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         if (pgtable)
 862                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 863         set_pmd_at(mm, haddr, pmd, entry);
 864         atomic_long_inc(&mm->nr_ptes);
 865         return true;
 866 }
 867
 868 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 869                                unsigned long address, pmd_t *pmd,
 870                                unsigned int flags)
 871 {
 872         gfp_t gfp;
 873         struct page *page;
 874         unsigned long haddr = address & HPAGE_PMD_MASK;
 875
 876         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 877                 return VM_FAULT_FALLBACK;
 878         if (unlikely(anon_vma_prepare(vma)))
 879                 return VM_FAULT_OOM;
 880         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 881                 return VM_FAULT_OOM;
 882         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 883                         transparent_hugepage_use_zero_page()) {
 884                 spinlock_t *ptl;
 885                 pgtable_t pgtable;
 886                 struct page *zero_page;
 887                 bool set;
 888                 int ret;
 889                 pgtable = pte_alloc_one(mm, haddr);
 890                 if (unlikely(!pgtable))
 891                         return VM_FAULT_OOM;
 892                 zero_page = get_huge_zero_page();
 893                 if (unlikely(!zero_page)) {
 894                         pte_free(mm, pgtable);
 895                         count_vm_event(THP_FAULT_FALLBACK);
 896                         return VM_FAULT_FALLBACK;
 897                 }
 898                 ptl = pmd_lock(mm, pmd);
 899                 ret = 0;
 900                 set = false;
 901                 if (pmd_none(*pmd)) {
 902                         if (userfaultfd_missing(vma)) {
 903                                 spin_unlock(ptl);
 904                                 ret = handle_userfault(vma, address, flags,
 905                                                        VM_UFFD_MISSING);
 906                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 907                         } else {
 908                                 set_huge_zero_page(pgtable, mm, vma,
 909                                                    haddr, pmd,
 910                                                    zero_page);
 911                                 spin_unlock(ptl);
 912                                 set = true;
 913                         }
 914                 } else
 915                         spin_unlock(ptl);
 916                 if (!set) {
 917                         pte_free(mm, pgtable);
 918                         put_huge_zero_page();
 919                 }
 920                 return ret;
 921         }
 922         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 923         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 924         if (unlikely(!page)) {
 925                 count_vm_event(THP_FAULT_FALLBACK);
 926                 return VM_FAULT_FALLBACK;
 927         }
 928         prep_transhuge_page(page);
 929         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 930                                             flags);
 931 }
 932
 933 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 934                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 935 {
 936         struct mm_struct *mm = vma->vm_mm;
 937         pmd_t entry;
 938         spinlock_t *ptl;
 939
 940         ptl = pmd_lock(mm, pmd);
 941         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 942         if (pfn_t_devmap(pfn))
 943                 entry = pmd_mkdevmap(entry);
 944         if (write) {
 945                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 946                 entry = maybe_pmd_mkwrite(entry, vma);
 947         }
 948         set_pmd_at(mm, addr, pmd, entry);
 949         update_mmu_cache_pmd(vma, addr, pmd);
 950         spin_unlock(ptl);
 951 }
 952
 953 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 954                         pmd_t *pmd, pfn_t pfn, bool write)
 955 {
 956         pgprot_t pgprot = vma->vm_page_prot;
 957         /*
 958          * If we had pmd_special, we could avoid all these restrictions,
 959          * but we need to be consistent with PTEs and architectures that
 960          * can't support a 'special' bit.
 961          */
 962         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 963         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 964                                                 (VM_PFNMAP|VM_MIXEDMAP));
 965         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 966         BUG_ON(!pfn_t_devmap(pfn));
 967
 968         if (addr < vma->vm_start || addr >= vma->vm_end)
 969                 return VM_FAULT_SIGBUS;
 970         if (track_pfn_insert(vma, &pgprot, pfn))
 971                 return VM_FAULT_SIGBUS;
 972         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 973         return VM_FAULT_NOPAGE;
 974 }
 975
 976 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 977                 pmd_t *pmd)
 978 {
 979         pmd_t _pmd;
 980
 981         /*
 982          * We should set the dirty bit only for FOLL_WRITE but for now
 983          * the dirty bit in the pmd is meaningless.  And if the dirty
 984          * bit will become meaningful and we'll only set it with
 985          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 986          * set the young bit, instead of the current set_pmd_at.
 987          */
 988         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 989         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 990                                 pmd, _pmd,  1))
 991                 update_mmu_cache_pmd(vma, addr, pmd);
 992 }
 993
 994 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 995                 pmd_t *pmd, int flags)
 996 {
 997         unsigned long pfn = pmd_pfn(*pmd);
 998         struct mm_struct *mm = vma->vm_mm;
 999         struct dev_pagemap *pgmap;
1000         struct page *page;
1001
1002         assert_spin_locked(pmd_lockptr(mm, pmd));
1003
1004         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1005                 return NULL;
1006
1007         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1008                 /* pass */;
1009         else
1010                 return NULL;
1011
1012         if (flags & FOLL_TOUCH)
1013                 touch_pmd(vma, addr, pmd);
1014
1015         /*
1016          * device mapped pages can only be returned if the
1017          * caller will manage the page reference count.
1018          */
1019         if (!(flags & FOLL_GET))
1020                 return ERR_PTR(-EEXIST);
1021
1022         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1023         pgmap = get_dev_pagemap(pfn, NULL);
1024         if (!pgmap)
1025                 return ERR_PTR(-EFAULT);
1026         page = pfn_to_page(pfn);
1027         get_page(page);
1028         put_dev_pagemap(pgmap);
1029
1030         return page;
1031 }
1032
1033 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1034                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1035                   struct vm_area_struct *vma)
1036 {
1037         spinlock_t *dst_ptl, *src_ptl;
1038         struct page *src_page;
1039         pmd_t pmd;
1040         pgtable_t pgtable = NULL;
1041         int ret;
1042
1043         if (!vma_is_dax(vma)) {
1044                 ret = -ENOMEM;
1045                 pgtable = pte_alloc_one(dst_mm, addr);
1046                 if (unlikely(!pgtable))
1047                         goto out;
1048         }
1049
1050         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1051         src_ptl = pmd_lockptr(src_mm, src_pmd);
1052         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1053
1054         ret = -EAGAIN;
1055         pmd = *src_pmd;
1056         if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1057                 pte_free(dst_mm, pgtable);
1058                 goto out_unlock;
1059         }
1060         /*
1061          * When page table lock is held, the huge zero pmd should not be
1062          * under splitting since we don't split the page itself, only pmd to
1063          * a page table.
1064          */
1065         if (is_huge_zero_pmd(pmd)) {
1066                 struct page *zero_page;
1067                 /*
1068                  * get_huge_zero_page() will never allocate a new page here,
1069                  * since we already have a zero page to copy. It just takes a
1070                  * reference.
1071                  */
1072                 zero_page = get_huge_zero_page();
1073                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1074                                 zero_page);
1075                 ret = 0;
1076                 goto out_unlock;
1077         }
1078
1079         if (!vma_is_dax(vma)) {
1080                 /* thp accounting separate from pmd_devmap accounting */
1081                 src_page = pmd_page(pmd);
1082                 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1083                 get_page(src_page);
1084                 page_dup_rmap(src_page, true);
1085                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1086                 atomic_long_inc(&dst_mm->nr_ptes);
1087                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1088         }
1089
1090         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1091         pmd = pmd_mkold(pmd_wrprotect(pmd));
1092         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1093
1094         ret = 0;
1095 out_unlock:
1096         spin_unlock(src_ptl);
1097         spin_unlock(dst_ptl);
1098 out:
1099         return ret;
1100 }
1101
1102 void huge_pmd_set_accessed(struct mm_struct *mm,
1103                            struct vm_area_struct *vma,
1104                            unsigned long address,
1105                            pmd_t *pmd, pmd_t orig_pmd,
1106                            int dirty)
1107 {
1108         spinlock_t *ptl;
1109         pmd_t entry;
1110         unsigned long haddr;
1111
1112         ptl = pmd_lock(mm, pmd);
1113         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1114                 goto unlock;
1115
1116         entry = pmd_mkyoung(orig_pmd);
1117         haddr = address & HPAGE_PMD_MASK;
1118         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1119                 update_mmu_cache_pmd(vma, address, pmd);
1120
1121 unlock:
1122         spin_unlock(ptl);
1123 }
1124
1125 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1126                                         struct vm_area_struct *vma,
1127                                         unsigned long address,
1128                                         pmd_t *pmd, pmd_t orig_pmd,
1129                                         struct page *page,
1130                                         unsigned long haddr)
1131 {
1132         struct mem_cgroup *memcg;
1133         spinlock_t *ptl;
1134         pgtable_t pgtable;
1135         pmd_t _pmd;
1136         int ret = 0, i;
1137         struct page **pages;
1138         unsigned long mmun_start;       /* For mmu_notifiers */
1139         unsigned long mmun_end;         /* For mmu_notifiers */
1140
1141         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1142                         GFP_KERNEL);
1143         if (unlikely(!pages)) {
1144                 ret |= VM_FAULT_OOM;
1145                 goto out;
1146         }
1147
1148         for (i = 0; i < HPAGE_PMD_NR; i++) {
1149                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1150                                                __GFP_OTHER_NODE,
1151                                                vma, address, page_to_nid(page));
1152                 if (unlikely(!pages[i] ||
1153                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1154                                                    &memcg, false))) {
1155                         if (pages[i])
1156                                 put_page(pages[i]);
1157                         while (--i >= 0) {
1158                                 memcg = (void *)page_private(pages[i]);
1159                                 set_page_private(pages[i], 0);
1160                                 mem_cgroup_cancel_charge(pages[i], memcg,
1161                                                 false);
1162                                 put_page(pages[i]);
1163                         }
1164                         kfree(pages);
1165                         ret |= VM_FAULT_OOM;
1166                         goto out;
1167                 }
1168                 set_page_private(pages[i], (unsigned long)memcg);
1169         }
1170
1171         for (i = 0; i < HPAGE_PMD_NR; i++) {
1172                 copy_user_highpage(pages[i], page + i,
1173                                    haddr + PAGE_SIZE * i, vma);
1174                 __SetPageUptodate(pages[i]);
1175                 cond_resched();
1176         }
1177
1178         mmun_start = haddr;
1179         mmun_end   = haddr + HPAGE_PMD_SIZE;
1180         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1181
1182         ptl = pmd_lock(mm, pmd);
1183         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1184                 goto out_free_pages;
1185         VM_BUG_ON_PAGE(!PageHead(page), page);
1186
1187         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1188         /* leave pmd empty until pte is filled */
1189
1190         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1191         pmd_populate(mm, &_pmd, pgtable);
1192
1193         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1194                 pte_t *pte, entry;
1195                 entry = mk_pte(pages[i], vma->vm_page_prot);
1196                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1197                 memcg = (void *)page_private(pages[i]);
1198                 set_page_private(pages[i], 0);
1199                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1200                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1201                 lru_cache_add_active_or_unevictable(pages[i], vma);
1202                 pte = pte_offset_map(&_pmd, haddr);
1203                 VM_BUG_ON(!pte_none(*pte));
1204                 set_pte_at(mm, haddr, pte, entry);
1205                 pte_unmap(pte);
1206         }
1207         kfree(pages);
1208
1209         smp_wmb(); /* make pte visible before pmd */
1210         pmd_populate(mm, pmd, pgtable);
1211         page_remove_rmap(page, true);
1212         spin_unlock(ptl);
1213
1214         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1215
1216         ret |= VM_FAULT_WRITE;
1217         put_page(page);
1218
1219 out:
1220         return ret;
1221
1222 out_free_pages:
1223         spin_unlock(ptl);
1224         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1225         for (i = 0; i < HPAGE_PMD_NR; i++) {
1226                 memcg = (void *)page_private(pages[i]);
1227                 set_page_private(pages[i], 0);
1228                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1229                 put_page(pages[i]);
1230         }
1231         kfree(pages);
1232         goto out;
1233 }
1234
1235 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1236                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1237 {
1238         spinlock_t *ptl;
1239         int ret = 0;
1240         struct page *page = NULL, *new_page;
1241         struct mem_cgroup *memcg;
1242         unsigned long haddr;
1243         unsigned long mmun_start;       /* For mmu_notifiers */
1244         unsigned long mmun_end;         /* For mmu_notifiers */
1245         gfp_t huge_gfp;                 /* for allocation and charge */
1246
1247         ptl = pmd_lockptr(mm, pmd);
1248         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1249         haddr = address & HPAGE_PMD_MASK;
1250         if (is_huge_zero_pmd(orig_pmd))
1251                 goto alloc;
1252         spin_lock(ptl);
1253         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1254                 goto out_unlock;
1255
1256         page = pmd_page(orig_pmd);
1257         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1258         /*
1259          * We can only reuse the page if nobody else maps the huge page or it's
1260          * part. We can do it by checking page_mapcount() on each sub-page, but
1261          * it's expensive.
