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         SCAN_EXCEED_SWAP_PTE
  62 };
  63
  64 #define CREATE_TRACE_POINTS
  65 #include <trace/events/huge_memory.h>
  66
  67 /*
  68  * By default transparent hugepage support is disabled in order that avoid
  69  * to risk increase the memory footprint of applications without a guaranteed
  70  * benefit. When transparent hugepage support is enabled, is for all mappings,
  71  * and khugepaged scans all mappings.
  72  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  73  * for all hugepage allocations.
  74  */
  75 unsigned long transparent_hugepage_flags __read_mostly =
  76 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  77         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  78 #endif
  79 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  80         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  81 #endif
  82         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  83         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  84         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  85
  86 /* default scan 8*512 pte (or vmas) every 30 second */
  87 static unsigned int khugepaged_pages_to_scan __read_mostly;
  88 static unsigned int khugepaged_pages_collapsed;
  89 static unsigned int khugepaged_full_scans;
  90 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  91 /* during fragmentation poll the hugepage allocator once every minute */
  92 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  93 static unsigned long khugepaged_sleep_expire;
  94 static struct task_struct *khugepaged_thread __read_mostly;
  95 static DEFINE_MUTEX(khugepaged_mutex);
  96 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  97 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  98 /*
  99  * default collapse hugepages if there is at least one pte mapped like
 100  * it would have happened if the vma was large enough during page
 101  * fault.
 102  */
 103 static unsigned int khugepaged_max_ptes_none __read_mostly;
 104 static unsigned int khugepaged_max_ptes_swap __read_mostly;
 105
 106 static int khugepaged(void *none);
 107 static int khugepaged_slab_init(void);
 108 static void khugepaged_slab_exit(void);
 109
 110 #define MM_SLOTS_HASH_BITS 10
 111 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 112
 113 static struct kmem_cache *mm_slot_cache __read_mostly;
 114
 115 /**
 116  * struct mm_slot - hash lookup from mm to mm_slot
 117  * @hash: hash collision list
 118  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 119  * @mm: the mm that this information is valid for
 120  */
 121 struct mm_slot {
 122         struct hlist_node hash;
 123         struct list_head mm_node;
 124         struct mm_struct *mm;
 125 };
 126
 127 /**
 128  * struct khugepaged_scan - cursor for scanning
 129  * @mm_head: the head of the mm list to scan
 130  * @mm_slot: the current mm_slot we are scanning
 131  * @address: the next address inside that to be scanned
 132  *
 133  * There is only the one khugepaged_scan instance of this cursor structure.
 134  */
 135 struct khugepaged_scan {
 136         struct list_head mm_head;
 137         struct mm_slot *mm_slot;
 138         unsigned long address;
 139 };
 140 static struct khugepaged_scan khugepaged_scan = {
 141         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 142 };
 143
 144 static struct shrinker deferred_split_shrinker;
 145
 146 static void set_recommended_min_free_kbytes(void)
 147 {
 148         struct zone *zone;
 149         int nr_zones = 0;
 150         unsigned long recommended_min;
 151
 152         for_each_populated_zone(zone)
 153                 nr_zones++;
 154
 155         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 156         recommended_min = pageblock_nr_pages * nr_zones * 2;
 157
 158         /*
 159          * Make sure that on average at least two pageblocks are almost free
 160          * of another type, one for a migratetype to fall back to and a
 161          * second to avoid subsequent fallbacks of other types There are 3
 162          * MIGRATE_TYPES we care about.
 163          */
 164         recommended_min += pageblock_nr_pages * nr_zones *
 165                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 166
 167         /* don't ever allow to reserve more than 5% of the lowmem */
 168         recommended_min = min(recommended_min,
 169                               (unsigned long) nr_free_buffer_pages() / 20);
 170         recommended_min <<= (PAGE_SHIFT-10);
 171
 172         if (recommended_min > min_free_kbytes) {
 173                 if (user_min_free_kbytes >= 0)
 174                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
 175                                 min_free_kbytes, recommended_min);
 176
 177                 min_free_kbytes = recommended_min;
 178         }
 179         setup_per_zone_wmarks();
 180 }
 181
 182 static int start_stop_khugepaged(void)
 183 {
 184         int err = 0;
 185         if (khugepaged_enabled()) {
 186                 if (!khugepaged_thread)
 187                         khugepaged_thread = kthread_run(khugepaged, NULL,
 188                                                         "khugepaged");
 189                 if (IS_ERR(khugepaged_thread)) {
 190                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 191                         err = PTR_ERR(khugepaged_thread);
 192                         khugepaged_thread = NULL;
 193                         goto fail;
 194                 }
 195
 196                 if (!list_empty(&khugepaged_scan.mm_head))
 197                         wake_up_interruptible(&khugepaged_wait);
 198
 199                 set_recommended_min_free_kbytes();
 200         } else if (khugepaged_thread) {
 201                 kthread_stop(khugepaged_thread);
 202                 khugepaged_thread = NULL;
 203         }
 204 fail:
 205         return err;
 206 }
 207
 208 static atomic_t huge_zero_refcount;
 209 struct page *huge_zero_page __read_mostly;
 210
 211 struct page *get_huge_zero_page(void)
 212 {
 213         struct page *zero_page;
 214 retry:
 215         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 216                 return READ_ONCE(huge_zero_page);
 217
 218         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 219                         HPAGE_PMD_ORDER);
 220         if (!zero_page) {
 221                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 222                 return NULL;
 223         }
 224         count_vm_event(THP_ZERO_PAGE_ALLOC);
 225         preempt_disable();
 226         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 227                 preempt_enable();
 228                 __free_pages(zero_page, compound_order(zero_page));
 229                 goto retry;
 230         }
 231
 232         /* We take additional reference here. It will be put back by shrinker */
 233         atomic_set(&huge_zero_refcount, 2);
 234         preempt_enable();
 235         return READ_ONCE(huge_zero_page);
 236 }
 237
 238 void put_huge_zero_page(void)
 239 {
 240         /*
 241          * Counter should never go to zero here. Only shrinker can put
 242          * last reference.
