfreeze_page(anon_vma, head);
3424                 ret = -EBUSY;
3425         }
3426
3427 out_unlock:
3428         anon_vma_unlock_write(anon_vma);
3429         put_anon_vma(anon_vma);
3430 out:
3431         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3432         return ret;
3433 }
3434
3435 void free_transhuge_page(struct page *page)
3436 {
3437         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3438         unsigned long flags;
3439
3440         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3441         if (!list_empty(page_deferred_list(page))) {
3442                 pgdata->split_queue_len--;
3443                 list_del(page_deferred_list(page));
3444         }
3445         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3446         free_compound_page(page);
3447 }
3448
3449 void deferred_split_huge_page(struct page *page)
3450 {
3451         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3452         unsigned long flags;
3453
3454         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3455
3456         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3457         if (list_empty(page_deferred_list(page))) {
3458                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3459                 pgdata->split_queue_len++;
3460         }
3461         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3462 }
3463
3464 static unsigned long deferred_split_count(struct shrinker *shrink,
3465                 struct shrink_control *sc)
3466 {
3467         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3468         /*
3469          * Split a page from split_queue will free up at least one page,
3470          * at most HPAGE_PMD_NR - 1. We don't track exact number.
3471          * Let's use HPAGE_PMD_NR / 2 as ballpark.
3472          */
3473         return ACCESS_ONCE(pgdata->split_queue_len) * HPAGE_PMD_NR / 2;
3474 }
3475
3476 static unsigned long deferred_split_scan(struct shrinker *shrink,
3477                 struct shrink_control *sc)
3478 {
3479         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3480         unsigned long flags;
3481         LIST_HEAD(list), *pos, *next;
3482         struct page *page;
3483         int split = 0;
3484
3485         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3486         list_splice_init(&pgdata->split_queue, &list);
3487
3488         /* Take pin on all head pages to avoid freeing them under us */
3489         list_for_each_safe(pos, next, &list) {
3490                 page = list_entry((void *)pos, struct page, mapping);
3491                 page = compound_head(page);
3492                 /* race with put_compound_page() */
3493                 if (!get_page_unless_zero(page)) {
3494                         list_del_init(page_deferred_list(page));
3495                         pgdata->split_queue_len--;
3496                 }
3497         }
3498         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3499
3500         list_for_each_safe(pos, next, &list) {
3501                 page = list_entry((void *)pos, struct page, mapping);
3502                 lock_page(page);
3503                 /* split_huge_page() removes page from list on success */
3504                 if (!split_huge_page(page))
3505                         split++;
3506                 unlock_page(page);
3507                 put_page(page);
3508         }
3509
3510         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3511         list_splice_tail(&list, &pgdata->split_queue);
3512         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3513
3514         return split * HPAGE_PMD_NR / 2;
3515 }
3516
3517 static struct shrinker deferred_split_shrinker = {
3518         .count_objects = deferred_split_count,
3519         .scan_objects = deferred_split_scan,
3520         .seeks = DEFAULT_SEEKS,
3521         .flags = SHRINKER_NUMA_AWARE,
3522 };
3523
3524 #ifdef CONFIG_DEBUG_FS
3525 static int split_huge_pages_set(void *data, u64 val)
3526 {
3527         struct zone *zone;
3528         struct page *page;
3529         unsigned long pfn, max_zone_pfn;
3530         unsigned long total = 0, split = 0;
3531
3532         if (val != 1)
3533                 return -EINVAL;
3534
3535         for_each_populated_zone(zone) {
3536                 max_zone_pfn = zone_end_pfn(zone);
3537                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3538                         if (!pfn_valid(pfn))
3539                                 continue;
3540
3541                         page = pfn_to_page(pfn);
3542                         if (!get_page_unless_zero(page))
3543                                 continue;
3544
3545                         if (zone != page_zone(page))
3546                                 goto next;
3547
3548                         if (!PageHead(page) || !PageAnon(page) ||
3549                                         PageHuge(page))
3550                                 goto next;
3551
3552                         total++;
3553                         lock_page(page);
3554                         if (!split_huge_page(page))
3555                                 split++;
3556                         unlock_page(page);
3557 next:
3558                         put_page(page);
3559                 }
3560         }
3561
3562         pr_info("%lu of %lu THP split", split, total);
3563
3564         return 0;
3565 }
3566 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3567                 "%llu\n");
3568
3569 static int __init split_huge_pages_debugfs(void)
3570 {
3571         void *ret;
3572
3573         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3574                         &split_huge_pages_fops);
3575         if (!ret)
3576                 pr_warn("Failed to create split_huge_pages in debugfs");
3577         return 0;
3578 }
3579 late_initcall(split_huge_pages_debugfs);
3580 #endif