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/dax.h>
  20 #include <linux/kthread.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/mman.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/migrate.h>
  26 #include <linux/hashtable.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/page_idle.h>
  29
  30 #include <asm/tlb.h>
  31 #include <asm/pgalloc.h>
  32 #include "internal.h"
  33
  34 enum scan_result {
  35         SCAN_FAIL,
  36         SCAN_SUCCEED,
  37         SCAN_PMD_NULL,
  38         SCAN_EXCEED_NONE_PTE,
  39         SCAN_PTE_NON_PRESENT,
  40         SCAN_PAGE_RO,
  41         SCAN_NO_REFERENCED_PAGE,
  42         SCAN_PAGE_NULL,
  43         SCAN_SCAN_ABORT,
  44         SCAN_PAGE_COUNT,
  45         SCAN_PAGE_LRU,
  46         SCAN_PAGE_LOCK,
  47         SCAN_PAGE_ANON,
  48         SCAN_ANY_PROCESS,
  49         SCAN_VMA_NULL,
  50         SCAN_VMA_CHECK,
  51         SCAN_ADDRESS_RANGE,
  52         SCAN_SWAP_CACHE_PAGE,
  53         SCAN_DEL_PAGE_LRU,
  54         SCAN_ALLOC_HUGE_PAGE_FAIL,
  55         SCAN_CGROUP_CHARGE_FAIL
  56 };
  57
  58 #define CREATE_TRACE_POINTS
  59 #include <trace/events/huge_memory.h>
  60
  61 /*
  62  * By default transparent hugepage support is disabled in order that avoid
  63  * to risk increase the memory footprint of applications without a guaranteed
  64  * benefit. When transparent hugepage support is enabled, is for all mappings,
  65  * and khugepaged scans all mappings.
  66  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  67  * for all hugepage allocations.
  68  */
  69 unsigned long transparent_hugepage_flags __read_mostly =
  70 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  71         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  72 #endif
  73 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  74         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  75 #endif
  76         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  77         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  78         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  79
  80 /* default scan 8*512 pte (or vmas) every 30 second */
  81 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  82 static unsigned int khugepaged_pages_collapsed;
  83 static unsigned int khugepaged_full_scans;
  84 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  85 /* during fragmentation poll the hugepage allocator once every minute */
  86 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  87 static struct task_struct *khugepaged_thread __read_mostly;
  88 static DEFINE_MUTEX(khugepaged_mutex);
  89 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  90 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  91 /*
  92  * default collapse hugepages if there is at least one pte mapped like
  93  * it would have happened if the vma was large enough during page
  94  * fault.
  95  */
  96 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  97
  98 static int khugepaged(void *none);
  99 static int khugepaged_slab_init(void);
 100 static void khugepaged_slab_exit(void);
 101
 102 #define MM_SLOTS_HASH_BITS 10
 103 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 104
 105 static struct kmem_cache *mm_slot_cache __read_mostly;
 106
 107 /**
 108  * struct mm_slot - hash lookup from mm to mm_slot
 109  * @hash: hash collision list
 110  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 111  * @mm: the mm that this information is valid for
 112  */
 113 struct mm_slot {
 114         struct hlist_node hash;
 115         struct list_head mm_node;
 116         struct mm_struct *mm;
 117 };
 118
 119 /**
 120  * struct khugepaged_scan - cursor for scanning
 121  * @mm_head: the head of the mm list to scan
 122  * @mm_slot: the current mm_slot we are scanning
 123  * @address: the next address inside that to be scanned
 124  *
 125  * There is only the one khugepaged_scan instance of this cursor structure.
 126  */
 127 struct khugepaged_scan {
 128         struct list_head mm_head;
 129         struct mm_slot *mm_slot;
 130         unsigned long address;
 131 };
 132 static struct khugepaged_scan khugepaged_scan = {
 133         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 134 };
 135
 136
 137 static void set_recommended_min_free_kbytes(void)
 138 {
 139         struct zone *zone;
 140         int nr_zones = 0;
 141         unsigned long recommended_min;
 142
 143         for_each_populated_zone(zone)
 144                 nr_zones++;
 145
 146         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 147         recommended_min = pageblock_nr_pages * nr_zones * 2;
 148
 149         /*
 150          * Make sure that on average at least two pageblocks are almost free
 151          * of another type, one for a migratetype to fall back to and a
 152          * second to avoid subsequent fallbacks of other types There are 3
 153          * MIGRATE_TYPES we care about.
 154          */
 155         recommended_min += pageblock_nr_pages * nr_zones *
 156                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 157
 158         /* don't ever allow to reserve more than 5% of the lowmem */
 159         recommended_min = min(recommended_min,
 160                               (unsigned long) nr_free_buffer_pages() / 20);
 161         recommended_min <<= (PAGE_SHIFT-10);
 162
 163         if (recommended_min > min_free_kbytes) {
 164                 if (user_min_free_kbytes >= 0)
 165                         pr_info("raising min_free_kbytes from %d to %lu "
 166                                 "to help transparent hugepage allocations\n",
 167                                 min_free_kbytes, recommended_min);
 168
 169                 min_free_kbytes = recommended_min;
 170         }
 171         setup_per_zone_wmarks();
 172 }
 173
 174 static int start_stop_khugepaged(void)
 175 {
 176         int err = 0;
 177         if (khugepaged_enabled()) {
 178                 if (!khugepaged_thread)
 179                         khugepaged_thread = kthread_run(khugepaged, NULL,
 180                                                         "khugepaged");
 181                 if (IS_ERR(khugepaged_thread)) {
 182                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 183                         err = PTR_ERR(khugepaged_thread);
 184                         khugepaged_thread = NULL;
 185                         goto fail;
 186                 }
 187
 188                 if (!list_empty(&khugepaged_scan.mm_head))
 189                         wake_up_interruptible(&khugepaged_wait);
 190
 191                 set_recommended_min_free_kbytes();
 192         } else if (khugepaged_thread) {
 193                 kthread_stop(khugepaged_thread);
 194                 khugepaged_thread = NULL;
 195         }
 196 fail:
 197         return err;
 198 }
 199
 200 static atomic_t huge_zero_refcount;
 201 struct page *huge_zero_page __read_mostly;
 202
 203 struct page *get_huge_zero_page(void)
 204 {
 205         struct page *zero_page;
 206 retry:
 207         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 208                 return READ_ONCE(huge_zero_page);
 209
 210         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 211                         HPAGE_PMD_ORDER);
 212         if (!zero_page) {
 213                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 214                 return NULL;
 215         }
 216         count_vm_event(THP_ZERO_PAGE_ALLOC);
 217         preempt_disable();
 218         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 219                 preempt_enable();
 220                 __free_pages(zero_page, compound_order(zero_page));
 221                 goto retry;
 222         }
 223
 224         /* We take additional reference here. It will be put back by shrinker */
 225         atomic_set(&huge_zero_refcount, 2);
 226         preempt_enable();
 227         return READ_ONCE(huge_zero_page);
 228 }
 229
 230 static void put_huge_zero_page(void)
 231 {
 232         /*
 233          * Counter should never go to zero here. Only shrinker can put
 234          * last reference.
 235          */
 236         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 237 }
 238
 239 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 240                                         struct shrink_control *sc)
 241 {
 242         /* we can free zero page only if last reference remains */
 243         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 244 }
 245
 246 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 247                                        struct shrink_control *sc)
 248 {
 249         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 250                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 251                 BUG_ON(zero_page == NULL);
 252                 __free_pages(zero_page, compound_order(zero_page));
 253                 return HPAGE_PMD_NR;
 254         }
 255
 256         return 0;
 257 }
 258
 259 static struct shrinker huge_zero_page_shrinker = {
 260         .count_objects = shrink_huge_zero_page_count,
 261         .scan_objects = shrink_huge_zero_page_scan,
 262         .seeks = DEFAULT_SEEKS,
 263 };
 264
 265 #ifdef CONFIG_SYSFS
 266
 267 static ssize_t double_flag_show(struct kobject *kobj,
 268                                 struct kobj_attribute *attr, char *buf,
 269                                 enum transparent_hugepage_flag enabled,
 270                                 enum transparent_hugepage_flag req_madv)
 271 {
 272         if (test_bit(enabled, &transparent_hugepage_flags)) {
 273                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 274                 return sprintf(buf, "[always] madvise never\n");
 275         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 276                 return sprintf(buf, "always [madvise] never\n");
 277         else
 278                 return sprintf(buf, "always madvise [never]\n");
 279 }
 280 static ssize_t double_flag_store(struct kobject *kobj,
 281                                  struct kobj_attribute *attr,
 282                                  const char *buf, size_t count,
 283                                  enum transparent_hugepage_flag enabled,
 284                                  enum transparent_hugepage_flag req_madv)
 285 {
 286         if (!memcmp("always", buf,
 287                     min(sizeof("always")-1, count))) {
 288                 set_bit(enabled, &transparent_hugepage_flags);
 289                 clear_bit(req_madv, &transparent_hugepage_flags);
 290         } else if (!memcmp("madvise", buf,
 291                            min(sizeof("madvise")-1, count))) {
 292                 clear_bit(enabled, &transparent_hugepage_flags);
 293                 set_bit(req_madv, &transparent_hugepage_flags);
 294         } else if (!memcmp("never", buf,
 295                            min(sizeof("never")-1, count))) {
 296                 clear_bit(enabled, &transparent_hugepage_flags);
 297                 clear_bit(req_madv, &transparent_hugepage_flags);
 298         } else
 299                 return -EINVAL;
 300
 301         return count;
 302 }
 303
 304 static ssize_t enabled_show(struct kobject *kobj,
 305                             struct kobj_attribute *attr, char *buf)
 306 {
 307         return double_flag_show(kobj, attr, buf,
 308                                 TRANSPARENT_HUGEPAGE_FLAG,
 309                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 310 }
 311 static ssize_t enabled_store(struct kobject *kobj,
 312                              struct kobj_attribute *attr,
 313                              const char *buf, size_t count)
 314 {
 315         ssize_t ret;
 316
 317         ret = double_flag_store(kobj, attr, buf, count,
 318                                 TRANSPARENT_HUGEPAGE_FLAG,
 319                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 320
 321         if (ret > 0) {
 322                 int err;
 323
 324                 mutex_lock(&khugepaged_mutex);
 325                 err = start_stop_khugepaged();
 326                 mutex_unlock(&khugepaged_mutex);
 327
 328                 if (err)
 329                         ret = err;
 330         }
 331
 332         return ret;
 333 }
 334 static struct kobj_attribute enabled_attr =
 335         __ATTR(enabled, 0644, enabled_show, enabled_store);
 336
 337 static ssize_t single_flag_show(struct kobject *kobj,
 338                                 struct kobj_attribute *attr, char *buf,
 339                                 enum transparent_hugepage_flag flag)
 340 {
 341         return sprintf(buf, "%d\n",
 342                        !!test_bit(flag, &transparent_hugepage_flags));
 343 }
 344
 345 static ssize_t single_flag_store(struct kobject *kobj,
 346                                  struct kobj_attribute *attr,
 347                                  const char *buf, size_t count,
 348                                  enum transparent_hugepage_flag flag)
 349 {
 350         unsigned long value;
 351         int ret;
 352
 353         ret = kstrtoul(buf, 10, &value);
 354         if (ret < 0)
 355                 return ret;
 356         if (value > 1)
 357                 return -EINVAL;
 358
 359         if (value)
 360                 set_bit(flag, &transparent_hugepage_flags);
 361         else
 362                 clear_bit(flag, &transparent_hugepage_flags);
 363
 364         return count;
 365 }
 366
 367 /*
 368  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 369  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 370  * memory just to allocate one more hugepage.