1262          * The cheaper way is to check page_count() to be equal 1: every
1263          * mapcount takes page reference reference, so this way we can
1264          * guarantee, that the PMD is the only mapping.
1265          * This can give false negative if somebody pinned the page, but that's
1266          * fine.
1267          */
1268         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1269                 pmd_t entry;
1270                 entry = pmd_mkyoung(orig_pmd);
1271                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1272                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1273                         update_mmu_cache_pmd(vma, address, pmd);
1274                 ret |= VM_FAULT_WRITE;
1275                 goto out_unlock;
1276         }
1277         get_page(page);
1278         spin_unlock(ptl);
1279 alloc:
1280         if (transparent_hugepage_enabled(vma) &&
1281             !transparent_hugepage_debug_cow()) {
1282                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1283                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1284         } else
1285                 new_page = NULL;
1286
1287         if (likely(new_page)) {
1288                 prep_transhuge_page(new_page);
1289         } else {
1290                 if (!page) {
1291                         split_huge_pmd(vma, pmd, address);
1292                         ret |= VM_FAULT_FALLBACK;
1293                 } else {
1294                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1295                                         pmd, orig_pmd, page, haddr);
1296                         if (ret & VM_FAULT_OOM) {
1297                                 split_huge_pmd(vma, pmd, address);
1298                                 ret |= VM_FAULT_FALLBACK;
1299                         }
1300                         put_page(page);
1301                 }
1302                 count_vm_event(THP_FAULT_FALLBACK);
1303                 goto out;
1304         }
1305
1306         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1307                                            true))) {
1308                 put_page(new_page);
1309                 if (page) {
1310                         split_huge_pmd(vma, pmd, address);
1311                         put_page(page);
1312                 } else
1313                         split_huge_pmd(vma, pmd, address);
1314                 ret |= VM_FAULT_FALLBACK;
1315                 count_vm_event(THP_FAULT_FALLBACK);
1316                 goto out;
1317         }
1318
1319         count_vm_event(THP_FAULT_ALLOC);
1320
1321         if (!page)
1322                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1323         else
1324                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1325         __SetPageUptodate(new_page);
1326
1327         mmun_start = haddr;
1328         mmun_end   = haddr + HPAGE_PMD_SIZE;
1329         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1330
1331         spin_lock(ptl);
1332         if (page)
1333                 put_page(page);
1334         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1335                 spin_unlock(ptl);
1336                 mem_cgroup_cancel_charge(new_page, memcg, true);
1337                 put_page(new_page);
1338                 goto out_mn;
1339         } else {
1340                 pmd_t entry;
1341                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1342                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1343                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1344                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1345                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1346                 lru_cache_add_active_or_unevictable(new_page, vma);
1347                 set_pmd_at(mm, haddr, pmd, entry);
1348                 update_mmu_cache_pmd(vma, address, pmd);
1349                 if (!page) {
1350                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1351                         put_huge_zero_page();
1352                 } else {
1353                         VM_BUG_ON_PAGE(!PageHead(page), page);
1354                         page_remove_rmap(page, true);
1355                         put_page(page);
1356                 }
1357                 ret |= VM_FAULT_WRITE;
1358         }
1359         spin_unlock(ptl);
1360 out_mn:
1361         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1362 out:
1363         return ret;
1364 out_unlock:
1365         spin_unlock(ptl);
1366         return ret;
1367 }
1368
1369 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1370                                    unsigned long addr,
1371                                    pmd_t *pmd,
1372                                    unsigned int flags)
1373 {
1374         struct mm_struct *mm = vma->vm_mm;
1375         struct page *page = NULL;
1376
1377         assert_spin_locked(pmd_lockptr(mm, pmd));
1378
1379         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1380                 goto out;
1381
1382         /* Avoid dumping huge zero page */
1383         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1384                 return ERR_PTR(-EFAULT);
1385
1386         /* Full NUMA hinting faults to serialise migration in fault paths */
1387         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1388                 goto out;
1389
1390         page = pmd_page(*pmd);
1391         VM_BUG_ON_PAGE(!PageHead(page), page);
1392         if (flags & FOLL_TOUCH)
1393                 touch_pmd(vma, addr, pmd);
1394         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1395                 /*
1396                  * We don't mlock() pte-mapped THPs. This way we can avoid
1397                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1398                  *
1399                  * In most cases the pmd is the only mapping of the page as we
1400                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1401                  * writable private mappings in populate_vma_page_range().
1402                  *
1403                  * The only scenario when we have the page shared here is if we
1404                  * mlocking read-only mapping shared over fork(). We skip
1405                  * mlocking such pages.
1406                  */
1407                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1408                                 page->mapping && trylock_page(page)) {
1409                         lru_add_drain();
1410                         if (page->mapping)
1411                                 mlock_vma_page(page);
1412                         unlock_page(page);
1413                 }
1414         }
1415         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1416         VM_BUG_ON_PAGE(!PageCompound(page), page);
1417         if (flags & FOLL_GET)
1418                 get_page(page);
1419
1420 out:
1421         return page;
1422 }
1423
1424 /* NUMA hinting page fault entry point for trans huge pmds */
1425 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1426                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1427 {
1428         spinlock_t *ptl;
1429         struct anon_vma *anon_vma = NULL;
1430         struct page *page;
1431         unsigned long haddr = addr & HPAGE_PMD_MASK;
1432         int page_nid = -1, this_nid = numa_node_id();
1433         int target_nid, last_cpupid = -1;
1434         bool page_locked;
1435         bool migrated = false;
1436         bool was_writable;
1437         int flags = 0;
1438
1439         /* A PROT_NONE fault should not end up here */
1440         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1441
1442         ptl = pmd_lock(mm, pmdp);
1443         if (unlikely(!pmd_same(pmd, *pmdp)))
1444                 goto out_unlock;
1445
1446         /*
1447          * If there are potential migrations, wait for completion and retry
1448          * without disrupting NUMA hinting information. Do not relock and
1449          * check_same as the page may no longer be mapped.
1450          */
1451         if (unlikely(pmd_trans_migrating(*pmdp))) {
1452                 page = pmd_page(*pmdp);
1453                 spin_unlock(ptl);
1454                 wait_on_page_locked(page);
1455                 goto out;
1456         }
1457
1458         page = pmd_page(pmd);
1459         BUG_ON(is_huge_zero_page(page));
1460         page_nid = page_to_nid(page);
1461         last_cpupid = page_cpupid_last(page);
1462         count_vm_numa_event(NUMA_HINT_FAULTS);
1463         if (page_nid == this_nid) {
1464                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1465                 flags |= TNF_FAULT_LOCAL;
1466         }
1467
1468         /* See similar comment in do_numa_page for explanation */
1469         if (!(vma->vm_flags & VM_WRITE))
1470                 flags |= TNF_NO_GROUP;
1471
1472         /*
1473          * Acquire the page lock to serialise THP migrations but avoid dropping
1474          * page_table_lock if at all possible
1475          */
1476         page_locked = trylock_page(page);
1477         target_nid = mpol_misplaced(page, vma, haddr);
1478         if (target_nid == -1) {
1479                 /* If the page was locked, there are no parallel migrations */
1480                 if (page_locked)
1481                         goto clear_pmdnuma;
1482         }
1483
1484         /* Migration could have started since the pmd_trans_migrating check */
1485         if (!page_locked) {
1486                 spin_unlock(ptl);
1487                 wait_on_page_locked(page);
1488                 page_nid = -1;
1489                 goto out;
1490         }
1491
1492         /*
1493          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1494          * to serialises splits
1495          */
1496         get_page(page);
1497         spin_unlock(ptl);
1498         anon_vma = page_lock_anon_vma_read(page);
1499
1500         /* Confirm the PMD did not change while page_table_lock was released */
1501         spin_lock(ptl);
1502         if (unlikely(!pmd_same(pmd, *pmdp))) {
1503                 unlock_page(page);
1504                 put_page(page);
1505                 page_nid = -1;
1506                 goto out_unlock;
1507         }
1508
1509         /* Bail if we fail to protect against THP splits for any reason */
1510         if (unlikely(!anon_vma)) {
1511                 put_page(page);
1512                 page_nid = -1;
1513                 goto clear_pmdnuma;
1514         }
1515
1516         /*
1517          * Migrate the THP to the requested node, returns with page unlocked
1518          * and access rights restored.