 243          */
 244         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 245 }
 246
 247 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 248                                         struct shrink_control *sc)
 249 {
 250         /* we can free zero page only if last reference remains */
 251         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 252 }
 253
 254 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 255                                        struct shrink_control *sc)
 256 {
 257         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 258                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 259                 BUG_ON(zero_page == NULL);
 260                 __free_pages(zero_page, compound_order(zero_page));
 261                 return HPAGE_PMD_NR;
 262         }
 263
 264         return 0;
 265 }
 266
 267 static struct shrinker huge_zero_page_shrinker = {
 268         .count_objects = shrink_huge_zero_page_count,
 269         .scan_objects = shrink_huge_zero_page_scan,
 270         .seeks = DEFAULT_SEEKS,
 271 };
 272
 273 #ifdef CONFIG_SYSFS
 274
 275 static ssize_t triple_flag_store(struct kobject *kobj,
 276                                  struct kobj_attribute *attr,
 277                                  const char *buf, size_t count,
 278                                  enum transparent_hugepage_flag enabled,
 279                                  enum transparent_hugepage_flag deferred,
 280                                  enum transparent_hugepage_flag req_madv)
 281 {
 282         if (!memcmp("defer", buf,
 283                     min(sizeof("defer")-1, count))) {
 284                 if (enabled == deferred)
 285                         return -EINVAL;
 286                 clear_bit(enabled, &transparent_hugepage_flags);
 287                 clear_bit(req_madv, &transparent_hugepage_flags);
 288                 set_bit(deferred, &transparent_hugepage_flags);
 289         } else if (!memcmp("always", buf,
 290                     min(sizeof("always")-1, count))) {
 291                 clear_bit(deferred, &transparent_hugepage_flags);
 292                 clear_bit(req_madv, &transparent_hugepage_flags);
 293                 set_bit(enabled, &transparent_hugepage_flags);
 294         } else if (!memcmp("madvise", buf,
 295                            min(sizeof("madvise")-1, count))) {
 296                 clear_bit(enabled, &transparent_hugepage_flags);
 297                 clear_bit(deferred, &transparent_hugepage_flags);
 298                 set_bit(req_madv, &transparent_hugepage_flags);
 299         } else if (!memcmp("never", buf,
 300                            min(sizeof("never")-1, count))) {
 301                 clear_bit(enabled, &transparent_hugepage_flags);
 302                 clear_bit(req_madv, &transparent_hugepage_flags);
 303                 clear_bit(deferred, &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         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 314                 return sprintf(buf, "[always] madvise never\n");
 315         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 316                 return sprintf(buf, "always [madvise] never\n");
 317         else
 318                 return sprintf(buf, "always madvise [never]\n");
 319 }
 320
 321 static ssize_t enabled_store(struct kobject *kobj,
 322                              struct kobj_attribute *attr,
 323                              const char *buf, size_t count)
 324 {
 325         ssize_t ret;
 326
 327         ret = triple_flag_store(kobj, attr, buf, count,
 328                                 TRANSPARENT_HUGEPAGE_FLAG,
 329                                 TRANSPARENT_HUGEPAGE_FLAG,
 330                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 331
 332         if (ret > 0) {
 333                 int err;
 334
 335                 mutex_lock(&khugepaged_mutex);
 336                 err = start_stop_khugepaged();
 337                 mutex_unlock(&khugepaged_mutex);
 338
 339                 if (err)
 340                         ret = err;
 341         }
 342
 343         return ret;
 344 }
 345 static struct kobj_attribute enabled_attr =
 346         __ATTR(enabled, 0644, enabled_show, enabled_store);
 347
 348 static ssize_t single_flag_show(struct kobject *kobj,
 349                                 struct kobj_attribute *attr, char *buf,
 350                                 enum transparent_hugepage_flag flag)
 351 {
 352         return sprintf(buf, "%d\n",
 353                        !!test_bit(flag, &transparent_hugepage_flags));
 354 }
 355
 356 static ssize_t single_flag_store(struct kobject *kobj,
 357                                  struct kobj_attribute *attr,
 358                                  const char *buf, size_t count,
 359                                  enum transparent_hugepage_flag flag)
 360 {
 361         unsigned long value;
 362         int ret;
 363
 364         ret = kstrtoul(buf, 10, &value);
 365         if (ret < 0)
 366                 return ret;
 367         if (value > 1)
 368                 return -EINVAL;
 369
 370         if (value)
 371                 set_bit(flag, &transparent_hugepage_flags);
 372         else
 373                 clear_bit(flag, &transparent_hugepage_flags);
 374
 375         return count;
 376 }
 377
 378 /*
 379  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 380  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 381  * memory just to allocate one more hugepage.