 371  */
 372 static ssize_t defrag_show(struct kobject *kobj,
 373                            struct kobj_attribute *attr, char *buf)
 374 {
 375         return double_flag_show(kobj, attr, buf,
 376                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 377                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 378 }
 379 static ssize_t defrag_store(struct kobject *kobj,
 380                             struct kobj_attribute *attr,
 381                             const char *buf, size_t count)
 382 {
 383         return double_flag_store(kobj, attr, buf, count,
 384                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 385                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 386 }
 387 static struct kobj_attribute defrag_attr =
 388         __ATTR(defrag, 0644, defrag_show, defrag_store);
 389
 390 static ssize_t use_zero_page_show(struct kobject *kobj,
 391                 struct kobj_attribute *attr, char *buf)
 392 {
 393         return single_flag_show(kobj, attr, buf,
 394                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 395 }
 396 static ssize_t use_zero_page_store(struct kobject *kobj,
 397                 struct kobj_attribute *attr, const char *buf, size_t count)
 398 {
 399         return single_flag_store(kobj, attr, buf, count,
 400                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 401 }
 402 static struct kobj_attribute use_zero_page_attr =
 403         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 404 #ifdef CONFIG_DEBUG_VM
 405 static ssize_t debug_cow_show(struct kobject *kobj,
 406                                 struct kobj_attribute *attr, char *buf)
 407 {
 408         return single_flag_show(kobj, attr, buf,
 409                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 410 }
 411 static ssize_t debug_cow_store(struct kobject *kobj,
 412                                struct kobj_attribute *attr,
 413                                const char *buf, size_t count)
 414 {
 415         return single_flag_store(kobj, attr, buf, count,
 416                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 417 }
 418 static struct kobj_attribute debug_cow_attr =
 419         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 420 #endif /* CONFIG_DEBUG_VM */
 421
 422 static struct attribute *hugepage_attr[] = {
 423         &enabled_attr.attr,
 424         &defrag_attr.attr,
 425         &use_zero_page_attr.attr,
 426 #ifdef CONFIG_DEBUG_VM
 427         &debug_cow_attr.attr,
 428 #endif
 429         NULL,
 430 };
 431
 432 static struct attribute_group hugepage_attr_group = {
 433         .attrs = hugepage_attr,
 434 };
 435
 436 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 437                                          struct kobj_attribute *attr,
 438                                          char *buf)
 439 {
 440         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 441 }
 442
 443 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 444                                           struct kobj_attribute *attr,
 445                                           const char *buf, size_t count)
 446 {
 447         unsigned long msecs;
 448         int err;
 449
 450         err = kstrtoul(buf, 10, &msecs);
 451         if (err || msecs > UINT_MAX)
 452                 return -EINVAL;
 453
 454         khugepaged_scan_sleep_millisecs = msecs;
 455         wake_up_interruptible(&khugepaged_wait);
 456
 457         return count;
 458 }
 459 static struct kobj_attribute scan_sleep_millisecs_attr =
 460         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 461                scan_sleep_millisecs_store);
 462
 463 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 464                                           struct kobj_attribute *attr,
 465                                           char *buf)
 466 {
 467         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 468 }
 469
 470 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 471                                            struct kobj_attribute *attr,
 472                                            const char *buf, size_t count)
 473 {
 474         unsigned long msecs;
 475         int err;
 476
 477         err = kstrtoul(buf, 10, &msecs);
 478         if (err || msecs > UINT_MAX)
 479                 return -EINVAL;
 480
 481         khugepaged_alloc_sleep_millisecs = msecs;
 482         wake_up_interruptible(&khugepaged_wait);
 483
 484         return count;
 485 }
 486 static struct kobj_attribute alloc_sleep_millisecs_attr =
 487         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 488                alloc_sleep_millisecs_store);
 489
 490 static ssize_t pages_to_scan_show(struct kobject *kobj,
 491                                   struct kobj_attribute *attr,
 492                                   char *buf)
 493 {
 494         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 495 }
 496 static ssize_t pages_to_scan_store(struct kobject *kobj,
 497                                    struct kobj_attribute *attr,
 498                                    const char *buf, size_t count)
 499 {
 500         int err;
 501         unsigned long pages;
 502
 503         err = kstrtoul(buf, 10, &pages);
 504         if (err || !pages || pages > UINT_MAX)
 505                 return -EINVAL;
 506
 507         khugepaged_pages_to_scan = pages;
 508
 509         return count;
 510 }
 511 static struct kobj_attribute pages_to_scan_attr =
 512         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 513                pages_to_scan_store);
 514
 515 static ssize_t pages_collapsed_show(struct kobject *kobj,
 516                                     struct kobj_attribute *attr,
 517                                     char *buf)
 518 {
 519         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 520 }
 521 static struct kobj_attribute pages_collapsed_attr =
 522         __ATTR_RO(pages_collapsed);
 523
 524 static ssize_t full_scans_show(struct kobject *kobj,
 525                                struct kobj_attribute *attr,
 526                                char *buf)
 527 {
 528         return sprintf(buf, "%u\n", khugepaged_full_scans);
 529 }
 530 static struct kobj_attribute full_scans_attr =
 531         __ATTR_RO(full_scans);
 532
 533 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 534                                       struct kobj_attribute *attr, char *buf)
 535 {
 536         return single_flag_show(kobj, attr, buf,
 537                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 538 }
 539 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 540                                        struct kobj_attribute *attr,
 541                                        const char *buf, size_t count)
 542 {
 543         return single_flag_store(kobj, attr, buf, count,
 544                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 545 }
 546 static struct kobj_attribute khugepaged_defrag_attr =
 547         __ATTR(defrag, 0644, khugepaged_defrag_show,
 548                khugepaged_defrag_store);
 549
 550 /*
 551  * max_ptes_none controls if khugepaged should collapse hugepages over
 552  * any unmapped ptes in turn potentially increasing the memory
 553  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 554  * reduce the available free memory in the system as it
 555  * runs. Increasing max_ptes_none will instead potentially reduce the
 556  * free memory in the system during the khugepaged scan.
 557  */
 558 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 559                                              struct kobj_attribute *attr,
 560                                              char *buf)
 561 {
 562         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 563 }
 564 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 565                                               struct kobj_attribute *attr,
 566                                               const char *buf, size_t count)
 567 {
 568         int err;
 569         unsigned long max_ptes_none;
 570
 571         err = kstrtoul(buf, 10, &max_ptes_none);
 572         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 573                 return -EINVAL;
 574
 575         khugepaged_max_ptes_none = max_ptes_none;
 576
 577         return count;
 578 }
 579 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 580         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 581                khugepaged_max_ptes_none_store);
 582
 583 static struct attribute *khugepaged_attr[] = {
 584         &khugepaged_defrag_attr.attr,
 585         &khugepaged_max_ptes_none_attr.attr,
 586         &pages_to_scan_attr.attr,
 587         &pages_collapsed_attr.attr,
 588         &full_scans_attr.attr,
 589         &scan_sleep_millisecs_attr.attr,
 590         &alloc_sleep_millisecs_attr.attr,
 591         NULL,
 592 };
 593
 594 static struct attribute_group khugepaged_attr_group = {
 595         .attrs = khugepaged_attr,
 596         .name = "khugepaged",
 597 };
 598
 599 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 600 {
 601         int err;
 602
 603         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 604         if (unlikely(!*hugepage_kobj)) {
 605                 pr_err("failed to create transparent hugepage kobject\n");
 606                 return -ENOMEM;
 607         }
 608
 609         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 610         if (err) {
 611                 pr_err("failed to register transparent hugepage group\n");
 612                 goto delete_obj;
 613         }
 614
 615         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 616         if (err) {
 617                 pr_err("failed to register transparent hugepage group\n");
 618                 goto remove_hp_group;
 619         }
 620
 621         return 0;
 622
 623 remove_hp_group:
 624         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 625 delete_obj:
 626         kobject_put(*hugepage_kobj);
 627         return err;
 628 }
 629
 630 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 631 {
 632         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 633         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 634         kobject_put(hugepage_kobj);
 635 }
 636 #else
 637 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 638 {
 639         return 0;
 640 }
 641
 642 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 643 {
 644 }
 645 #endif /* CONFIG_SYSFS */
 646
 647 static int __init hugepage_init(void)
 648 {
 649         int err;
 650         struct kobject *hugepage_kobj;
 651
 652         if (!has_transparent_hugepage()) {
 653                 transparent_hugepage_flags = 0;
 654                 return -EINVAL;
 655         }
 656
 657         err = hugepage_init_sysfs(&hugepage_kobj);
 658         if (err)
 659                 goto err_sysfs;
 660
 661         err = khugepaged_slab_init();
 662         if (err)
 663                 goto err_slab;
 664
 665         err = register_shrinker(&huge_zero_page_shrinker);
 666         if (err)
 667                 goto err_hzp_shrinker;
 668
 669         /*
 670          * By default disable transparent hugepages on smaller systems,
 671          * where the extra memory used could hurt more than TLB overhead
 672          * is likely to save.  The admin can still enable it through /sys.
 673          */
 674         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 675                 transparent_hugepage_flags = 0;
 676                 return 0;
 677         }
 678
 679         err = start_stop_khugepaged();
 680         if (err)
 681                 goto err_khugepaged;
 682
 683         return 0;
 684 err_khugepaged:
 685         unregister_shrinker(&huge_zero_page_shrinker);
 686 err_hzp_shrinker:
 687         khugepaged_slab_exit();
 688 err_slab:
 689         hugepage_exit_sysfs(hugepage_kobj);
 690 err_sysfs:
 691         return err;
 692 }
 693 subsys_initcall(hugepage_init);
 694
 695 static int __init setup_transparent_hugepage(char *str)
 696 {
 697         int ret = 0;
 698         if (!str)
 699                 goto out;
 700         if (!strcmp(str, "always")) {
 701                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 702                         &transparent_hugepage_flags);
 703                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 704                           &transparent_hugepage_flags);
 705                 ret = 1;
 706         } else if (!strcmp(str, "madvise")) {
 707                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 708                           &transparent_hugepage_flags);
 709                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 710                         &transparent_hugepage_flags);
 711                 ret = 1;
 712         } else if (!strcmp(str, "never")) {
 713                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 714                           &transparent_hugepage_flags);
 715                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 716                           &transparent_hugepage_flags);
 717                 ret = 1;
 718         }
 719 out:
 720         if (!ret)
 721                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 722         return ret;
 723 }
 724 __setup("transparent_hugepage=", setup_transparent_hugepage);
 725
 726 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 727 {
 728         if (likely(vma->vm_flags & VM_WRITE))
 729                 pmd = pmd_mkwrite(pmd);
 730         return pmd;
 731 }
 732
 733 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 734 {
 735         pmd_t entry;
 736         entry = mk_pmd(page, prot);
 737         entry = pmd_mkhuge(entry);
 738         return entry;
 739 }
 740
 741 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 742                                         struct vm_area_struct *vma,
 743                                         unsigned long address, pmd_t *pmd,
 744                                         struct page *page, gfp_t gfp,
 745                                         unsigned int flags)
 746 {
 747         struct mem_cgroup *memcg;
 748         pgtable_t pgtable;
 749         spinlock_t *ptl;
 750         unsigned long haddr = address & HPAGE_PMD_MASK;
 751
 752         VM_BUG_ON_PAGE(!PageCompound(page), page);
 753
 754         if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
 755                 put_page(page);
 756                 count_vm_event(THP_FAULT_FALLBACK);
 757                 return VM_FAULT_FALLBACK;
 758         }
 759
 760         pgtable = pte_alloc_one(mm, haddr);
 761         if (unlikely(!pgtable)) {
 762                 mem_cgroup_cancel_charge(page, memcg);
 763                 put_page(page);
 764                 return VM_FAULT_OOM;
 765         }
 766
 767         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 768         /*
 769          * The memory barrier inside __SetPageUptodate makes sure that
 770          * clear_huge_page writes become visible before the set_pmd_at()
 771          * write.
