mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/kthread.h>
  22 #include <linux/khugepaged.h>
  23 #include <linux/freezer.h>
  24 #include <linux/pfn_t.h>
  25 #include <linux/mman.h>
  26 #include <linux/memremap.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/debugfs.h>
  29 #include <linux/migrate.h>
  30 #include <linux/hashtable.h>
  31 #include <linux/userfaultfd_k.h>
  32 #include <linux/page_idle.h>
  33
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 enum scan_result {
  39         SCAN_FAIL,
  40         SCAN_SUCCEED,
  41         SCAN_PMD_NULL,
  42         SCAN_EXCEED_NONE_PTE,
  43         SCAN_PTE_NON_PRESENT,
  44         SCAN_PAGE_RO,
  45         SCAN_NO_REFERENCED_PAGE,
  46         SCAN_PAGE_NULL,
  47         SCAN_SCAN_ABORT,
  48         SCAN_PAGE_COUNT,
  49         SCAN_PAGE_LRU,
  50         SCAN_PAGE_LOCK,
  51         SCAN_PAGE_ANON,
  52         SCAN_PAGE_COMPOUND,
  53         SCAN_ANY_PROCESS,
  54         SCAN_VMA_NULL,
  55         SCAN_VMA_CHECK,
  56         SCAN_ADDRESS_RANGE,
  57         SCAN_SWAP_CACHE_PAGE,
  58         SCAN_DEL_PAGE_LRU,
  59         SCAN_ALLOC_HUGE_PAGE_FAIL,
  60         SCAN_CGROUP_CHARGE_FAIL
  61 };
  62
  63 #define CREATE_TRACE_POINTS
  64 #include <trace/events/huge_memory.h>
  65
  66 /*
  67  * By default transparent hugepage support is disabled in order that avoid
  68  * to risk increase the memory footprint of applications without a guaranteed
  69  * benefit. When transparent hugepage support is enabled, is for all mappings,
  70  * and khugepaged scans all mappings.
  71  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  72  * for all hugepage allocations.
  73  */
  74 unsigned long transparent_hugepage_flags __read_mostly =
  75 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  76         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  77 #endif
  78 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  79         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  80 #endif
  81         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  82         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  83         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  84
  85 /* default scan 8*512 pte (or vmas) every 30 second */
  86 static unsigned int khugepaged_pages_to_scan __read_mostly;
  87 static unsigned int khugepaged_pages_collapsed;
  88 static unsigned int khugepaged_full_scans;
  89 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  90 /* during fragmentation poll the hugepage allocator once every minute */
  91 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  92 static struct task_struct *khugepaged_thread __read_mostly;
  93 static DEFINE_MUTEX(khugepaged_mutex);
  94 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  95 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  96 /*
  97  * default collapse hugepages if there is at least one pte mapped like
  98  * it would have happened if the vma was large enough during page
  99  * fault.
 100  */
 101 static unsigned int khugepaged_max_ptes_none __read_mostly;
 102
 103 static int khugepaged(void *none);
 104 static int khugepaged_slab_init(void);
 105 static void khugepaged_slab_exit(void);
 106
 107 #define MM_SLOTS_HASH_BITS 10
 108 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 109
 110 static struct kmem_cache *mm_slot_cache __read_mostly;
 111
 112 /**
 113  * struct mm_slot - hash lookup from mm to mm_slot
 114  * @hash: hash collision list
 115  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 116  * @mm: the mm that this information is valid for
 117  */
 118 struct mm_slot {
 119         struct hlist_node hash;
 120         struct list_head mm_node;
 121         struct mm_struct *mm;
 122 };
 123
 124 /**
 125  * struct khugepaged_scan - cursor for scanning
 126  * @mm_head: the head of the mm list to scan
 127  * @mm_slot: the current mm_slot we are scanning
 128  * @address: the next address inside that to be scanned
 129  *
 130  * There is only the one khugepaged_scan instance of this cursor structure.
 131  */
 132 struct khugepaged_scan {
 133         struct list_head mm_head;
 134         struct mm_slot *mm_slot;
 135         unsigned long address;
 136 };
 137 static struct khugepaged_scan khugepaged_scan = {
 138         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 139 };
 140
 141 static struct shrinker deferred_split_shrinker;
 142
 143 static void set_recommended_min_free_kbytes(void)
 144 {
 145         struct zone *zone;
 146         int nr_zones = 0;
 147         unsigned long recommended_min;
 148
 149         for_each_populated_zone(zone)
 150                 nr_zones++;
 151
 152         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 153         recommended_min = pageblock_nr_pages * nr_zones * 2;
 154
 155         /*
 156          * Make sure that on average at least two pageblocks are almost free
 157          * of another type, one for a migratetype to fall back to and a
 158          * second to avoid subsequent fallbacks of other types There are 3
 159          * MIGRATE_TYPES we care about.
 160          */
 161         recommended_min += pageblock_nr_pages * nr_zones *
 162                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 163
 164         /* don't ever allow to reserve more than 5% of the lowmem */
 165         recommended_min = min(recommended_min,
 166                               (unsigned long) nr_free_buffer_pages() / 20);
 167         recommended_min <<= (PAGE_SHIFT-10);
 168
 169         if (recommended_min > min_free_kbytes) {
 170                 if (user_min_free_kbytes >= 0)
 171                         pr_info("raising min_free_kbytes from %d to %lu "
 172                                 "to help transparent hugepage allocations\n",
 173                                 min_free_kbytes, recommended_min);
 174
 175                 min_free_kbytes = recommended_min;
 176         }
 177         setup_per_zone_wmarks();
 178 }
 179
 180 static int start_stop_khugepaged(void)
 181 {
 182         int err = 0;
 183         if (khugepaged_enabled()) {
 184                 if (!khugepaged_thread)
 185                         khugepaged_thread = kthread_run(khugepaged, NULL,
 186                                                         "khugepaged");
 187                 if (IS_ERR(khugepaged_thread)) {
 188                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 189                         err = PTR_ERR(khugepaged_thread);
 190                         khugepaged_thread = NULL;
 191                         goto fail;
 192                 }
 193
 194                 if (!list_empty(&khugepaged_scan.mm_head))
 195                         wake_up_interruptible(&khugepaged_wait);
 196
 197                 set_recommended_min_free_kbytes();
 198         } else if (khugepaged_thread) {
 199                 kthread_stop(khugepaged_thread);
 200                 khugepaged_thread = NULL;
 201         }
 202 fail:
 203         return err;
 204 }
 205
 206 static atomic_t huge_zero_refcount;
 207 struct page *huge_zero_page __read_mostly;
 208
 209 struct page *get_huge_zero_page(void)
 210 {
 211         struct page *zero_page;
 212 retry:
 213         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 214                 return READ_ONCE(huge_zero_page);
 215
 216         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 217                         HPAGE_PMD_ORDER);
 218         if (!zero_page) {
 219                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 220                 return NULL;
 221         }
 222         count_vm_event(THP_ZERO_PAGE_ALLOC);
 223         preempt_disable();
 224         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 225                 preempt_enable();
 226                 __free_pages(zero_page, compound_order(zero_page));
 227                 goto retry;
 228         }
 229
 230         /* We take additional reference here. It will be put back by shrinker */
 231         atomic_set(&huge_zero_refcount, 2);
 232         preempt_enable();
 233         return READ_ONCE(huge_zero_page);
 234 }
 235
 236 static void put_huge_zero_page(void)
 237 {
 238         /*
 239          * Counter should never go to zero here. Only shrinker can put
 240          * last reference.
 241          */
 242         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 243 }
 244
 245 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 246                                         struct shrink_control *sc)
 247 {
 248         /* we can free zero page only if last reference remains */
 249         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 250 }
 251
 252 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 253                                        struct shrink_control *sc)
 254 {
 255         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 256                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 257                 BUG_ON(zero_page == NULL);
 258                 __free_pages(zero_page, compound_order(zero_page));
 259                 return HPAGE_PMD_NR;
 260         }
 261
 262         return 0;
 263 }
 264
 265 static struct shrinker huge_zero_page_shrinker = {
 266         .count_objects = shrink_huge_zero_page_count,
 267         .scan_objects = shrink_huge_zero_page_scan,
 268         .seeks = DEFAULT_SEEKS,
 269 };
 270
 271 #ifdef CONFIG_SYSFS
 272
 273 static ssize_t double_flag_show(struct kobject *kobj,
 274                                 struct kobj_attribute *attr, char *buf,
 275                                 enum transparent_hugepage_flag enabled,
 276                                 enum transparent_hugepage_flag req_madv)
 277 {
 278         if (test_bit(enabled, &transparent_hugepage_flags)) {
 279                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 280                 return sprintf(buf, "[always] madvise never\n");
 281         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 282                 return sprintf(buf, "always [madvise] never\n");
 283         else
 284                 return sprintf(buf, "always madvise [never]\n");
 285 }
 286 static ssize_t double_flag_store(struct kobject *kobj,
 287                                  struct kobj_attribute *attr,
 288                                  const char *buf, size_t count,
 289                                  enum transparent_hugepage_flag enabled,
 290                                  enum transparent_hugepage_flag req_madv)
 291 {
 292         if (!memcmp("always", buf,
 293                     min(sizeof("always")-1, count))) {
 294                 set_bit(enabled, &transparent_hugepage_flags);
 295                 clear_bit(req_madv, &transparent_hugepage_flags);
 296         } else if (!memcmp("madvise", buf,
 297                            min(sizeof("madvise")-1, count))) {
 298                 clear_bit(enabled, &transparent_hugepage_flags);
 299                 set_bit(req_madv, &transparent_hugepage_flags);
 300         } else if (!memcmp("never", buf,
 301                            min(sizeof("never")-1, count))) {
 302                 clear_bit(enabled, &transparent_hugepage_flags);
 303                 clear_bit(req_madv, &transparent_hugepage_flags);
 304         } else
 305                 return -EINVAL;
 306
 307         return count;
 308 }
 309
 310 static ssize_t enabled_show(struct kobject *kobj,
 311                             struct kobj_attribute *attr, char *buf)
 312 {
 313         return double_flag_show(kobj, attr, buf,
 314                                 TRANSPARENT_HUGEPAGE_FLAG,
 315                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 316 }
 317 static ssize_t enabled_store(struct kobject *kobj,
 318                              struct kobj_attribute *attr,
 319                              const char *buf, size_t count)
 320 {
 321         ssize_t ret;
 322
 323         ret = double_flag_store(kobj, attr, buf, count,
 324                                 TRANSPARENT_HUGEPAGE_FLAG,
 325                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 326
 327         if (ret > 0) {
 328                 int err;
 329
 330                 mutex_lock(&khugepaged_mutex);
 331                 err = start_stop_khugepaged();
 332                 mutex_unlock(&khugepaged_mutex);
 333
 334                 if (err)
 335                         ret = err;
 336         }
 337
 338         return ret;
 339 }
 340 static struct kobj_attribute enabled_attr =
 341         __ATTR(enabled, 0644, enabled_show, enabled_store);
 342
 343 static ssize_t single_flag_show(struct kobject *kobj,
 344                                 struct kobj_attribute *attr, char *buf,
 345                                 enum transparent_hugepage_flag flag)
 346 {
 347         return sprintf(buf, "%d\n",
 348                        !!test_bit(flag, &transparent_hugepage_flags));
 349 }
 350
 351 static ssize_t single_flag_store(struct kobject *kobj,
 352                                  struct kobj_attribute *attr,
 353                                  const char *buf, size_t count,
 354                                  enum transparent_hugepage_flag flag)
 355 {
 356         unsigned long value;
 357         int ret;
 358
 359         ret = kstrtoul(buf, 10, &value);
 360         if (ret < 0)
 361                 return ret;
 362         if (value > 1)
 363                 return -EINVAL;
 364
 365         if (value)
 366                 set_bit(flag, &transparent_hugepage_flags);
 367         else
 368                 clear_bit(flag, &transparent_hugepage_flags);
 369
 370         return count;
 371 }
 372
 373 /*
 374  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 375  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 376  * memory just to allocate one more hugepage.