 382  */
 383 static ssize_t defrag_show(struct kobject *kobj,
 384                            struct kobj_attribute *attr, char *buf)
 385 {
 386         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 387                 return sprintf(buf, "[always] defer madvise never\n");
 388         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 389                 return sprintf(buf, "always [defer] madvise never\n");
 390         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 391                 return sprintf(buf, "always defer [madvise] never\n");
 392         else
 393                 return sprintf(buf, "always defer madvise [never]\n");
 394
 395 }
 396 static ssize_t defrag_store(struct kobject *kobj,
 397                             struct kobj_attribute *attr,
 398                             const char *buf, size_t count)
 399 {
 400         return triple_flag_store(kobj, attr, buf, count,
 401                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 402                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 403                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 404 }
 405 static struct kobj_attribute defrag_attr =
 406         __ATTR(defrag, 0644, defrag_show, defrag_store);
 407
 408 static ssize_t use_zero_page_show(struct kobject *kobj,
 409                 struct kobj_attribute *attr, char *buf)
 410 {
 411         return single_flag_show(kobj, attr, buf,
 412                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 413 }
 414 static ssize_t use_zero_page_store(struct kobject *kobj,
 415                 struct kobj_attribute *attr, const char *buf, size_t count)
 416 {
 417         return single_flag_store(kobj, attr, buf, count,
 418                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 419 }
 420 static struct kobj_attribute use_zero_page_attr =
 421         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 422 #ifdef CONFIG_DEBUG_VM
 423 static ssize_t debug_cow_show(struct kobject *kobj,
 424                                 struct kobj_attribute *attr, char *buf)
 425 {
 426         return single_flag_show(kobj, attr, buf,
 427                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 428 }
 429 static ssize_t debug_cow_store(struct kobject *kobj,
 430                                struct kobj_attribute *attr,
 431                                const char *buf, size_t count)
 432 {
 433         return single_flag_store(kobj, attr, buf, count,
 434                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 435 }
 436 static struct kobj_attribute debug_cow_attr =
 437         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 438 #endif /* CONFIG_DEBUG_VM */
 439
 440 static struct attribute *hugepage_attr[] = {
 441         &enabled_attr.attr,
 442         &defrag_attr.attr,
 443         &use_zero_page_attr.attr,
 444 #ifdef CONFIG_DEBUG_VM
 445         &debug_cow_attr.attr,
 446 #endif
 447         NULL,
 448 };
 449
 450 static struct attribute_group hugepage_attr_group = {
 451         .attrs = hugepage_attr,
 452 };
 453
 454 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 455                                          struct kobj_attribute *attr,
 456                                          char *buf)
 457 {
 458         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 459 }
 460
 461 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 462                                           struct kobj_attribute *attr,
 463                                           const char *buf, size_t count)
 464 {
 465         unsigned long msecs;
 466         int err;
 467
 468         err = kstrtoul(buf, 10, &msecs);
 469         if (err || msecs > UINT_MAX)
 470                 return -EINVAL;
 471
 472         khugepaged_scan_sleep_millisecs = msecs;
 473         khugepaged_sleep_expire = 0;
 474         wake_up_interruptible(&khugepaged_wait);
 475
 476         return count;
 477 }
 478 static struct kobj_attribute scan_sleep_millisecs_attr =
 479         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 480                scan_sleep_millisecs_store);
 481
 482 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 483                                           struct kobj_attribute *attr,
 484                                           char *buf)
 485 {
 486         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 487 }
 488
 489 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 490                                            struct kobj_attribute *attr,
 491                                            const char *buf, size_t count)
 492 {
 493         unsigned long msecs;
 494         int err;
 495
 496         err = kstrtoul(buf, 10, &msecs);
 497         if (err || msecs > UINT_MAX)
 498                 return -EINVAL;
 499
 500         khugepaged_alloc_sleep_millisecs = msecs;
 501         khugepaged_sleep_expire = 0;
 502         wake_up_interruptible(&khugepaged_wait);
 503
 504         return count;
 505 }
 506 static struct kobj_attribute alloc_sleep_millisecs_attr =
 507         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 508                alloc_sleep_millisecs_store);
 509
 510 static ssize_t pages_to_scan_show(struct kobject *kobj,
 511                                   struct kobj_attribute *attr,
 512                                   char *buf)
 513 {
 514         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 515 }
 516 static ssize_t pages_to_scan_store(struct kobject *kobj,
 517                                    struct kobj_attribute *attr,
 518                                    const char *buf, size_t count)
 519 {
 520         int err;
 521         unsigned long pages;
 522
 523         err = kstrtoul(buf, 10, &pages);
 524         if (err || !pages || pages > UINT_MAX)
 525                 return -EINVAL;
 526
 527         khugepaged_pages_to_scan = pages;
 528
 529         return count;
 530 }
 531 static struct kobj_attribute pages_to_scan_attr =
 532         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 533                pages_to_scan_store);
 534
 535 static ssize_t pages_collapsed_show(struct kobject *kobj,
 536                                     struct kobj_attribute *attr,
 537                                     char *buf)
 538 {
 539         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 540 }
 541 static struct kobj_attribute pages_collapsed_attr =
 542         __ATTR_RO(pages_collapsed);
 543
 544 static ssize_t full_scans_show(struct kobject *kobj,
 545                                struct kobj_attribute *attr,
 546                                char *buf)
 547 {
 548         return sprintf(buf, "%u\n", khugepaged_full_scans);
 549 }
 550 static struct kobj_attribute full_scans_attr =
 551         __ATTR_RO(full_scans);
 552
 553 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 554                                       struct kobj_attribute *attr, char *buf)
 555 {
 556         return single_flag_show(kobj, attr, buf,
 557                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 558 }
 559 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 560                                        struct kobj_attribute *attr,
 561                                        const char *buf, size_t count)
 562 {
 563         return single_flag_store(kobj, attr, buf, count,
 564                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 565 }
 566 static struct kobj_attribute khugepaged_defrag_attr =
 567         __ATTR(defrag, 0644, khugepaged_defrag_show,
 568                khugepaged_defrag_store);
 569
 570 /*
 571  * max_ptes_none controls if khugepaged should collapse hugepages over
 572  * any unmapped ptes in turn potentially increasing the memory
 573  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 574  * reduce the available free memory in the system as it
 575  * runs. Increasing max_ptes_none will instead potentially reduce the
 576  * free memory in the system during the khugepaged scan.