 772          */
 773         __SetPageUptodate(page);
 774
 775         ptl = pmd_lock(mm, pmd);
 776         if (unlikely(!pmd_none(*pmd))) {
 777                 spin_unlock(ptl);
 778                 mem_cgroup_cancel_charge(page, memcg);
 779                 put_page(page);
 780                 pte_free(mm, pgtable);
 781         } else {
 782                 pmd_t entry;
 783
 784                 /* Deliver the page fault to userland */
 785                 if (userfaultfd_missing(vma)) {
 786                         int ret;
 787
 788                         spin_unlock(ptl);
 789                         mem_cgroup_cancel_charge(page, memcg);
 790                         put_page(page);
 791                         pte_free(mm, pgtable);
 792                         ret = handle_userfault(vma, address, flags,
 793                                                VM_UFFD_MISSING);
 794                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 795                         return ret;
 796                 }
 797
 798                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 799                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 800                 page_add_new_anon_rmap(page, vma, haddr);
 801                 mem_cgroup_commit_charge(page, memcg, false);
 802                 lru_cache_add_active_or_unevictable(page, vma);
 803                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 804                 set_pmd_at(mm, haddr, pmd, entry);
 805                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 806                 atomic_long_inc(&mm->nr_ptes);
 807                 spin_unlock(ptl);
 808                 count_vm_event(THP_FAULT_ALLOC);
 809         }
 810
 811         return 0;
 812 }
 813
 814 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 815 {
 816         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 817 }
 818
 819 /* Caller must hold page table lock. */
 820 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 821                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 822                 struct page *zero_page)
 823 {
 824         pmd_t entry;
 825         if (!pmd_none(*pmd))
 826                 return false;
 827         entry = mk_pmd(zero_page, vma->vm_page_prot);
 828         entry = pmd_mkhuge(entry);
 829         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 830         set_pmd_at(mm, haddr, pmd, entry);
 831         atomic_long_inc(&mm->nr_ptes);
 832         return true;
 833 }
 834
 835 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 836                                unsigned long address, pmd_t *pmd,
 837                                unsigned int flags)
 838 {
 839         gfp_t gfp;
 840         struct page *page;
 841         unsigned long haddr = address & HPAGE_PMD_MASK;
 842
 843         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 844                 return VM_FAULT_FALLBACK;
 845         if (unlikely(anon_vma_prepare(vma)))
 846                 return VM_FAULT_OOM;
 847         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 848                 return VM_FAULT_OOM;
 849         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 850                         transparent_hugepage_use_zero_page()) {
 851                 spinlock_t *ptl;
 852                 pgtable_t pgtable;
 853                 struct page *zero_page;
 854                 bool set;
 855                 int ret;
 856                 pgtable = pte_alloc_one(mm, haddr);
 857                 if (unlikely(!pgtable))
 858                         return VM_FAULT_OOM;
 859                 zero_page = get_huge_zero_page();
 860                 if (unlikely(!zero_page)) {
 861                         pte_free(mm, pgtable);
 862                         count_vm_event(THP_FAULT_FALLBACK);
 863                         return VM_FAULT_FALLBACK;
 864                 }
 865                 ptl = pmd_lock(mm, pmd);
 866                 ret = 0;
 867                 set = false;
 868                 if (pmd_none(*pmd)) {
 869                         if (userfaultfd_missing(vma)) {
 870                                 spin_unlock(ptl);
 871                                 ret = handle_userfault(vma, address, flags,
 872                                                        VM_UFFD_MISSING);
 873                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 874                         } else {
 875                                 set_huge_zero_page(pgtable, mm, vma,
 876                                                    haddr, pmd,
 877                                                    zero_page);
 878                                 spin_unlock(ptl);
 879                                 set = true;
 880                         }
 881                 } else
 882                         spin_unlock(ptl);
 883                 if (!set) {
 884                         pte_free(mm, pgtable);
 885                         put_huge_zero_page();
 886                 }
 887                 return ret;
 888         }
 889         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 890         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 891         if (unlikely(!page)) {
 892                 count_vm_event(THP_FAULT_FALLBACK);
 893                 return VM_FAULT_FALLBACK;
 894         }
 895         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 896                                             flags);
 897 }
 898
 899 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 900                 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
 901 {
 902         struct mm_struct *mm = vma->vm_mm;
 903         pmd_t entry;
 904         spinlock_t *ptl;
 905
 906         ptl = pmd_lock(mm, pmd);
 907         if (pmd_none(*pmd)) {
 908                 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
 909                 if (write) {
 910                         entry = pmd_mkyoung(pmd_mkdirty(entry));
 911                         entry = maybe_pmd_mkwrite(entry, vma);
 912                 }
 913                 set_pmd_at(mm, addr, pmd, entry);
 914                 update_mmu_cache_pmd(vma, addr, pmd);
 915         }
 916         spin_unlock(ptl);
 917 }
 918
 919 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 920                         pmd_t *pmd, unsigned long pfn, bool write)
 921 {
 922         pgprot_t pgprot = vma->vm_page_prot;
 923         /*
 924          * If we had pmd_special, we could avoid all these restrictions,
 925          * but we need to be consistent with PTEs and architectures that
 926          * can't support a 'special' bit.
 927          */
 928         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 929         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 930                                                 (VM_PFNMAP|VM_MIXEDMAP));
 931         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 932         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
 933
 934         if (addr < vma->vm_start || addr >= vma->vm_end)
 935                 return VM_FAULT_SIGBUS;
 936         if (track_pfn_insert(vma, &pgprot, pfn))
 937                 return VM_FAULT_SIGBUS;
 938         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 939         return VM_FAULT_NOPAGE;
 940 }
 941
 942 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 943                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 944                   struct vm_area_struct *vma)
 945 {
 946         spinlock_t *dst_ptl, *src_ptl;
 947         struct page *src_page;
 948         pmd_t pmd;
 949         pgtable_t pgtable;
 950         int ret;
 951
 952         ret = -ENOMEM;
 953         pgtable = pte_alloc_one(dst_mm, addr);
 954         if (unlikely(!pgtable))
 955                 goto out;
 956
 957         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 958         src_ptl = pmd_lockptr(src_mm, src_pmd);
 959         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 960
 961         ret = -EAGAIN;
 962         pmd = *src_pmd;
 963         if (unlikely(!pmd_trans_huge(pmd))) {
 964                 pte_free(dst_mm, pgtable);
 965                 goto out_unlock;
 966         }
 967         /*
 968          * When page table lock is held, the huge zero pmd should not be
 969          * under splitting since we don't split the page itself, only pmd to
 970          * a page table.
 971          */
 972         if (is_huge_zero_pmd(pmd)) {
 973                 struct page *zero_page;
 974                 /*
 975                  * get_huge_zero_page() will never allocate a new page here,
 976                  * since we already have a zero page to copy. It just takes a
 977                  * reference.
 978                  */
 979                 zero_page = get_huge_zero_page();
 980                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 981                                 zero_page);
 982                 ret = 0;
 983                 goto out_unlock;
 984         }
 985
 986         if (unlikely(pmd_trans_splitting(pmd))) {
 987                 /* split huge page running from under us */
 988                 spin_unlock(src_ptl);
 989                 spin_unlock(dst_ptl);
 990                 pte_free(dst_mm, pgtable);
 991
 992                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 993                 goto out;
 994         }
 995         src_page = pmd_page(pmd);
 996         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 997         get_page(src_page);
 998         page_dup_rmap(src_page);
 999         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1000
1001         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1002         pmd = pmd_mkold(pmd_wrprotect(pmd));
1003         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1004         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1005         atomic_long_inc(&dst_mm->nr_ptes);
1006
1007         ret = 0;
1008 out_unlock:
1009         spin_unlock(src_ptl);
1010         spin_unlock(dst_ptl);
1011 out:
1012         return ret;
1013 }
1014
1015 void huge_pmd_set_accessed(struct mm_struct *mm,
1016                            struct vm_area_struct *vma,
1017                            unsigned long address,
1018                            pmd_t *pmd, pmd_t orig_pmd,
1019                            int dirty)
1020 {
1021         spinlock_t *ptl;
1022         pmd_t entry;
1023         unsigned long haddr;
1024
1025         ptl = pmd_lock(mm, pmd);
1026         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1027                 goto unlock;
1028
1029         entry = pmd_mkyoung(orig_pmd);
1030         haddr = address & HPAGE_PMD_MASK;
1031         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1032                 update_mmu_cache_pmd(vma, address, pmd);
1033
1034 unlock:
1035         spin_unlock(ptl);
1036 }
1037
1038 /*
1039  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1040  * during copy_user_huge_page()'s copy_page_rep(): in the case when
1041  * the source page gets split and a tail freed before copy completes.
1042  * Called under pmd_lock of checked pmd, so safe from splitting itself.
1043  */
1044 static void get_user_huge_page(struct page *page)
1045 {
1046         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1047                 struct page *endpage = page + HPAGE_PMD_NR;
1048
1049                 atomic_add(HPAGE_PMD_NR, &page->_count);
1050                 while (++page < endpage)
1051                         get_huge_page_tail(page);
1052         } else {
1053                 get_page(page);
1054         }
1055 }
1056
1057 static void put_user_huge_page(struct page *page)
1058 {
1059         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1060                 struct page *endpage = page + HPAGE_PMD_NR;
1061
1062                 while (page < endpage)
1063                         put_page(page++);
1064         } else {
1065                 put_page(page);
1066         }
1067 }
1068
1069 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1070                                         struct vm_area_struct *vma,
1071                                         unsigned long address,
1072                                         pmd_t *pmd, pmd_t orig_pmd,
1073                                         struct page *page,
1074                                         unsigned long haddr)
1075 {
1076         struct mem_cgroup *memcg;
1077         spinlock_t *ptl;
1078         pgtable_t pgtable;
1079         pmd_t _pmd;
1080         int ret = 0, i;
1081         struct page **pages;
1082         unsigned long mmun_start;       /* For mmu_notifiers */
1083         unsigned long mmun_end;         /* For mmu_notifiers */
1084
1085         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1086                         GFP_KERNEL);
1087         if (unlikely(!pages)) {
1088                 ret |= VM_FAULT_OOM;
1089                 goto out;
1090         }
1091
1092         for (i = 0; i < HPAGE_PMD_NR; i++) {
1093                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1094                                                __GFP_OTHER_NODE,
1095                                                vma, address, page_to_nid(page));
1096                 if (unlikely(!pages[i] ||
1097                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1098                                                    &memcg))) {
1099                         if (pages[i])
1100                                 put_page(pages[i]);
1101                         while (--i >= 0) {
1102                                 memcg = (void *)page_private(pages[i]);
1103                                 set_page_private(pages[i], 0);
1104                                 mem_cgroup_cancel_charge(pages[i], memcg);
1105                                 put_page(pages[i]);
1106                         }
1107                         kfree(pages);
1108                         ret |= VM_FAULT_OOM;
1109                         goto out;
1110                 }
1111                 set_page_private(pages[i], (unsigned long)memcg);
1112         }
1113
1114         for (i = 0; i < HPAGE_PMD_NR; i++) {
1115                 copy_user_highpage(pages[i], page + i,
1116                                    haddr + PAGE_SIZE * i, vma);
1117                 __SetPageUptodate(pages[i]);
1118                 cond_resched();
1119         }
1120
1121         mmun_start = haddr;
1122         mmun_end   = haddr + HPAGE_PMD_SIZE;
1123         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1124
1125         ptl = pmd_lock(mm, pmd);
1126         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1127                 goto out_free_pages;
1128         VM_BUG_ON_PAGE(!PageHead(page), page);
1129
1130         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1131         /* leave pmd empty until pte is filled */
1132
1133         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1134         pmd_populate(mm, &_pmd, pgtable);
1135
1136         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1137                 pte_t *pte, entry;
1138                 entry = mk_pte(pages[i], vma->vm_page_prot);
1139                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1140                 memcg = (void *)page_private(pages[i]);
1141                 set_page_private(pages[i], 0);
1142                 page_add_new_anon_rmap(pages[i], vma, haddr);
1143                 mem_cgroup_commit_charge(pages[i], memcg, false);
1144                 lru_cache_add_active_or_unevictable(pages[i], vma);
1145                 pte = pte_offset_map(&_pmd, haddr);
1146                 VM_BUG_ON(!pte_none(*pte));
1147                 set_pte_at(mm, haddr, pte, entry);
1148                 pte_unmap(pte);
1149         }
1150         kfree(pages);
1151
1152         smp_wmb(); /* make pte visible before pmd */
1153         pmd_populate(mm, pmd, pgtable);
1154         page_remove_rmap(page);
1155         spin_unlock(ptl);
1156
1157         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1158
1159         ret |= VM_FAULT_WRITE;
1160         put_page(page);
1161
1162 out:
1163         return ret;
1164
1165 out_free_pages:
1166         spin_unlock(ptl);
1167         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1168         for (i = 0; i < HPAGE_PMD_NR; i++) {
1169                 memcg = (void *)page_private(pages[i]);
1170                 set_page_private(pages[i], 0);
1171                 mem_cgroup_cancel_charge(pages[i], memcg);
1172                 put_page(pages[i]);
1173         }
1174         kfree(pages);
1175         goto out;
1176 }
1177
1178 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1179                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1180 {
1181         spinlock_t *ptl;
1182         int ret = 0;
1183         struct page *page = NULL, *new_page;
1184         struct mem_cgroup *memcg;
1185         unsigned long haddr;
1186         unsigned long mmun_start;       /* For mmu_notifiers */
1187         unsigned long mmun_end;         /* For mmu_notifiers */
1188         gfp_t huge_gfp;                 /* for allocation and charge */
1189
1190         ptl = pmd_lockptr(mm, pmd);
1191         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1192         haddr = address & HPAGE_PMD_MASK;
1193         if (is_huge_zero_pmd(orig_pmd))
1194                 goto alloc;
1195         spin_lock(ptl);
1196         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1197                 goto out_unlock;
1198
1199         page = pmd_page(orig_pmd);
1200         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1201         if (page_mapcount(page) == 1) {
1202                 pmd_t entry;
1203                 entry = pmd_mkyoung(orig_pmd);
1204                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1205                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1206                         update_mmu_cache_pmd(vma, address, pmd);
1207                 ret |= VM_FAULT_WRITE;
1208                 goto out_unlock;
1209         }
1210         get_user_huge_page(page);
1211         spin_unlock(ptl);
1212 alloc:
1213         if (transparent_hugepage_enabled(vma) &&
1214             !transparent_hugepage_debug_cow()) {
1215                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1216                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1217         } else
1218                 new_page = NULL;
1219
1220         if (unlikely(!new_page)) {
1221                 if (!page) {
1222                         split_huge_page_pmd(vma, address, pmd);
1223                         ret |= VM_FAULT_FALLBACK;
1224                 } else {
1225                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1226                                         pmd, orig_pmd, page, haddr);
1227                         if (ret & VM_FAULT_OOM) {
1228                                 split_huge_page(page);
1229                                 ret |= VM_FAULT_FALLBACK;
1230                         }
1231                         put_user_huge_page(page);
1232                 }
1233                 count_vm_event(THP_FAULT_FALLBACK);
1234                 goto out;
1235         }
1236
1237         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1238                 put_page(new_page);
1239                 if (page) {
1240                         split_huge_page(page);
1241                         put_user_huge_page(page);
1242                 } else
1243                         split_huge_page_pmd(vma, address, pmd);
1244                 ret |= VM_FAULT_FALLBACK;
1245                 count_vm_event(THP_FAULT_FALLBACK);
1246                 goto out;
1247         }
1248
1249         count_vm_event(THP_FAULT_ALLOC);
1250
1251         if (!page)
1252                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1253         else
1254                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1255         __SetPageUptodate(new_page);
1256
1257         mmun_start = haddr;
1258         mmun_end   = haddr + HPAGE_PMD_SIZE;
1259         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1260
1261         spin_lock(ptl);
1262         if (page)
1263                 put_user_huge_page(page);
1264         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1265                 spin_unlock(ptl);
1266                 mem_cgroup_cancel_charge(new_page, memcg);
1267                 put_page(new_page);
1268                 goto out_mn;
1269         } else {
1270                 pmd_t entry;
1271                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1272                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1273                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1274                 page_add_new_anon_rmap(new_page, vma, haddr);
1275                 mem_cgroup_commit_charge(new_page, memcg, false);
1276                 lru_cache_add_active_or_unevictable(new_page, vma);
1277                 set_pmd_at(mm, haddr, pmd, entry);
1278                 update_mmu_cache_pmd(vma, address, pmd);
1279                 if (!page) {
1280                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1281                         put_huge_zero_page();
1282                 } else {
1283                         VM_BUG_ON_PAGE(!PageHead(page), page);
1284                         page_remove_rmap(page);
1285                         put_page(page);
1286                 }
1287                 ret |= VM_FAULT_WRITE;
1288         }
1289         spin_unlock(ptl);
1290 out_mn:
1291         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1292 out:
1293         return ret;
1294 out_unlock:
1295         spin_unlock(ptl);
1296         return ret;
1297 }
1298
1299 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1300                                    unsigned long addr,
1301                                    pmd_t *pmd,
1302                                    unsigned int flags)
1303 {
1304         struct mm_struct *mm = vma->vm_mm;
1305         struct page *page = NULL;
1306
1307         assert_spin_locked(pmd_lockptr(mm, pmd));
1308
1309         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1310                 goto out;
1311
1312         /* Avoid dumping huge zero page */
1313         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1314                 return ERR_PTR(-EFAULT);
1315
1316         /* Full NUMA hinting faults to serialise migration in fault paths */
1317         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1318                 goto out;
1319
1320         page = pmd_page(*pmd);
1321         VM_BUG_ON_PAGE(!PageHead(page), page);
1322         if (flags & FOLL_TOUCH) {
1323                 pmd_t _pmd;
1324                 /*
1325                  * We should set the dirty bit only for FOLL_WRITE but
1326                  * for now the dirty bit in the pmd is meaningless.