1519          */
1520         spin_unlock(ptl);
1521         migrated = migrate_misplaced_transhuge_page(mm, vma,
1522                                 pmdp, pmd, addr, page, target_nid);
1523         if (migrated) {
1524                 flags |= TNF_MIGRATED;
1525                 page_nid = target_nid;
1526         } else
1527                 flags |= TNF_MIGRATE_FAIL;
1528
1529         goto out;
1530 clear_pmdnuma:
1531         BUG_ON(!PageLocked(page));
1532         was_writable = pmd_write(pmd);
1533         pmd = pmd_modify(pmd, vma->vm_page_prot);
1534         pmd = pmd_mkyoung(pmd);
1535         if (was_writable)
1536                 pmd = pmd_mkwrite(pmd);
1537         set_pmd_at(mm, haddr, pmdp, pmd);
1538         update_mmu_cache_pmd(vma, addr, pmdp);
1539         unlock_page(page);
1540 out_unlock:
1541         spin_unlock(ptl);
1542
1543 out:
1544         if (anon_vma)
1545                 page_unlock_anon_vma_read(anon_vma);
1546
1547         if (page_nid != -1)
1548                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1549
1550         return 0;
1551 }
1552
1553 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1554                 pmd_t *pmd, unsigned long addr, unsigned long next)
1555
1556 {
1557         spinlock_t *ptl;
1558         pmd_t orig_pmd;
1559         struct page *page;
1560         struct mm_struct *mm = tlb->mm;
1561         int ret = 0;
1562
1563         ptl = pmd_trans_huge_lock(pmd, vma);
1564         if (!ptl)
1565                 goto out_unlocked;
1566
1567         orig_pmd = *pmd;
1568         if (is_huge_zero_pmd(orig_pmd)) {
1569                 ret = 1;
1570                 goto out;
1571         }
1572
1573         page = pmd_page(orig_pmd);
1574         /*
1575          * If other processes are mapping this page, we couldn't discard
1576          * the page unless they all do MADV_FREE so let's skip the page.
1577          */
1578         if (page_mapcount(page) != 1)
1579                 goto out;
1580
1581         if (!trylock_page(page))
1582                 goto out;
1583
1584         /*
1585          * If user want to discard part-pages of THP, split it so MADV_FREE
1586          * will deactivate only them.
1587          */
1588         if (next - addr != HPAGE_PMD_SIZE) {
1589                 get_page(page);
1590                 spin_unlock(ptl);
1591                 if (split_huge_page(page)) {
1592                         put_page(page);
1593                         unlock_page(page);
1594                         goto out_unlocked;
1595                 }
1596                 put_page(page);
1597                 unlock_page(page);
1598                 ret = 1;
1599                 goto out_unlocked;
1600         }
1601
1602         if (PageDirty(page))
1603                 ClearPageDirty(page);
1604         unlock_page(page);
1605
1606         if (PageActive(page))
1607                 deactivate_page(page);
1608
1609         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1610                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1611                         tlb->fullmm);
1612                 orig_pmd = pmd_mkold(orig_pmd);
1613                 orig_pmd = pmd_mkclean(orig_pmd);
1614
1615                 set_pmd_at(mm, addr, pmd, orig_pmd);
1616                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1617         }
1618         ret = 1;
1619 out:
1620         spin_unlock(ptl);
1621 out_unlocked:
1622         return ret;
1623 }
1624
1625 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1626                  pmd_t *pmd, unsigned long addr)
1627 {
1628         pmd_t orig_pmd;
1629         spinlock_t *ptl;
1630
1631         ptl = __pmd_trans_huge_lock(pmd, vma);
1632         if (!ptl)
1633                 return 0;
1634         /*
1635          * For architectures like ppc64 we look at deposited pgtable
1636          * when calling pmdp_huge_get_and_clear. So do the
1637          * pgtable_trans_huge_withdraw after finishing pmdp related
1638          * operations.
1639          */
1640         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1641                         tlb->fullmm);
1642         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1643         if (vma_is_dax(vma)) {
1644                 spin_unlock(ptl);
1645                 if (is_huge_zero_pmd(orig_pmd))
1646                         put_huge_zero_page();
1647         } else if (is_huge_zero_pmd(orig_pmd)) {
1648                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1649                 atomic_long_dec(&tlb->mm->nr_ptes);
1650                 spin_unlock(ptl);
1651                 put_huge_zero_page();
1652         } else {
1653                 struct page *page = pmd_page(orig_pmd);
1654                 page_remove_rmap(page, true);
1655                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1656                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1657                 VM_BUG_ON_PAGE(!PageHead(page), page);
1658                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1659                 atomic_long_dec(&tlb->mm->nr_ptes);
1660                 spin_unlock(ptl);
1661                 tlb_remove_page(tlb, page);
1662         }
1663         return 1;
1664 }
1665
1666 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1667                   unsigned long old_addr,
1668                   unsigned long new_addr, unsigned long old_end,
1669                   pmd_t *old_pmd, pmd_t *new_pmd)
1670 {
1671         spinlock_t *old_ptl, *new_ptl;
1672         pmd_t pmd;
1673
1674         struct mm_struct *mm = vma->vm_mm;
1675
1676         if ((old_addr & ~HPAGE_PMD_MASK) ||
1677             (new_addr & ~HPAGE_PMD_MASK) ||
1678             old_end - old_addr < HPAGE_PMD_SIZE ||
1679             (new_vma->vm_flags & VM_NOHUGEPAGE))
1680                 return false;
1681
1682         /*
1683          * The destination pmd shouldn't be established, free_pgtables()
1684          * should have release it.
1685          */
1686         if (WARN_ON(!pmd_none(*new_pmd))) {
1687                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1688                 return false;
1689         }
1690
1691         /*
1692          * We don't have to worry about the ordering of src and dst
1693          * ptlocks because exclusive mmap_sem prevents deadlock.
1694          */
1695         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1696         if (old_ptl) {
1697                 new_ptl = pmd_lockptr(mm, new_pmd);
1698                 if (new_ptl != old_ptl)
1699                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1700                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1701                 VM_BUG_ON(!pmd_none(*new_pmd));
1702
1703                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1704                         pgtable_t pgtable;
1705                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1706                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1707                 }
1708                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1709                 if (new_ptl != old_ptl)
1710                         spin_unlock(new_ptl);
1711                 spin_unlock(old_ptl);
1712                 return true;
1713         }
1714         return false;
1715 }
1716
1717 /*
1718  * Returns
1719  *  - 0 if PMD could not be locked
1720  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1721  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1722  */
1723 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1724                 unsigned long addr, pgprot_t newprot, int prot_numa)
1725 {
1726         struct mm_struct *mm = vma->vm_mm;
1727         spinlock_t *ptl;
1728         int ret = 0;
1729
1730         ptl = __pmd_trans_huge_lock(pmd, vma);
1731         if (ptl) {
1732                 pmd_t entry;
1733                 bool preserve_write = prot_numa && pmd_write(*pmd);
1734                 ret = 1;
1735
1736                 /*
1737                  * Avoid trapping faults against the zero page. The read-only
1738                  * data is likely to be read-cached on the local CPU and
1739                  * local/remote hits to the zero page are not interesting.
1740                  */
1741                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1742                         spin_unlock(ptl);
1743                         return ret;
1744                 }
1745
1746                 if (!prot_numa || !pmd_protnone(*pmd)) {
1747                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1748                         entry = pmd_modify(entry, newprot);
1749                         if (preserve_write)
1750                                 entry = pmd_mkwrite(entry);
1751                         ret = HPAGE_PMD_NR;
1752                         set_pmd_at(mm, addr, pmd, entry);
1753                         BUG_ON(!preserve_write && pmd_write(entry));
1754                 }
1755                 spin_unlock(ptl);
1756         }
1757
1758         return ret;
1759 }
1760
1761 /*
1762  * Returns true if a given pmd maps a thp, false otherwise.
1763  *
1764  * Note that if it returns true, this routine returns without unlocking page
1765  * table lock. So callers must unlock it.
1766  */
1767 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1768 {
1769         spinlock_t *ptl;
1770         ptl = pmd_lock(vma->vm_mm, pmd);
1771         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1772                 return ptl;
1773         spin_unlock(ptl);
1774         return NULL;
1775 }
1776
1777 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1778
1779 int hugepage_madvise(struct vm_area_struct *vma,
1780                      unsigned long *vm_flags, int advice)
1781 {
1782         switch (advice) {
1783         case MADV_HUGEPAGE:
1784 #ifdef CONFIG_S390
1785                 /*
1786                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1787                  * can't handle this properly after s390_enable_sie, so we simply
1788                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1789                  */
1790                 if (mm_has_pgste(vma->vm_mm))
1791                         return 0;
1792 #endif
1793                 /*
1794                  * Be somewhat over-protective like KSM for now!
1795                  */
1796                 if (*vm_flags & VM_NO_THP)
1797                         return -EINVAL;
1798                 *vm_flags &= ~VM_NOHUGEPAGE;
1799                 *vm_flags |= VM_HUGEPAGE;
1800                 /*
1801                  * If the vma become good for khugepaged to scan,
1802                  * register it here without waiting a page fault that
1803                  * may not happen any time soon.
1804                  */
1805                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1806                         return -ENOMEM;
1807                 break;
1808         case MADV_NOHUGEPAGE:
1809                 /*
1810                  * Be somewhat over-protective like KSM for now!
1811                  */
1812                 if (*vm_flags & VM_NO_THP)
1813                         return -EINVAL;
1814                 *vm_flags &= ~VM_HUGEPAGE;
1815                 *vm_flags |= VM_NOHUGEPAGE;
1816                 /*
1817                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1818                  * this vma even if we leave the mm registered in khugepaged if
1819                  * it got registered before VM_NOHUGEPAGE was set.
1820                  */
1821                 break;
1822         }
1823
1824         return 0;
1825 }
1826
1827 static int __init khugepaged_slab_init(void)
1828 {
1829         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1830                                           sizeof(struct mm_slot),
1831                                           __alignof__(struct mm_slot), 0, NULL);
1832         if (!mm_slot_cache)
1833                 return -ENOMEM;
1834
1835         return 0;
1836 }
1837
1838 static void __init khugepaged_slab_exit(void)
1839 {
1840         kmem_cache_destroy(mm_slot_cache);
1841 }
1842
1843 static inline struct mm_slot *alloc_mm_slot(void)
1844 {
1845         if (!mm_slot_cache)     /* initialization failed */
1846                 return NULL;
1847         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1848 }
1849
1850 static inline void free_mm_slot(struct mm_slot *mm_slot)
1851 {
1852         kmem_cache_free(mm_slot_cache, mm_slot);
1853 }
1854
1855 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1856 {
1857         struct mm_slot *mm_slot;
1858
1859         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1860                 if (mm == mm_slot->mm)
1861                         return mm_slot;
1862
1863         return NULL;
1864 }
1865
1866 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1867                                     struct mm_slot *mm_slot)
1868 {
1869         mm_slot->mm = mm;
1870         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1871 }
1872
1873 static inline int khugepaged_test_exit(struct mm_struct *mm)
1874 {
1875         return atomic_read(&mm->mm_users) == 0;
1876 }
1877
1878 int __khugepaged_enter(struct mm_struct *mm)
1879 {
1880         struct mm_slot *mm_slot;
1881         int wakeup;
1882
1883         mm_slot = alloc_mm_slot();
1884         if (!mm_slot)
1885                 return -ENOMEM;
1886
1887         /* __khugepaged_exit() must not run from under us */
1888         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1889         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1890                 free_mm_slot(mm_slot);
1891                 return 0;
1892         }
1893
1894         spin_lock(&khugepaged_mm_lock);
1895         insert_to_mm_slots_hash(mm, mm_slot);
1896         /*
1897          * Insert just behind the scanning cursor, to let the area settle
1898          * down a little.