 577  */
 578 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 579                                              struct kobj_attribute *attr,
 580                                              char *buf)
 581 {
 582         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 583 }
 584 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 585                                               struct kobj_attribute *attr,
 586                                               const char *buf, size_t count)
 587 {
 588         int err;
 589         unsigned long max_ptes_none;
 590
 591         err = kstrtoul(buf, 10, &max_ptes_none);
 592         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 593                 return -EINVAL;
 594
 595         khugepaged_max_ptes_none = max_ptes_none;
 596
 597         return count;
 598 }
 599 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 600         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 601                khugepaged_max_ptes_none_store);
 602
 603 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 604                                              struct kobj_attribute *attr,
 605                                              char *buf)
 606 {
 607         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 608 }
 609
 610 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 611                                               struct kobj_attribute *attr,
 612                                               const char *buf, size_t count)
 613 {
 614         int err;
 615         unsigned long max_ptes_swap;
 616
 617         err  = kstrtoul(buf, 10, &max_ptes_swap);
 618         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 619                 return -EINVAL;
 620
 621         khugepaged_max_ptes_swap = max_ptes_swap;
 622
 623         return count;
 624 }
 625
 626 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 627         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 628                khugepaged_max_ptes_swap_store);
 629
 630 static struct attribute *khugepaged_attr[] = {
 631         &khugepaged_defrag_attr.attr,
 632         &khugepaged_max_ptes_none_attr.attr,
 633         &pages_to_scan_attr.attr,
 634         &pages_collapsed_attr.attr,
 635         &full_scans_attr.attr,
 636         &scan_sleep_millisecs_attr.attr,
 637         &alloc_sleep_millisecs_attr.attr,
 638         &khugepaged_max_ptes_swap_attr.attr,
 639         NULL,
 640 };
 641
 642 static struct attribute_group khugepaged_attr_group = {
 643         .attrs = khugepaged_attr,
 644         .name = "khugepaged",
 645 };
 646
 647 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 648 {
 649         int err;
 650
 651         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 652         if (unlikely(!*hugepage_kobj)) {
 653                 pr_err("failed to create transparent hugepage kobject\n");
 654                 return -ENOMEM;
 655         }
 656
 657         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 658         if (err) {
 659                 pr_err("failed to register transparent hugepage group\n");
 660                 goto delete_obj;
 661         }
 662
 663         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 664         if (err) {
 665                 pr_err("failed to register transparent hugepage group\n");
 666                 goto remove_hp_group;
 667         }
 668
 669         return 0;
 670
 671 remove_hp_group:
 672         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 673 delete_obj:
 674         kobject_put(*hugepage_kobj);
 675         return err;
 676 }
 677
 678 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 679 {
 680         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 681         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 682         kobject_put(hugepage_kobj);
 683 }
 684 #else
 685 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 686 {
 687         return 0;
 688 }
 689
 690 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 691 {
 692 }
 693 #endif /* CONFIG_SYSFS */
 694
 695 static int __init hugepage_init(void)
 696 {
 697         int err;
 698         struct kobject *hugepage_kobj;
 699
 700         if (!has_transparent_hugepage()) {
 701                 transparent_hugepage_flags = 0;
 702                 return -EINVAL;
 703         }
 704
 705         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 706         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 707         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 708         /*
 709          * hugepages can't be allocated by the buddy allocator
 710          */
 711         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 712         /*
 713          * we use page->mapping and page->index in second tail page
 714          * as list_head: assuming THP order >= 2
 715          */
 716         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 717
 718         err = hugepage_init_sysfs(&hugepage_kobj);
 719         if (err)
 720                 goto err_sysfs;
 721
 722         err = khugepaged_slab_init();
 723         if (err)
 724                 goto err_slab;
 725
 726         err = register_shrinker(&huge_zero_page_shrinker);
 727         if (err)
 728                 goto err_hzp_shrinker;
 729         err = register_shrinker(&deferred_split_shrinker);
 730         if (err)
 731                 goto err_split_shrinker;
 732
 733         /*
 734          * By default disable transparent hugepages on smaller systems,
 735          * where the extra memory used could hurt more than TLB overhead
 736          * is likely to save.  The admin can still enable it through /sys.