1327                  * And if the dirty bit will become meaningful and
1328                  * we'll only set it with FOLL_WRITE, an atomic
1329                  * set_bit will be required on the pmd to set the
1330                  * young bit, instead of the current set_pmd_at.
1331                  */
1332                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1333                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1334                                           pmd, _pmd,  1))
1335                         update_mmu_cache_pmd(vma, addr, pmd);
1336         }
1337         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1338                 if (page->mapping && trylock_page(page)) {
1339                         lru_add_drain();
1340                         if (page->mapping)
1341                                 mlock_vma_page(page);
1342                         unlock_page(page);
1343                 }
1344         }
1345         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1346         VM_BUG_ON_PAGE(!PageCompound(page), page);
1347         if (flags & FOLL_GET)
1348                 get_page_foll(page);
1349
1350 out:
1351         return page;
1352 }
1353
1354 /* NUMA hinting page fault entry point for trans huge pmds */
1355 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1356                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1357 {
1358         spinlock_t *ptl;
1359         struct anon_vma *anon_vma = NULL;
1360         struct page *page;
1361         unsigned long haddr = addr & HPAGE_PMD_MASK;
1362         int page_nid = -1, this_nid = numa_node_id();
1363         int target_nid, last_cpupid = -1;
1364         bool page_locked;
1365         bool migrated = false;
1366         bool was_writable;
1367         int flags = 0;
1368
1369         /* A PROT_NONE fault should not end up here */
1370         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1371
1372         ptl = pmd_lock(mm, pmdp);
1373         if (unlikely(!pmd_same(pmd, *pmdp)))
1374                 goto out_unlock;
1375
1376         /*
1377          * If there are potential migrations, wait for completion and retry
1378          * without disrupting NUMA hinting information. Do not relock and
1379          * check_same as the page may no longer be mapped.
1380          */
1381         if (unlikely(pmd_trans_migrating(*pmdp))) {
1382                 page = pmd_page(*pmdp);
1383                 spin_unlock(ptl);
1384                 wait_on_page_locked(page);
1385                 goto out;
1386         }
1387
1388         page = pmd_page(pmd);
1389         BUG_ON(is_huge_zero_page(page));
1390         page_nid = page_to_nid(page);
1391         last_cpupid = page_cpupid_last(page);
1392         count_vm_numa_event(NUMA_HINT_FAULTS);
1393         if (page_nid == this_nid) {
1394                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1395                 flags |= TNF_FAULT_LOCAL;
1396         }
1397
1398         /* See similar comment in do_numa_page for explanation */
1399         if (!(vma->vm_flags & VM_WRITE))
1400                 flags |= TNF_NO_GROUP;
1401
1402         /*
1403          * Acquire the page lock to serialise THP migrations but avoid dropping
1404          * page_table_lock if at all possible
1405          */
1406         page_locked = trylock_page(page);
1407         target_nid = mpol_misplaced(page, vma, haddr);
1408         if (target_nid == -1) {
1409                 /* If the page was locked, there are no parallel migrations */
1410                 if (page_locked)
1411                         goto clear_pmdnuma;
1412         }
1413
1414         /* Migration could have started since the pmd_trans_migrating check */
1415         if (!page_locked) {
1416                 spin_unlock(ptl);
1417                 wait_on_page_locked(page);
1418                 page_nid = -1;
1419                 goto out;
1420         }
1421
1422         /*
1423          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1424          * to serialises splits
1425          */
1426         get_page(page);
1427         spin_unlock(ptl);
1428         anon_vma = page_lock_anon_vma_read(page);
1429
1430         /* Confirm the PMD did not change while page_table_lock was released */
1431         spin_lock(ptl);
1432         if (unlikely(!pmd_same(pmd, *pmdp))) {
1433                 unlock_page(page);
1434                 put_page(page);
1435                 page_nid = -1;
1436                 goto out_unlock;
1437         }
1438
1439         /* Bail if we fail to protect against THP splits for any reason */
1440         if (unlikely(!anon_vma)) {
1441                 put_page(page);
1442                 page_nid = -1;
1443                 goto clear_pmdnuma;
1444         }
1445
1446         /*
1447          * Migrate the THP to the requested node, returns with page unlocked
1448          * and access rights restored.
1449          */
1450         spin_unlock(ptl);
1451         migrated = migrate_misplaced_transhuge_page(mm, vma,
1452                                 pmdp, pmd, addr, page, target_nid);
1453         if (migrated) {
1454                 flags |= TNF_MIGRATED;
1455                 page_nid = target_nid;
1456         } else
1457                 flags |= TNF_MIGRATE_FAIL;
1458
1459         goto out;
1460 clear_pmdnuma:
1461         BUG_ON(!PageLocked(page));
1462         was_writable = pmd_write(pmd);
1463         pmd = pmd_modify(pmd, vma->vm_page_prot);
1464         pmd = pmd_mkyoung(pmd);
1465         if (was_writable)
1466                 pmd = pmd_mkwrite(pmd);
1467         set_pmd_at(mm, haddr, pmdp, pmd);
1468         update_mmu_cache_pmd(vma, addr, pmdp);
1469         unlock_page(page);
1470 out_unlock:
1471         spin_unlock(ptl);
1472
1473 out:
1474         if (anon_vma)
1475                 page_unlock_anon_vma_read(anon_vma);
1476
1477         if (page_nid != -1)
1478                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1479
1480         return 0;
1481 }
1482
1483 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1484                  pmd_t *pmd, unsigned long addr)
1485 {
1486         pmd_t orig_pmd;
1487         spinlock_t *ptl;
1488
1489         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1490                 return 0;
1491         /*
1492          * For architectures like ppc64 we look at deposited pgtable
1493          * when calling pmdp_huge_get_and_clear. So do the
1494          * pgtable_trans_huge_withdraw after finishing pmdp related
1495          * operations.
1496          */
1497         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1498                         tlb->fullmm);
1499         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1500         if (vma_is_dax(vma)) {
1501                 spin_unlock(ptl);
1502                 if (is_huge_zero_pmd(orig_pmd))
1503                         put_huge_zero_page();
1504         } else if (is_huge_zero_pmd(orig_pmd)) {
1505                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1506                 atomic_long_dec(&tlb->mm->nr_ptes);
1507                 spin_unlock(ptl);
1508                 put_huge_zero_page();
1509         } else {
1510                 struct page *page = pmd_page(orig_pmd);
1511                 page_remove_rmap(page);
1512                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1513                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1514                 VM_BUG_ON_PAGE(!PageHead(page), page);
1515                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1516                 atomic_long_dec(&tlb->mm->nr_ptes);
1517                 spin_unlock(ptl);
1518                 tlb_remove_page(tlb, page);
1519         }
1520         return 1;
1521 }
1522
1523 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1524                   unsigned long old_addr,
1525                   unsigned long new_addr, unsigned long old_end,
1526                   pmd_t *old_pmd, pmd_t *new_pmd)
1527 {
1528         spinlock_t *old_ptl, *new_ptl;
1529         int ret = 0;
1530         pmd_t pmd;
1531
1532         struct mm_struct *mm = vma->vm_mm;
1533
1534         if ((old_addr & ~HPAGE_PMD_MASK) ||
1535             (new_addr & ~HPAGE_PMD_MASK) ||
1536             old_end - old_addr < HPAGE_PMD_SIZE ||
1537             (new_vma->vm_flags & VM_NOHUGEPAGE))
1538                 goto out;
1539
1540         /*
1541          * The destination pmd shouldn't be established, free_pgtables()
1542          * should have release it.
1543          */
1544         if (WARN_ON(!pmd_none(*new_pmd))) {
1545                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1546                 goto out;
1547         }
1548
1549         /*
1550          * We don't have to worry about the ordering of src and dst
1551          * ptlocks because exclusive mmap_sem prevents deadlock.
1552          */
1553         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1554         if (ret == 1) {
1555                 new_ptl = pmd_lockptr(mm, new_pmd);
1556                 if (new_ptl != old_ptl)
1557                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1558                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1559                 VM_BUG_ON(!pmd_none(*new_pmd));
1560
1561                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1562                         pgtable_t pgtable;
1563                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1564                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1565                 }
1566                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1567                 if (new_ptl != old_ptl)
1568                         spin_unlock(new_ptl);
1569                 spin_unlock(old_ptl);
1570         }
1571 out:
1572         return ret;
1573 }
1574
1575 /*
1576  * Returns
1577  *  - 0 if PMD could not be locked
1578  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1579  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1580  */
1581 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1582                 unsigned long addr, pgprot_t newprot, int prot_numa)
1583 {
1584         struct mm_struct *mm = vma->vm_mm;
1585         spinlock_t *ptl;
1586         int ret = 0;
1587
1588         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1589                 pmd_t entry;
1590                 bool preserve_write = prot_numa && pmd_write(*pmd);
1591                 ret = 1;
1592
1593                 /*
1594                  * Avoid trapping faults against the zero page. The read-only
1595                  * data is likely to be read-cached on the local CPU and
1596                  * local/remote hits to the zero page are not interesting.
1597                  */
1598                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1599                         spin_unlock(ptl);
1600                         return ret;
1601                 }
1602
1603                 if (!prot_numa || !pmd_protnone(*pmd)) {
1604                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1605                         entry = pmd_modify(entry, newprot);
1606                         if (preserve_write)
1607                                 entry = pmd_mkwrite(entry);
1608                         ret = HPAGE_PMD_NR;
1609                         set_pmd_at(mm, addr, pmd, entry);
1610                         BUG_ON(!preserve_write && pmd_write(entry));
1611                 }
1612                 spin_unlock(ptl);
1613         }
1614
1615         return ret;
1616 }
1617
1618 /*
1619  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1620  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1621  *
1622  * Note that if it returns 1, this routine returns without unlocking page
1623  * table locks. So callers must unlock them.