 377  */
 378 static ssize_t defrag_show(struct kobject *kobj,
 379                            struct kobj_attribute *attr, char *buf)
 380 {
 381         return double_flag_show(kobj, attr, buf,
 382                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 383                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 384 }
 385 static ssize_t defrag_store(struct kobject *kobj,
 386                             struct kobj_attribute *attr,
 387                             const char *buf, size_t count)
 388 {
 389         return double_flag_store(kobj, attr, buf, count,
 390                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 391                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 392 }
 393 static struct kobj_attribute defrag_attr =
 394         __ATTR(defrag, 0644, defrag_show, defrag_store);
 395
 396 static ssize_t use_zero_page_show(struct kobject *kobj,
 397                 struct kobj_attribute *attr, char *buf)
 398 {
 399         return single_flag_show(kobj, attr, buf,
 400                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 401 }
 402 static ssize_t use_zero_page_store(struct kobject *kobj,
 403                 struct kobj_attribute *attr, const char *buf, size_t count)
 404 {
 405         return single_flag_store(kobj, attr, buf, count,
 406                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 407 }
 408 static struct kobj_attribute use_zero_page_attr =
 409         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 410 #ifdef CONFIG_DEBUG_VM
 411 static ssize_t debug_cow_show(struct kobject *kobj,
 412                                 struct kobj_attribute *attr, char *buf)
 413 {
 414         return single_flag_show(kobj, attr, buf,
 415                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 416 }
 417 static ssize_t debug_cow_store(struct kobject *kobj,
 418                                struct kobj_attribute *attr,
 419                                const char *buf, size_t count)
 420 {
 421         return single_flag_store(kobj, attr, buf, count,
 422                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 423 }
 424 static struct kobj_attribute debug_cow_attr =
 425         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 426 #endif /* CONFIG_DEBUG_VM */
 427
 428 static struct attribute *hugepage_attr[] = {
 429         &enabled_attr.attr,
 430         &defrag_attr.attr,
 431         &use_zero_page_attr.attr,
 432 #ifdef CONFIG_DEBUG_VM
 433         &debug_cow_attr.attr,
 434 #endif
 435         NULL,
 436 };
 437
 438 static struct attribute_group hugepage_attr_group = {
 439         .attrs = hugepage_attr,
 440 };
 441
 442 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 443                                          struct kobj_attribute *attr,
 444                                          char *buf)
 445 {
 446         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 447 }
 448
 449 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 450                                           struct kobj_attribute *attr,
 451                                           const char *buf, size_t count)
 452 {
 453         unsigned long msecs;
 454         int err;
 455
 456         err = kstrtoul(buf, 10, &msecs);
 457         if (err || msecs > UINT_MAX)
 458                 return -EINVAL;
 459
 460         khugepaged_scan_sleep_millisecs = msecs;
 461         wake_up_interruptible(&khugepaged_wait);
 462
 463         return count;
 464 }
 465 static struct kobj_attribute scan_sleep_millisecs_attr =
 466         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 467                scan_sleep_millisecs_store);
 468
 469 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 470                                           struct kobj_attribute *attr,
 471                                           char *buf)
 472 {
 473         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 474 }
 475
 476 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 477                                            struct kobj_attribute *attr,
 478                                            const char *buf, size_t count)
 479 {
 480         unsigned long msecs;
 481         int err;
 482
 483         err = kstrtoul(buf, 10, &msecs);
 484         if (err || msecs > UINT_MAX)
 485                 return -EINVAL;
 486
 487         khugepaged_alloc_sleep_millisecs = msecs;
 488         wake_up_interruptible(&khugepaged_wait);
 489
 490         return count;
 491 }
 492 static struct kobj_attribute alloc_sleep_millisecs_attr =
 493         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 494                alloc_sleep_millisecs_store);
 495
 496 static ssize_t pages_to_scan_show(struct kobject *kobj,
 497                                   struct kobj_attribute *attr,
 498                                   char *buf)
 499 {
 500         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 501 }
 502 static ssize_t pages_to_scan_store(struct kobject *kobj,
 503                                    struct kobj_attribute *attr,
 504                                    const char *buf, size_t count)
 505 {
 506         int err;
 507         unsigned long pages;
 508
 509         err = kstrtoul(buf, 10, &pages);
 510         if (err || !pages || pages > UINT_MAX)
 511                 return -EINVAL;
 512
 513         khugepaged_pages_to_scan = pages;
 514
 515         return count;
 516 }
 517 static struct kobj_attribute pages_to_scan_attr =
 518         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 519                pages_to_scan_store);
 520
 521 static ssize_t pages_collapsed_show(struct kobject *kobj,
 522                                     struct kobj_attribute *attr,
 523                                     char *buf)
 524 {
 525         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 526 }
 527 static struct kobj_attribute pages_collapsed_attr =
 528         __ATTR_RO(pages_collapsed);
 529
 530 static ssize_t full_scans_show(struct kobject *kobj,
 531                                struct kobj_attribute *attr,
 532                                char *buf)
 533 {
 534         return sprintf(buf, "%u\n", khugepaged_full_scans);
 535 }
 536 static struct kobj_attribute full_scans_attr =
 537         __ATTR_RO(full_scans);
 538
 539 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 540                                       struct kobj_attribute *attr, char *buf)
 541 {
 542         return single_flag_show(kobj, attr, buf,
 543                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 544 }
 545 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 546                                        struct kobj_attribute *attr,
 547                                        const char *buf, size_t count)
 548 {
 549         return single_flag_store(kobj, attr, buf, count,
 550                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 551 }
 552 static struct kobj_attribute khugepaged_defrag_attr =
 553         __ATTR(defrag, 0644, khugepaged_defrag_show,
 554                khugepaged_defrag_store);
 555
 556 /*
 557  * max_ptes_none controls if khugepaged should collapse hugepages over
 558  * any unmapped ptes in turn potentially increasing the memory
 559  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 560  * reduce the available free memory in the system as it
 561  * runs. Increasing max_ptes_none will instead potentially reduce the
 562  * free memory in the system during the khugepaged scan.
 563  */
 564 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 565                                              struct kobj_attribute *attr,
 566                                              char *buf)
 567 {
 568         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 569 }
 570 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 571                                               struct kobj_attribute *attr,
 572                                               const char *buf, size_t count)
 573 {
 574         int err;
 575         unsigned long max_ptes_none;
 576
 577         err = kstrtoul(buf, 10, &max_ptes_none);
 578         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 579                 return -EINVAL;
 580
 581         khugepaged_max_ptes_none = max_ptes_none;
 582
 583         return count;
 584 }
 585 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 586         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 587                khugepaged_max_ptes_none_store);
 588
 589 static struct attribute *khugepaged_attr[] = {
 590         &khugepaged_defrag_attr.attr,
 591         &khugepaged_max_ptes_none_attr.attr,
 592         &pages_to_scan_attr.attr,
 593         &pages_collapsed_attr.attr,
 594         &full_scans_attr.attr,
 595         &scan_sleep_millisecs_attr.attr,
 596         &alloc_sleep_millisecs_attr.attr,
 597         NULL,
 598 };
 599
 600 static struct attribute_group khugepaged_attr_group = {
 601         .attrs = khugepaged_attr,
 602         .name = "khugepaged",
 603 };
 604
 605 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 606 {
 607         int err;
 608
 609         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 610         if (unlikely(!*hugepage_kobj)) {
 611                 pr_err("failed to create transparent hugepage kobject\n");
 612                 return -ENOMEM;
 613         }
 614
 615         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 616         if (err) {
 617                 pr_err("failed to register transparent hugepage group\n");
 618                 goto delete_obj;
 619         }
 620
 621         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 622         if (err) {
 623                 pr_err("failed to register transparent hugepage group\n");
 624                 goto remove_hp_group;
 625         }
 626
 627         return 0;
 628
 629 remove_hp_group:
 630         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 631 delete_obj:
 632         kobject_put(*hugepage_kobj);
 633         return err;
 634 }
 635
 636 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 637 {
 638         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 639         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 640         kobject_put(hugepage_kobj);
 641 }
 642 #else
 643 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 644 {
 645         return 0;
 646 }
 647
 648 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 649 {
 650 }
 651 #endif /* CONFIG_SYSFS */
 652
 653 static int __init hugepage_init(void)
 654 {
 655         int err;
 656         struct kobject *hugepage_kobj;
 657
 658         if (!has_transparent_hugepage()) {
 659                 transparent_hugepage_flags = 0;
 660                 return -EINVAL;
 661         }
 662
 663         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 664         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 665         /*
 666          * hugepages can't be allocated by the buddy allocator
 667          */
 668         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 669         /*
 670          * we use page->mapping and page->index in second tail page
 671          * as list_head: assuming THP order >= 2
 672          */
 673         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 674
 675         err = hugepage_init_sysfs(&hugepage_kobj);
 676         if (err)
 677                 goto err_sysfs;
 678
 679         err = khugepaged_slab_init();
 680         if (err)
 681                 goto err_slab;
 682
 683         err = register_shrinker(&huge_zero_page_shrinker);
 684         if (err)
 685                 goto err_hzp_shrinker;
 686         err = register_shrinker(&deferred_split_shrinker);
 687         if (err)
 688                 goto err_split_shrinker;
 689
 690         /*
 691          * By default disable transparent hugepages on smaller systems,
 692          * where the extra memory used could hurt more than TLB overhead
 693          * is likely to save.  The admin can still enable it through /sys.
 694          */
 695         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 696                 transparent_hugepage_flags = 0;
 697                 return 0;
 698         }
 699
 700         err = start_stop_khugepaged();
 701         if (err)
 702                 goto err_khugepaged;
 703
 704         return 0;
 705 err_khugepaged:
 706         unregister_shrinker(&deferred_split_shrinker);
 707 err_split_shrinker:
 708         unregister_shrinker(&huge_zero_page_shrinker);
 709 err_hzp_shrinker:
 710         khugepaged_slab_exit();
 711 err_slab:
 712         hugepage_exit_sysfs(hugepage_kobj);
 713 err_sysfs:
 714         return err;
 715 }
 716 subsys_initcall(hugepage_init);
 717
 718 static int __init setup_transparent_hugepage(char *str)
 719 {
 720         int ret = 0;
 721         if (!str)
 722                 goto out;
 723         if (!strcmp(str, "always")) {
 724                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 725                         &transparent_hugepage_flags);
 726                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 727                           &transparent_hugepage_flags);
 728                 ret = 1;
 729         } else if (!strcmp(str, "madvise")) {
 730                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 731                           &transparent_hugepage_flags);
 732                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 733                         &transparent_hugepage_flags);
 734                 ret = 1;
 735         } else if (!strcmp(str, "never")) {
 736                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 737                           &transparent_hugepage_flags);
 738                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 739                           &transparent_hugepage_flags);
 740                 ret = 1;
 741         }
 742 out:
 743         if (!ret)
 744                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 745         return ret;
 746 }
 747 __setup("transparent_hugepage=", setup_transparent_hugepage);
 748
 749 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 750 {
 751         if (likely(vma->vm_flags & VM_WRITE))
 752                 pmd = pmd_mkwrite(pmd);
 753         return pmd;
 754 }
 755
 756 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 757 {
 758         pmd_t entry;
 759         entry = mk_pmd(page, prot);
 760         entry = pmd_mkhuge(entry);
 761         return entry;
 762 }
 763
 764 static inline struct list_head *page_deferred_list(struct page *page)
 765 {
 766         /*
 767          * ->lru in the tail pages is occupied by compound_head.
 768          * Let's use ->mapping + ->index in the second tail page as list_head.
 769          */
 770         return (struct list_head *)&page[2].mapping;
 771 }
 772
 773 void prep_transhuge_page(struct page *page)
 774 {
 775         /*
 776          * we use page->mapping and page->indexlru in second tail page
 777          * as list_head: assuming THP order >= 2
 778          */
 779
 780         INIT_LIST_HEAD(page_deferred_list(page));
 781         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 782 }
 783
 784 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 785                                         struct vm_area_struct *vma,
 786                                         unsigned long address, pmd_t *pmd,
 787                                         struct page *page, gfp_t gfp,
 788                                         unsigned int flags)
 789 {
 790         struct mem_cgroup *memcg;
 791         pgtable_t pgtable;
 792         spinlock_t *ptl;
 793         unsigned long haddr = address & HPAGE_PMD_MASK;
 794
 795         VM_BUG_ON_PAGE(!PageCompound(page), page);
 796
 797         if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
 798                 put_page(page);
 799                 count_vm_event(THP_FAULT_FALLBACK);
 800                 return VM_FAULT_FALLBACK;
 801         }
 802
 803         pgtable = pte_alloc_one(mm, haddr);
 804         if (unlikely(!pgtable)) {
 805                 mem_cgroup_cancel_charge(page, memcg, true);
 806                 put_page(page);
 807                 return VM_FAULT_OOM;
 808         }
 809
 810         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 811         /*
 812          * The memory barrier inside __SetPageUptodate makes sure that
 813          * clear_huge_page writes become visible before the set_pmd_at()
 814          * write.
 815          */
 816         __SetPageUptodate(page);
 817
 818         ptl = pmd_lock(mm, pmd);
 819         if (unlikely(!pmd_none(*pmd))) {
 820                 spin_unlock(ptl);
 821                 mem_cgroup_cancel_charge(page, memcg, true);
 822                 put_page(page);
 823                 pte_free(mm, pgtable);
 824         } else {
 825                 pmd_t entry;
 826
 827                 /* Deliver the page fault to userland */
 828                 if (userfaultfd_missing(vma)) {
 829                         int ret;
 830
 831                         spin_unlock(ptl);
 832                         mem_cgroup_cancel_charge(page, memcg, true);
 833                         put_page(page);
 834                         pte_free(mm, pgtable);
 835                         ret = handle_userfault(vma, address, flags,
 836                                                VM_UFFD_MISSING);
 837                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 838                         return ret;
 839                 }
 840
 841                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 842                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 843                 page_add_new_anon_rmap(page, vma, haddr, true);
 844                 mem_cgroup_commit_charge(page, memcg, false, true);
 845                 lru_cache_add_active_or_unevictable(page, vma);
 846                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 847                 set_pmd_at(mm, haddr, pmd, entry);
 848                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 849                 atomic_long_inc(&mm->nr_ptes);
 850                 spin_unlock(ptl);
 851                 count_vm_event(THP_FAULT_ALLOC);
 852         }
 853
 854         return 0;
 855 }
 856
 857 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 858 {
 859         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 860 }
 861
 862 /* Caller must hold page table lock. */
 863 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 864                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 865                 struct page *zero_page)
 866 {
 867         pmd_t entry;
 868         if (!pmd_none(*pmd))
 869                 return false;
 870         entry = mk_pmd(zero_page, vma->vm_page_prot);
 871         entry = pmd_mkhuge(entry);
 872         if (pgtable)
 873                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 874         set_pmd_at(mm, haddr, pmd, entry);
 875         atomic_long_inc(&mm->nr_ptes);
 876         return true;
 877 }
 878
 879 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 880                                unsigned long address, pmd_t *pmd,
 881                                unsigned int flags)
 882 {
 883         gfp_t gfp;
 884         struct page *page;
 885         unsigned long haddr = address & HPAGE_PMD_MASK;
 886
 887         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 888                 return VM_FAULT_FALLBACK;
 889         if (unlikely(anon_vma_prepare(vma)))
 890                 return VM_FAULT_OOM;
 891         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 892                 return VM_FAULT_OOM;
 893         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 894                         transparent_hugepage_use_zero_page()) {
 895                 spinlock_t *ptl;
 896                 pgtable_t pgtable;
 897                 struct page *zero_page;
 898                 bool set;
 899                 int ret;
 900                 pgtable = pte_alloc_one(mm, haddr);
 901                 if (unlikely(!pgtable))
 902                         return VM_FAULT_OOM;
 903                 zero_page = get_huge_zero_page();
 904                 if (unlikely(!zero_page)) {
 905                         pte_free(mm, pgtable);
 906                         count_vm_event(THP_FAULT_FALLBACK);
 907                         return VM_FAULT_FALLBACK;
 908                 }
 909                 ptl = pmd_lock(mm, pmd);
 910                 ret = 0;
 911                 set = false;
 912                 if (pmd_none(*pmd)) {
 913                         if (userfaultfd_missing(vma)) {
 914                                 spin_unlock(ptl);
 915                                 ret = handle_userfault(vma, address, flags,
 916                                                        VM_UFFD_MISSING);
 917                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 918                         } else {
 919                                 set_huge_zero_page(pgtable, mm, vma,
 920                                                    haddr, pmd,
 921                                                    zero_page);
 922                                 spin_unlock(ptl);
 923                                 set = true;
 924                         }
 925                 } else
 926                         spin_unlock(ptl);
 927                 if (!set) {
 928                         pte_free(mm, pgtable);
 929                         put_huge_zero_page();
 930                 }
 931                 return ret;
 932         }
 933         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 934         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 935         if (unlikely(!page)) {
 936                 count_vm_event(THP_FAULT_FALLBACK);
 937                 return VM_FAULT_FALLBACK;
 938         }
 939         prep_transhuge_page(page);
 940         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 941                                             flags);
 942 }
 943
 944 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 945                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 946 {
 947         struct mm_struct *mm = vma->vm_mm;
 948         pmd_t entry;
 949         spinlock_t *ptl;
 950
 951         ptl = pmd_lock(mm, pmd);
 952         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 953         if (pfn_t_devmap(pfn))
 954                 entry = pmd_mkdevmap(entry);
 955         if (write) {
 956                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 957                 entry = maybe_pmd_mkwrite(entry, vma);
 958         }
 959         set_pmd_at(mm, addr, pmd, entry);
 960         update_mmu_cache_pmd(vma, addr, pmd);
 961         spin_unlock(ptl);
 962 }
 963
 964 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 965                         pmd_t *pmd, pfn_t pfn, bool write)
 966 {
 967         pgprot_t pgprot = vma->vm_page_prot;
 968         /*
 969          * If we had pmd_special, we could avoid all these restrictions,
 970          * but we need to be consistent with PTEs and architectures that
 971          * can't support a 'special' bit.