 737          */
 738         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 739                 transparent_hugepage_flags = 0;
 740                 return 0;
 741         }
 742
 743         err = start_stop_khugepaged();
 744         if (err)
 745                 goto err_khugepaged;
 746
 747         return 0;
 748 err_khugepaged:
 749         unregister_shrinker(&deferred_split_shrinker);
 750 err_split_shrinker:
 751         unregister_shrinker(&huge_zero_page_shrinker);
 752 err_hzp_shrinker:
 753         khugepaged_slab_exit();
 754 err_slab:
 755         hugepage_exit_sysfs(hugepage_kobj);
 756 err_sysfs:
 757         return err;
 758 }
 759 subsys_initcall(hugepage_init);
 760
 761 static int __init setup_transparent_hugepage(char *str)
 762 {
 763         int ret = 0;
 764         if (!str)
 765                 goto out;
 766         if (!strcmp(str, "always")) {
 767                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 768                         &transparent_hugepage_flags);
 769                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 770                           &transparent_hugepage_flags);
 771                 ret = 1;
 772         } else if (!strcmp(str, "madvise")) {
 773                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 774                           &transparent_hugepage_flags);
 775                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 776                         &transparent_hugepage_flags);
 777                 ret = 1;
 778         } else if (!strcmp(str, "never")) {
 779                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 780                           &transparent_hugepage_flags);
 781                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 782                           &transparent_hugepage_flags);
 783                 ret = 1;
 784         }
 785 out:
 786         if (!ret)
 787                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 788         return ret;
 789 }
 790 __setup("transparent_hugepage=", setup_transparent_hugepage);
 791
 792 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 793 {
 794         if (likely(vma->vm_flags & VM_WRITE))
 795                 pmd = pmd_mkwrite(pmd);
 796         return pmd;
 797 }
 798
 799 static inline struct list_head *page_deferred_list(struct page *page)
 800 {
 801         /*
 802          * ->lru in the tail pages is occupied by compound_head.
 803          * Let's use ->mapping + ->index in the second tail page as list_head.
 804          */
 805         return (struct list_head *)&page[2].mapping;
 806 }
 807
 808 void prep_transhuge_page(struct page *page)
 809 {
 810         /*
 811          * we use page->mapping and page->indexlru in second tail page
 812          * as list_head: assuming THP order >= 2
 813          */
 814
 815         INIT_LIST_HEAD(page_deferred_list(page));
 816         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 817 }
 818
 819 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
 820                 gfp_t gfp)
 821 {
 822         struct vm_area_struct *vma = fe->vma;
 823         struct mem_cgroup *memcg;
 824         pgtable_t pgtable;
 825         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 826
 827         VM_BUG_ON_PAGE(!PageCompound(page), page);
 828
 829         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 830                 put_page(page);
 831                 count_vm_event(THP_FAULT_FALLBACK);
 832                 return VM_FAULT_FALLBACK;
 833         }
 834
 835         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 836         if (unlikely(!pgtable)) {
 837                 mem_cgroup_cancel_charge(page, memcg, true);
 838                 put_page(page);
 839                 return VM_FAULT_OOM;
 840         }
 841
 842         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 843         /*
 844          * The memory barrier inside __SetPageUptodate makes sure that
 845          * clear_huge_page writes become visible before the set_pmd_at()
 846          * write.
 847          */
 848         __SetPageUptodate(page);
 849
 850         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 851         if (unlikely(!pmd_none(*fe->pmd))) {
 852                 spin_unlock(fe->ptl);
 853                 mem_cgroup_cancel_charge(page, memcg, true);
 854                 put_page(page);
 855                 pte_free(vma->vm_mm, pgtable);
 856         } else {
 857                 pmd_t entry;
 858
 859                 /* Deliver the page fault to userland */
 860                 if (userfaultfd_missing(vma)) {
 861                         int ret;
 862
 863                         spin_unlock(fe->ptl);
 864                         mem_cgroup_cancel_charge(page, memcg, true);
 865                         put_page(page);
 866                         pte_free(vma->vm_mm, pgtable);
 867                         ret = handle_userfault(fe, VM_UFFD_MISSING);
 868                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 869                         return ret;
 870                 }
 871
 872                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 873                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 874                 page_add_new_anon_rmap(page, vma, haddr, true);
 875                 mem_cgroup_commit_charge(page, memcg, false, true);
 876                 lru_cache_add_active_or_unevictable(page, vma);
 877                 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
 878                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
 879                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 880                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 881                 spin_unlock(fe->ptl);
 882                 count_vm_event(THP_FAULT_ALLOC);
 883         }
 884
 885         return 0;
 886 }
 887
 888 /*
 889  * If THP is set to always then directly reclaim/compact as necessary
 890  * If set to defer then do no reclaim and defer to khugepaged
 891  * If set to madvise and the VMA is flagged then directly reclaim/compact
 892  */
 893 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 894 {
 895         gfp_t reclaim_flags = 0;
 896
 897         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags) &&
 898             (vma->vm_flags & VM_HUGEPAGE))
 899                 reclaim_flags = __GFP_DIRECT_RECLAIM;
 900         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 901                 reclaim_flags = __GFP_KSWAPD_RECLAIM;
 902         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 903                 reclaim_flags = __GFP_DIRECT_RECLAIM;
 904
 905         return GFP_TRANSHUGE | reclaim_flags;
 906 }
 907
 908 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 909 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 910 {
 911         return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