1899          */
1900         wakeup = list_empty(&khugepaged_scan.mm_head);
1901         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1902         spin_unlock(&khugepaged_mm_lock);
1903
1904         atomic_inc(&mm->mm_count);
1905         if (wakeup)
1906                 wake_up_interruptible(&khugepaged_wait);
1907
1908         return 0;
1909 }
1910
1911 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1912                                unsigned long vm_flags)
1913 {
1914         unsigned long hstart, hend;
1915         if (!vma->anon_vma)
1916                 /*
1917                  * Not yet faulted in so we will register later in the
1918                  * page fault if needed.
1919                  */
1920                 return 0;
1921         if (vma->vm_ops)
1922                 /* khugepaged not yet working on file or special mappings */
1923                 return 0;
1924         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1925         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1926         hend = vma->vm_end & HPAGE_PMD_MASK;
1927         if (hstart < hend)
1928                 return khugepaged_enter(vma, vm_flags);
1929         return 0;
1930 }
1931
1932 void __khugepaged_exit(struct mm_struct *mm)
1933 {
1934         struct mm_slot *mm_slot;
1935         int free = 0;
1936
1937         spin_lock(&khugepaged_mm_lock);
1938         mm_slot = get_mm_slot(mm);
1939         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1940                 hash_del(&mm_slot->hash);
1941                 list_del(&mm_slot->mm_node);
1942                 free = 1;
1943         }
1944         spin_unlock(&khugepaged_mm_lock);
1945
1946         if (free) {
1947                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1948                 free_mm_slot(mm_slot);
1949                 mmdrop(mm);
1950         } else if (mm_slot) {
1951                 /*
1952                  * This is required to serialize against
1953                  * khugepaged_test_exit() (which is guaranteed to run
1954                  * under mmap sem read mode). Stop here (after we
1955                  * return all pagetables will be destroyed) until
1956                  * khugepaged has finished working on the pagetables
1957                  * under the mmap_sem.
1958                  */
1959                 down_write(&mm->mmap_sem);
1960                 up_write(&mm->mmap_sem);
1961         }
1962 }
1963
1964 static void release_pte_page(struct page *page)
1965 {
1966         /* 0 stands for page_is_file_cache(page) == false */
1967         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1968         unlock_page(page);
1969         putback_lru_page(page);
1970 }
1971
1972 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1973 {
1974         while (--_pte >= pte) {
1975                 pte_t pteval = *_pte;
1976                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1977                         release_pte_page(pte_page(pteval));
1978         }
1979 }
1980
1981 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1982                                         unsigned long address,
1983                                         pte_t *pte)
1984 {
1985         struct page *page = NULL;
1986         pte_t *_pte;
1987         int none_or_zero = 0, result = 0;
1988         bool referenced = false, writable = false;
1989
1990         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1991              _pte++, address += PAGE_SIZE) {
1992                 pte_t pteval = *_pte;
1993                 if (pte_none(pteval) || (pte_present(pteval) &&
1994                                 is_zero_pfn(pte_pfn(pteval)))) {
1995                         if (!userfaultfd_armed(vma) &&
1996                             ++none_or_zero <= khugepaged_max_ptes_none) {
1997                                 continue;
1998                         } else {
1999                                 result = SCAN_EXCEED_NONE_PTE;
2000                                 goto out;
2001                         }
2002                 }
2003                 if (!pte_present(pteval)) {
2004                         result = SCAN_PTE_NON_PRESENT;
2005                         goto out;
2006                 }
2007                 page = vm_normal_page(vma, address, pteval);
2008                 if (unlikely(!page)) {
2009                         result = SCAN_PAGE_NULL;
2010                         goto out;
2011                 }
2012
2013                 VM_BUG_ON_PAGE(PageCompound(page), page);
2014                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2015                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2016
2017                 /*
2018                  * We can do it before isolate_lru_page because the
2019                  * page can't be freed from under us. NOTE: PG_lock
2020                  * is needed to serialize against split_huge_page
2021                  * when invoked from the VM.
2022                  */
2023                 if (!trylock_page(page)) {
2024                         result = SCAN_PAGE_LOCK;
2025                         goto out;
2026                 }
2027
2028                 /*
2029                  * cannot use mapcount: can't collapse if there's a gup pin.
2030                  * The page must only be referenced by the scanned process
2031                  * and page swap cache.
2032                  */
2033                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2034                         unlock_page(page);
2035                         result = SCAN_PAGE_COUNT;
2036                         goto out;
2037                 }
2038                 if (pte_write(pteval)) {
2039                         writable = true;
2040                 } else {
2041                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2042                                 unlock_page(page);
2043                                 result = SCAN_SWAP_CACHE_PAGE;
2044                                 goto out;
2045                         }
2046                         /*
2047                          * Page is not in the swap cache. It can be collapsed
2048                          * into a THP.
2049                          */
2050                 }
2051
2052                 /*
2053                  * Isolate the page to avoid collapsing an hugepage
2054                  * currently in use by the VM.
2055                  */
2056                 if (isolate_lru_page(page)) {
2057                         unlock_page(page);
2058                         result = SCAN_DEL_PAGE_LRU;
2059                         goto out;
2060                 }
2061                 /* 0 stands for page_is_file_cache(page) == false */
2062                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2063                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2064                 VM_BUG_ON_PAGE(PageLRU(page), page);
2065
2066                 /* If there is no mapped pte young don't collapse the page */
2067                 if (pte_young(pteval) ||
2068                     page_is_young(page) || PageReferenced(page) ||
2069                     mmu_notifier_test_young(vma->vm_mm, address))
2070                         referenced = true;
2071         }
2072         if (likely(writable)) {
2073                 if (likely(referenced)) {
2074                         result = SCAN_SUCCEED;
2075                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2076                                                             referenced, writable, result);
2077                         return 1;
2078                 }
2079         } else {
2080                 result = SCAN_PAGE_RO;
2081         }
2082
2083 out:
2084         release_pte_pages(pte, _pte);
2085         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2086                                             referenced, writable, result);
2087         return 0;
2088 }
2089
2090 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2091                                       struct vm_area_struct *vma,
2092                                       unsigned long address,
2093                                       spinlock_t *ptl)
2094 {
2095         pte_t *_pte;
2096         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2097                 pte_t pteval = *_pte;
2098                 struct page *src_page;
2099
2100                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2101                         clear_user_highpage(page, address);
2102                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2103                         if (is_zero_pfn(pte_pfn(pteval))) {
2104                                 /*
2105                                  * ptl mostly unnecessary.
2106                                  */
2107                                 spin_lock(ptl);
2108                                 /*
2109                                  * paravirt calls inside pte_clear here are
2110                                  * superfluous.
2111                                  */
2112                                 pte_clear(vma->vm_mm, address, _pte);
2113                                 spin_unlock(ptl);
2114                         }
2115                 } else {
2116                         src_page = pte_page(pteval);
2117                         copy_user_highpage(page, src_page, address, vma);
2118                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2119                         release_pte_page(src_page);
2120                         /*
2121                          * ptl mostly unnecessary, but preempt has to
2122                          * be disabled to update the per-cpu stats
2123                          * inside page_remove_rmap().
2124                          */
2125                         spin_lock(ptl);
2126                         /*
2127                          * paravirt calls inside pte_clear here are
2128                          * superfluous.
2129                          */
2130                         pte_clear(vma->vm_mm, address, _pte);
2131                         page_remove_rmap(src_page, false);
2132                         spin_unlock(ptl);
2133                         free_page_and_swap_cache(src_page);
2134                 }
2135
2136                 address += PAGE_SIZE;
2137                 page++;
2138         }
2139 }
2140
2141 static void khugepaged_alloc_sleep(void)
2142 {
2143         DEFINE_WAIT(wait);
2144
2145         add_wait_queue(&khugepaged_wait, &wait);
2146         freezable_schedule_timeout_interruptible(
2147                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2148         remove_wait_queue(&khugepaged_wait, &wait);
2149 }
2150
2151 static int khugepaged_node_load[MAX_NUMNODES];
2152
2153 static bool khugepaged_scan_abort(int nid)
2154 {
2155         int i;
2156
2157         /*
2158          * If zone_reclaim_mode is disabled, then no extra effort is made to
2159          * allocate memory locally.
2160          */
2161         if (!zone_reclaim_mode)
2162                 return false;
2163
2164         /* If there is a count for this node already, it must be acceptable */
2165         if (khugepaged_node_load[nid])
2166                 return false;
2167
2168         for (i = 0; i < MAX_NUMNODES; i++) {
2169                 if (!khugepaged_node_load[i])
2170                         continue;
2171                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2172                         return true;
2173         }
2174         return false;
2175 }
2176
2177 #ifdef CONFIG_NUMA
2178 static int khugepaged_find_target_node(void)
2179 {
2180         static int last_khugepaged_target_node = NUMA_NO_NODE;
2181         int nid, target_node = 0, max_value = 0;
2182
2183         /* find first node with max normal pages hit */
2184         for (nid = 0; nid < MAX_NUMNODES; nid++)
2185                 if (khugepaged_node_load[nid] > max_value) {
2186                         max_value = khugepaged_node_load[nid];
2187                         target_node = nid;
2188                 }
2189
2190         /* do some balance if several nodes have the same hit record */
2191         if (target_node <= last_khugepaged_target_node)
2192                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2193                                 nid++)
2194                         if (max_value == khugepaged_node_load[nid]) {
2195                                 target_node = nid;
2196                                 break;
2197                         }
2198
2199         last_khugepaged_target_node = target_node;
2200         return target_node;
2201 }
2202
2203 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2204 {
2205         if (IS_ERR(*hpage)) {
2206                 if (!*wait)
2207                         return false;
2208
2209                 *wait = false;
2210                 *hpage = NULL;
2211                 khugepaged_alloc_sleep();
2212         } else if (*hpage) {
2213                 put_page(*hpage);
2214                 *hpage = NULL;
2215         }
2216
2217         return true;
2218 }
2219
2220 static struct page *
2221 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2222                        unsigned long address, int node)
2223 {
2224         VM_BUG_ON_PAGE(*hpage, *hpage);
2225
2226         /*
2227          * Before allocating the hugepage, release the mmap_sem read lock.