1624  */
1625 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1626                 spinlock_t **ptl)
1627 {
1628         *ptl = pmd_lock(vma->vm_mm, pmd);
1629         if (likely(pmd_trans_huge(*pmd))) {
1630                 if (unlikely(pmd_trans_splitting(*pmd))) {
1631                         spin_unlock(*ptl);
1632                         wait_split_huge_page(vma->anon_vma, pmd);
1633                         return -1;
1634                 } else {
1635                         /* Thp mapped by 'pmd' is stable, so we can
1636                          * handle it as it is. */
1637                         return 1;
1638                 }
1639         }
1640         spin_unlock(*ptl);
1641         return 0;
1642 }
1643
1644 /*
1645  * This function returns whether a given @page is mapped onto the @address
1646  * in the virtual space of @mm.
1647  *
1648  * When it's true, this function returns *pmd with holding the page table lock
1649  * and passing it back to the caller via @ptl.
1650  * If it's false, returns NULL without holding the page table lock.
1651  */
1652 pmd_t *page_check_address_pmd(struct page *page,
1653                               struct mm_struct *mm,
1654                               unsigned long address,
1655                               enum page_check_address_pmd_flag flag,
1656                               spinlock_t **ptl)
1657 {
1658         pgd_t *pgd;
1659         pud_t *pud;
1660         pmd_t *pmd;
1661
1662         if (address & ~HPAGE_PMD_MASK)
1663                 return NULL;
1664
1665         pgd = pgd_offset(mm, address);
1666         if (!pgd_present(*pgd))
1667                 return NULL;
1668         pud = pud_offset(pgd, address);
1669         if (!pud_present(*pud))
1670                 return NULL;
1671         pmd = pmd_offset(pud, address);
1672
1673         *ptl = pmd_lock(mm, pmd);
1674         if (!pmd_present(*pmd))
1675                 goto unlock;
1676         if (pmd_page(*pmd) != page)
1677                 goto unlock;
1678         /*
1679          * split_vma() may create temporary aliased mappings. There is
1680          * no risk as long as all huge pmd are found and have their
1681          * splitting bit set before __split_huge_page_refcount
1682          * runs. Finding the same huge pmd more than once during the
1683          * same rmap walk is not a problem.
1684          */
1685         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1686             pmd_trans_splitting(*pmd))
1687                 goto unlock;
1688         if (pmd_trans_huge(*pmd)) {
1689                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1690                           !pmd_trans_splitting(*pmd));
1691                 return pmd;
1692         }
1693 unlock:
1694         spin_unlock(*ptl);
1695         return NULL;
1696 }
1697
1698 static int __split_huge_page_splitting(struct page *page,
1699                                        struct vm_area_struct *vma,
1700                                        unsigned long address)
1701 {
1702         struct mm_struct *mm = vma->vm_mm;
1703         spinlock_t *ptl;
1704         pmd_t *pmd;
1705         int ret = 0;
1706         /* For mmu_notifiers */
1707         const unsigned long mmun_start = address;
1708         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1709
1710         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1711         pmd = page_check_address_pmd(page, mm, address,
1712                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1713         if (pmd) {
1714                 /*
1715                  * We can't temporarily set the pmd to null in order
1716                  * to split it, the pmd must remain marked huge at all
1717                  * times or the VM won't take the pmd_trans_huge paths
1718                  * and it won't wait on the anon_vma->root->rwsem to
1719                  * serialize against split_huge_page*.
1720                  */
1721                 pmdp_splitting_flush(vma, address, pmd);
1722
1723                 ret = 1;
1724                 spin_unlock(ptl);
1725         }
1726         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1727
1728         return ret;
1729 }
1730
1731 static void __split_huge_page_refcount(struct page *page,
1732                                        struct list_head *list)
1733 {
1734         int i;
1735         struct zone *zone = page_zone(page);
1736         struct lruvec *lruvec;
1737         int tail_count = 0;
1738
1739         /* prevent PageLRU to go away from under us, and freeze lru stats */
1740         spin_lock_irq(&zone->lru_lock);
1741         lruvec = mem_cgroup_page_lruvec(page, zone);
1742
1743         compound_lock(page);
1744         /* complete memcg works before add pages to LRU */
1745         mem_cgroup_split_huge_fixup(page);
1746
1747         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1748                 struct page *page_tail = page + i;
1749
1750                 /* tail_page->_mapcount cannot change */
1751                 BUG_ON(page_mapcount(page_tail) < 0);
1752                 tail_count += page_mapcount(page_tail);
1753                 /* check for overflow */
1754                 BUG_ON(tail_count < 0);
1755                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1756                 /*
1757                  * tail_page->_count is zero and not changing from
1758                  * under us. But get_page_unless_zero() may be running
1759                  * from under us on the tail_page. If we used
1760                  * atomic_set() below instead of atomic_add(), we
1761                  * would then run atomic_set() concurrently with
1762                  * get_page_unless_zero(), and atomic_set() is
1763                  * implemented in C not using locked ops. spin_unlock
1764                  * on x86 sometime uses locked ops because of PPro
1765                  * errata 66, 92, so unless somebody can guarantee
1766                  * atomic_set() here would be safe on all archs (and
1767                  * not only on x86), it's safer to use atomic_add().
1768                  */
1769                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1770                            &page_tail->_count);
1771
1772                 /* after clearing PageTail the gup refcount can be released */
1773                 smp_mb__after_atomic();
1774
1775                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1776                 page_tail->flags |= (page->flags &
1777                                      ((1L << PG_referenced) |
1778                                       (1L << PG_swapbacked) |
1779                                       (1L << PG_mlocked) |
1780                                       (1L << PG_uptodate) |
1781                                       (1L << PG_active) |
1782                                       (1L << PG_unevictable)));
1783                 page_tail->flags |= (1L << PG_dirty);
1784
1785                 clear_compound_head(page_tail);
1786
1787                 if (page_is_young(page))
1788                         set_page_young(page_tail);
1789                 if (page_is_idle(page))
1790                         set_page_idle(page_tail);
1791
1792                 /*
1793                  * __split_huge_page_splitting() already set the
1794                  * splitting bit in all pmd that could map this
1795                  * hugepage, that will ensure no CPU can alter the
1796                  * mapcount on the head page. The mapcount is only
1797                  * accounted in the head page and it has to be
1798                  * transferred to all tail pages in the below code. So
1799                  * for this code to be safe, the split the mapcount
1800                  * can't change. But that doesn't mean userland can't
1801                  * keep changing and reading the page contents while
1802                  * we transfer the mapcount, so the pmd splitting
1803                  * status is achieved setting a reserved bit in the
1804                  * pmd, not by clearing the present bit.
1805                 */
1806                 page_tail->_mapcount = page->_mapcount;
1807
1808                 BUG_ON(page_tail->mapping);
1809                 page_tail->mapping = page->mapping;
1810
1811                 page_tail->index = page->index + i;
1812                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1813
1814                 BUG_ON(!PageAnon(page_tail));
1815                 BUG_ON(!PageUptodate(page_tail));
1816                 BUG_ON(!PageDirty(page_tail));
1817                 BUG_ON(!PageSwapBacked(page_tail));
1818
1819                 lru_add_page_tail(page, page_tail, lruvec, list);
1820         }
1821         atomic_sub(tail_count, &page->_count);
1822         BUG_ON(atomic_read(&page->_count) <= 0);
1823
1824         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1825
1826         ClearPageCompound(page);
1827         compound_unlock(page);
1828         spin_unlock_irq(&zone->lru_lock);
1829
1830         for (i = 1; i < HPAGE_PMD_NR; i++) {
1831                 struct page *page_tail = page + i;
1832                 BUG_ON(page_count(page_tail) <= 0);
1833                 /*
1834                  * Tail pages may be freed if there wasn't any mapping
1835                  * like if add_to_swap() is running on a lru page that
1836                  * had its mapping zapped. And freeing these pages
1837                  * requires taking the lru_lock so we do the put_page
1838                  * of the tail pages after the split is complete.
1839                  */
1840                 put_page(page_tail);
1841         }
1842
1843         /*
1844          * Only the head page (now become a regular page) is required
1845          * to be pinned by the caller.
1846          */
1847         BUG_ON(page_count(page) <= 0);
1848 }
1849
1850 static int __split_huge_page_map(struct page *page,
1851                                  struct vm_area_struct *vma,
1852                                  unsigned long address)
1853 {
1854         struct mm_struct *mm = vma->vm_mm;
1855         spinlock_t *ptl;
1856         pmd_t *pmd, _pmd;
1857         int ret = 0, i;
1858         pgtable_t pgtable;
1859         unsigned long haddr;
1860
1861         pmd = page_check_address_pmd(page, mm, address,
1862                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1863         if (pmd) {
1864                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1865                 pmd_populate(mm, &_pmd, pgtable);
1866                 if (pmd_write(*pmd))
1867                         BUG_ON(page_mapcount(page) != 1);
1868
1869                 haddr = address;
1870                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1871                         pte_t *pte, entry;
1872                         BUG_ON(PageCompound(page+i));
1873                         /*
1874                          * Note that NUMA hinting access restrictions are not
1875                          * transferred to avoid any possibility of altering
1876                          * permissions across VMAs.
1877                          */
1878                         entry = mk_pte(page + i, vma->vm_page_prot);
1879                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1880                         if (!pmd_write(*pmd))
1881                                 entry = pte_wrprotect(entry);
1882                         if (!pmd_young(*pmd))
1883                                 entry = pte_mkold(entry);
1884                         pte = pte_offset_map(&_pmd, haddr);
1885                         BUG_ON(!pte_none(*pte));
1886                         set_pte_at(mm, haddr, pte, entry);
1887                         pte_unmap(pte);
1888                 }
1889
1890                 smp_wmb(); /* make pte visible before pmd */
1891                 /*
1892                  * Up to this point the pmd is present and huge and
1893                  * userland has the whole access to the hugepage
1894                  * during the split (which happens in place). If we
1895                  * overwrite the pmd with the not-huge version
1896                  * pointing to the pte here (which of course we could
1897                  * if all CPUs were bug free), userland could trigger
1898                  * a small page size TLB miss on the small sized TLB
1899                  * while the hugepage TLB entry is still established
1900                  * in the huge TLB. Some CPU doesn't like that. See
1901                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1902                  * Erratum 383 on page 93. Intel should be safe but is
1903                  * also warns that it's only safe if the permission
1904                  * and cache attributes of the two entries loaded in
1905                  * the two TLB is identical (which should be the case
1906                  * here). But it is generally safer to never allow
1907                  * small and huge TLB entries for the same virtual
1908                  * address to be loaded simultaneously. So instead of
1909                  * doing "pmd_populate(); flush_pmd_tlb_range();" we first
1910                  * mark the current pmd notpresent (atomically because
1911                  * here the pmd_trans_huge and pmd_trans_splitting
1912                  * must remain set at all times on the pmd until the
1913                  * split is complete for this pmd), then we flush the
1914                  * SMP TLB and finally we write the non-huge version
1915                  * of the pmd entry with pmd_populate.
1916                  */
1917                 pmdp_invalidate(vma, address, pmd);
1918                 pmd_populate(mm, pmd, pgtable);
1919                 ret = 1;
1920                 spin_unlock(ptl);
1921         }
1922
1923         return ret;
1924 }
1925
1926 /* must be called with anon_vma->root->rwsem held */
1927 static void __split_huge_page(struct page *page,
1928                               struct anon_vma *anon_vma,
1929                               struct list_head *list)
1930 {
1931         int mapcount, mapcount2;
1932         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1933         struct anon_vma_chain *avc;
1934
1935         BUG_ON(!PageHead(page));
1936         BUG_ON(PageTail(page));
1937
1938         mapcount = 0;
1939         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1940                 struct vm_area_struct *vma = avc->vma;
1941                 unsigned long addr = vma_address(page, vma);
1942                 BUG_ON(is_vma_temporary_stack(vma));
1943                 mapcount += __split_huge_page_splitting(page, vma, addr);
1944         }
1945         /*
1946          * It is critical that new vmas are added to the tail of the
1947          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1948          * and establishes a child pmd before
1949          * __split_huge_page_splitting() freezes the parent pmd (so if
1950          * we fail to prevent copy_huge_pmd() from running until the
1951          * whole __split_huge_page() is complete), we will still see
1952          * the newly established pmd of the child later during the
1953          * walk, to be able to set it as pmd_trans_splitting too.
1954          */
1955         if (mapcount != page_mapcount(page)) {
1956                 pr_err("mapcount %d page_mapcount %d\n",
1957                         mapcount, page_mapcount(page));
1958                 BUG();
1959         }
1960
1961         __split_huge_page_refcount(page, list);
1962
1963         mapcount2 = 0;
1964         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1965                 struct vm_area_struct *vma = avc->vma;
1966                 unsigned long addr = vma_address(page, vma);
1967                 BUG_ON(is_vma_temporary_stack(vma));
1968                 mapcount2 += __split_huge_page_map(page, vma, addr);
1969         }
1970         if (mapcount != mapcount2) {
1971                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1972                         mapcount, mapcount2, page_mapcount(page));
1973                 BUG();
1974         }
1975 }
1976
1977 /*
1978  * Split a hugepage into normal pages. This doesn't change the position of head
1979  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1980  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1981  * from the hugepage.