 912 }
 913
 914 /* Caller must hold page table lock. */
 915 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 916                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 917                 struct page *zero_page)
 918 {
 919         pmd_t entry;
 920         if (!pmd_none(*pmd))
 921                 return false;
 922         entry = mk_pmd(zero_page, vma->vm_page_prot);
 923         entry = pmd_mkhuge(entry);
 924         if (pgtable)
 925                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 926         set_pmd_at(mm, haddr, pmd, entry);
 927         atomic_long_inc(&mm->nr_ptes);
 928         return true;
 929 }
 930
 931 int do_huge_pmd_anonymous_page(struct fault_env *fe)
 932 {
 933         struct vm_area_struct *vma = fe->vma;
 934         gfp_t gfp;
 935         struct page *page;
 936         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 937
 938         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 939                 return VM_FAULT_FALLBACK;
 940         if (unlikely(anon_vma_prepare(vma)))
 941                 return VM_FAULT_OOM;
 942         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 943                 return VM_FAULT_OOM;
 944         if (!(fe->flags & FAULT_FLAG_WRITE) &&
 945                         !mm_forbids_zeropage(vma->vm_mm) &&
 946                         transparent_hugepage_use_zero_page()) {
 947                 pgtable_t pgtable;
 948                 struct page *zero_page;
 949                 bool set;
 950                 int ret;
 951                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 952                 if (unlikely(!pgtable))
 953                         return VM_FAULT_OOM;
 954                 zero_page = get_huge_zero_page();
 955                 if (unlikely(!zero_page)) {
 956                         pte_free(vma->vm_mm, pgtable);
 957                         count_vm_event(THP_FAULT_FALLBACK);
 958                         return VM_FAULT_FALLBACK;
 959                 }
 960                 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 961                 ret = 0;
 962                 set = false;
 963                 if (pmd_none(*fe->pmd)) {
 964                         if (userfaultfd_missing(vma)) {
 965                                 spin_unlock(fe->ptl);
 966                                 ret = handle_userfault(fe, VM_UFFD_MISSING);
 967                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 968                         } else {
 969                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 970                                                    haddr, fe->pmd, zero_page);
 971                                 spin_unlock(fe->ptl);
 972                                 set = true;
 973                         }
 974                 } else
 975                         spin_unlock(fe->ptl);
 976                 if (!set) {
 977                         pte_free(vma->vm_mm, pgtable);
 978                         put_huge_zero_page();
 979                 }
 980                 return ret;
 981         }
 982         gfp = alloc_hugepage_direct_gfpmask(vma);
 983         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 984         if (unlikely(!page)) {
 985                 count_vm_event(THP_FAULT_FALLBACK);
 986                 return VM_FAULT_FALLBACK;
 987         }
 988         prep_transhuge_page(page);
 989         return __do_huge_pmd_anonymous_page(fe, page, gfp);
 990 }
 991
 992 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 993                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 994 {
 995         struct mm_struct *mm = vma->vm_mm;
 996         pmd_t entry;
 997         spinlock_t *ptl;
 998
 999         ptl = pmd_lock(mm, pmd);
1000         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
1001         if (pfn_t_devmap(pfn))
1002                 entry = pmd_mkdevmap(entry);
1003         if (write) {
1004                 entry = pmd_mkyoung(pmd_mkdirty(entry));
1005                 entry = maybe_pmd_mkwrite(entry, vma);
1006         }
1007         set_pmd_at(mm, addr, pmd, entry);
1008         update_mmu_cache_pmd(vma, addr, pmd);
1009         spin_unlock(ptl);
1010 }
1011
1012 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
1013                         pmd_t *pmd, pfn_t pfn, bool write)
1014 {
1015         pgprot_t pgprot = vma->vm_page_prot;
1016         /*
1017          * If we had pmd_special, we could avoid all these restrictions,
1018          * but we need to be consistent with PTEs and architectures that
1019          * can't support a 'special' bit.
1020          */
1021         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1022         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1023                                                 (VM_PFNMAP|VM_MIXEDMAP));
1024         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1025         BUG_ON(!pfn_t_devmap(pfn));
1026
1027         if (addr < vma->vm_start || addr >= vma->vm_end)
1028                 return VM_FAULT_SIGBUS;
1029         if (track_pfn_insert(vma, &pgprot, pfn))
1030                 return VM_FAULT_SIGBUS;
1031         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
1032         return VM_FAULT_NOPAGE;
1033 }
1034 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
1035
1036 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1037                 pmd_t *pmd)
1038 {
1039         pmd_t _pmd;
1040
1041         /*
1042          * We should set the dirty bit only for FOLL_WRITE but for now
1043          * the dirty bit in the pmd is meaningless.  And if the dirty
1044          * bit will become meaningful and we'll only set it with
1045          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1046          * set the young bit, instead of the current set_pmd_at.
1047          */
1048         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1049         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1050                                 pmd, _pmd,  1))
1051                 update_mmu_cache_pmd(vma, addr, pmd);
1052 }
1053
1054 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1055                 pmd_t *pmd, int flags)
1056 {
1057         unsigned long pfn = pmd_pfn(*pmd);
1058         struct mm_struct *mm = vma->vm_mm;
1059         struct dev_pagemap *pgmap;
1060         struct page *page;
1061
1062         assert_spin_locked(pmd_lockptr(mm, pmd));
1063
1064         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1065                 return NULL;
1066
1067         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1068                 /* pass */;
1069         else
1070                 return NULL;
1071
1072         if (flags & FOLL_TOUCH)
1073                 touch_pmd(vma, addr, pmd);
1074
1075         /*
1076          * device mapped pages can only be returned if the
1077          * caller will manage the page reference count.