2228          * The allocation can take potentially a long time if it involves
2229          * sync compaction, and we do not need to hold the mmap_sem during
2230          * that. We will recheck the vma after taking it again in write mode.
2231          */
2232         up_read(&mm->mmap_sem);
2233
2234         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2235         if (unlikely(!*hpage)) {
2236                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2237                 *hpage = ERR_PTR(-ENOMEM);
2238                 return NULL;
2239         }
2240
2241         prep_transhuge_page(*hpage);
2242         count_vm_event(THP_COLLAPSE_ALLOC);
2243         return *hpage;
2244 }
2245 #else
2246 static int khugepaged_find_target_node(void)
2247 {
2248         return 0;
2249 }
2250
2251 static inline struct page *alloc_hugepage(int defrag)
2252 {
2253         struct page *page;
2254
2255         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2256         if (page)
2257                 prep_transhuge_page(page);
2258         return page;
2259 }
2260
2261 static struct page *khugepaged_alloc_hugepage(bool *wait)
2262 {
2263         struct page *hpage;
2264
2265         do {
2266                 hpage = alloc_hugepage(khugepaged_defrag());
2267                 if (!hpage) {
2268                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2269                         if (!*wait)
2270                                 return NULL;
2271
2272                         *wait = false;
2273                         khugepaged_alloc_sleep();
2274                 } else
2275                         count_vm_event(THP_COLLAPSE_ALLOC);
2276         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2277
2278         return hpage;
2279 }
2280
2281 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2282 {
2283         if (!*hpage)
2284                 *hpage = khugepaged_alloc_hugepage(wait);
2285
2286         if (unlikely(!*hpage))
2287                 return false;
2288
2289         return true;
2290 }
2291
2292 static struct page *
2293 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2294                        unsigned long address, int node)
2295 {
2296         up_read(&mm->mmap_sem);
2297         VM_BUG_ON(!*hpage);
2298
2299         return  *hpage;
2300 }
2301 #endif
2302
2303 static bool hugepage_vma_check(struct vm_area_struct *vma)
2304 {
2305         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2306             (vma->vm_flags & VM_NOHUGEPAGE))
2307                 return false;
2308         if (!vma->anon_vma || vma->vm_ops)
2309                 return false;
2310         if (is_vma_temporary_stack(vma))
2311                 return false;
2312         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2313         return true;
2314 }
2315
2316 static void collapse_huge_page(struct mm_struct *mm,
2317                                    unsigned long address,
2318                                    struct page **hpage,
2319                                    struct vm_area_struct *vma,
2320                                    int node)
2321 {
2322         pmd_t *pmd, _pmd;
2323         pte_t *pte;
2324         pgtable_t pgtable;
2325         struct page *new_page;
2326         spinlock_t *pmd_ptl, *pte_ptl;
2327         int isolated = 0, result = 0;
2328         unsigned long hstart, hend;
2329         struct mem_cgroup *memcg;
2330         unsigned long mmun_start;       /* For mmu_notifiers */
2331         unsigned long mmun_end;         /* For mmu_notifiers */
2332         gfp_t gfp;
2333
2334         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2335
2336         /* Only allocate from the target node */
2337         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2338                 __GFP_THISNODE;
2339
2340         /* release the mmap_sem read lock. */
2341         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2342         if (!new_page) {
2343                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2344                 goto out_nolock;
2345         }
2346
2347         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2348                 result = SCAN_CGROUP_CHARGE_FAIL;
2349                 goto out_nolock;
2350         }
2351
2352         /*
2353          * Prevent all access to pagetables with the exception of
2354          * gup_fast later hanlded by the ptep_clear_flush and the VM
2355          * handled by the anon_vma lock + PG_lock.
2356          */
2357         down_write(&mm->mmap_sem);
2358         if (unlikely(khugepaged_test_exit(mm))) {
2359                 result = SCAN_ANY_PROCESS;
2360                 goto out;
2361         }
2362
2363         vma = find_vma(mm, address);
2364         if (!vma) {
2365                 result = SCAN_VMA_NULL;
2366                 goto out;
2367         }
2368         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2369         hend = vma->vm_end & HPAGE_PMD_MASK;
2370         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2371                 result = SCAN_ADDRESS_RANGE;
2372                 goto out;
2373         }
2374         if (!hugepage_vma_check(vma)) {
2375                 result = SCAN_VMA_CHECK;
2376                 goto out;
2377         }
2378         pmd = mm_find_pmd(mm, address);
2379         if (!pmd) {
2380                 result = SCAN_PMD_NULL;
2381                 goto out;
2382         }
2383
2384         anon_vma_lock_write(vma->anon_vma);
2385
2386         pte = pte_offset_map(pmd, address);
2387         pte_ptl = pte_lockptr(mm, pmd);
2388
2389         mmun_start = address;
2390         mmun_end   = address + HPAGE_PMD_SIZE;
2391         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2392         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2393         /*
2394          * After this gup_fast can't run anymore. This also removes
2395          * any huge TLB entry from the CPU so we won't allow
2396          * huge and small TLB entries for the same virtual address
2397          * to avoid the risk of CPU bugs in that area.
2398          */
2399         _pmd = pmdp_collapse_flush(vma, address, pmd);
2400         spin_unlock(pmd_ptl);
2401         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2402
2403         spin_lock(pte_ptl);
2404         isolated = __collapse_huge_page_isolate(vma, address, pte);
2405         spin_unlock(pte_ptl);
2406
2407         if (unlikely(!isolated)) {
2408                 pte_unmap(pte);
2409                 spin_lock(pmd_ptl);
2410                 BUG_ON(!pmd_none(*pmd));
2411                 /*
2412                  * We can only use set_pmd_at when establishing
2413                  * hugepmds and never for establishing regular pmds that
2414                  * points to regular pagetables. Use pmd_populate for that
2415                  */
2416                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2417                 spin_unlock(pmd_ptl);
2418                 anon_vma_unlock_write(vma->anon_vma);
2419                 result = SCAN_FAIL;
2420                 goto out;
2421         }
2422
2423         /*
2424          * All pages are isolated and locked so anon_vma rmap
2425          * can't run anymore.
2426          */
2427         anon_vma_unlock_write(vma->anon_vma);
2428
2429         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2430         pte_unmap(pte);
2431         __SetPageUptodate(new_page);
2432         pgtable = pmd_pgtable(_pmd);
2433
2434         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2435         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2436
2437         /*
2438          * spin_lock() below is not the equivalent of smp_wmb(), so
2439          * this is needed to avoid the copy_huge_page writes to become
2440          * visible after the set_pmd_at() write.
2441          */
2442         smp_wmb();
2443
2444         spin_lock(pmd_ptl);
2445         BUG_ON(!pmd_none(*pmd));
2446         page_add_new_anon_rmap(new_page, vma, address, true);
2447         mem_cgroup_commit_charge(new_page, memcg, false, true);
2448         lru_cache_add_active_or_unevictable(new_page, vma);
2449         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2450         set_pmd_at(mm, address, pmd, _pmd);
2451         update_mmu_cache_pmd(vma, address, pmd);
2452         spin_unlock(pmd_ptl);
2453
2454         *hpage = NULL;
2455
2456         khugepaged_pages_collapsed++;
2457         result = SCAN_SUCCEED;
2458 out_up_write:
2459         up_write(&mm->mmap_sem);
2460         trace_mm_collapse_huge_page(mm, isolated, result);
2461         return;
2462
2463 out_nolock:
2464         trace_mm_collapse_huge_page(mm, isolated, result);
2465         return;
2466 out:
2467         mem_cgroup_cancel_charge(new_page, memcg, true);
2468         goto out_up_write;
2469 }
2470
2471 static int khugepaged_scan_pmd(struct mm_struct *mm,
2472                                struct vm_area_struct *vma,
2473                                unsigned long address,
2474                                struct page **hpage)
2475 {
2476         pmd_t *pmd;
2477         pte_t *pte, *_pte;
2478         int ret = 0, none_or_zero = 0, result = 0;
2479         struct page *page = NULL;
2480         unsigned long _address;
2481         spinlock_t *ptl;
2482         int node = NUMA_NO_NODE;
2483         bool writable = false, referenced = false;
2484
2485         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2486
2487         pmd = mm_find_pmd(mm, address);
2488         if (!pmd) {
2489                 result = SCAN_PMD_NULL;
2490                 goto out;
2491         }
2492
2493         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2494         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2495         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2496              _pte++, _address += PAGE_SIZE) {
2497                 pte_t pteval = *_pte;
2498                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2499                         if (!userfaultfd_armed(vma) &&
2500                             ++none_or_zero <= khugepaged_max_ptes_none) {
2501                                 continue;
2502                         } else {
2503                                 result = SCAN_EXCEED_NONE_PTE;
2504                                 goto out_unmap;
2505                         }
2506                 }
2507                 if (!pte_present(pteval)) {
2508                         result = SCAN_PTE_NON_PRESENT;
2509                         goto out_unmap;
2510                 }
2511                 if (pte_write(pteval))
2512                         writable = true;
2513
2514                 page = vm_normal_page(vma, _address, pteval);
2515                 if (unlikely(!page)) {
2516                         result = SCAN_PAGE_NULL;
2517                         goto out_unmap;
2518                 }
2519
2520                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2521                 if (PageCompound(page)) {
2522                         result = SCAN_PAGE_COMPOUND;
2523                         goto out_unmap;
2524                 }
2525
2526                 /*
2527                  * Record which node the original page is from and save this
2528                  * information to khugepaged_node_load[].
2529                  * Khupaged will allocate hugepage from the node has the max
2530                  * hit record.
2531                  */
2532                 node = page_to_nid(page);
2533                 if (khugepaged_scan_abort(node)) {
2534                         result = SCAN_SCAN_ABORT;
2535                         goto out_unmap;
2536                 }
2537                 khugepaged_node_load[node]++;
2538                 if (!PageLRU(page)) {
2539                         result = SCAN_SCAN_ABORT;
2540                         goto out_unmap;
2541                 }
2542                 if (PageLocked(page)) {
2543                         result = SCAN_PAGE_LOCK;
2544                         goto out_unmap;
2545                 }
2546                 if (!PageAnon(page)) {
2547                         result = SCAN_PAGE_ANON;
2548                         goto out_unmap;
2549                 }
2550
2551                 /*
2552                  * cannot use mapcount: can't collapse if there's a gup pin.
2553                  * The page must only be referenced by the scanned process
2554                  * and page swap cache.