 972          */
 973         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 974         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 975                                                 (VM_PFNMAP|VM_MIXEDMAP));
 976         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 977         BUG_ON(!pfn_t_devmap(pfn));
 978
 979         if (addr < vma->vm_start || addr >= vma->vm_end)
 980                 return VM_FAULT_SIGBUS;
 981         if (track_pfn_insert(vma, &pgprot, pfn))
 982                 return VM_FAULT_SIGBUS;
 983         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 984         return VM_FAULT_NOPAGE;
 985 }
 986
 987 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 988                 pmd_t *pmd)
 989 {
 990         pmd_t _pmd;
 991
 992         /*
 993          * We should set the dirty bit only for FOLL_WRITE but for now
 994          * the dirty bit in the pmd is meaningless.  And if the dirty
 995          * bit will become meaningful and we'll only set it with
 996          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 997          * set the young bit, instead of the current set_pmd_at.
 998          */
 999         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1000         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1001                                 pmd, _pmd,  1))
1002                 update_mmu_cache_pmd(vma, addr, pmd);
1003 }
1004
1005 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1006                 pmd_t *pmd, int flags)
1007 {
1008         unsigned long pfn = pmd_pfn(*pmd);
1009         struct mm_struct *mm = vma->vm_mm;
1010         struct dev_pagemap *pgmap;
1011         struct page *page;
1012
1013         assert_spin_locked(pmd_lockptr(mm, pmd));
1014
1015         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1016                 return NULL;
1017
1018         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1019                 /* pass */;
1020         else
1021                 return NULL;
1022
1023         if (flags & FOLL_TOUCH)
1024                 touch_pmd(vma, addr, pmd);
1025
1026         /*
1027          * device mapped pages can only be returned if the
1028          * caller will manage the page reference count.
1029          */
1030         if (!(flags & FOLL_GET))
1031                 return ERR_PTR(-EEXIST);
1032
1033         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1034         pgmap = get_dev_pagemap(pfn, NULL);
1035         if (!pgmap)
1036                 return ERR_PTR(-EFAULT);
1037         page = pfn_to_page(pfn);
1038         get_page(page);
1039         put_dev_pagemap(pgmap);
1040
1041         return page;
1042 }
1043
1044 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1046                   struct vm_area_struct *vma)
1047 {
1048         spinlock_t *dst_ptl, *src_ptl;
1049         struct page *src_page;
1050         pmd_t pmd;
1051         pgtable_t pgtable = NULL;
1052         int ret;
1053
1054         if (!vma_is_dax(vma)) {
1055                 ret = -ENOMEM;
1056                 pgtable = pte_alloc_one(dst_mm, addr);
1057                 if (unlikely(!pgtable))
1058                         goto out;
1059         }
1060
1061         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1062         src_ptl = pmd_lockptr(src_mm, src_pmd);
1063         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1064
1065         ret = -EAGAIN;
1066         pmd = *src_pmd;
1067         if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1068                 pte_free(dst_mm, pgtable);
1069                 goto out_unlock;
1070         }
1071         /*
1072          * When page table lock is held, the huge zero pmd should not be
1073          * under splitting since we don't split the page itself, only pmd to
1074          * a page table.
1075          */
1076         if (is_huge_zero_pmd(pmd)) {
1077                 struct page *zero_page;
1078                 /*
1079                  * get_huge_zero_page() will never allocate a new page here,
1080                  * since we already have a zero page to copy. It just takes a
1081                  * reference.
1082                  */
1083                 zero_page = get_huge_zero_page();
1084                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1085                                 zero_page);
1086                 ret = 0;
1087                 goto out_unlock;
1088         }
1089
1090         if (!vma_is_dax(vma)) {
1091                 /* thp accounting separate from pmd_devmap accounting */
1092                 src_page = pmd_page(pmd);
1093                 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1094                 get_page(src_page);
1095                 page_dup_rmap(src_page, true);
1096                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1097                 atomic_long_inc(&dst_mm->nr_ptes);
1098                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1099         }
1100
1101         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1102         pmd = pmd_mkold(pmd_wrprotect(pmd));
1103         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1104
1105         ret = 0;
1106 out_unlock:
1107         spin_unlock(src_ptl);
1108         spin_unlock(dst_ptl);
1109 out:
1110         return ret;
1111 }
1112
1113 void huge_pmd_set_accessed(struct mm_struct *mm,
1114                            struct vm_area_struct *vma,
1115                            unsigned long address,
1116                            pmd_t *pmd, pmd_t orig_pmd,
1117                            int dirty)
1118 {
1119         spinlock_t *ptl;
1120         pmd_t entry;
1121         unsigned long haddr;
1122
1123         ptl = pmd_lock(mm, pmd);
1124         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1125                 goto unlock;
1126
1127         entry = pmd_mkyoung(orig_pmd);
1128         haddr = address & HPAGE_PMD_MASK;
1129         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1130                 update_mmu_cache_pmd(vma, address, pmd);
1131
1132 unlock:
1133         spin_unlock(ptl);
1134 }
1135
1136 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1137                                         struct vm_area_struct *vma,
1138                                         unsigned long address,
1139                                         pmd_t *pmd, pmd_t orig_pmd,
1140                                         struct page *page,
1141                                         unsigned long haddr)
1142 {
1143         struct mem_cgroup *memcg;
1144         spinlock_t *ptl;
1145         pgtable_t pgtable;
1146         pmd_t _pmd;
1147         int ret = 0, i;
1148         struct page **pages;
1149         unsigned long mmun_start;       /* For mmu_notifiers */
1150         unsigned long mmun_end;         /* For mmu_notifiers */
1151
1152         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1153                         GFP_KERNEL);
1154         if (unlikely(!pages)) {
1155                 ret |= VM_FAULT_OOM;
1156                 goto out;
1157         }
1158
1159         for (i = 0; i < HPAGE_PMD_NR; i++) {
1160                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1161                                                __GFP_OTHER_NODE,
1162                                                vma, address, page_to_nid(page));
1163                 if (unlikely(!pages[i] ||
1164                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1165                                                    &memcg, false))) {
1166                         if (pages[i])
1167                                 put_page(pages[i]);
1168                         while (--i >= 0) {
1169                                 memcg = (void *)page_private(pages[i]);
1170                                 set_page_private(pages[i], 0);
1171                                 mem_cgroup_cancel_charge(pages[i], memcg,
1172                                                 false);
1173                                 put_page(pages[i]);
1174                         }
1175                         kfree(pages);
1176                         ret |= VM_FAULT_OOM;
1177                         goto out;
1178                 }
1179                 set_page_private(pages[i], (unsigned long)memcg);
1180         }
1181
1182         for (i = 0; i < HPAGE_PMD_NR; i++) {
1183                 copy_user_highpage(pages[i], page + i,
1184                                    haddr + PAGE_SIZE * i, vma);
1185                 __SetPageUptodate(pages[i]);
1186                 cond_resched();
1187         }
1188
1189         mmun_start = haddr;
1190         mmun_end   = haddr + HPAGE_PMD_SIZE;
1191         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1192
1193         ptl = pmd_lock(mm, pmd);
1194         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1195                 goto out_free_pages;
1196         VM_BUG_ON_PAGE(!PageHead(page), page);
1197
1198         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1199         /* leave pmd empty until pte is filled */
1200
1201         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1202         pmd_populate(mm, &_pmd, pgtable);
1203
1204         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1205                 pte_t *pte, entry;
1206                 entry = mk_pte(pages[i], vma->vm_page_prot);
1207                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1208                 memcg = (void *)page_private(pages[i]);
1209                 set_page_private(pages[i], 0);
1210                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1211                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1212                 lru_cache_add_active_or_unevictable(pages[i], vma);
1213                 pte = pte_offset_map(&_pmd, haddr);
1214                 VM_BUG_ON(!pte_none(*pte));
1215                 set_pte_at(mm, haddr, pte, entry);
1216                 pte_unmap(pte);
1217         }
1218         kfree(pages);
1219
1220         smp_wmb(); /* make pte visible before pmd */
1221         pmd_populate(mm, pmd, pgtable);
1222         page_remove_rmap(page, true);
1223         spin_unlock(ptl);
1224
1225         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1226
1227         ret |= VM_FAULT_WRITE;
1228         put_page(page);
1229
1230 out:
1231         return ret;
1232
1233 out_free_pages:
1234         spin_unlock(ptl);
1235         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1236         for (i = 0; i < HPAGE_PMD_NR; i++) {
1237                 memcg = (void *)page_private(pages[i]);
1238                 set_page_private(pages[i], 0);
1239                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1240                 put_page(pages[i]);
1241         }
1242         kfree(pages);
1243         goto out;
1244 }
1245
1246 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1247                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1248 {
1249         spinlock_t *ptl;
1250         int ret = 0;
1251         struct page *page = NULL, *new_page;
1252         struct mem_cgroup *memcg;
1253         unsigned long haddr;
1254         unsigned long mmun_start;       /* For mmu_notifiers */
1255         unsigned long mmun_end;         /* For mmu_notifiers */
1256         gfp_t huge_gfp;                 /* for allocation and charge */
1257
1258         ptl = pmd_lockptr(mm, pmd);
1259         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1260         haddr = address & HPAGE_PMD_MASK;
1261         if (is_huge_zero_pmd(orig_pmd))
1262                 goto alloc;
1263         spin_lock(ptl);
1264         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1265                 goto out_unlock;
1266
1267         page = pmd_page(orig_pmd);
1268         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1269         /*
1270          * We can only reuse the page if nobody else maps the huge page or it's
1271          * part. We can do it by checking page_mapcount() on each sub-page, but
1272          * it's expensive.
1273          * The cheaper way is to check page_count() to be equal 1: every
1274          * mapcount takes page reference reference, so this way we can
1275          * guarantee, that the PMD is the only mapping.
1276          * This can give false negative if somebody pinned the page, but that's
1277          * fine.
1278          */
1279         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1280                 pmd_t entry;
1281                 entry = pmd_mkyoung(orig_pmd);
1282                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1283                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1284                         update_mmu_cache_pmd(vma, address, pmd);
1285                 ret |= VM_FAULT_WRITE;
1286                 goto out_unlock;
1287         }
1288         get_page(page);
1289         spin_unlock(ptl);
1290 alloc:
1291         if (transparent_hugepage_enabled(vma) &&
1292             !transparent_hugepage_debug_cow()) {
1293                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1294                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1295         } else
1296                 new_page = NULL;
1297
1298         if (likely(new_page)) {
1299                 prep_transhuge_page(new_page);
1300         } else {
1301                 if (!page) {
1302                         split_huge_pmd(vma, pmd, address);
1303                         ret |= VM_FAULT_FALLBACK;
1304                 } else {
1305                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1306                                         pmd, orig_pmd, page, haddr);
1307                         if (ret & VM_FAULT_OOM) {
1308                                 split_huge_pmd(vma, pmd, address);
1309                                 ret |= VM_FAULT_FALLBACK;
1310                         }
1311                         put_page(page);
1312                 }
1313                 count_vm_event(THP_FAULT_FALLBACK);
1314                 goto out;
1315         }
1316
1317         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1318                                            true))) {
1319                 put_page(new_page);
1320                 if (page) {
1321                         split_huge_pmd(vma, pmd, address);
1322                         put_page(page);
1323                 } else
1324                         split_huge_pmd(vma, pmd, address);
1325                 ret |= VM_FAULT_FALLBACK;
1326                 count_vm_event(THP_FAULT_FALLBACK);
1327                 goto out;
1328         }
1329
1330         count_vm_event(THP_FAULT_ALLOC);
1331
1332         if (!page)
1333                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1334         else
1335                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1336         __SetPageUptodate(new_page);
1337
1338         mmun_start = haddr;
1339         mmun_end   = haddr + HPAGE_PMD_SIZE;
1340         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1341
1342         spin_lock(ptl);
1343         if (page)
1344                 put_page(page);
1345         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1346                 spin_unlock(ptl);
1347                 mem_cgroup_cancel_charge(new_page, memcg, true);
1348                 put_page(new_page);
1349                 goto out_mn;
1350         } else {
1351                 pmd_t entry;
1352                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1353                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1354                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1355                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1356                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1357                 lru_cache_add_active_or_unevictable(new_page, vma);
1358                 set_pmd_at(mm, haddr, pmd, entry);
1359                 update_mmu_cache_pmd(vma, address, pmd);
1360                 if (!page) {
1361                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1362                         put_huge_zero_page();
1363                 } else {
1364                         VM_BUG_ON_PAGE(!PageHead(page), page);
1365                         page_remove_rmap(page, true);
1366                         put_page(page);
1367                 }
1368                 ret |= VM_FAULT_WRITE;
1369         }
1370         spin_unlock(ptl);
1371 out_mn:
1372         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1373 out:
1374         return ret;
1375 out_unlock:
1376         spin_unlock(ptl);
1377         return ret;
1378 }
1379
1380 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1381                                    unsigned long addr,
1382                                    pmd_t *pmd,
1383                                    unsigned int flags)
1384 {
1385         struct mm_struct *mm = vma->vm_mm;
1386         struct page *page = NULL;
1387
1388         assert_spin_locked(pmd_lockptr(mm, pmd));
1389
1390         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1391                 goto out;
1392
1393         /* Avoid dumping huge zero page */
1394         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1395                 return ERR_PTR(-EFAULT);
1396
1397         /* Full NUMA hinting faults to serialise migration in fault paths */
1398         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1399                 goto out;
1400
1401         page = pmd_page(*pmd);
1402         VM_BUG_ON_PAGE(!PageHead(page), page);
1403         if (flags & FOLL_TOUCH)
1404                 touch_pmd(vma, addr, pmd);
1405         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1406                 /*
1407                  * We don't mlock() pte-mapped THPs. This way we can avoid
1408                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1409                  *
1410                  * In most cases the pmd is the only mapping of the page as we
1411                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1412                  * writable private mappings in populate_vma_page_range().