1078          */
1079         if (!(flags & FOLL_GET))
1080                 return ERR_PTR(-EEXIST);
1081
1082         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1083         pgmap = get_dev_pagemap(pfn, NULL);
1084         if (!pgmap)
1085                 return ERR_PTR(-EFAULT);
1086         page = pfn_to_page(pfn);
1087         get_page(page);
1088         put_dev_pagemap(pgmap);
1089
1090         return page;
1091 }
1092
1093 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1094                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1095                   struct vm_area_struct *vma)
1096 {
1097         spinlock_t *dst_ptl, *src_ptl;
1098         struct page *src_page;
1099         pmd_t pmd;
1100         pgtable_t pgtable = NULL;
1101         int ret;
1102
1103         if (!vma_is_dax(vma)) {
1104                 ret = -ENOMEM;
1105                 pgtable = pte_alloc_one(dst_mm, addr);
1106                 if (unlikely(!pgtable))
1107                         goto out;
1108         }
1109
1110         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1111         src_ptl = pmd_lockptr(src_mm, src_pmd);
1112         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1113
1114         ret = -EAGAIN;
1115         pmd = *src_pmd;
1116         if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1117                 pte_free(dst_mm, pgtable);
1118                 goto out_unlock;
1119         }
1120         /*
1121          * When page table lock is held, the huge zero pmd should not be
1122          * under splitting since we don't split the page itself, only pmd to
1123          * a page table.
1124          */
1125         if (is_huge_zero_pmd(pmd)) {
1126                 struct page *zero_page;
1127                 /*
1128                  * get_huge_zero_page() will never allocate a new page here,
1129                  * since we already have a zero page to copy. It just takes a
1130                  * reference.
1131                  */
1132                 zero_page = get_huge_zero_page();
1133                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1134                                 zero_page);
1135                 ret = 0;
1136                 goto out_unlock;
1137         }
1138
1139         if (!vma_is_dax(vma)) {
1140                 /* thp accounting separate from pmd_devmap accounting */
1141                 src_page = pmd_page(pmd);
1142                 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1143                 get_page(src_page);
1144                 page_dup_rmap(src_page, true);
1145                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1146                 atomic_long_inc(&dst_mm->nr_ptes);
1147                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1148         }
1149
1150         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1151         pmd = pmd_mkold(pmd_wrprotect(pmd));
1152         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1153
1154         ret = 0;
1155 out_unlock:
1156         spin_unlock(src_ptl);
1157         spin_unlock(dst_ptl);
1158 out:
1159         return ret;
1160 }
1161
1162 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
1163 {
1164         pmd_t entry;
1165         unsigned long haddr;
1166
1167         fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
1168         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1169                 goto unlock;
1170
1171         entry = pmd_mkyoung(orig_pmd);
1172         haddr = fe->address & HPAGE_PMD_MASK;
1173         if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry,
1174                                 fe->flags & FAULT_FLAG_WRITE))
1175                 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
1176
1177 unlock:
1178         spin_unlock(fe->ptl);
1179 }
1180
1181 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
1182                 struct page *page)
1183 {
1184         struct vm_area_struct *vma = fe->vma;
1185         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1186         struct mem_cgroup *memcg;
1187         pgtable_t pgtable;
1188         pmd_t _pmd;
1189         int ret = 0, i;
1190         struct page **pages;
1191         unsigned long mmun_start;       /* For mmu_notifiers */
1192         unsigned long mmun_end;         /* For mmu_notifiers */
1193
1194         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1195                         GFP_KERNEL);
1196         if (unlikely(!pages)) {
1197                 ret |= VM_FAULT_OOM;
1198                 goto out;
1199         }
1200
1201         for (i = 0; i < HPAGE_PMD_NR; i++) {
1202                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1203                                                __GFP_OTHER_NODE, vma,
1204                                                fe->address, page_to_nid(page));
1205                 if (unlikely(!pages[i] ||
1206                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1207                                      GFP_KERNEL, &memcg, false))) {
1208                         if (pages[i])
1209                                 put_page(pages[i]);
1210                         while (--i >= 0) {
1211                                 memcg = (void *)page_private(pages[i]);
1212                                 set_page_private(pages[i], 0);
1213                                 mem_cgroup_cancel_charge(pages[i], memcg,
1214                                                 false);
1215                                 put_page(pages[i]);
1216                         }
1217                         kfree(pages);
1218                         ret |= VM_FAULT_OOM;
1219                         goto out;
1220                 }
1221                 set_page_private(pages[i], (unsigned long)memcg);
1222         }
1223
1224         for (i = 0; i < HPAGE_PMD_NR; i++) {
1225                 copy_user_highpage(pages[i], page + i,
1226                                    haddr + PAGE_SIZE * i, vma);
1227                 __SetPageUptodate(pages[i]);
1228                 cond_resched();
1229         }
1230
1231         mmun_start = haddr;
1232         mmun_end   = haddr + HPAGE_PMD_SIZE;
1233         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1234
1235         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1236         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1237                 goto out_free_pages;
1238         VM_BUG_ON_PAGE(!PageHead(page), page);
1239
1240         pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1241         /* leave pmd empty until pte is filled */
1242
1243         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
1244         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1245
1246         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1247                 pte_t entry;
1248                 entry = mk_pte(pages[i], vma->vm_page_prot);
1249                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1250                 memcg = (void *)page_private(pages[i]);
1251                 set_page_private(pages[i], 0);
1252                 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
1253                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1254                 lru_cache_add_active_or_unevictable(pages[i], vma);
1255                 fe->pte = pte_offset_map(&_pmd, haddr);
1256                 VM_BUG_ON(!pte_none(*fe->pte));
1257                 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
1258                 pte_unmap(fe->pte);
1259         }
1260         kfree(pages);
1261
1262         smp_wmb(); /* make pte visible before pmd */
1263         pmd_populate(vma->vm_mm, fe->pmd, pgtable);
1264         page_remove_rmap(page, true);
1265         spin_unlock(fe->ptl);
1266
1267         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1268
1269         ret |= VM_FAULT_WRITE;
1270         put_page(page);
1271
1272 out:
1273         return ret;
1274
1275 out_free_pages:
1276         spin_unlock(fe->ptl);
1277         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1278         for (i = 0; i < HPAGE_PMD_NR; i++) {
1279                 memcg = (void *)page_private(pages[i]);
1280                 set_page_private(pages[i], 0);
1281                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1282                 put_page(pages[i]);
1283         }
1284         kfree(pages);
1285         goto out;
1286 }
1287
1288 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
1289 {
1290         struct vm_area_struct *vma = fe->vma;
1291         struct page *page = NULL, *new_page;
1292         struct mem_cgroup *memcg;
1293         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1294         unsigned long mmun_start;       /* For mmu_notifiers */
1295         unsigned long mmun_end;         /* For mmu_notifiers */
1296         gfp_t huge_gfp;                 /* for allocation and charge */
1297         int ret = 0;
1298
1299         fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1300         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1301         if (is_huge_zero_pmd(orig_pmd))
1302                 goto alloc;
1303         spin_lock(fe->ptl);
1304         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1305                 goto out_unlock;
1306
1307         page = pmd_page(orig_pmd);
1308         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1309         /*
1310          * We can only reuse the page if nobody else maps the huge page or it's
1311          * part.