2555                  */
2556                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2557                         result = SCAN_PAGE_COUNT;
2558                         goto out_unmap;
2559                 }
2560                 if (pte_young(pteval) ||
2561                     page_is_young(page) || PageReferenced(page) ||
2562                     mmu_notifier_test_young(vma->vm_mm, address))
2563                         referenced = true;
2564         }
2565         if (writable) {
2566                 if (referenced) {
2567                         result = SCAN_SUCCEED;
2568                         ret = 1;
2569                 } else {
2570                         result = SCAN_NO_REFERENCED_PAGE;
2571                 }
2572         } else {
2573                 result = SCAN_PAGE_RO;
2574         }
2575 out_unmap:
2576         pte_unmap_unlock(pte, ptl);
2577         if (ret) {
2578                 node = khugepaged_find_target_node();
2579                 /* collapse_huge_page will return with the mmap_sem released */
2580                 collapse_huge_page(mm, address, hpage, vma, node);
2581         }
2582 out:
2583         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2584                                      none_or_zero, result);
2585         return ret;
2586 }
2587
2588 static void collect_mm_slot(struct mm_slot *mm_slot)
2589 {
2590         struct mm_struct *mm = mm_slot->mm;
2591
2592         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2593
2594         if (khugepaged_test_exit(mm)) {
2595                 /* free mm_slot */
2596                 hash_del(&mm_slot->hash);
2597                 list_del(&mm_slot->mm_node);
2598
2599                 /*
2600                  * Not strictly needed because the mm exited already.
2601                  *
2602                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2603                  */
2604
2605                 /* khugepaged_mm_lock actually not necessary for the below */
2606                 free_mm_slot(mm_slot);
2607                 mmdrop(mm);
2608         }
2609 }
2610
2611 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2612                                             struct page **hpage)
2613         __releases(&khugepaged_mm_lock)
2614         __acquires(&khugepaged_mm_lock)
2615 {
2616         struct mm_slot *mm_slot;
2617         struct mm_struct *mm;
2618         struct vm_area_struct *vma;
2619         int progress = 0;
2620
2621         VM_BUG_ON(!pages);
2622         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2623
2624         if (khugepaged_scan.mm_slot)
2625                 mm_slot = khugepaged_scan.mm_slot;
2626         else {
2627                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2628                                      struct mm_slot, mm_node);
2629                 khugepaged_scan.address = 0;
2630                 khugepaged_scan.mm_slot = mm_slot;
2631         }
2632         spin_unlock(&khugepaged_mm_lock);
2633
2634         mm = mm_slot->mm;
2635         down_read(&mm->mmap_sem);
2636         if (unlikely(khugepaged_test_exit(mm)))
2637                 vma = NULL;
2638         else
2639                 vma = find_vma(mm, khugepaged_scan.address);
2640
2641         progress++;
2642         for (; vma; vma = vma->vm_next) {
2643                 unsigned long hstart, hend;
2644
2645                 cond_resched();
2646                 if (unlikely(khugepaged_test_exit(mm))) {
2647                         progress++;
2648                         break;
2649                 }
2650                 if (!hugepage_vma_check(vma)) {
2651 skip:
2652                         progress++;
2653                         continue;
2654                 }
2655                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2656                 hend = vma->vm_end & HPAGE_PMD_MASK;
2657                 if (hstart >= hend)
2658                         goto skip;
2659                 if (khugepaged_scan.address > hend)
2660                         goto skip;
2661                 if (khugepaged_scan.address < hstart)
2662                         khugepaged_scan.address = hstart;
2663                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2664
2665                 while (khugepaged_scan.address < hend) {
2666                         int ret;
2667                         cond_resched();
2668                         if (unlikely(khugepaged_test_exit(mm)))
2669                                 goto breakouterloop;
2670
2671                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2672                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2673                                   hend);
2674                         ret = khugepaged_scan_pmd(mm, vma,
2675                                                   khugepaged_scan.address,
2676                                                   hpage);
2677                         /* move to next address */
2678                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2679                         progress += HPAGE_PMD_NR;
2680                         if (ret)
2681                                 /* we released mmap_sem so break loop */
2682                                 goto breakouterloop_mmap_sem;
2683                         if (progress >= pages)
2684                                 goto breakouterloop;
2685                 }
2686         }
2687 breakouterloop:
2688         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2689 breakouterloop_mmap_sem:
2690
2691         spin_lock(&khugepaged_mm_lock);
2692         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2693         /*
2694          * Release the current mm_slot if this mm is about to die, or
2695          * if we scanned all vmas of this mm.
2696          */
2697         if (khugepaged_test_exit(mm) || !vma) {
2698                 /*
2699                  * Make sure that if mm_users is reaching zero while
2700                  * khugepaged runs here, khugepaged_exit will find
2701                  * mm_slot not pointing to the exiting mm.
2702                  */
2703                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2704                         khugepaged_scan.mm_slot = list_entry(
2705                                 mm_slot->mm_node.next,
2706                                 struct mm_slot, mm_node);
2707                         khugepaged_scan.address = 0;
2708                 } else {
2709                         khugepaged_scan.mm_slot = NULL;
2710                         khugepaged_full_scans++;
2711                 }
2712
2713                 collect_mm_slot(mm_slot);
2714         }
2715
2716         return progress;
2717 }
2718
2719 static int khugepaged_has_work(void)
2720 {
2721         return !list_empty(&khugepaged_scan.mm_head) &&
2722                 khugepaged_enabled();
2723 }
2724
2725 static int khugepaged_wait_event(void)
2726 {
2727         return !list_empty(&khugepaged_scan.mm_head) ||
2728                 kthread_should_stop();
2729 }
2730
2731 static void khugepaged_do_scan(void)
2732 {
2733         struct page *hpage = NULL;
2734         unsigned int progress = 0, pass_through_head = 0;
2735         unsigned int pages = khugepaged_pages_to_scan;
2736         bool wait = true;
2737
2738         barrier(); /* write khugepaged_pages_to_scan to local stack */
2739
2740         while (progress < pages) {
2741                 if (!khugepaged_prealloc_page(&hpage, &wait))
2742                         break;
2743
2744                 cond_resched();
2745
2746                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2747                         break;
2748
2749                 spin_lock(&khugepaged_mm_lock);
2750                 if (!khugepaged_scan.mm_slot)
2751                         pass_through_head++;
2752                 if (khugepaged_has_work() &&
2753                     pass_through_head < 2)
2754                         progress += khugepaged_scan_mm_slot(pages - progress,
2755                                                             &hpage);
2756                 else
2757                         progress = pages;
2758                 spin_unlock(&khugepaged_mm_lock);
2759         }
2760
2761         if (!IS_ERR_OR_NULL(hpage))
2762                 put_page(hpage);
2763 }
2764
2765 static void khugepaged_wait_work(void)
2766 {
2767         if (khugepaged_has_work()) {
2768                 if (!khugepaged_scan_sleep_millisecs)
2769                         return;
2770
2771                 wait_event_freezable_timeout(khugepaged_wait,
2772                                              kthread_should_stop(),
2773                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2774                 return;
2775         }
2776
2777         if (khugepaged_enabled())
2778                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2779 }
2780
2781 static int khugepaged(void *none)
2782 {
2783         struct mm_slot *mm_slot;
2784
2785         set_freezable();
2786         set_user_nice(current, MAX_NICE);
2787
2788         while (!kthread_should_stop()) {
2789                 khugepaged_do_scan();
2790                 khugepaged_wait_work();
2791         }
2792
2793         spin_lock(&khugepaged_mm_lock);
2794         mm_slot = khugepaged_scan.mm_slot;
2795         khugepaged_scan.mm_slot = NULL;
2796         if (mm_slot)
2797                 collect_mm_slot(mm_slot);
2798         spin_unlock(&khugepaged_mm_lock);
2799         return 0;
2800 }
2801
2802 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2803                 unsigned long haddr, pmd_t *pmd)
2804 {
2805         struct mm_struct *mm = vma->vm_mm;
2806         pgtable_t pgtable;
2807         pmd_t _pmd;
2808         int i;
2809
2810         /* leave pmd empty until pte is filled */
2811         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2812
2813         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2814         pmd_populate(mm, &_pmd, pgtable);
2815
2816         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2817                 pte_t *pte, entry;
2818                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2819                 entry = pte_mkspecial(entry);
2820                 pte = pte_offset_map(&_pmd, haddr);
2821                 VM_BUG_ON(!pte_none(*pte));
2822                 set_pte_at(mm, haddr, pte, entry);
2823                 pte_unmap(pte);
2824         }
2825         smp_wmb(); /* make pte visible before pmd */
2826         pmd_populate(mm, pmd, pgtable);
2827         put_huge_zero_page();
2828 }
2829
2830 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2831                 unsigned long haddr, bool freeze)
2832 {
2833         struct mm_struct *mm = vma->vm_mm;
2834         struct page *page;
2835         pgtable_t pgtable;
2836         pmd_t _pmd;
2837         bool young, write, dirty;
2838         int i;
2839
2840         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2841         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2842         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2843         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2844
2845         count_vm_event(THP_SPLIT_PMD);
2846
2847         if (vma_is_dax(vma)) {
2848                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2849                 if (is_huge_zero_pmd(_pmd))
2850                         put_huge_zero_page();
2851                 return;
2852         } else if (is_huge_zero_pmd(*pmd)) {
2853                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2854         }
2855
2856         page = pmd_page(*pmd);
2857         VM_BUG_ON_PAGE(!page_count(page), page);
2858         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2859         write = pmd_write(*pmd);
2860         young = pmd_young(*pmd);
2861         dirty = pmd_dirty(*pmd);
2862
2863         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2864         pmd_populate(mm, &_pmd, pgtable);
2865
2866         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2867                 pte_t entry, *pte;
2868                 /*
2869                  * Note that NUMA hinting access restrictions are not
2870                  * transferred to avoid any possibility of altering
2871                  * permissions across VMAs.
2872                  */
2873                 if (freeze) {
2874                         swp_entry_t swp_entry;
2875                         swp_entry = make_migration_entry(page + i, write);
2876                         entry = swp_entry_to_pte(swp_entry);
2877                 } else {
2878                         entry = mk_pte(page + i, vma->vm_page_prot);
2879                         entry = maybe_mkwrite(entry, vma);
2880                         if (!write)
2881                                 entry = pte_wrprotect(entry);
2882                         if (!young)
2883                                 entry = pte_mkold(entry);
2884                 }
2885                 if (dirty)
2886                         SetPageDirty(page + i);
2887                 pte = pte_offset_map(&_pmd, haddr);
2888                 BUG_ON(!pte_none(*pte));
2889                 set_pte_at(mm, haddr, pte, entry);
2890                 atomic_inc(&page[i]._mapcount);
2891                 pte_unmap(pte);
2892         }
2893
2894         /*
2895          * Set PG_double_map before dropping compound_mapcount to avoid
2896          * false-negative page_mapped().