1413                  *
1414                  * The only scenario when we have the page shared here is if we
1415                  * mlocking read-only mapping shared over fork(). We skip
1416                  * mlocking such pages.
1417                  */
1418                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1419                                 page->mapping && trylock_page(page)) {
1420                         lru_add_drain();
1421                         if (page->mapping)
1422                                 mlock_vma_page(page);
1423                         unlock_page(page);
1424                 }
1425         }
1426         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1427         VM_BUG_ON_PAGE(!PageCompound(page), page);
1428         if (flags & FOLL_GET)
1429                 get_page(page);
1430
1431 out:
1432         return page;
1433 }
1434
1435 /* NUMA hinting page fault entry point for trans huge pmds */
1436 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1437                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1438 {
1439         spinlock_t *ptl;
1440         struct anon_vma *anon_vma = NULL;
1441         struct page *page;
1442         unsigned long haddr = addr & HPAGE_PMD_MASK;
1443         int page_nid = -1, this_nid = numa_node_id();
1444         int target_nid, last_cpupid = -1;
1445         bool page_locked;
1446         bool migrated = false;
1447         bool was_writable;
1448         int flags = 0;
1449
1450         /* A PROT_NONE fault should not end up here */
1451         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1452
1453         ptl = pmd_lock(mm, pmdp);
1454         if (unlikely(!pmd_same(pmd, *pmdp)))
1455                 goto out_unlock;
1456
1457         /*
1458          * If there are potential migrations, wait for completion and retry
1459          * without disrupting NUMA hinting information. Do not relock and
1460          * check_same as the page may no longer be mapped.
1461          */
1462         if (unlikely(pmd_trans_migrating(*pmdp))) {
1463                 page = pmd_page(*pmdp);
1464                 spin_unlock(ptl);
1465                 wait_on_page_locked(page);
1466                 goto out;
1467         }
1468
1469         page = pmd_page(pmd);
1470         BUG_ON(is_huge_zero_page(page));
1471         page_nid = page_to_nid(page);
1472         last_cpupid = page_cpupid_last(page);
1473         count_vm_numa_event(NUMA_HINT_FAULTS);
1474         if (page_nid == this_nid) {
1475                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1476                 flags |= TNF_FAULT_LOCAL;
1477         }
1478
1479         /* See similar comment in do_numa_page for explanation */
1480         if (!(vma->vm_flags & VM_WRITE))
1481                 flags |= TNF_NO_GROUP;
1482
1483         /*
1484          * Acquire the page lock to serialise THP migrations but avoid dropping
1485          * page_table_lock if at all possible
1486          */
1487         page_locked = trylock_page(page);
1488         target_nid = mpol_misplaced(page, vma, haddr);
1489         if (target_nid == -1) {
1490                 /* If the page was locked, there are no parallel migrations */
1491                 if (page_locked)
1492                         goto clear_pmdnuma;
1493         }
1494
1495         /* Migration could have started since the pmd_trans_migrating check */
1496         if (!page_locked) {
1497                 spin_unlock(ptl);
1498                 wait_on_page_locked(page);
1499                 page_nid = -1;
1500                 goto out;
1501         }
1502
1503         /*
1504          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1505          * to serialises splits
1506          */
1507         get_page(page);
1508         spin_unlock(ptl);
1509         anon_vma = page_lock_anon_vma_read(page);
1510
1511         /* Confirm the PMD did not change while page_table_lock was released */
1512         spin_lock(ptl);
1513         if (unlikely(!pmd_same(pmd, *pmdp))) {
1514                 unlock_page(page);
1515                 put_page(page);
1516                 page_nid = -1;
1517                 goto out_unlock;
1518         }
1519
1520         /* Bail if we fail to protect against THP splits for any reason */
1521         if (unlikely(!anon_vma)) {
1522                 put_page(page);
1523                 page_nid = -1;
1524                 goto clear_pmdnuma;
1525         }
1526
1527         /*
1528          * Migrate the THP to the requested node, returns with page unlocked
1529          * and access rights restored.
1530          */
1531         spin_unlock(ptl);
1532         migrated = migrate_misplaced_transhuge_page(mm, vma,
1533                                 pmdp, pmd, addr, page, target_nid);
1534         if (migrated) {
1535                 flags |= TNF_MIGRATED;
1536                 page_nid = target_nid;
1537         } else
1538                 flags |= TNF_MIGRATE_FAIL;
1539
1540         goto out;
1541 clear_pmdnuma:
1542         BUG_ON(!PageLocked(page));
1543         was_writable = pmd_write(pmd);
1544         pmd = pmd_modify(pmd, vma->vm_page_prot);
1545         pmd = pmd_mkyoung(pmd);
1546         if (was_writable)
1547                 pmd = pmd_mkwrite(pmd);
1548         set_pmd_at(mm, haddr, pmdp, pmd);
1549         update_mmu_cache_pmd(vma, addr, pmdp);
1550         unlock_page(page);
1551 out_unlock:
1552         spin_unlock(ptl);
1553
1554 out:
1555         if (anon_vma)
1556                 page_unlock_anon_vma_read(anon_vma);
1557
1558         if (page_nid != -1)
1559                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1560
1561         return 0;
1562 }
1563
1564 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1565                 pmd_t *pmd, unsigned long addr, unsigned long next)
1566
1567 {
1568         spinlock_t *ptl;
1569         pmd_t orig_pmd;
1570         struct page *page;
1571         struct mm_struct *mm = tlb->mm;
1572         int ret = 0;
1573
1574         ptl = pmd_trans_huge_lock(pmd, vma);
1575         if (!ptl)
1576                 goto out_unlocked;
1577
1578         orig_pmd = *pmd;
1579         if (is_huge_zero_pmd(orig_pmd)) {
1580                 ret = 1;
1581                 goto out;
1582         }
1583
1584         page = pmd_page(orig_pmd);
1585         /*
1586          * If other processes are mapping this page, we couldn't discard
1587          * the page unless they all do MADV_FREE so let's skip the page.
1588          */
1589         if (page_mapcount(page) != 1)
1590                 goto out;
1591
1592         if (!trylock_page(page))
1593                 goto out;
1594
1595         /*
1596          * If user want to discard part-pages of THP, split it so MADV_FREE
1597          * will deactivate only them.
1598          */
1599         if (next - addr != HPAGE_PMD_SIZE) {
1600                 get_page(page);
1601                 spin_unlock(ptl);
1602                 if (split_huge_page(page)) {
1603                         put_page(page);
1604                         unlock_page(page);
1605                         goto out_unlocked;
1606                 }
1607                 put_page(page);
1608                 unlock_page(page);
1609                 ret = 1;
1610                 goto out_unlocked;
1611         }
1612
1613         if (PageDirty(page))
1614                 ClearPageDirty(page);
1615         unlock_page(page);
1616
1617         if (PageActive(page))
1618                 deactivate_page(page);
1619
1620         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1621                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1622                         tlb->fullmm);
1623                 orig_pmd = pmd_mkold(orig_pmd);
1624                 orig_pmd = pmd_mkclean(orig_pmd);
1625
1626                 set_pmd_at(mm, addr, pmd, orig_pmd);
1627                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1628         }
1629         ret = 1;
1630 out:
1631         spin_unlock(ptl);
1632 out_unlocked:
1633         return ret;
1634 }
1635
1636 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1637                  pmd_t *pmd, unsigned long addr)
1638 {
1639         pmd_t orig_pmd;
1640         spinlock_t *ptl;
1641
1642         ptl = __pmd_trans_huge_lock(pmd, vma);
1643         if (!ptl)
1644                 return 0;
1645         /*
1646          * For architectures like ppc64 we look at deposited pgtable
1647          * when calling pmdp_huge_get_and_clear. So do the
1648          * pgtable_trans_huge_withdraw after finishing pmdp related
1649          * operations.
1650          */
1651         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1652                         tlb->fullmm);
1653         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1654         if (vma_is_dax(vma)) {
1655                 spin_unlock(ptl);
1656                 if (is_huge_zero_pmd(orig_pmd))
1657                         put_huge_zero_page();
1658         } else if (is_huge_zero_pmd(orig_pmd)) {
1659                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1660                 atomic_long_dec(&tlb->mm->nr_ptes);
1661                 spin_unlock(ptl);
1662                 put_huge_zero_page();
1663         } else {
1664                 struct page *page = pmd_page(orig_pmd);
1665                 page_remove_rmap(page, true);
1666                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1667                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1668                 VM_BUG_ON_PAGE(!PageHead(page), page);
1669                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1670                 atomic_long_dec(&tlb->mm->nr_ptes);
1671                 spin_unlock(ptl);
1672                 tlb_remove_page(tlb, page);
1673         }
1674         return 1;
1675 }
1676
1677 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1678                   unsigned long old_addr,
1679                   unsigned long new_addr, unsigned long old_end,
1680                   pmd_t *old_pmd, pmd_t *new_pmd)
1681 {
1682         spinlock_t *old_ptl, *new_ptl;
1683         pmd_t pmd;
1684
1685         struct mm_struct *mm = vma->vm_mm;
1686
1687         if ((old_addr & ~HPAGE_PMD_MASK) ||
1688             (new_addr & ~HPAGE_PMD_MASK) ||
1689             old_end - old_addr < HPAGE_PMD_SIZE ||
1690             (new_vma->vm_flags & VM_NOHUGEPAGE))
1691                 return false;
1692
1693         /*
1694          * The destination pmd shouldn't be established, free_pgtables()
1695          * should have release it.
1696          */
1697         if (WARN_ON(!pmd_none(*new_pmd))) {
1698                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1699                 return false;
1700         }
1701
1702         /*
1703          * We don't have to worry about the ordering of src and dst
1704          * ptlocks because exclusive mmap_sem prevents deadlock.
1705          */
1706         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1707         if (old_ptl) {
1708                 new_ptl = pmd_lockptr(mm, new_pmd);
1709                 if (new_ptl != old_ptl)
1710                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1711                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1712                 VM_BUG_ON(!pmd_none(*new_pmd));
1713
1714                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1715                         pgtable_t pgtable;
1716                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1717                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1718                 }
1719                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1720                 if (new_ptl != old_ptl)
1721                         spin_unlock(new_ptl);
1722                 spin_unlock(old_ptl);
1723                 return true;
1724         }
1725         return false;
1726 }
1727
1728 /*
1729  * Returns
1730  *  - 0 if PMD could not be locked
1731  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1732  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1733  */
1734 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1735                 unsigned long addr, pgprot_t newprot, int prot_numa)
1736 {
1737         struct mm_struct *mm = vma->vm_mm;
1738         spinlock_t *ptl;
1739         int ret = 0;
1740
1741         ptl = __pmd_trans_huge_lock(pmd, vma);
1742         if (ptl) {
1743                 pmd_t entry;
1744                 bool preserve_write = prot_numa && pmd_write(*pmd);
1745                 ret = 1;
1746
1747                 /*
1748                  * Avoid trapping faults against the zero page. The read-only
1749                  * data is likely to be read-cached on the local CPU and
1750                  * local/remote hits to the zero page are not interesting.
1751                  */
1752                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1753                         spin_unlock(ptl);
1754                         return ret;
1755                 }
1756
1757                 if (!prot_numa || !pmd_protnone(*pmd)) {
1758                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1759                         entry = pmd_modify(entry, newprot);
1760                         if (preserve_write)
1761                                 entry = pmd_mkwrite(entry);
1762                         ret = HPAGE_PMD_NR;
1763                         set_pmd_at(mm, addr, pmd, entry);
1764                         BUG_ON(!preserve_write && pmd_write(entry));
1765                 }
1766                 spin_unlock(ptl);
1767         }
1768
1769         return ret;
1770 }
1771
1772 /*
1773  * Returns true if a given pmd maps a thp, false otherwise.
1774  *
1775  * Note that if it returns true, this routine returns without unlocking page
1776  * table lock. So callers must unlock it.
1777  */
1778 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1779 {
1780         spinlock_t *ptl;
1781         ptl = pmd_lock(vma->vm_mm, pmd);
1782         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1783                 return ptl;
1784         spin_unlock(ptl);
1785         return NULL;
1786 }
1787
1788 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1789
1790 int hugepage_madvise(struct vm_area_struct *vma,
1791                      unsigned long *vm_flags, int advice)
1792 {
1793         switch (advice) {
1794         case MADV_HUGEPAGE:
1795 #ifdef CONFIG_S390
1796                 /*
1797                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1798                  * can't handle this properly after s390_enable_sie, so we simply
1799                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1800                  */
1801                 if (mm_has_pgste(vma->vm_mm))
1802                         return 0;
1803 #endif
1804                 /*
1805                  * Be somewhat over-protective like KSM for now!