1982  * Return 0 if the hugepage is split successfully otherwise return 1.
1983  */
1984 int split_huge_page_to_list(struct page *page, struct list_head *list)
1985 {
1986         struct anon_vma *anon_vma;
1987         int ret = 1;
1988
1989         BUG_ON(is_huge_zero_page(page));
1990         BUG_ON(!PageAnon(page));
1991
1992         /*
1993          * The caller does not necessarily hold an mmap_sem that would prevent
1994          * the anon_vma disappearing so we first we take a reference to it
1995          * and then lock the anon_vma for write. This is similar to
1996          * page_lock_anon_vma_read except the write lock is taken to serialise
1997          * against parallel split or collapse operations.
1998          */
1999         anon_vma = page_get_anon_vma(page);
2000         if (!anon_vma)
2001                 goto out;
2002         anon_vma_lock_write(anon_vma);
2003
2004         ret = 0;
2005         if (!PageCompound(page))
2006                 goto out_unlock;
2007
2008         BUG_ON(!PageSwapBacked(page));
2009         __split_huge_page(page, anon_vma, list);
2010         count_vm_event(THP_SPLIT);
2011
2012         BUG_ON(PageCompound(page));
2013 out_unlock:
2014         anon_vma_unlock_write(anon_vma);
2015         put_anon_vma(anon_vma);
2016 out:
2017         return ret;
2018 }
2019
2020 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
2021
2022 int hugepage_madvise(struct vm_area_struct *vma,
2023                      unsigned long *vm_flags, int advice)
2024 {
2025         switch (advice) {
2026         case MADV_HUGEPAGE:
2027 #ifdef CONFIG_S390
2028                 /*
2029                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2030                  * can't handle this properly after s390_enable_sie, so we simply
2031                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
2032                  */
2033                 if (mm_has_pgste(vma->vm_mm))
2034                         return 0;
2035 #endif
2036                 /*
2037                  * Be somewhat over-protective like KSM for now!
2038                  */
2039                 if (*vm_flags & VM_NO_THP)
2040                         return -EINVAL;
2041                 *vm_flags &= ~VM_NOHUGEPAGE;
2042                 *vm_flags |= VM_HUGEPAGE;
2043                 /*
2044                  * If the vma become good for khugepaged to scan,
2045                  * register it here without waiting a page fault that
2046                  * may not happen any time soon.
2047                  */
2048                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2049                         return -ENOMEM;
2050                 break;
2051         case MADV_NOHUGEPAGE:
2052                 /*
2053                  * Be somewhat over-protective like KSM for now!
2054                  */
2055                 if (*vm_flags & VM_NO_THP)
2056                         return -EINVAL;
2057                 *vm_flags &= ~VM_HUGEPAGE;
2058                 *vm_flags |= VM_NOHUGEPAGE;
2059                 /*
2060                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2061                  * this vma even if we leave the mm registered in khugepaged if
2062                  * it got registered before VM_NOHUGEPAGE was set.
2063                  */
2064                 break;
2065         }
2066
2067         return 0;
2068 }
2069
2070 static int __init khugepaged_slab_init(void)
2071 {
2072         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2073                                           sizeof(struct mm_slot),
2074                                           __alignof__(struct mm_slot), 0, NULL);
2075         if (!mm_slot_cache)
2076                 return -ENOMEM;
2077
2078         return 0;
2079 }
2080
2081 static void __init khugepaged_slab_exit(void)
2082 {
2083         kmem_cache_destroy(mm_slot_cache);
2084 }
2085
2086 static inline struct mm_slot *alloc_mm_slot(void)
2087 {
2088         if (!mm_slot_cache)     /* initialization failed */
2089                 return NULL;
2090         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2091 }
2092
2093 static inline void free_mm_slot(struct mm_slot *mm_slot)
2094 {
2095         kmem_cache_free(mm_slot_cache, mm_slot);
2096 }
2097
2098 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2099 {
2100         struct mm_slot *mm_slot;
2101
2102         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2103                 if (mm == mm_slot->mm)
2104                         return mm_slot;
2105
2106         return NULL;
2107 }
2108
2109 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2110                                     struct mm_slot *mm_slot)
2111 {
2112         mm_slot->mm = mm;
2113         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2114 }
2115
2116 static inline int khugepaged_test_exit(struct mm_struct *mm)
2117 {
2118         return atomic_read(&mm->mm_users) == 0;
2119 }
2120
2121 int __khugepaged_enter(struct mm_struct *mm)
2122 {
2123         struct mm_slot *mm_slot;
2124         int wakeup;
2125
2126         mm_slot = alloc_mm_slot();
2127         if (!mm_slot)
2128                 return -ENOMEM;
2129
2130         /* __khugepaged_exit() must not run from under us */
2131         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2132         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2133                 free_mm_slot(mm_slot);
2134                 return 0;
2135         }
2136
2137         spin_lock(&khugepaged_mm_lock);
2138         insert_to_mm_slots_hash(mm, mm_slot);
2139         /*
2140          * Insert just behind the scanning cursor, to let the area settle
2141          * down a little.
2142          */
2143         wakeup = list_empty(&khugepaged_scan.mm_head);
2144         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2145         spin_unlock(&khugepaged_mm_lock);
2146
2147         atomic_inc(&mm->mm_count);
2148         if (wakeup)
2149                 wake_up_interruptible(&khugepaged_wait);
2150
2151         return 0;
2152 }
2153
2154 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2155                                unsigned long vm_flags)
2156 {
2157         unsigned long hstart, hend;
2158         if (!vma->anon_vma)
2159                 /*
2160                  * Not yet faulted in so we will register later in the
2161                  * page fault if needed.
2162                  */
2163                 return 0;
2164         if (vma->vm_ops)
2165                 /* khugepaged not yet working on file or special mappings */
2166                 return 0;
2167         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2168         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2169         hend = vma->vm_end & HPAGE_PMD_MASK;
2170         if (hstart < hend)
2171                 return khugepaged_enter(vma, vm_flags);
2172         return 0;
2173 }
2174
2175 void __khugepaged_exit(struct mm_struct *mm)
2176 {
2177         struct mm_slot *mm_slot;
2178         int free = 0;
2179
2180         spin_lock(&khugepaged_mm_lock);
2181         mm_slot = get_mm_slot(mm);
2182         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2183                 hash_del(&mm_slot->hash);
2184                 list_del(&mm_slot->mm_node);
2185                 free = 1;
2186         }
2187         spin_unlock(&khugepaged_mm_lock);
2188
2189         if (free) {
2190                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2191                 free_mm_slot(mm_slot);
2192                 mmdrop(mm);
2193         } else if (mm_slot) {
2194                 /*
2195                  * This is required to serialize against
2196                  * khugepaged_test_exit() (which is guaranteed to run
2197                  * under mmap sem read mode). Stop here (after we
2198                  * return all pagetables will be destroyed) until
2199                  * khugepaged has finished working on the pagetables
2200                  * under the mmap_sem.
2201                  */
2202                 down_write(&mm->mmap_sem);
2203                 up_write(&mm->mmap_sem);
2204         }
2205 }
2206
2207 static void release_pte_page(struct page *page)
2208 {
2209         /* 0 stands for page_is_file_cache(page) == false */
2210         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2211         unlock_page(page);
2212         putback_lru_page(page);
2213 }
2214
2215 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2216 {
2217         while (--_pte >= pte) {
2218                 pte_t pteval = *_pte;
2219                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2220                         release_pte_page(pte_page(pteval));
2221         }
2222 }
2223
2224 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2225                                         unsigned long address,
2226                                         pte_t *pte)
2227 {
2228         struct page *page = NULL;
2229         pte_t *_pte;
2230         int none_or_zero = 0, result = 0;
2231         bool referenced = false, writable = false;
2232
2233         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2234              _pte++, address += PAGE_SIZE) {
2235                 pte_t pteval = *_pte;
2236                 if (pte_none(pteval) || (pte_present(pteval) &&
2237                                 is_zero_pfn(pte_pfn(pteval)))) {
2238                         if (!userfaultfd_armed(vma) &&
2239                             ++none_or_zero <= khugepaged_max_ptes_none) {
2240                                 continue;
2241                         } else {
2242                                 result = SCAN_EXCEED_NONE_PTE;
2243                                 goto out;
2244                         }
2245                 }
2246                 if (!pte_present(pteval)) {
2247                         result = SCAN_PTE_NON_PRESENT;
2248                         goto out;
2249                 }
2250                 page = vm_normal_page(vma, address, pteval);
2251                 if (unlikely(!page)) {
2252                         result = SCAN_PAGE_NULL;
2253                         goto out;
2254                 }
2255
2256                 VM_BUG_ON_PAGE(PageCompound(page), page);
2257                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2258                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2259
2260                 /*
2261                  * We can do it before isolate_lru_page because the
2262                  * page can't be freed from under us. NOTE: PG_lock
2263                  * is needed to serialize against split_huge_page
2264                  * when invoked from the VM.
2265                  */
2266                 if (!trylock_page(page)) {
2267                         result = SCAN_PAGE_LOCK;
2268                         goto out;
2269                 }
2270
2271                 /*
2272                  * cannot use mapcount: can't collapse if there's a gup pin.
2273                  * The page must only be referenced by the scanned process
2274                  * and page swap cache.
2275                  */
2276                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2277                         unlock_page(page);
2278                         result = SCAN_PAGE_COUNT;
2279                         goto out;
2280                 }
2281                 if (pte_write(pteval)) {
2282                         writable = true;
2283                 } else {
2284                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2285                                 unlock_page(page);
2286                                 result = SCAN_SWAP_CACHE_PAGE;
2287                                 goto out;
2288                         }
2289                         /*
2290                          * Page is not in the swap cache. It can be collapsed
2291                          * into a THP.
2292                          */
2293                 }
2294
2295                 /*
2296                  * Isolate the page to avoid collapsing an hugepage
2297                  * currently in use by the VM.
2298                  */
2299                 if (isolate_lru_page(page)) {
2300                         unlock_page(page);
2301                         result = SCAN_DEL_PAGE_LRU;
2302                         goto out;
2303                 }
2304                 /* 0 stands for page_is_file_cache(page) == false */
2305                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2306                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2307                 VM_BUG_ON_PAGE(PageLRU(page), page);
2308
2309                 /* If there is no mapped pte young don't collapse the page */
2310                 if (pte_young(pteval) ||
2311                     page_is_young(page) || PageReferenced(page) ||
2312                     mmu_notifier_test_young(vma->vm_mm, address))
2313                         referenced = true;
2314         }
2315         if (likely(writable)) {
2316                 if (likely(referenced)) {
2317                         result = SCAN_SUCCEED;
2318                         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
2319                                                             referenced, writable, result);
2320                         return 1;
2321                 }
2322         } else {
2323                 result = SCAN_PAGE_RO;
2324         }
2325
2326 out:
2327         release_pte_pages(pte, _pte);
2328         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
2329                                             referenced, writable, result);
2330         return 0;
2331 }
2332
2333 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2334                                       struct vm_area_struct *vma,
2335                                       unsigned long address,
2336                                       spinlock_t *ptl)
2337 {
2338         pte_t *_pte;
2339         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2340                 pte_t pteval = *_pte;
2341                 struct page *src_page;
2342
2343                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2344                         clear_user_highpage(page, address);
2345                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2346                         if (is_zero_pfn(pte_pfn(pteval))) {
2347                                 /*
2348                                  * ptl mostly unnecessary.
2349                                  */
2350                                 spin_lock(ptl);
2351                                 /*
2352                                  * paravirt calls inside pte_clear here are
2353                                  * superfluous.
2354                                  */
2355                                 pte_clear(vma->vm_mm, address, _pte);
2356                                 spin_unlock(ptl);
2357                         }
2358                 } else {
2359                         src_page = pte_page(pteval);
2360                         copy_user_highpage(page, src_page, address, vma);
2361                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2362                         release_pte_page(src_page);
2363                         /*
2364                          * ptl mostly unnecessary, but preempt has to
2365                          * be disabled to update the per-cpu stats
2366                          * inside page_remove_rmap().
2367                          */
2368                         spin_lock(ptl);
2369                         /*
2370                          * paravirt calls inside pte_clear here are
2371                          * superfluous.