2897          */
2898         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2899                 for (i = 0; i < HPAGE_PMD_NR; i++)
2900                         atomic_inc(&page[i]._mapcount);
2901         }
2902
2903         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2904                 /* Last compound_mapcount is gone. */
2905                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2906                 if (TestClearPageDoubleMap(page)) {
2907                         /* No need in mapcount reference anymore */
2908                         for (i = 0; i < HPAGE_PMD_NR; i++)
2909                                 atomic_dec(&page[i]._mapcount);
2910                 }
2911         }
2912
2913         smp_wmb(); /* make pte visible before pmd */
2914         /*
2915          * Up to this point the pmd is present and huge and userland has the
2916          * whole access to the hugepage during the split (which happens in
2917          * place). If we overwrite the pmd with the not-huge version pointing
2918          * to the pte here (which of course we could if all CPUs were bug
2919          * free), userland could trigger a small page size TLB miss on the
2920          * small sized TLB while the hugepage TLB entry is still established in
2921          * the huge TLB. Some CPU doesn't like that.
2922          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2923          * 383 on page 93. Intel should be safe but is also warns that it's
2924          * only safe if the permission and cache attributes of the two entries
2925          * loaded in the two TLB is identical (which should be the case here).
2926          * But it is generally safer to never allow small and huge TLB entries
2927          * for the same virtual address to be loaded simultaneously. So instead
2928          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2929          * current pmd notpresent (atomically because here the pmd_trans_huge
2930          * and pmd_trans_splitting must remain set at all times on the pmd
2931          * until the split is complete for this pmd), then we flush the SMP TLB
2932          * and finally we write the non-huge version of the pmd entry with
2933          * pmd_populate.
2934          */
2935         pmdp_invalidate(vma, haddr, pmd);
2936         pmd_populate(mm, pmd, pgtable);
2937
2938         if (freeze) {
2939                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2940                         page_remove_rmap(page + i, false);
2941                         put_page(page + i);
2942                 }
2943         }
2944 }
2945
2946 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2947                 unsigned long address)
2948 {
2949         spinlock_t *ptl;
2950         struct mm_struct *mm = vma->vm_mm;
2951         struct page *page = NULL;
2952         unsigned long haddr = address & HPAGE_PMD_MASK;
2953
2954         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2955         ptl = pmd_lock(mm, pmd);
2956         if (pmd_trans_huge(*pmd)) {
2957                 page = pmd_page(*pmd);
2958                 if (PageMlocked(page))
2959                         get_page(page);
2960                 else
2961                         page = NULL;
2962         } else if (!pmd_devmap(*pmd))
2963                 goto out;
2964         __split_huge_pmd_locked(vma, pmd, haddr, false);
2965 out:
2966         spin_unlock(ptl);
2967         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2968         if (page) {
2969                 lock_page(page);
2970                 munlock_vma_page(page);
2971                 unlock_page(page);
2972                 put_page(page);
2973         }
2974 }
2975
2976 static void split_huge_pmd_address(struct vm_area_struct *vma,
2977                                     unsigned long address)
2978 {
2979         pgd_t *pgd;
2980         pud_t *pud;
2981         pmd_t *pmd;
2982
2983         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2984
2985         pgd = pgd_offset(vma->vm_mm, address);
2986         if (!pgd_present(*pgd))
2987                 return;
2988
2989         pud = pud_offset(pgd, address);
2990         if (!pud_present(*pud))
2991                 return;
2992
2993         pmd = pmd_offset(pud, address);
2994         if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
2995                 return;
2996         /*
2997          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2998          * materialize from under us.
2999          */
3000         split_huge_pmd(vma, pmd, address);
3001 }
3002
3003 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3004                              unsigned long start,
3005                              unsigned long end,
3006                              long adjust_next)
3007 {
3008         /*
3009          * If the new start address isn't hpage aligned and it could
3010          * previously contain an hugepage: check if we need to split
3011          * an huge pmd.
3012          */
3013         if (start & ~HPAGE_PMD_MASK &&
3014             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3015             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3016                 split_huge_pmd_address(vma, start);
3017
3018         /*
3019          * If the new end address isn't hpage aligned and it could
3020          * previously contain an hugepage: check if we need to split
3021          * an huge pmd.
3022          */
3023         if (end & ~HPAGE_PMD_MASK &&
3024             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3025             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3026                 split_huge_pmd_address(vma, end);
3027
3028         /*
3029          * If we're also updating the vma->vm_next->vm_start, if the new
3030          * vm_next->vm_start isn't page aligned and it could previously
3031          * contain an hugepage: check if we need to split an huge pmd.
3032          */
3033         if (adjust_next > 0) {
3034                 struct vm_area_struct *next = vma->vm_next;
3035                 unsigned long nstart = next->vm_start;
3036                 nstart += adjust_next << PAGE_SHIFT;
3037                 if (nstart & ~HPAGE_PMD_MASK &&
3038                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3039                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3040                         split_huge_pmd_address(next, nstart);
3041         }
3042 }
3043
3044 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
3045                 unsigned long address)
3046 {
3047         unsigned long haddr = address & HPAGE_PMD_MASK;
3048         spinlock_t *ptl;
3049         pgd_t *pgd;
3050         pud_t *pud;
3051         pmd_t *pmd;
3052         pte_t *pte;
3053         int i, nr = HPAGE_PMD_NR;
3054
3055         /* Skip pages which doesn't belong to the VMA */
3056         if (address < vma->vm_start) {
3057                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3058                 page += off;
3059                 nr -= off;
3060                 address = vma->vm_start;
3061         }
3062
3063         pgd = pgd_offset(vma->vm_mm, address);
3064         if (!pgd_present(*pgd))
3065                 return;
3066         pud = pud_offset(pgd, address);
3067         if (!pud_present(*pud))
3068                 return;
3069         pmd = pmd_offset(pud, address);
3070         ptl = pmd_lock(vma->vm_mm, pmd);
3071         if (!pmd_present(*pmd)) {
3072                 spin_unlock(ptl);
3073                 return;
3074         }
3075         if (pmd_trans_huge(*pmd)) {
3076                 if (page == pmd_page(*pmd))
3077                         __split_huge_pmd_locked(vma, pmd, haddr, true);
3078                 spin_unlock(ptl);
3079                 return;
3080         }
3081         spin_unlock(ptl);
3082
3083         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3084         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3085                 pte_t entry, swp_pte;
3086                 swp_entry_t swp_entry;
3087
3088                 /*
3089                  * We've just crossed page table boundary: need to map next one.
3090                  * It can happen if THP was mremaped to non PMD-aligned address.
3091                  */
3092                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3093                         pte_unmap_unlock(pte - 1, ptl);
3094                         pmd = mm_find_pmd(vma->vm_mm, address);
3095                         if (!pmd)
3096                                 return;
3097                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3098                                         address, &ptl);
3099                 }
3100
3101                 if (!pte_present(*pte))
3102                         continue;
3103                 if (page_to_pfn(page) != pte_pfn(*pte))
3104                         continue;
3105                 flush_cache_page(vma, address, page_to_pfn(page));
3106                 entry = ptep_clear_flush(vma, address, pte);
3107                 if (pte_dirty(entry))
3108                         SetPageDirty(page);
3109                 swp_entry = make_migration_entry(page, pte_write(entry));
3110                 swp_pte = swp_entry_to_pte(swp_entry);
3111                 if (pte_soft_dirty(entry))
3112                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
3113                 set_pte_at(vma->vm_mm, address, pte, swp_pte);
3114                 page_remove_rmap(page, false);
3115                 put_page(page);
3116         }
3117         pte_unmap_unlock(pte - 1, ptl);
3118 }
3119
3120 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
3121 {
3122         struct anon_vma_chain *avc;
3123         pgoff_t pgoff = page_to_pgoff(page);
3124
3125         VM_BUG_ON_PAGE(!PageHead(page), page);
3126
3127         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
3128                         pgoff + HPAGE_PMD_NR - 1) {
3129                 unsigned long address = __vma_address(page, avc->vma);
3130
3131                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3132                                 address, address + HPAGE_PMD_SIZE);
3133                 freeze_page_vma(avc->vma, page, address);
3134                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3135                                 address, address + HPAGE_PMD_SIZE);
3136         }
3137 }
3138
3139 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
3140                 unsigned long address)
3141 {
3142         spinlock_t *ptl;
3143         pmd_t *pmd;
3144         pte_t *pte, entry;
3145         swp_entry_t swp_entry;
3146         unsigned long haddr = address & HPAGE_PMD_MASK;
3147         int i, nr = HPAGE_PMD_NR;
3148
3149         /* Skip pages which doesn't belong to the VMA */
3150         if (address < vma->vm_start) {
3151                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3152                 page += off;
3153                 nr -= off;
3154                 address = vma->vm_start;
3155         }
3156
3157         pmd = mm_find_pmd(vma->vm_mm, address);
3158         if (!pmd)
3159                 return;
3160
3161         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3162         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3163                 /*
3164                  * We've just crossed page table boundary: need to map next one.
3165                  * It can happen if THP was mremaped to non-PMD aligned address.
3166                  */
3167                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3168                         pte_unmap_unlock(pte - 1, ptl);
3169                         pmd = mm_find_pmd(vma->vm_mm, address);
3170                         if (!pmd)
3171                                 return;
3172                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3173                                         address, &ptl);
3174                 }
3175
3176                 if (!is_swap_pte(*pte))
3177                         continue;
3178
3179                 swp_entry = pte_to_swp_entry(*pte);
3180                 if (!is_migration_entry(swp_entry))
3181                         continue;
3182                 if (migration_entry_to_page(swp_entry) != page)
3183                         continue;
3184
3185                 get_page(page);
3186                 page_add_anon_rmap(page, vma, address, false);
3187
3188                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3189                 if (PageDirty(page))
3190                         entry = pte_mkdirty(entry);
3191                 if (is_write_migration_entry(swp_entry))
3192                         entry = maybe_mkwrite(entry, vma);
3193
3194                 flush_dcache_page(page);
3195                 set_pte_at(vma->vm_mm, address, pte, entry);
3196
3197                 /* No need to invalidate - it was non-present before */
3198                 update_mmu_cache(vma, address, pte);
3199         }
3200         pte_unmap_unlock(pte - 1, ptl);
3201 }
3202
3203 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3204 {
3205         struct anon_vma_chain *avc;
3206         pgoff_t pgoff = page_to_pgoff(page);
3207
3208         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3209                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3210                 unsigned long address = __vma_address(page, avc->vma);
3211
3212                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3213                                 address, address + HPAGE_PMD_SIZE);
3214                 unfreeze_page_vma(avc->vma, page, address);
3215                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3216                                 address, address + HPAGE_PMD_SIZE);
3217         }
3218 }
3219
3220 static int __split_huge_page_tail(struct page *head, int tail,
3221                 struct lruvec *lruvec, struct list_head *list)
3222 {
3223         int mapcount;
3224         struct page *page_tail = head + tail;
3225
3226         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3227         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3228
3229         /*
3230          * tail_page->_count is zero and not changing from under us. But
3231          * get_page_unless_zero() may be running from under us on the
3232          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3233          * would then run atomic_set() concurrently with
3234          * get_page_unless_zero(), and atomic_set() is implemented in C not
3235          * using locked ops. spin_unlock on x86 sometime uses locked ops
3236          * because of PPro errata 66, 92, so unless somebody can guarantee
3237          * atomic_set() here would be safe on all archs (and not only on x86),
3238          * it's safer to use atomic_add().