1806                  */
1807                 if (*vm_flags & VM_NO_THP)
1808                         return -EINVAL;
1809                 *vm_flags &= ~VM_NOHUGEPAGE;
1810                 *vm_flags |= VM_HUGEPAGE;
1811                 /*
1812                  * If the vma become good for khugepaged to scan,
1813                  * register it here without waiting a page fault that
1814                  * may not happen any time soon.
1815                  */
1816                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1817                         return -ENOMEM;
1818                 break;
1819         case MADV_NOHUGEPAGE:
1820                 /*
1821                  * Be somewhat over-protective like KSM for now!
1822                  */
1823                 if (*vm_flags & VM_NO_THP)
1824                         return -EINVAL;
1825                 *vm_flags &= ~VM_HUGEPAGE;
1826                 *vm_flags |= VM_NOHUGEPAGE;
1827                 /*
1828                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1829                  * this vma even if we leave the mm registered in khugepaged if
1830                  * it got registered before VM_NOHUGEPAGE was set.
1831                  */
1832                 break;
1833         }
1834
1835         return 0;
1836 }
1837
1838 static int __init khugepaged_slab_init(void)
1839 {
1840         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1841                                           sizeof(struct mm_slot),
1842                                           __alignof__(struct mm_slot), 0, NULL);
1843         if (!mm_slot_cache)
1844                 return -ENOMEM;
1845
1846         return 0;
1847 }
1848
1849 static void __init khugepaged_slab_exit(void)
1850 {
1851         kmem_cache_destroy(mm_slot_cache);
1852 }
1853
1854 static inline struct mm_slot *alloc_mm_slot(void)
1855 {
1856         if (!mm_slot_cache)     /* initialization failed */
1857                 return NULL;
1858         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1859 }
1860
1861 static inline void free_mm_slot(struct mm_slot *mm_slot)
1862 {
1863         kmem_cache_free(mm_slot_cache, mm_slot);
1864 }
1865
1866 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1867 {
1868         struct mm_slot *mm_slot;
1869
1870         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1871                 if (mm == mm_slot->mm)
1872                         return mm_slot;
1873
1874         return NULL;
1875 }
1876
1877 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1878                                     struct mm_slot *mm_slot)
1879 {
1880         mm_slot->mm = mm;
1881         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1882 }
1883
1884 static inline int khugepaged_test_exit(struct mm_struct *mm)
1885 {
1886         return atomic_read(&mm->mm_users) == 0;
1887 }
1888
1889 int __khugepaged_enter(struct mm_struct *mm)
1890 {
1891         struct mm_slot *mm_slot;
1892         int wakeup;
1893
1894         mm_slot = alloc_mm_slot();
1895         if (!mm_slot)
1896                 return -ENOMEM;
1897
1898         /* __khugepaged_exit() must not run from under us */
1899         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1900         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1901                 free_mm_slot(mm_slot);
1902                 return 0;
1903         }
1904
1905         spin_lock(&khugepaged_mm_lock);
1906         insert_to_mm_slots_hash(mm, mm_slot);
1907         /*
1908          * Insert just behind the scanning cursor, to let the area settle
1909          * down a little.
1910          */
1911         wakeup = list_empty(&khugepaged_scan.mm_head);
1912         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1913         spin_unlock(&khugepaged_mm_lock);
1914
1915         atomic_inc(&mm->mm_count);
1916         if (wakeup)
1917                 wake_up_interruptible(&khugepaged_wait);
1918
1919         return 0;
1920 }
1921
1922 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1923                                unsigned long vm_flags)
1924 {
1925         unsigned long hstart, hend;
1926         if (!vma->anon_vma)
1927                 /*
1928                  * Not yet faulted in so we will register later in the
1929                  * page fault if needed.
1930                  */
1931                 return 0;
1932         if (vma->vm_ops)
1933                 /* khugepaged not yet working on file or special mappings */
1934                 return 0;
1935         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1936         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1937         hend = vma->vm_end & HPAGE_PMD_MASK;
1938         if (hstart < hend)
1939                 return khugepaged_enter(vma, vm_flags);
1940         return 0;
1941 }
1942
1943 void __khugepaged_exit(struct mm_struct *mm)
1944 {
1945         struct mm_slot *mm_slot;
1946         int free = 0;
1947
1948         spin_lock(&khugepaged_mm_lock);
1949         mm_slot = get_mm_slot(mm);
1950         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1951                 hash_del(&mm_slot->hash);
1952                 list_del(&mm_slot->mm_node);
1953                 free = 1;
1954         }
1955         spin_unlock(&khugepaged_mm_lock);
1956
1957         if (free) {
1958                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1959                 free_mm_slot(mm_slot);
1960                 mmdrop(mm);
1961         } else if (mm_slot) {
1962                 /*
1963                  * This is required to serialize against
1964                  * khugepaged_test_exit() (which is guaranteed to run
1965                  * under mmap sem read mode). Stop here (after we
1966                  * return all pagetables will be destroyed) until
1967                  * khugepaged has finished working on the pagetables
1968                  * under the mmap_sem.
1969                  */
1970                 down_write(&mm->mmap_sem);
1971                 up_write(&mm->mmap_sem);
1972         }
1973 }
1974
1975 static void release_pte_page(struct page *page)
1976 {
1977         /* 0 stands for page_is_file_cache(page) == false */
1978         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1979         unlock_page(page);
1980         putback_lru_page(page);
1981 }
1982
1983 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1984 {
1985         while (--_pte >= pte) {
1986                 pte_t pteval = *_pte;
1987                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1988                         release_pte_page(pte_page(pteval));
1989         }
1990 }
1991
1992 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1993                                         unsigned long address,
1994                                         pte_t *pte)
1995 {
1996         struct page *page = NULL;
1997         pte_t *_pte;
1998         int none_or_zero = 0, result = 0;
1999         bool referenced = false, writable = false;
2000
2001         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2002              _pte++, address += PAGE_SIZE) {
2003                 pte_t pteval = *_pte;
2004                 if (pte_none(pteval) || (pte_present(pteval) &&
2005                                 is_zero_pfn(pte_pfn(pteval)))) {
2006                         if (!userfaultfd_armed(vma) &&
2007                             ++none_or_zero <= khugepaged_max_ptes_none) {
2008                                 continue;
2009                         } else {
2010                                 result = SCAN_EXCEED_NONE_PTE;
2011                                 goto out;
2012                         }
2013                 }
2014                 if (!pte_present(pteval)) {
2015                         result = SCAN_PTE_NON_PRESENT;
2016                         goto out;
2017                 }
2018                 page = vm_normal_page(vma, address, pteval);
2019                 if (unlikely(!page)) {
2020                         result = SCAN_PAGE_NULL;
2021                         goto out;
2022                 }
2023
2024                 VM_BUG_ON_PAGE(PageCompound(page), page);
2025                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2026                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2027
2028                 /*
2029                  * We can do it before isolate_lru_page because the
2030                  * page can't be freed from under us. NOTE: PG_lock
2031                  * is needed to serialize against split_huge_page
2032                  * when invoked from the VM.
2033                  */
2034                 if (!trylock_page(page)) {
2035                         result = SCAN_PAGE_LOCK;
2036                         goto out;
2037                 }
2038
2039                 /*
2040                  * cannot use mapcount: can't collapse if there's a gup pin.
2041                  * The page must only be referenced by the scanned process
2042                  * and page swap cache.
2043                  */
2044                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2045                         unlock_page(page);
2046                         result = SCAN_PAGE_COUNT;
2047                         goto out;
2048                 }
2049                 if (pte_write(pteval)) {
2050                         writable = true;
2051                 } else {
2052                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2053                                 unlock_page(page);
2054                                 result = SCAN_SWAP_CACHE_PAGE;
2055                                 goto out;
2056                         }
2057                         /*
2058                          * Page is not in the swap cache. It can be collapsed
2059                          * into a THP.
2060                          */
2061                 }
2062
2063                 /*
2064                  * Isolate the page to avoid collapsing an hugepage
2065                  * currently in use by the VM.
2066                  */
2067                 if (isolate_lru_page(page)) {
2068                         unlock_page(page);
2069                         result = SCAN_DEL_PAGE_LRU;
2070                         goto out;
2071                 }
2072                 /* 0 stands for page_is_file_cache(page) == false */
2073                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2074                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2075                 VM_BUG_ON_PAGE(PageLRU(page), page);
2076
2077                 /* If there is no mapped pte young don't collapse the page */
2078                 if (pte_young(pteval) ||
2079                     page_is_young(page) || PageReferenced(page) ||
2080                     mmu_notifier_test_young(vma->vm_mm, address))
2081                         referenced = true;
2082         }
2083         if (likely(writable)) {
2084                 if (likely(referenced)) {
2085                         result = SCAN_SUCCEED;
2086                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2087                                                             referenced, writable, result);
2088                         return 1;
2089                 }
2090         } else {
2091                 result = SCAN_PAGE_RO;
2092         }
2093
2094 out:
2095         release_pte_pages(pte, _pte);
2096         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2097                                             referenced, writable, result);
2098         return 0;
2099 }
2100
2101 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2102                                       struct vm_area_struct *vma,
2103                                       unsigned long address,
2104                                       spinlock_t *ptl)
2105 {
2106         pte_t *_pte;
2107         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2108                 pte_t pteval = *_pte;
2109                 struct page *src_page;
2110
2111                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2112                         clear_user_highpage(page, address);
2113                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2114                         if (is_zero_pfn(pte_pfn(pteval))) {
2115                                 /*
2116                                  * ptl mostly unnecessary.
2117                                  */
2118                                 spin_lock(ptl);
2119                                 /*
2120                                  * paravirt calls inside pte_clear here are
2121                                  * superfluous.
2122                                  */
2123                                 pte_clear(vma->vm_mm, address, _pte);
2124                                 spin_unlock(ptl);
2125                         }
2126                 } else {
2127                         src_page = pte_page(pteval);
2128                         copy_user_highpage(page, src_page, address, vma);
2129                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2130                         release_pte_page(src_page);
2131                         /*
2132                          * ptl mostly unnecessary, but preempt has to
2133                          * be disabled to update the per-cpu stats
2134                          * inside page_remove_rmap().
2135                          */
2136                         spin_lock(ptl);
2137                         /*
2138                          * paravirt calls inside pte_clear here are
2139                          * superfluous.
2140                          */
2141                         pte_clear(vma->vm_mm, address, _pte);
2142                         page_remove_rmap(src_page, false);
2143                         spin_unlock(ptl);
2144                         free_page_and_swap_cache(src_page);
2145                 }
2146
2147                 address += PAGE_SIZE;
2148                 page++;
2149         }
2150 }
2151
2152 static void khugepaged_alloc_sleep(void)
2153 {
2154         DEFINE_WAIT(wait);
2155
2156         add_wait_queue(&khugepaged_wait, &wait);
2157         freezable_schedule_timeout_interruptible(
2158                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2159         remove_wait_queue(&khugepaged_wait, &wait);
2160 }
2161
2162 static int khugepaged_node_load[MAX_NUMNODES];
2163
2164 static bool khugepaged_scan_abort(int nid)
2165 {
2166         int i;
2167
2168         /*
2169          * If zone_reclaim_mode is disabled, then no extra effort is made to
2170          * allocate memory locally.
2171          */
2172         if (!zone_reclaim_mode)
2173                 return false;
2174
2175         /* If there is a count for this node already, it must be acceptable */
2176         if (khugepaged_node_load[nid])
2177                 return false;
2178
2179         for (i = 0; i < MAX_NUMNODES; i++) {
2180                 if (!khugepaged_node_load[i])
2181                         continue;
2182                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2183                         return true;
2184         }
2185         return false;
2186 }
2187
2188 #ifdef CONFIG_NUMA
2189 static int khugepaged_find_target_node(void)
2190 {
2191         static int last_khugepaged_target_node = NUMA_NO_NODE;
2192         int nid, target_node = 0, max_value = 0;
2193
2194         /* find first node with max normal pages hit */
2195         for (nid = 0; nid < MAX_NUMNODES; nid++)
2196                 if (khugepaged_node_load[nid] > max_value) {
2197                         max_value = khugepaged_node_load[nid];
2198                         target_node = nid;
2199                 }
2200
2201         /* do some balance if several nodes have the same hit record */
2202         if (target_node <= last_khugepaged_target_node)
2203                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2204                                 nid++)
2205                         if (max_value == khugepaged_node_load[nid]) {
2206                                 target_node = nid;
2207                                 break;
2208                         }
2209
2210         last_khugepaged_target_node = target_node;
2211         return target_node;
2212 }
2213
2214 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2215 {
2216         if (IS_ERR(*hpage)) {
2217                 if (!*wait)
2218                         return false;
2219
2220                 *wait = false;
2221                 *hpage = NULL;
2222                 khugepaged_alloc_sleep();
2223         } else if (*hpage) {
2224                 put_page(*hpage);
2225                 *hpage = NULL;
2226         }
2227
2228         return true;
2229 }
2230
2231 static struct page *
2232 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2233                        unsigned long address, int node)
2234 {
2235         VM_BUG_ON_PAGE(*hpage, *hpage);
2236
2237         /*
2238          * Before allocating the hugepage, release the mmap_sem read lock.
2239          * The allocation can take potentially a long time if it involves
2240          * sync compaction, and we do not need to hold the mmap_sem during
2241          * that. We will recheck the vma after taking it again in write mode.