2372                          */
2373                         pte_clear(vma->vm_mm, address, _pte);
2374                         page_remove_rmap(src_page);
2375                         spin_unlock(ptl);
2376                         free_page_and_swap_cache(src_page);
2377                 }
2378
2379                 address += PAGE_SIZE;
2380                 page++;
2381         }
2382 }
2383
2384 static void khugepaged_alloc_sleep(void)
2385 {
2386         DEFINE_WAIT(wait);
2387
2388         add_wait_queue(&khugepaged_wait, &wait);
2389         freezable_schedule_timeout_interruptible(
2390                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2391         remove_wait_queue(&khugepaged_wait, &wait);
2392 }
2393
2394 static int khugepaged_node_load[MAX_NUMNODES];
2395
2396 static bool khugepaged_scan_abort(int nid)
2397 {
2398         int i;
2399
2400         /*
2401          * If zone_reclaim_mode is disabled, then no extra effort is made to
2402          * allocate memory locally.
2403          */
2404         if (!zone_reclaim_mode)
2405                 return false;
2406
2407         /* If there is a count for this node already, it must be acceptable */
2408         if (khugepaged_node_load[nid])
2409                 return false;
2410
2411         for (i = 0; i < MAX_NUMNODES; i++) {
2412                 if (!khugepaged_node_load[i])
2413                         continue;
2414                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2415                         return true;
2416         }
2417         return false;
2418 }
2419
2420 #ifdef CONFIG_NUMA
2421 static int khugepaged_find_target_node(void)
2422 {
2423         static int last_khugepaged_target_node = NUMA_NO_NODE;
2424         int nid, target_node = 0, max_value = 0;
2425
2426         /* find first node with max normal pages hit */
2427         for (nid = 0; nid < MAX_NUMNODES; nid++)
2428                 if (khugepaged_node_load[nid] > max_value) {
2429                         max_value = khugepaged_node_load[nid];
2430                         target_node = nid;
2431                 }
2432
2433         /* do some balance if several nodes have the same hit record */
2434         if (target_node <= last_khugepaged_target_node)
2435                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2436                                 nid++)
2437                         if (max_value == khugepaged_node_load[nid]) {
2438                                 target_node = nid;
2439                                 break;
2440                         }
2441
2442         last_khugepaged_target_node = target_node;
2443         return target_node;
2444 }
2445
2446 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2447 {
2448         if (IS_ERR(*hpage)) {
2449                 if (!*wait)
2450                         return false;
2451
2452                 *wait = false;
2453                 *hpage = NULL;
2454                 khugepaged_alloc_sleep();
2455         } else if (*hpage) {
2456                 put_page(*hpage);
2457                 *hpage = NULL;
2458         }
2459
2460         return true;
2461 }
2462
2463 static struct page *
2464 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2465                        unsigned long address, int node)
2466 {
2467         VM_BUG_ON_PAGE(*hpage, *hpage);
2468
2469         /*
2470          * Before allocating the hugepage, release the mmap_sem read lock.
2471          * The allocation can take potentially a long time if it involves
2472          * sync compaction, and we do not need to hold the mmap_sem during
2473          * that. We will recheck the vma after taking it again in write mode.
2474          */
2475         up_read(&mm->mmap_sem);
2476
2477         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2478         if (unlikely(!*hpage)) {
2479                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2480                 *hpage = ERR_PTR(-ENOMEM);
2481                 return NULL;
2482         }
2483
2484         count_vm_event(THP_COLLAPSE_ALLOC);
2485         return *hpage;
2486 }
2487 #else
2488 static int khugepaged_find_target_node(void)
2489 {
2490         return 0;
2491 }
2492
2493 static inline struct page *alloc_hugepage(int defrag)
2494 {
2495         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2496                            HPAGE_PMD_ORDER);
2497 }
2498
2499 static struct page *khugepaged_alloc_hugepage(bool *wait)
2500 {
2501         struct page *hpage;
2502
2503         do {
2504                 hpage = alloc_hugepage(khugepaged_defrag());
2505                 if (!hpage) {
2506                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2507                         if (!*wait)
2508                                 return NULL;
2509
2510                         *wait = false;
2511                         khugepaged_alloc_sleep();
2512                 } else
2513                         count_vm_event(THP_COLLAPSE_ALLOC);
2514         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2515
2516         return hpage;
2517 }
2518
2519 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2520 {
2521         if (!*hpage)
2522                 *hpage = khugepaged_alloc_hugepage(wait);
2523
2524         if (unlikely(!*hpage))
2525                 return false;
2526
2527         return true;
2528 }
2529
2530 static struct page *
2531 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2532                        unsigned long address, int node)
2533 {
2534         up_read(&mm->mmap_sem);
2535         VM_BUG_ON(!*hpage);
2536
2537         return  *hpage;
2538 }
2539 #endif
2540
2541 static bool hugepage_vma_check(struct vm_area_struct *vma)
2542 {
2543         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2544             (vma->vm_flags & VM_NOHUGEPAGE))
2545                 return false;
2546
2547         if (!vma->anon_vma || vma->vm_ops)
2548                 return false;
2549         if (is_vma_temporary_stack(vma))
2550                 return false;
2551         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2552         return true;
2553 }
2554
2555 static void collapse_huge_page(struct mm_struct *mm,
2556                                    unsigned long address,
2557                                    struct page **hpage,
2558                                    struct vm_area_struct *vma,
2559                                    int node)
2560 {
2561         pmd_t *pmd, _pmd;
2562         pte_t *pte;
2563         pgtable_t pgtable;
2564         struct page *new_page;
2565         spinlock_t *pmd_ptl, *pte_ptl;
2566         int isolated, result = 0;
2567         unsigned long hstart, hend;
2568         struct mem_cgroup *memcg;
2569         unsigned long mmun_start;       /* For mmu_notifiers */
2570         unsigned long mmun_end;         /* For mmu_notifiers */
2571         gfp_t gfp;
2572
2573         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2574
2575         /* Only allocate from the target node */
2576         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2577                 __GFP_THISNODE;
2578
2579         /* release the mmap_sem read lock. */
2580         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2581         if (!new_page) {
2582                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2583                 goto out_nolock;
2584         }
2585
2586         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg))) {
2587                 result = SCAN_CGROUP_CHARGE_FAIL;
2588                 goto out_nolock;
2589         }
2590
2591         /*
2592          * Prevent all access to pagetables with the exception of
2593          * gup_fast later hanlded by the ptep_clear_flush and the VM
2594          * handled by the anon_vma lock + PG_lock.
2595          */
2596         down_write(&mm->mmap_sem);
2597         if (unlikely(khugepaged_test_exit(mm))) {
2598                 result = SCAN_ANY_PROCESS;
2599                 goto out;
2600         }
2601
2602         vma = find_vma(mm, address);
2603         if (!vma) {
2604                 result = SCAN_VMA_NULL;
2605                 goto out;
2606         }
2607         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2608         hend = vma->vm_end & HPAGE_PMD_MASK;
2609         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2610                 result = SCAN_ADDRESS_RANGE;
2611                 goto out;
2612         }
2613         if (!hugepage_vma_check(vma)) {
2614                 result = SCAN_VMA_CHECK;
2615                 goto out;
2616         }
2617         pmd = mm_find_pmd(mm, address);
2618         if (!pmd) {
2619                 result = SCAN_PMD_NULL;
2620                 goto out;
2621         }
2622
2623         anon_vma_lock_write(vma->anon_vma);
2624
2625         pte = pte_offset_map(pmd, address);
2626         pte_ptl = pte_lockptr(mm, pmd);
2627
2628         mmun_start = address;
2629         mmun_end   = address + HPAGE_PMD_SIZE;
2630         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2631         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2632         /*
2633          * After this gup_fast can't run anymore. This also removes
2634          * any huge TLB entry from the CPU so we won't allow
2635          * huge and small TLB entries for the same virtual address
2636          * to avoid the risk of CPU bugs in that area.
2637          */
2638         _pmd = pmdp_collapse_flush(vma, address, pmd);
2639         spin_unlock(pmd_ptl);
2640         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2641
2642         spin_lock(pte_ptl);
2643         isolated = __collapse_huge_page_isolate(vma, address, pte);
2644         spin_unlock(pte_ptl);
2645
2646         if (unlikely(!isolated)) {
2647                 pte_unmap(pte);
2648                 spin_lock(pmd_ptl);
2649                 BUG_ON(!pmd_none(*pmd));
2650                 /*
2651                  * We can only use set_pmd_at when establishing
2652                  * hugepmds and never for establishing regular pmds that
2653                  * points to regular pagetables. Use pmd_populate for that
2654                  */
2655                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2656                 spin_unlock(pmd_ptl);
2657                 anon_vma_unlock_write(vma->anon_vma);
2658                 result = SCAN_FAIL;
2659                 goto out;
2660         }
2661
2662         /*
2663          * All pages are isolated and locked so anon_vma rmap
2664          * can't run anymore.
2665          */
2666         anon_vma_unlock_write(vma->anon_vma);
2667
2668         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2669         pte_unmap(pte);
2670         __SetPageUptodate(new_page);
2671         pgtable = pmd_pgtable(_pmd);
2672
2673         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2674         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2675
2676         /*
2677          * spin_lock() below is not the equivalent of smp_wmb(), so
2678          * this is needed to avoid the copy_huge_page writes to become
2679          * visible after the set_pmd_at() write.
2680          */
2681         smp_wmb();
2682
2683         spin_lock(pmd_ptl);
2684         BUG_ON(!pmd_none(*pmd));
2685         page_add_new_anon_rmap(new_page, vma, address);
2686         mem_cgroup_commit_charge(new_page, memcg, false);
2687         lru_cache_add_active_or_unevictable(new_page, vma);
2688         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2689         set_pmd_at(mm, address, pmd, _pmd);
2690         update_mmu_cache_pmd(vma, address, pmd);
2691         spin_unlock(pmd_ptl);
2692
2693         *hpage = NULL;
2694
2695         khugepaged_pages_collapsed++;
2696         result = SCAN_SUCCEED;
2697 out_up_write:
2698         up_write(&mm->mmap_sem);
2699         trace_mm_collapse_huge_page(mm, isolated, result);
2700         return;
2701
2702 out_nolock:
2703         trace_mm_collapse_huge_page(mm, isolated, result);
2704         return;
2705 out:
2706         mem_cgroup_cancel_charge(new_page, memcg);
2707         goto out_up_write;
2708 }
2709
2710 static int khugepaged_scan_pmd(struct mm_struct *mm,
2711                                struct vm_area_struct *vma,
2712                                unsigned long address,
2713                                struct page **hpage)
2714 {
2715         pmd_t *pmd;
2716         pte_t *pte, *_pte;
2717         int ret = 0, none_or_zero = 0, result = 0;
2718         struct page *page = NULL;
2719         unsigned long _address;
2720         spinlock_t *ptl;
2721         int node = NUMA_NO_NODE;
2722         bool writable = false, referenced = false;
2723
2724         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2725
2726         pmd = mm_find_pmd(mm, address);
2727         if (!pmd) {
2728                 result = SCAN_PMD_NULL;
2729                 goto out;
2730         }
2731
2732         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2733         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2734         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2735              _pte++, _address += PAGE_SIZE) {
2736                 pte_t pteval = *_pte;
2737                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2738                         if (!userfaultfd_armed(vma) &&
2739                             ++none_or_zero <= khugepaged_max_ptes_none) {
2740                                 continue;
2741                         } else {
2742                                 result = SCAN_EXCEED_NONE_PTE;
2743                                 goto out_unmap;
2744                         }
2745                 }
2746                 if (!pte_present(pteval)) {
2747                         result = SCAN_PTE_NON_PRESENT;
2748                         goto out_unmap;
2749                 }
2750                 if (pte_write(pteval))
2751                         writable = true;
2752
2753                 page = vm_normal_page(vma, _address, pteval);
2754                 if (unlikely(!page)) {
2755                         result = SCAN_PAGE_NULL;
2756                         goto out_unmap;
2757                 }
2758                 /*
2759                  * Record which node the original page is from and save this
2760                  * information to khugepaged_node_load[].
2761                  * Khupaged will allocate hugepage from the node has the max
2762                  * hit record.
2763                  */
2764                 node = page_to_nid(page);
2765                 if (khugepaged_scan_abort(node)) {
2766                         result = SCAN_SCAN_ABORT;
2767                         goto out_unmap;
2768                 }
2769                 khugepaged_node_load[node]++;
2770                 VM_BUG_ON_PAGE(PageCompound(page), page);
2771                 if (!PageLRU(page)) {
2772                         result = SCAN_SCAN_ABORT;
2773                         goto out_unmap;
2774                 }
2775                 if (PageLocked(page)) {
2776                         result = SCAN_PAGE_LOCK;
2777                         goto out_unmap;
2778                 }
2779                 if (!PageAnon(page)) {
2780                         result = SCAN_PAGE_ANON;
2781                         goto out_unmap;
2782                 }
2783
2784                 /*
2785                  * cannot use mapcount: can't collapse if there's a gup pin.