3239          */
3240         atomic_add(mapcount + 1, &page_tail->_count);
3241
3242
3243         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3244         page_tail->flags |= (head->flags &
3245                         ((1L << PG_referenced) |
3246                          (1L << PG_swapbacked) |
3247                          (1L << PG_mlocked) |
3248                          (1L << PG_uptodate) |
3249                          (1L << PG_active) |
3250                          (1L << PG_locked) |
3251                          (1L << PG_unevictable) |
3252                          (1L << PG_dirty)));
3253
3254         /*
3255          * After clearing PageTail the gup refcount can be released.
3256          * Page flags also must be visible before we make the page non-compound.
3257          */
3258         smp_wmb();
3259
3260         clear_compound_head(page_tail);
3261
3262         if (page_is_young(head))
3263                 set_page_young(page_tail);
3264         if (page_is_idle(head))
3265                 set_page_idle(page_tail);
3266
3267         /* ->mapping in first tail page is compound_mapcount */
3268         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3269                         page_tail);
3270         page_tail->mapping = head->mapping;
3271
3272         page_tail->index = head->index + tail;
3273         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3274         lru_add_page_tail(head, page_tail, lruvec, list);
3275
3276         return mapcount;
3277 }
3278
3279 static void __split_huge_page(struct page *page, struct list_head *list)
3280 {
3281         struct page *head = compound_head(page);
3282         struct zone *zone = page_zone(head);
3283         struct lruvec *lruvec;
3284         int i, tail_mapcount;
3285
3286         /* prevent PageLRU to go away from under us, and freeze lru stats */
3287         spin_lock_irq(&zone->lru_lock);
3288         lruvec = mem_cgroup_page_lruvec(head, zone);
3289
3290         /* complete memcg works before add pages to LRU */
3291         mem_cgroup_split_huge_fixup(head);
3292
3293         tail_mapcount = 0;
3294         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3295                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3296         atomic_sub(tail_mapcount, &head->_count);
3297
3298         ClearPageCompound(head);
3299         spin_unlock_irq(&zone->lru_lock);
3300
3301         unfreeze_page(page_anon_vma(head), head);
3302
3303         for (i = 0; i < HPAGE_PMD_NR; i++) {
3304                 struct page *subpage = head + i;
3305                 if (subpage == page)
3306                         continue;
3307                 unlock_page(subpage);
3308
3309                 /*
3310                  * Subpages may be freed if there wasn't any mapping
3311                  * like if add_to_swap() is running on a lru page that
3312                  * had its mapping zapped. And freeing these pages
3313                  * requires taking the lru_lock so we do the put_page
3314                  * of the tail pages after the split is complete.
3315                  */
3316                 put_page(subpage);
3317         }
3318 }
3319
3320 int total_mapcount(struct page *page)
3321 {
3322         int i, ret;
3323
3324         VM_BUG_ON_PAGE(PageTail(page), page);
3325
3326         if (likely(!PageCompound(page)))
3327                 return atomic_read(&page->_mapcount) + 1;
3328
3329         ret = compound_mapcount(page);
3330         if (PageHuge(page))
3331                 return ret;
3332         for (i = 0; i < HPAGE_PMD_NR; i++)
3333                 ret += atomic_read(&page[i]._mapcount) + 1;
3334         if (PageDoubleMap(page))
3335                 ret -= HPAGE_PMD_NR;
3336         return ret;
3337 }
3338
3339 /*
3340  * This function splits huge page into normal pages. @page can point to any
3341  * subpage of huge page to split. Split doesn't change the position of @page.
3342  *
3343  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3344  * The huge page must be locked.
3345  *
3346  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3347  *
3348  * Both head page and tail pages will inherit mapping, flags, and so on from
3349  * the hugepage.
3350  *
3351  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3352  * they are not mapped.
3353  *
3354  * Returns 0 if the hugepage is split successfully.
3355  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3356  * us.
3357  */
3358 int split_huge_page_to_list(struct page *page, struct list_head *list)
3359 {
3360         struct page *head = compound_head(page);
3361         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3362         struct anon_vma *anon_vma;
3363         int count, mapcount, ret;
3364         bool mlocked;
3365         unsigned long flags;
3366
3367         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3368         VM_BUG_ON_PAGE(!PageAnon(page), page);
3369         VM_BUG_ON_PAGE(!PageLocked(page), page);
3370         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3371         VM_BUG_ON_PAGE(!PageCompound(page), page);
3372
3373         /*
3374          * The caller does not necessarily hold an mmap_sem that would prevent
3375          * the anon_vma disappearing so we first we take a reference to it
3376          * and then lock the anon_vma for write. This is similar to
3377          * page_lock_anon_vma_read except the write lock is taken to serialise
3378          * against parallel split or collapse operations.
3379          */
3380         anon_vma = page_get_anon_vma(head);
3381         if (!anon_vma) {
3382                 ret = -EBUSY;
3383                 goto out;
3384         }
3385         anon_vma_lock_write(anon_vma);
3386
3387         /*
3388          * Racy check if we can split the page, before freeze_page() will
3389          * split PMDs
3390          */
3391         if (total_mapcount(head) != page_count(head) - 1) {
3392                 ret = -EBUSY;
3393                 goto out_unlock;
3394         }
3395
3396         mlocked = PageMlocked(page);
3397         freeze_page(anon_vma, head);
3398         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3399
3400         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3401         if (mlocked)
3402                 lru_add_drain();
3403
3404         /* Prevent deferred_split_scan() touching ->_count */
3405         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3406         count = page_count(head);
3407         mapcount = total_mapcount(head);
3408         if (!mapcount && count == 1) {
3409                 if (!list_empty(page_deferred_list(head))) {
3410                         pgdata->split_queue_len--;
3411                         list_del(page_deferred_list(head));
3412                 }
3413                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3414                 __split_huge_page(page, list);
3415                 ret = 0;
3416         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3417                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3418                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3419                                 mapcount, count);
3420                 if (PageTail(page))
3421                         dump_page(head, NULL);
3422                 dump_page(page, "total_mapcount(head) > 0");
3423                 BUG();
3424         } else {
3425                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3426                 unfreeze_page(anon_vma, head);
3427                 ret = -EBUSY;
3428         }
3429
3430 out_unlock:
3431         anon_vma_unlock_write(anon_vma);
3432         put_anon_vma(anon_vma);
3433 out:
3434         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3435         return ret;
3436 }
3437
3438 void free_transhuge_page(struct page *page)
3439 {
3440         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3441         unsigned long flags;
3442
3443         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3444         if (!list_empty(page_deferred_list(page))) {
3445                 pgdata->split_queue_len--;
3446                 list_del(page_deferred_list(page));
3447         }
3448         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3449         free_compound_page(page);
3450 }
3451
3452 void deferred_split_huge_page(struct page *page)
3453 {
3454         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3455         unsigned long flags;
3456
3457         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3458
3459         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3460         if (list_empty(page_deferred_list(page))) {
3461                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3462                 pgdata->split_queue_len++;
3463         }
3464         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3465 }
3466
3467 static unsigned long deferred_split_count(struct shrinker *shrink,
3468                 struct shrink_control *sc)
3469 {
3470         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3471         return ACCESS_ONCE(pgdata->split_queue_len);
3472 }
3473
3474 static unsigned long deferred_split_scan(struct shrinker *shrink,
3475                 struct shrink_control *sc)
3476 {
3477         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3478         unsigned long flags;
3479         LIST_HEAD(list), *pos, *next;
3480         struct page *page;
3481         int split = 0;
3482
3483         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3484         /* Take pin on all head pages to avoid freeing them under us */
3485         list_for_each_safe(pos, next, &pgdata->split_queue) {
3486                 page = list_entry((void *)pos, struct page, mapping);
3487                 page = compound_head(page);
3488                 if (get_page_unless_zero(page)) {
3489                         list_move(page_deferred_list(page), &list);
3490                 } else {
3491                         /* We lost race with put_compound_page() */
3492                         list_del_init(page_deferred_list(page));
3493                         pgdata->split_queue_len--;
3494                 }
3495                 if (!--sc->nr_to_scan)
3496                         break;
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         /*
3515          * Stop shrinker if we didn't split any page, but the queue is empty.
3516          * This can happen if pages were freed under us.
3517          */
3518         if (!split && list_empty(&pgdata->split_queue))
3519                 return SHRINK_STOP;
3520         return split;
3521 }
3522
3523 static struct shrinker deferred_split_shrinker = {
3524         .count_objects = deferred_split_count,
3525         .scan_objects = deferred_split_scan,
3526         .seeks = DEFAULT_SEEKS,
3527         .flags = SHRINKER_NUMA_AWARE,
3528 };
3529
3530 #ifdef CONFIG_DEBUG_FS
3531 static int split_huge_pages_set(void *data, u64 val)
3532 {
3533         struct zone *zone;
3534         struct page *page;
3535         unsigned long pfn, max_zone_pfn;
3536         unsigned long total = 0, split = 0;
3537
3538         if (val != 1)
3539                 return -EINVAL;
3540
3541         for_each_populated_zone(zone) {
3542                 max_zone_pfn = zone_end_pfn(zone);
3543                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3544                         if (!pfn_valid(pfn))
3545                                 continue;
3546
3547                         page = pfn_to_page(pfn);
3548                         if (!get_page_unless_zero(page))
3549                                 continue;
3550
3551                         if (zone != page_zone(page))
3552                                 goto next;
3553
3554                         if (!PageHead(page) || !PageAnon(page) ||
3555                                         PageHuge(page))
3556                                 goto next;
3557
3558                         total++;
3559                         lock_page(page);
3560                         if (!split_huge_page(page))
3561                                 split++;
3562                         unlock_page(page);
3563 next:
3564                         put_page(page);
3565                 }
3566         }
3567
3568         pr_info("%lu of %lu THP split", split, total);
3569
3570         return 0;
3571 }
3572 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3573                 "%llu\n");
3574
3575 static int __init split_huge_pages_debugfs(void)
3576 {
3577         void *ret;
3578
3579         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3580                         &split_huge_pages_fops);
3581         if (!ret)
3582                 pr_warn("Failed to create split_huge_pages in debugfs");
3583         return 0;
3584 }
3585 late_initcall(split_huge_pages_debugfs);
3586 #endif