2242          */
2243         up_read(&mm->mmap_sem);
2244
2245         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2246         if (unlikely(!*hpage)) {
2247                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2248                 *hpage = ERR_PTR(-ENOMEM);
2249                 return NULL;
2250         }
2251
2252         prep_transhuge_page(*hpage);
2253         count_vm_event(THP_COLLAPSE_ALLOC);
2254         return *hpage;
2255 }
2256 #else
2257 static int khugepaged_find_target_node(void)
2258 {
2259         return 0;
2260 }
2261
2262 static inline struct page *alloc_hugepage(int defrag)
2263 {
2264         struct page *page;
2265
2266         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2267         if (page)
2268                 prep_transhuge_page(page);
2269         return page;
2270 }
2271
2272 static struct page *khugepaged_alloc_hugepage(bool *wait)
2273 {
2274         struct page *hpage;
2275
2276         do {
2277                 hpage = alloc_hugepage(khugepaged_defrag());
2278                 if (!hpage) {
2279                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2280                         if (!*wait)
2281                                 return NULL;
2282
2283                         *wait = false;
2284                         khugepaged_alloc_sleep();
2285                 } else
2286                         count_vm_event(THP_COLLAPSE_ALLOC);
2287         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2288
2289         return hpage;
2290 }
2291
2292 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2293 {
2294         if (!*hpage)
2295                 *hpage = khugepaged_alloc_hugepage(wait);
2296
2297         if (unlikely(!*hpage))
2298                 return false;
2299
2300         return true;
2301 }
2302
2303 static struct page *
2304 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2305                        unsigned long address, int node)
2306 {
2307         up_read(&mm->mmap_sem);
2308         VM_BUG_ON(!*hpage);
2309
2310         return  *hpage;
2311 }
2312 #endif
2313
2314 static bool hugepage_vma_check(struct vm_area_struct *vma)
2315 {
2316         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2317             (vma->vm_flags & VM_NOHUGEPAGE))
2318                 return false;
2319         if (!vma->anon_vma || vma->vm_ops)
2320                 return false;
2321         if (is_vma_temporary_stack(vma))
2322                 return false;
2323         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2324         return true;
2325 }
2326
2327 static void collapse_huge_page(struct mm_struct *mm,
2328                                    unsigned long address,
2329                                    struct page **hpage,
2330                                    struct vm_area_struct *vma,
2331                                    int node)
2332 {
2333         pmd_t *pmd, _pmd;
2334         pte_t *pte;
2335         pgtable_t pgtable;
2336         struct page *new_page;
2337         spinlock_t *pmd_ptl, *pte_ptl;
2338         int isolated = 0, result = 0;
2339         unsigned long hstart, hend;
2340         struct mem_cgroup *memcg;
2341         unsigned long mmun_start;       /* For mmu_notifiers */
2342         unsigned long mmun_end;         /* For mmu_notifiers */
2343         gfp_t gfp;
2344
2345         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2346
2347         /* Only allocate from the target node */
2348         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2349                 __GFP_THISNODE;
2350
2351         /* release the mmap_sem read lock. */
2352         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2353         if (!new_page) {
2354                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2355                 goto out_nolock;
2356         }
2357
2358         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2359                 result = SCAN_CGROUP_CHARGE_FAIL;
2360                 goto out_nolock;
2361         }
2362
2363         /*
2364          * Prevent all access to pagetables with the exception of
2365          * gup_fast later hanlded by the ptep_clear_flush and the VM
2366          * handled by the anon_vma lock + PG_lock.
2367          */
2368         down_write(&mm->mmap_sem);
2369         if (unlikely(khugepaged_test_exit(mm))) {
2370                 result = SCAN_ANY_PROCESS;
2371                 goto out;
2372         }
2373
2374         vma = find_vma(mm, address);
2375         if (!vma) {
2376                 result = SCAN_VMA_NULL;
2377                 goto out;
2378         }
2379         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2380         hend = vma->vm_end & HPAGE_PMD_MASK;
2381         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2382                 result = SCAN_ADDRESS_RANGE;
2383                 goto out;
2384         }
2385         if (!hugepage_vma_check(vma)) {
2386                 result = SCAN_VMA_CHECK;
2387                 goto out;
2388         }
2389         pmd = mm_find_pmd(mm, address);
2390         if (!pmd) {
2391                 result = SCAN_PMD_NULL;
2392                 goto out;
2393         }
2394
2395         anon_vma_lock_write(vma->anon_vma);
2396
2397         pte = pte_offset_map(pmd, address);
2398         pte_ptl = pte_lockptr(mm, pmd);
2399
2400         mmun_start = address;
2401         mmun_end   = address + HPAGE_PMD_SIZE;
2402         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2403         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2404         /*
2405          * After this gup_fast can't run anymore. This also removes
2406          * any huge TLB entry from the CPU so we won't allow
2407          * huge and small TLB entries for the same virtual address
2408          * to avoid the risk of CPU bugs in that area.
2409          */
2410         _pmd = pmdp_collapse_flush(vma, address, pmd);
2411         spin_unlock(pmd_ptl);
2412         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2413
2414         spin_lock(pte_ptl);
2415         isolated = __collapse_huge_page_isolate(vma, address, pte);
2416         spin_unlock(pte_ptl);
2417
2418         if (unlikely(!isolated)) {
2419                 pte_unmap(pte);
2420                 spin_lock(pmd_ptl);
2421                 BUG_ON(!pmd_none(*pmd));
2422                 /*
2423                  * We can only use set_pmd_at when establishing
2424                  * hugepmds and never for establishing regular pmds that
2425                  * points to regular pagetables. Use pmd_populate for that
2426                  */
2427                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2428                 spin_unlock(pmd_ptl);
2429                 anon_vma_unlock_write(vma->anon_vma);
2430                 result = SCAN_FAIL;
2431                 goto out;
2432         }
2433
2434         /*
2435          * All pages are isolated and locked so anon_vma rmap
2436          * can't run anymore.
2437          */
2438         anon_vma_unlock_write(vma->anon_vma);
2439
2440         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2441         pte_unmap(pte);
2442         __SetPageUptodate(new_page);
2443         pgtable = pmd_pgtable(_pmd);
2444
2445         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2446         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2447
2448         /*
2449          * spin_lock() below is not the equivalent of smp_wmb(), so
2450          * this is needed to avoid the copy_huge_page writes to become
2451          * visible after the set_pmd_at() write.
2452          */
2453         smp_wmb();
2454
2455         spin_lock(pmd_ptl);
2456         BUG_ON(!pmd_none(*pmd));
2457         page_add_new_anon_rmap(new_page, vma, address, true);
2458         mem_cgroup_commit_charge(new_page, memcg, false, true);
2459         lru_cache_add_active_or_unevictable(new_page, vma);
2460         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2461         set_pmd_at(mm, address, pmd, _pmd);
2462         update_mmu_cache_pmd(vma, address, pmd);
2463         spin_unlock(pmd_ptl);
2464
2465         *hpage = NULL;
2466
2467         khugepaged_pages_collapsed++;
2468         result = SCAN_SUCCEED;
2469 out_up_write:
2470         up_write(&mm->mmap_sem);
2471         trace_mm_collapse_huge_page(mm, isolated, result);
2472         return;
2473
2474 out_nolock:
2475         trace_mm_collapse_huge_page(mm, isolated, result);
2476         return;
2477 out:
2478         mem_cgroup_cancel_charge(new_page, memcg, true);
2479         goto out_up_write;
2480 }
2481
2482 static int khugepaged_scan_pmd(struct mm_struct *mm,
2483                                struct vm_area_struct *vma,
2484                                unsigned long address,
2485                                struct page **hpage)
2486 {
2487         pmd_t *pmd;
2488         pte_t *pte, *_pte;
2489         int ret = 0, none_or_zero = 0, result = 0;
2490         struct page *page = NULL;
2491         unsigned long _address;
2492         spinlock_t *ptl;
2493         int node = NUMA_NO_NODE;
2494         bool writable = false, referenced = false;
2495
2496         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2497
2498         pmd = mm_find_pmd(mm, address);
2499         if (!pmd) {
2500                 result = SCAN_PMD_NULL;
2501                 goto out;
2502         }
2503
2504         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2505         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2506         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2507              _pte++, _address += PAGE_SIZE) {
2508                 pte_t pteval = *_pte;
2509                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2510                         if (!userfaultfd_armed(vma) &&
2511                             ++none_or_zero <= khugepaged_max_ptes_none) {
2512                                 continue;
2513                         } else {
2514                                 result = SCAN_EXCEED_NONE_PTE;
2515                                 goto out_unmap;
2516                         }
2517                 }
2518                 if (!pte_present(pteval)) {
2519                         result = SCAN_PTE_NON_PRESENT;
2520                         goto out_unmap;
2521                 }
2522                 if (pte_write(pteval))
2523                         writable = true;
2524
2525                 page = vm_normal_page(vma, _address, pteval);
2526                 if (unlikely(!page)) {
2527                         result = SCAN_PAGE_NULL;
2528                         goto out_unmap;
2529                 }
2530
2531                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2532                 if (PageCompound(page)) {
2533                         result = SCAN_PAGE_COMPOUND;
2534                         goto out_unmap;
2535                 }
2536
2537                 /*
2538                  * Record which node the original page is from and save this
2539                  * information to khugepaged_node_load[].
2540                  * Khupaged will allocate hugepage from the node has the max
2541                  * hit record.
2542                  */
2543                 node = page_to_nid(page);
2544                 if (khugepaged_scan_abort(node)) {
2545                         result = SCAN_SCAN_ABORT;
2546                         goto out_unmap;
2547                 }
2548                 khugepaged_node_load[node]++;
2549                 if (!PageLRU(page)) {
2550                         result = SCAN_SCAN_ABORT;
2551                         goto out_unmap;
2552                 }
2553                 if (PageLocked(page)) {
2554                         result = SCAN_PAGE_LOCK;
2555                         goto out_unmap;
2556                 }
2557                 if (!PageAnon(page)) {
2558                         result = SCAN_PAGE_ANON;
2559                         goto out_unmap;
2560                 }
2561
2562                 /*
2563                  * cannot use mapcount: can't collapse if there's a gup pin.
2564                  * The page must only be referenced by the scanned process
2565                  * and page swap cache.
2566                  */
2567                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2568                         result = SCAN_PAGE_COUNT;
2569                         goto out_unmap;
2570                 }
2571                 if (pte_young(pteval) ||
2572                     page_is_young(page) || PageReferenced(page) ||
2573                     mmu_notifier_test_young(vma->vm_mm, address))
2574                         referenced = true;
2575         }
2576         if (writable) {
2577                 if (referenced) {
2578                         result = SCAN_SUCCEED;
2579                         ret = 1;
2580                 } else {
2581                         result = SCAN_NO_REFERENCED_PAGE;
2582                 }
2583         } else {
2584                 result = SCAN_PAGE_RO;
2585         }
2586 out_unmap:
2587         pte_unmap_unlock(pte, ptl);
2588         if (ret) {
2589                 node = khugepaged_find_target_node();
2590                 /* collapse_huge_page will return with the mmap_sem released */
2591                 collapse_huge_page(mm, address, hpage, vma, node);
2592         }
2593 out:
2594         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2595                                      none_or_zero, result);
2596         return ret;
2597 }
2598
2599 static void collect_mm_slot(struct mm_slot *mm_slot)
2600 {
2601         struct mm_struct *mm = mm_slot->mm;
2602
2603         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2604
2605         if (khugepaged_test_exit(mm)) {
2606                 /* free mm_slot */
2607                 hash_del(&mm_slot->hash);
2608                 list_del(&mm_slot->mm_node);
2609
2610                 /*
2611                  * Not strictly needed because the mm exited already.
2612                  *
2613                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2614                  */
2615
2616                 /* khugepaged_mm_lock actually not necessary for the below */
2617                 free_mm_slot(mm_slot);
2618                 mmdrop(mm);
2619         }
2620 }
2621
2622 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2623                                             struct page **hpage)
2624         __releases(&khugepaged_mm_lock)
2625         __acquires(&khugepaged_mm_lock)
2626 {
2627         struct mm_slot *mm_slot;
2628         struct mm_struct *mm;
2629         struct vm_area_struct *vma;
2630         int progress = 0;
2631
2632         VM_BUG_ON(!pages);
2633         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2634
2635         if (khugepaged_scan.mm_slot)
2636                 mm_slot = khugepaged_scan.mm_slot;
2637         else {
2638                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2639                                      struct mm_slot, mm_node);
2640                 khugepaged_scan.address = 0;
2641                 khugepaged_scan.mm_slot = mm_slot;
2642         }
2643         spin_unlock(&khugepaged_mm_lock);
2644
2645         mm = mm_slot->mm;
2646         down_read(&mm->mmap_sem);
2647         if (unlikely(khugepaged_test_exit(mm)))
2648                 vma = NULL;
2649         else
2650                 vma = find_vma(mm, khugepaged_scan.address);
2651
2652         progress++;
2653         for (; vma; vma = vma->vm_next) {
2654                 unsigned long hstart, hend;
2655
2656                 cond_resched();
2657                 if (unlikely(khugepaged_test_exit(mm))) {
2658                         progress++;
2659                         break;
2660                 }
2661                 if (!hugepage_vma_check(vma)) {
2662 skip:
2663                         progress++;
2664                         continue;
2665                 }
2666                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2667                 hend = vma->vm_end & HPAGE_PMD_MASK;
2668                 if (hstart >= hend)
2669                         goto skip;
2670                 if (khugepaged_scan.address > hend)
2671                         goto skip;
2672                 if (khugepaged_scan.address < hstart)
2673                         khugepaged_scan.address = hstart;
2674                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2675
2676                 while (khugepaged_scan.address < hend) {
2677                         int ret;
2678                         cond_resched();
2679                         if (unlikely(khugepaged_test_exit(mm)))
2680                                 goto breakouterloop;
2681
2682                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2683                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2684                                   hend);
2685                         ret = khugepaged_scan_pmd(mm, vma,
2686                                                   khugepaged_scan.address,
2687                                                   hpage);
2688                         /* move to next address */
2689                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2690                         progress += HPAGE_PMD_NR;
2691                         if (ret)
2692                                 /* we released mmap_sem so break loop */
2693                                 goto breakouterloop_mmap_sem;
2694                         if (progress >= pages)
2695                                 goto breakouterloop;
2696                 }
2697         }
2698 breakouterloop:
2699         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2700 breakouterloop_mmap_sem:
2701
2702         spin_lock(&khugepaged_mm_lock);
2703         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2704         /*
2705          * Release the current mm_slot if this mm is about to die, or
2706          * if we scanned all vmas of this mm.
2707          */
2708         if (khugepaged_test_exit(mm) || !vma) {
2709                 /*
2710                  * Make sure that if mm_users is reaching zero while
2711                  * khugepaged runs here, khugepaged_exit will find
2712                  * mm_slot not pointing to the exiting mm.