2786                  * The page must only be referenced by the scanned process
2787                  * and page swap cache.
2788                  */
2789                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2790                         result = SCAN_PAGE_COUNT;
2791                         goto out_unmap;
2792                 }
2793                 if (pte_young(pteval) ||
2794                     page_is_young(page) || PageReferenced(page) ||
2795                     mmu_notifier_test_young(vma->vm_mm, address))
2796                         referenced = true;
2797         }
2798         if (writable) {
2799                 if (referenced) {
2800                         result = SCAN_SUCCEED;
2801                         ret = 1;
2802                 } else {
2803                         result = SCAN_NO_REFERENCED_PAGE;
2804                 }
2805         } else {
2806                 result = SCAN_PAGE_RO;
2807         }
2808 out_unmap:
2809         pte_unmap_unlock(pte, ptl);
2810         if (ret) {
2811                 node = khugepaged_find_target_node();
2812                 /* collapse_huge_page will return with the mmap_sem released */
2813                 collapse_huge_page(mm, address, hpage, vma, node);
2814         }
2815 out:
2816         trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced,
2817                                      none_or_zero, result);
2818         return ret;
2819 }
2820
2821 static void collect_mm_slot(struct mm_slot *mm_slot)
2822 {
2823         struct mm_struct *mm = mm_slot->mm;
2824
2825         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2826
2827         if (khugepaged_test_exit(mm)) {
2828                 /* free mm_slot */
2829                 hash_del(&mm_slot->hash);
2830                 list_del(&mm_slot->mm_node);
2831
2832                 /*
2833                  * Not strictly needed because the mm exited already.
2834                  *
2835                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2836                  */
2837
2838                 /* khugepaged_mm_lock actually not necessary for the below */
2839                 free_mm_slot(mm_slot);
2840                 mmdrop(mm);
2841         }
2842 }
2843
2844 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2845                                             struct page **hpage)
2846         __releases(&khugepaged_mm_lock)
2847         __acquires(&khugepaged_mm_lock)
2848 {
2849         struct mm_slot *mm_slot;
2850         struct mm_struct *mm;
2851         struct vm_area_struct *vma;
2852         int progress = 0;
2853
2854         VM_BUG_ON(!pages);
2855         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2856
2857         if (khugepaged_scan.mm_slot)
2858                 mm_slot = khugepaged_scan.mm_slot;
2859         else {
2860                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2861                                      struct mm_slot, mm_node);
2862                 khugepaged_scan.address = 0;
2863                 khugepaged_scan.mm_slot = mm_slot;
2864         }
2865         spin_unlock(&khugepaged_mm_lock);
2866
2867         mm = mm_slot->mm;
2868         down_read(&mm->mmap_sem);
2869         if (unlikely(khugepaged_test_exit(mm)))
2870                 vma = NULL;
2871         else
2872                 vma = find_vma(mm, khugepaged_scan.address);
2873
2874         progress++;
2875         for (; vma; vma = vma->vm_next) {
2876                 unsigned long hstart, hend;
2877
2878                 cond_resched();
2879                 if (unlikely(khugepaged_test_exit(mm))) {
2880                         progress++;
2881                         break;
2882                 }
2883                 if (!hugepage_vma_check(vma)) {
2884 skip:
2885                         progress++;
2886                         continue;
2887                 }
2888                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2889                 hend = vma->vm_end & HPAGE_PMD_MASK;
2890                 if (hstart >= hend)
2891                         goto skip;
2892                 if (khugepaged_scan.address > hend)
2893                         goto skip;
2894                 if (khugepaged_scan.address < hstart)
2895                         khugepaged_scan.address = hstart;
2896                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2897
2898                 while (khugepaged_scan.address < hend) {
2899                         int ret;
2900                         cond_resched();
2901                         if (unlikely(khugepaged_test_exit(mm)))
2902                                 goto breakouterloop;
2903
2904                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2905                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2906                                   hend);
2907                         ret = khugepaged_scan_pmd(mm, vma,
2908                                                   khugepaged_scan.address,
2909                                                   hpage);
2910                         /* move to next address */
2911                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2912                         progress += HPAGE_PMD_NR;
2913                         if (ret)
2914                                 /* we released mmap_sem so break loop */
2915                                 goto breakouterloop_mmap_sem;
2916                         if (progress >= pages)
2917                                 goto breakouterloop;
2918                 }
2919         }
2920 breakouterloop:
2921         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2922 breakouterloop_mmap_sem:
2923
2924         spin_lock(&khugepaged_mm_lock);
2925         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2926         /*
2927          * Release the current mm_slot if this mm is about to die, or
2928          * if we scanned all vmas of this mm.
2929          */
2930         if (khugepaged_test_exit(mm) || !vma) {
2931                 /*
2932                  * Make sure that if mm_users is reaching zero while
2933                  * khugepaged runs here, khugepaged_exit will find
2934                  * mm_slot not pointing to the exiting mm.
2935                  */
2936                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2937                         khugepaged_scan.mm_slot = list_entry(
2938                                 mm_slot->mm_node.next,
2939                                 struct mm_slot, mm_node);
2940                         khugepaged_scan.address = 0;
2941                 } else {
2942                         khugepaged_scan.mm_slot = NULL;
2943                         khugepaged_full_scans++;
2944                 }
2945
2946                 collect_mm_slot(mm_slot);
2947         }
2948
2949         return progress;
2950 }
2951
2952 static int khugepaged_has_work(void)
2953 {
2954         return !list_empty(&khugepaged_scan.mm_head) &&
2955                 khugepaged_enabled();
2956 }
2957
2958 static int khugepaged_wait_event(void)
2959 {
2960         return !list_empty(&khugepaged_scan.mm_head) ||
2961                 kthread_should_stop();
2962 }
2963
2964 static void khugepaged_do_scan(void)
2965 {
2966         struct page *hpage = NULL;
2967         unsigned int progress = 0, pass_through_head = 0;
2968         unsigned int pages = khugepaged_pages_to_scan;
2969         bool wait = true;
2970
2971         barrier(); /* write khugepaged_pages_to_scan to local stack */
2972
2973         while (progress < pages) {
2974                 if (!khugepaged_prealloc_page(&hpage, &wait))
2975                         break;
2976
2977                 cond_resched();
2978
2979                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2980                         break;
2981
2982                 spin_lock(&khugepaged_mm_lock);
2983                 if (!khugepaged_scan.mm_slot)
2984                         pass_through_head++;
2985                 if (khugepaged_has_work() &&
2986                     pass_through_head < 2)
2987                         progress += khugepaged_scan_mm_slot(pages - progress,
2988                                                             &hpage);
2989                 else
2990                         progress = pages;
2991                 spin_unlock(&khugepaged_mm_lock);
2992         }
2993
2994         if (!IS_ERR_OR_NULL(hpage))
2995                 put_page(hpage);
2996 }
2997
2998 static void khugepaged_wait_work(void)
2999 {
3000         if (khugepaged_has_work()) {
3001                 if (!khugepaged_scan_sleep_millisecs)
3002                         return;
3003
3004                 wait_event_freezable_timeout(khugepaged_wait,
3005                                              kthread_should_stop(),
3006                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
3007                 return;
3008         }
3009
3010         if (khugepaged_enabled())
3011                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
3012 }
3013
3014 static int khugepaged(void *none)
3015 {
3016         struct mm_slot *mm_slot;
3017
3018         set_freezable();
3019         set_user_nice(current, MAX_NICE);
3020
3021         while (!kthread_should_stop()) {
3022                 khugepaged_do_scan();
3023                 khugepaged_wait_work();
3024         }
3025
3026         spin_lock(&khugepaged_mm_lock);
3027         mm_slot = khugepaged_scan.mm_slot;
3028         khugepaged_scan.mm_slot = NULL;
3029         if (mm_slot)
3030                 collect_mm_slot(mm_slot);
3031         spin_unlock(&khugepaged_mm_lock);
3032         return 0;
3033 }
3034
3035 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
3036                 unsigned long haddr, pmd_t *pmd)
3037 {
3038         struct mm_struct *mm = vma->vm_mm;
3039         pgtable_t pgtable;
3040         pmd_t _pmd;
3041         int i;
3042
3043         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
3044         /* leave pmd empty until pte is filled */
3045
3046         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
3047         pmd_populate(mm, &_pmd, pgtable);
3048
3049         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
3050                 pte_t *pte, entry;
3051                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
3052                 entry = pte_mkspecial(entry);
3053                 pte = pte_offset_map(&_pmd, haddr);
3054                 VM_BUG_ON(!pte_none(*pte));
3055                 set_pte_at(mm, haddr, pte, entry);
3056                 pte_unmap(pte);
3057         }
3058         smp_wmb(); /* make pte visible before pmd */
3059         pmd_populate(mm, pmd, pgtable);
3060         put_huge_zero_page();
3061 }
3062
3063 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
3064                 pmd_t *pmd)
3065 {
3066         spinlock_t *ptl;
3067         struct page *page = NULL;
3068         struct mm_struct *mm = vma->vm_mm;
3069         unsigned long haddr = address & HPAGE_PMD_MASK;
3070         unsigned long mmun_start;       /* For mmu_notifiers */
3071         unsigned long mmun_end;         /* For mmu_notifiers */
3072
3073         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
3074
3075         mmun_start = haddr;
3076         mmun_end   = haddr + HPAGE_PMD_SIZE;
3077 again:
3078         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3079         ptl = pmd_lock(mm, pmd);
3080         if (unlikely(!pmd_trans_huge(*pmd)))
3081                 goto unlock;
3082         if (vma_is_dax(vma)) {
3083                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
3084                 if (is_huge_zero_pmd(_pmd))
3085                         put_huge_zero_page();
3086         } else if (is_huge_zero_pmd(*pmd)) {
3087                 __split_huge_zero_page_pmd(vma, haddr, pmd);
3088         } else {
3089                 page = pmd_page(*pmd);
3090                 VM_BUG_ON_PAGE(!page_count(page), page);
3091                 get_page(page);
3092         }
3093  unlock:
3094         spin_unlock(ptl);
3095         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3096
3097         if (!page)
3098                 return;
3099
3100         split_huge_page(page);
3101         put_page(page);
3102
3103         /*
3104          * We don't always have down_write of mmap_sem here: a racing
3105          * do_huge_pmd_wp_page() might have copied-on-write to another
3106          * huge page before our split_huge_page() got the anon_vma lock.
3107          */
3108         if (unlikely(pmd_trans_huge(*pmd)))
3109                 goto again;
3110 }
3111
3112 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3113                 pmd_t *pmd)
3114 {
3115         struct vm_area_struct *vma;
3116
3117         vma = find_vma(mm, address);
3118         BUG_ON(vma == NULL);
3119         split_huge_page_pmd(vma, address, pmd);
3120 }
3121
3122 static void split_huge_page_address(struct mm_struct *mm,
3123                                     unsigned long address)
3124 {
3125         pgd_t *pgd;
3126         pud_t *pud;
3127         pmd_t *pmd;
3128
3129         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3130
3131         pgd = pgd_offset(mm, address);
3132         if (!pgd_present(*pgd))
3133                 return;
3134
3135         pud = pud_offset(pgd, address);
3136         if (!pud_present(*pud))
3137                 return;
3138
3139         pmd = pmd_offset(pud, address);
3140         if (!pmd_present(*pmd))
3141                 return;
3142         /*
3143          * Caller holds the mmap_sem write mode, so a huge pmd cannot
3144          * materialize from under us.
3145          */
3146         split_huge_page_pmd_mm(mm, address, pmd);
3147 }
3148
3149 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3150                              unsigned long start,
3151                              unsigned long end,
3152                              long adjust_next)
3153 {
3154         /*
3155          * If the new start address isn't hpage aligned and it could
3156          * previously contain an hugepage: check if we need to split
3157          * an huge pmd.
3158          */
3159         if (start & ~HPAGE_PMD_MASK &&
3160             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3161             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3162                 split_huge_page_address(vma->vm_mm, start);
3163
3164         /*
3165          * If the new end address isn't hpage aligned and it could
3166          * previously contain an hugepage: check if we need to split
3167          * an huge pmd.
3168          */
3169         if (end & ~HPAGE_PMD_MASK &&
3170             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3171             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3172                 split_huge_page_address(vma->vm_mm, end);
3173
3174         /*
3175          * If we're also updating the vma->vm_next->vm_start, if the new
3176          * vm_next->vm_start isn't page aligned and it could previously
3177          * contain an hugepage: check if we need to split an huge pmd.
3178          */
3179         if (adjust_next > 0) {
3180                 struct vm_area_struct *next = vma->vm_next;
3181                 unsigned long nstart = next->vm_start;
3182                 nstart += adjust_next << PAGE_SHIFT;
3183                 if (nstart & ~HPAGE_PMD_MASK &&
3184                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3185                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3186                         split_huge_page_address(next->vm_mm, nstart);
3187         }
3188 }