2713                  */
2714                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2715                         khugepaged_scan.mm_slot = list_entry(
2716                                 mm_slot->mm_node.next,
2717                                 struct mm_slot, mm_node);
2718                         khugepaged_scan.address = 0;
2719                 } else {
2720                         khugepaged_scan.mm_slot = NULL;
2721                         khugepaged_full_scans++;
2722                 }
2723
2724                 collect_mm_slot(mm_slot);
2725         }
2726
2727         return progress;
2728 }
2729
2730 static int khugepaged_has_work(void)
2731 {
2732         return !list_empty(&khugepaged_scan.mm_head) &&
2733                 khugepaged_enabled();
2734 }
2735
2736 static int khugepaged_wait_event(void)
2737 {
2738         return !list_empty(&khugepaged_scan.mm_head) ||
2739                 kthread_should_stop();
2740 }
2741
2742 static void khugepaged_do_scan(void)
2743 {
2744         struct page *hpage = NULL;
2745         unsigned int progress = 0, pass_through_head = 0;
2746         unsigned int pages = khugepaged_pages_to_scan;
2747         bool wait = true;
2748
2749         barrier(); /* write khugepaged_pages_to_scan to local stack */
2750
2751         while (progress < pages) {
2752                 if (!khugepaged_prealloc_page(&hpage, &wait))
2753                         break;
2754
2755                 cond_resched();
2756
2757                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2758                         break;
2759
2760                 spin_lock(&khugepaged_mm_lock);
2761                 if (!khugepaged_scan.mm_slot)
2762                         pass_through_head++;
2763                 if (khugepaged_has_work() &&
2764                     pass_through_head < 2)
2765                         progress += khugepaged_scan_mm_slot(pages - progress,
2766                                                             &hpage);
2767                 else
2768                         progress = pages;
2769                 spin_unlock(&khugepaged_mm_lock);
2770         }
2771
2772         if (!IS_ERR_OR_NULL(hpage))
2773                 put_page(hpage);
2774 }
2775
2776 static void khugepaged_wait_work(void)
2777 {
2778         if (khugepaged_has_work()) {
2779                 if (!khugepaged_scan_sleep_millisecs)
2780                         return;
2781
2782                 wait_event_freezable_timeout(khugepaged_wait,
2783                                              kthread_should_stop(),
2784                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2785                 return;
2786         }
2787
2788         if (khugepaged_enabled())
2789                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2790 }
2791
2792 static int khugepaged(void *none)
2793 {
2794         struct mm_slot *mm_slot;
2795
2796         set_freezable();
2797         set_user_nice(current, MAX_NICE);
2798
2799         while (!kthread_should_stop()) {
2800                 khugepaged_do_scan();
2801                 khugepaged_wait_work();
2802         }
2803
2804         spin_lock(&khugepaged_mm_lock);
2805         mm_slot = khugepaged_scan.mm_slot;
2806         khugepaged_scan.mm_slot = NULL;
2807         if (mm_slot)
2808                 collect_mm_slot(mm_slot);
2809         spin_unlock(&khugepaged_mm_lock);
2810         return 0;
2811 }
2812
2813 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2814                 unsigned long haddr, pmd_t *pmd)
2815 {
2816         struct mm_struct *mm = vma->vm_mm;
2817         pgtable_t pgtable;
2818         pmd_t _pmd;
2819         int i;
2820
2821         /* leave pmd empty until pte is filled */
2822         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2823
2824         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2825         pmd_populate(mm, &_pmd, pgtable);
2826
2827         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2828                 pte_t *pte, entry;
2829                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2830                 entry = pte_mkspecial(entry);
2831                 pte = pte_offset_map(&_pmd, haddr);
2832                 VM_BUG_ON(!pte_none(*pte));
2833                 set_pte_at(mm, haddr, pte, entry);
2834                 pte_unmap(pte);
2835         }
2836         smp_wmb(); /* make pte visible before pmd */
2837         pmd_populate(mm, pmd, pgtable);
2838         put_huge_zero_page();
2839 }
2840
2841 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2842                 unsigned long haddr, bool freeze)
2843 {
2844         struct mm_struct *mm = vma->vm_mm;
2845         struct page *page;
2846         pgtable_t pgtable;
2847         pmd_t _pmd;
2848         bool young, write, dirty;
2849         int i;
2850
2851         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2852         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2853         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2854         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2855
2856         count_vm_event(THP_SPLIT_PMD);
2857
2858         if (vma_is_dax(vma)) {
2859                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2860                 if (is_huge_zero_pmd(_pmd))
2861                         put_huge_zero_page();
2862                 return;
2863         } else if (is_huge_zero_pmd(*pmd)) {
2864                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2865         }
2866
2867         page = pmd_page(*pmd);
2868         VM_BUG_ON_PAGE(!page_count(page), page);
2869         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2870         write = pmd_write(*pmd);
2871         young = pmd_young(*pmd);
2872         dirty = pmd_dirty(*pmd);
2873
2874         pmdp_huge_split_prepare(vma, haddr, pmd);
2875         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2876         pmd_populate(mm, &_pmd, pgtable);
2877
2878         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2879                 pte_t entry, *pte;
2880                 /*
2881                  * Note that NUMA hinting access restrictions are not
2882                  * transferred to avoid any possibility of altering
2883                  * permissions across VMAs.
2884                  */
2885                 if (freeze) {
2886                         swp_entry_t swp_entry;
2887                         swp_entry = make_migration_entry(page + i, write);
2888                         entry = swp_entry_to_pte(swp_entry);
2889                 } else {
2890                         entry = mk_pte(page + i, vma->vm_page_prot);
2891                         entry = maybe_mkwrite(entry, vma);
2892                         if (!write)
2893                                 entry = pte_wrprotect(entry);
2894                         if (!young)
2895                                 entry = pte_mkold(entry);
2896                 }
2897                 if (dirty)
2898                         SetPageDirty(page + i);
2899                 pte = pte_offset_map(&_pmd, haddr);
2900                 BUG_ON(!pte_none(*pte));
2901                 set_pte_at(mm, haddr, pte, entry);
2902                 atomic_inc(&page[i]._mapcount);
2903                 pte_unmap(pte);
2904         }
2905
2906         /*
2907          * Set PG_double_map before dropping compound_mapcount to avoid
2908          * false-negative page_mapped().
2909          */
2910         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2911                 for (i = 0; i < HPAGE_PMD_NR; i++)
2912                         atomic_inc(&page[i]._mapcount);
2913         }
2914
2915         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2916                 /* Last compound_mapcount is gone. */
2917                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2918                 if (TestClearPageDoubleMap(page)) {
2919                         /* No need in mapcount reference anymore */
2920                         for (i = 0; i < HPAGE_PMD_NR; i++)
2921                                 atomic_dec(&page[i]._mapcount);
2922                 }
2923         }
2924
2925         smp_wmb(); /* make pte visible before pmd */
2926         /*
2927          * Up to this point the pmd is present and huge and userland has the
2928          * whole access to the hugepage during the split (which happens in
2929          * place). If we overwrite the pmd with the not-huge version pointing
2930          * to the pte here (which of course we could if all CPUs were bug
2931          * free), userland could trigger a small page size TLB miss on the
2932          * small sized TLB while the hugepage TLB entry is still established in
2933          * the huge TLB. Some CPU doesn't like that.
2934          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2935          * 383 on page 93. Intel should be safe but is also warns that it's
2936          * only safe if the permission and cache attributes of the two entries
2937          * loaded in the two TLB is identical (which should be the case here).
2938          * But it is generally safer to never allow small and huge TLB entries
2939          * for the same virtual address to be loaded simultaneously. So instead
2940          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2941          * current pmd notpresent (atomically because here the pmd_trans_huge
2942          * and pmd_trans_splitting must remain set at all times on the pmd
2943          * until the split is complete for this pmd), then we flush the SMP TLB
2944          * and finally we write the non-huge version of the pmd entry with
2945          * pmd_populate.
2946          */
2947         pmdp_invalidate(vma, haddr, pmd);
2948         pmd_populate(mm, pmd, pgtable);
2949
2950         if (freeze) {
2951                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2952                         page_remove_rmap(page + i, false);
2953                         put_page(page + i);
2954                 }
2955         }
2956 }
2957
2958 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2959                 unsigned long address)
2960 {
2961         spinlock_t *ptl;
2962         struct mm_struct *mm = vma->vm_mm;
2963         struct page *page = NULL;
2964         unsigned long haddr = address & HPAGE_PMD_MASK;
2965
2966         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2967         ptl = pmd_lock(mm, pmd);
2968         if (pmd_trans_huge(*pmd)) {
2969                 page = pmd_page(*pmd);
2970                 if (PageMlocked(page))
2971                         get_page(page);
2972                 else
2973                         page = NULL;
2974         } else if (!pmd_devmap(*pmd))
2975                 goto out;
2976         __split_huge_pmd_locked(vma, pmd, haddr, false);
2977 out:
2978         spin_unlock(ptl);
2979         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2980         if (page) {
2981                 lock_page(page);
2982                 munlock_vma_page(page);
2983                 unlock_page(page);
2984                 put_page(page);
2985         }
2986 }
2987
2988 static void split_huge_pmd_address(struct vm_area_struct *vma,
2989                                     unsigned long address)
2990 {
2991         pgd_t *pgd;
2992         pud_t *pud;
2993         pmd_t *pmd;
2994
2995         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2996
2997         pgd = pgd_offset(vma->vm_mm, address);
2998         if (!pgd_present(*pgd))
2999                 return;
3000
3001         pud = pud_offset(pgd, address);
3002         if (!pud_present(*pud))
3003                 return;
3004
3005         pmd = pmd_offset(pud, address);
3006         if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
3007                 return;
3008         /*
3009          * Caller holds the mmap_sem write mode, so a huge pmd cannot
3010          * materialize from under us.
3011          */
3012         split_huge_pmd(vma, pmd, address);
3013 }
3014
3015 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3016                              unsigned long start,
3017                              unsigned long end,
3018                              long adjust_next)
3019 {
3020         /*
3021          * If the new start address isn't hpage aligned and it could
3022          * previously contain an hugepage: check if we need to split
3023          * an huge pmd.
3024          */
3025         if (start & ~HPAGE_PMD_MASK &&
3026             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3027             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3028                 split_huge_pmd_address(vma, start);
3029
3030         /*
3031          * If the new end address isn't hpage aligned and it could
3032          * previously contain an hugepage: check if we need to split
3033          * an huge pmd.
3034          */
3035         if (end & ~HPAGE_PMD_MASK &&
3036             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3037             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3038                 split_huge_pmd_address(vma, end);
3039
3040         /*
3041          * If we're also updating the vma->vm_next->vm_start, if the new
3042          * vm_next->vm_start isn't page aligned and it could previously
3043          * contain an hugepage: check if we need to split an huge pmd.
3044          */
3045         if (adjust_next > 0) {
3046                 struct vm_area_struct *next = vma->vm_next;
3047                 unsigned long nstart = next->vm_start;
3048                 nstart += adjust_next << PAGE_SHIFT;
3049                 if (nstart & ~HPAGE_PMD_MASK &&
3050                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3051                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3052                         split_huge_pmd_address(next, nstart);
3053         }
3054 }
3055
3056 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
3057                 unsigned long address)
3058 {
3059         unsigned long haddr = address & HPAGE_PMD_MASK;
3060         spinlock_t *ptl;
3061         pgd_t *pgd;
3062         pud_t *pud;
3063         pmd_t *pmd;
3064         pte_t *pte;
3065         int i, nr = HPAGE_PMD_NR;
3066
3067         /* Skip pages which doesn't belong to the VMA */
3068         if (address < vma->vm_start) {
3069                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3070                 page += off;
3071                 nr -= off;
3072                 address = vma->vm_start;
3073         }
3074
3075         pgd = pgd_offset(vma->vm_mm, address);
3076         if (!pgd_present(*pgd))
3077                 return;
3078         pud = pud_offset(pgd, address);
3079         if (!pud_present(*pud))
3080                 return;
3081         pmd = pmd_offset(pud, address);
3082         ptl = pmd_lock(vma->vm_mm, pmd);
3083         if (!pmd_present(*pmd)) {
3084                 spin_unlock(ptl);
3085                 return;
3086         }
3087         if (pmd_trans_huge(*pmd)) {
3088                 if (page == pmd_page(*pmd))
3089                         __split_huge_pmd_locked(vma, pmd, haddr, true);
3090                 spin_unlock(ptl);
3091                 return;
3092         }
3093         spin_unlock(ptl);
3094
3095         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3096         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3097                 pte_t entry, swp_pte;
3098                 swp_entry_t swp_entry;
3099
3100                 /*
3101                  * We've just crossed page table boundary: need to map next one.
3102                  * It can happen if THP was mremaped to non PMD-aligned address.
3103                  */
3104                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3105                         pte_unmap_unlock(pte - 1, ptl);
3106                         pmd = mm_find_pmd(vma->vm_mm, address);
3107                         if (!pmd)
3108                                 return;
3109                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3110                                         address, &ptl);
3111                 }
3112
3113                 if (!pte_present(*pte))
3114                         continue;
3115                 if (page_to_pfn(page) != pte_pfn(*pte))
3116                         continue;
3117                 flush_cache_page(vma, address, page_to_pfn(page));
3118                 entry = ptep_clear_flush(vma, address, pte);
3119                 if (pte_dirty(entry))
3120                         SetPageDirty(page);
3121                 swp_entry = make_migration_entry(page, pte_write(entry));
3122                 swp_pte = swp_entry_to_pte(swp_entry);
3123                 if (pte_soft_dirty(entry))
3124                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
3125                 set_pte_at(vma->vm_mm, address, pte, swp_pte);
3126                 page_remove_rmap(page, false);
3127                 put_page(page);
3128         }
3129         pte_unmap_unlock(pte - 1, ptl);
3130 }
3131
3132 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
3133 {
3134         struct anon_vma_chain *avc;
3135         pgoff_t pgoff = page_to_pgoff(page);
3136
3137         VM_BUG_ON_PAGE(!PageHead(page), page);
3138
3139         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
3140                         pgoff + HPAGE_PMD_NR - 1) {
3141                 unsigned long address = __vma_address(page, avc->vma);
3142
3143                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3144                                 address, address + HPAGE_PMD_SIZE);
3145                 freeze_page_vma(avc->vma, page, address);
3146                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3147                                 address, address + HPAGE_PMD_SIZE);
3148         }
3149 }
3150
3151 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
3152                 unsigned long address)
3153 {
3154         spinlock_t *ptl;
3155         pmd_t *pmd;
3156         pte_t *pte, entry;
3157         swp_entry_t swp_entry;
3158         unsigned long haddr = address & HPAGE_PMD_MASK;
3159         int i, nr = HPAGE_PMD_NR;
3160
3161         /* Skip pages which doesn't belong to the VMA */
3162         if (address < vma->vm_start) {
3163                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3164                 page += off;
3165                 nr -= off;
3166                 address = vma->vm_start;
3167         }
3168
3169         pmd = mm_find_pmd(vma->vm_mm, address);
3170         if (!pmd)
3171                 return;
3172
3173         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3174         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3175                 /*
3176                  * We've just crossed page table boundary: need to map next one.
3177                  * It can happen if THP was mremaped to non-PMD aligned address.
3178                  */
3179                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3180                         pte_unmap_unlock(pte - 1, ptl);
3181                         pmd = mm_find_pmd(vma->vm_mm, address);
3182                         if (!pmd)
3183                                 return;
3184                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3185                                         address, &ptl);
3186                 }
3187
3188                 if (!is_swap_pte(*pte))
3189                         continue;
3190
3191                 swp_entry = pte_to_swp_entry(*pte);
3192                 if (!is_migration_entry(swp_entry))
3193                         continue;
3194                 if (migration_entry_to_page(swp_entry) != page)
3195                         continue;
3196
3197                 get_page(page);
3198                 page_add_anon_rmap(page, vma, address, false);
3199
3200                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3201                 if (PageDirty(page))
3202                         entry = pte_mkdirty(entry);
3203                 if (is_write_migration_entry(swp_entry))
3204                         entry = maybe_mkwrite(entry, vma);
3205
3206                 flush_dcache_page(page);
3207                 set_pte_at(vma->vm_mm, address, pte, entry);
3208
3209                 /* No need to invalidate - it was non-present before */
3210                 update_mmu_cache(vma, address, pte);
3211         }
3212         pte_unmap_unlock(pte - 1, ptl);
3213 }
3214
3215 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3216 {
3217         struct anon_vma_chain *avc;
3218         pgoff_t pgoff = page_to_pgoff(page);
3219
3220         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3221                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3222                 unsigned long address = __vma_address(page, avc->vma);
3223
3224                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3225                                 address, address + HPAGE_PMD_SIZE);
3226                 unfreeze_page_vma(avc->vma, page, address);
3227                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3228                                 address, address + HPAGE_PMD_SIZE);
3229         }
3230 }
3231
3232 static int __split_huge_page_tail(struct page *head, int tail,
3233                 struct lruvec *lruvec, struct list_head *list)
3234 {
3235         int mapcount;
3236         struct page *page_tail = head + tail;
3237
3238         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3239         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3240
3241         /*
3242          * tail_page->_count is zero and not changing from under us. But
3243          * get_page_unless_zero() may be running from under us on the
3244          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3245          * would then run atomic_set() concurrently with
3246          * get_page_unless_zero(), and atomic_set() is implemented in C not
3247          * using locked ops. spin_unlock on x86 sometime uses locked ops
3248          * because of PPro errata 66, 92, so unless somebody can guarantee
3249          * atomic_set() here would be safe on all archs (and not only on x86),
3250          * it's safer to use atomic_add().
3251          */
3252         atomic_add(mapcount + 1, &page_tail->_count);
3253
3254
3255         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3256         page_tail->flags |= (head->flags &
3257                         ((1L << PG_referenced) |
3258                          (1L << PG_swapbacked) |
3259                          (1L << PG_mlocked) |
3260                          (1L << PG_uptodate) |
3261                          (1L << PG_active) |
3262                          (1L << PG_locked) |
3263                          (1L << PG_unevictable) |
3264                          (1L << PG_dirty)));
3265
3266         /*
3267          * After clearing PageTail the gup refcount can be released.
3268          * Page flags also must be visible before we make the page non-compound.
3269          */
3270         smp_wmb();
3271
3272         clear_compound_head(page_tail);
3273
3274         if (page_is_young(head))
3275                 set_page_young(page_tail);
3276         if (page_is_idle(head))
3277                 set_page_idle(page_tail);
3278
3279         /* ->mapping in first tail page is compound_mapcount */
3280         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3281                         page_tail);
3282         page_tail->mapping = head->mapping;
3283
3284         page_tail->index = head->index + tail;
3285         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3286         lru_add_page_tail(head, page_tail, lruvec, list);
3287
3288         return mapcount;
3289 }
3290
3291 static void __split_huge_page(struct page *page, struct list_head *list)
3292 {
3293         struct page *head = compound_head(page);
3294         struct zone *zone = page_zone(head);
3295         struct lruvec *lruvec;
3296         int i, tail_mapcount;
3297
3298         /* prevent PageLRU to go away from under us, and freeze lru stats */
3299         spin_lock_irq(&zone->lru_lock);
3300         lruvec = mem_cgroup_page_lruvec(head, zone);
3301
3302         /* complete memcg works before add pages to LRU */
3303         mem_cgroup_split_huge_fixup(head);
3304
3305         tail_mapcount = 0;
3306         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3307                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3308         atomic_sub(tail_mapcount, &head->_count);
3309
3310         ClearPageCompound(head);
3311         spin_unlock_irq(&zone->lru_lock);
3312
3313         unfreeze_page(page_anon_vma(head), head);
3314
3315         for (i = 0; i < HPAGE_PMD_NR; i++) {
3316                 struct page *subpage = head + i;
3317                 if (subpage == page)
3318                         continue;
3319                 unlock_page(subpage);
3320
3321                 /*
3322                  * Subpages may be freed if there wasn't any mapping
3323                  * like if add_to_swap() is running on a lru page that
3324                  * had its mapping zapped. And freeing these pages
3325                  * requires taking the lru_lock so we do the put_page
3326                  * of the tail pages after the split is complete.
3327                  */
3328                 put_page(subpage);
3329         }
3330 }
3331
3332 int total_mapcount(struct page *page)
3333 {
3334         int i, ret;
3335
3336         VM_BUG_ON_PAGE(PageTail(page), page);
3337
3338         if (likely(!PageCompound(page)))
3339                 return atomic_read(&page->_mapcount) + 1;
3340
3341         ret = compound_mapcount(page);
3342         if (PageHuge(page))
3343                 return ret;
3344         for (i = 0; i < HPAGE_PMD_NR; i++)
3345                 ret += atomic_read(&page[i]._mapcount) + 1;
3346         if (PageDoubleMap(page))
3347                 ret -= HPAGE_PMD_NR;
3348         return ret;
3349 }
3350
3351 /*
3352  * This function splits huge page into normal pages. @page can point to any
3353  * subpage of huge page to split. Split doesn't change the position of @page.
3354  *
3355  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3356  * The huge page must be locked.
3357  *
3358  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3359  *
3360  * Both head page and tail pages will inherit mapping, flags, and so on from
3361  * the hugepage.
3362  *
3363  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3364  * they are not mapped.
3365  *
3366  * Returns 0 if the hugepage is split successfully.
3367  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3368  * us.
3369  */
3370 int split_huge_page_to_list(struct page *page, struct list_head *list)
3371 {
3372         struct page *head = compound_head(page);
3373         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3374         struct anon_vma *anon_vma;
3375         int count, mapcount, ret;
3376         bool mlocked;
3377         unsigned long flags;
3378
3379         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3380         VM_BUG_ON_PAGE(!PageAnon(page), page);
3381         VM_BUG_ON_PAGE(!PageLocked(page), page);
3382         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3383         VM_BUG_ON_PAGE(!PageCompound(page), page);
3384
3385         /*
3386          * The caller does not necessarily hold an mmap_sem that would prevent
3387          * the anon_vma disappearing so we first we take a reference to it
3388          * and then lock the anon_vma for write. This is similar to
3389          * page_lock_anon_vma_read except the write lock is taken to serialise
3390          * against parallel split or collapse operations.
3391          */
3392         anon_vma = page_get_anon_vma(head);
3393         if (!anon_vma) {
3394                 ret = -EBUSY;
3395                 goto out;
3396         }
3397         anon_vma_lock_write(anon_vma);
3398
3399         /*
3400          * Racy check if we can split the page, before freeze_page() will
3401          * split PMDs
3402          */
3403         if (total_mapcount(head) != page_count(head) - 1) {
3404                 ret = -EBUSY;
3405                 goto out_unlock;
3406         }
3407
3408         mlocked = PageMlocked(page);
3409         freeze_page(anon_vma, head);
3410         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3411
3412         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3413         if (mlocked)
3414                 lru_add_drain();
3415
3416         /* Prevent deferred_split_scan() touching ->_count */
3417         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3418         count = page_count(head);
3419         mapcount = total_mapcount(head);
3420         if (!mapcount && count == 1) {
3421                 if (!list_empty(page_deferred_list(head))) {
3422                         pgdata->split_queue_len--;
3423                         list_del(page_deferred_list(head));
3424                 }
3425                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3426                 __split_huge_page(page, list);
3427                 ret = 0;
3428         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3429                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3430                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3431                                 mapcount, count);
3432                 if (PageTail(page))
3433                         dump_page(head, NULL);
3434                 dump_page(page, "total_mapcount(head) > 0");
3435                 BUG();
3436         } else {
3437                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3438                 unfreeze_page(anon_vma, head);
3439                 ret = -EBUSY;
3440         }
3441
3442 out_unlock:
3443         anon_vma_unlock_write(anon_vma);
3444         put_anon_vma(anon_vma);
3445 out:
3446         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3447         return ret;
3448 }
3449
3450 void free_transhuge_page(struct page *page)
3451 {
3452         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3453         unsigned long flags;
3454
3455         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3456         if (!list_empty(page_deferred_list(page))) {
3457                 pgdata->split_queue_len--;
3458                 list_del(page_deferred_list(page));
3459         }
3460         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3461         free_compound_page(page);
3462 }
3463
3464 void deferred_split_huge_page(struct page *page)
3465 {
3466         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3467         unsigned long flags;
3468
3469         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3470
3471         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3472         if (list_empty(page_deferred_list(page))) {
3473                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3474                 pgdata->split_queue_len++;
3475         }
3476         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3477 }
3478
3479 static unsigned long deferred_split_count(struct shrinker *shrink,
3480                 struct shrink_control *sc)
3481 {
3482         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3483         return ACCESS_ONCE(pgdata->split_queue_len);
3484 }
3485
3486 static unsigned long deferred_split_scan(struct shrinker *shrink,
3487                 struct shrink_control *sc)
3488 {
3489         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3490         unsigned long flags;
3491         LIST_HEAD(list), *pos, *next;
3492         struct page *page;
3493         int split = 0;
3494
3495         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3496         /* Take pin on all head pages to avoid freeing them under us */
3497         list_for_each_safe(pos, next, &pgdata->split_queue) {
3498                 page = list_entry((void *)pos, struct page, mapping);
3499                 page = compound_head(page);
3500                 if (get_page_unless_zero(page)) {
3501                         list_move(page_deferred_list(page), &list);
3502                 } else {
3503                         /* We lost race with put_compound_page() */
3504                         list_del_init(page_deferred_list(page));
3505                         pgdata->split_queue_len--;
3506                 }
3507                 if (!--sc->nr_to_scan)
3508                         break;
3509         }
3510         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3511
3512         list_for_each_safe(pos, next, &list) {
3513                 page = list_entry((void *)pos, struct page, mapping);
3514                 lock_page(page);
3515                 /* split_huge_page() removes page from list on success */
3516                 if (!split_huge_page(page))
3517                         split++;
3518                 unlock_page(page);
3519                 put_page(page);
3520         }
3521
3522         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3523         list_splice_tail(&list, &pgdata->split_queue);
3524         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3525
3526         /*
3527          * Stop shrinker if we didn't split any page, but the queue is empty.
3528          * This can happen if pages were freed under us.
3529          */
3530         if (!split && list_empty(&pgdata->split_queue))
3531                 return SHRINK_STOP;
3532         return split;
3533 }
3534
3535 static struct shrinker deferred_split_shrinker = {
3536         .count_objects = deferred_split_count,
3537         .scan_objects = deferred_split_scan,
3538         .seeks = DEFAULT_SEEKS,
3539         .flags = SHRINKER_NUMA_AWARE,
3540 };
3541
3542 #ifdef CONFIG_DEBUG_FS
3543 static int split_huge_pages_set(void *data, u64 val)
3544 {
3545         struct zone *zone;
3546         struct page *page;
3547         unsigned long pfn, max_zone_pfn;
3548         unsigned long total = 0, split = 0;
3549
3550         if (val != 1)
3551                 return -EINVAL;
3552
3553         for_each_populated_zone(zone) {
3554                 max_zone_pfn = zone_end_pfn(zone);
3555                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3556                         if (!pfn_valid(pfn))
3557                                 continue;
3558
3559                         page = pfn_to_page(pfn);
3560                         if (!get_page_unless_zero(page))
3561                                 continue;
3562
3563                         if (zone != page_zone(page))
3564                                 goto next;
3565
3566                         if (!PageHead(page) || !PageAnon(page) ||
3567                                         PageHuge(page))
3568                                 goto next;
3569
3570                         total++;
3571                         lock_page(page);
3572                         if (!split_huge_page(page))
3573                                 split++;
3574                         unlock_page(page);
3575 next:
3576                         put_page(page);
3577                 }
3578         }
3579
3580         pr_info("%lu of %lu THP split", split, total);
3581
3582         return 0;
3583 }
3584 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3585                 "%llu\n");
3586
3587 static int __init split_huge_pages_debugfs(void)
3588 {
3589         void *ret;
3590
3591         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3592                         &split_huge_pages_fops);
3593         if (!ret)
3594                 pr_warn("Failed to create split_huge_pages in debugfs");
3595         return 0;
3596 }
3597 late_initcall(split_huge_pages_debugfs);
3598 #endif