1312          */
1313         if (page_trans_huge_mapcount(page, NULL) == 1) {
1314                 pmd_t entry;
1315                 entry = pmd_mkyoung(orig_pmd);
1316                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1317                 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry,  1))
1318                         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1319                 ret |= VM_FAULT_WRITE;
1320                 goto out_unlock;
1321         }
1322         get_page(page);
1323         spin_unlock(fe->ptl);
1324 alloc:
1325         if (transparent_hugepage_enabled(vma) &&
1326             !transparent_hugepage_debug_cow()) {
1327                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1328                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1329         } else
1330                 new_page = NULL;
1331
1332         if (likely(new_page)) {
1333                 prep_transhuge_page(new_page);
1334         } else {
1335                 if (!page) {
1336                         split_huge_pmd(vma, fe->pmd, fe->address);
1337                         ret |= VM_FAULT_FALLBACK;
1338                 } else {
1339                         ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1340                         if (ret & VM_FAULT_OOM) {
1341                                 split_huge_pmd(vma, fe->pmd, fe->address);
1342                                 ret |= VM_FAULT_FALLBACK;
1343                         }
1344                         put_page(page);
1345                 }
1346                 count_vm_event(THP_FAULT_FALLBACK);
1347                 goto out;
1348         }
1349
1350         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1351                                         huge_gfp, &memcg, true))) {
1352                 put_page(new_page);
1353                 split_huge_pmd(vma, fe->pmd, fe->address);
1354                 if (page)
1355                         put_page(page);
1356                 ret |= VM_FAULT_FALLBACK;
1357                 count_vm_event(THP_FAULT_FALLBACK);
1358                 goto out;
1359         }
1360
1361         count_vm_event(THP_FAULT_ALLOC);
1362
1363         if (!page)
1364                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1365         else
1366                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1367         __SetPageUptodate(new_page);
1368
1369         mmun_start = haddr;
1370         mmun_end   = haddr + HPAGE_PMD_SIZE;
1371         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1372
1373         spin_lock(fe->ptl);
1374         if (page)
1375                 put_page(page);
1376         if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1377                 spin_unlock(fe->ptl);
1378                 mem_cgroup_cancel_charge(new_page, memcg, true);
1379                 put_page(new_page);
1380                 goto out_mn;
1381         } else {
1382                 pmd_t entry;
1383                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1384                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1385                 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1386                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1387                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1388                 lru_cache_add_active_or_unevictable(new_page, vma);
1389                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1390                 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1391                 if (!page) {
1392                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1393                         put_huge_zero_page();
1394                 } else {
1395                         VM_BUG_ON_PAGE(!PageHead(page), page);
1396                         page_remove_rmap(page, true);
1397                         put_page(page);
1398                 }
1399                 ret |= VM_FAULT_WRITE;
1400         }
1401         spin_unlock(fe->ptl);
1402 out_mn:
1403         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1404 out:
1405         return ret;
1406 out_unlock:
1407         spin_unlock(fe->ptl);
1408         return ret;
1409 }
1410
1411 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1412                                    unsigned long addr,
1413                                    pmd_t *pmd,
1414                                    unsigned int flags)
1415 {
1416         struct mm_struct *mm = vma->vm_mm;
1417         struct page *page = NULL;
1418
1419         assert_spin_locked(pmd_lockptr(mm, pmd));
1420
1421         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1422                 goto out;
1423
1424         /* Avoid dumping huge zero page */
1425         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1426                 return ERR_PTR(-EFAULT);
1427
1428         /* Full NUMA hinting faults to serialise migration in fault paths */
1429         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1430                 goto out;
1431
1432         page = pmd_page(*pmd);
1433         VM_BUG_ON_PAGE(!PageHead(page), page);
1434         if (flags & FOLL_TOUCH)
1435                 touch_pmd(vma, addr, pmd);
1436         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1437                 /*
1438                  * We don't mlock() pte-mapped THPs. This way we can avoid
1439                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1440                  *
1441                  * In most cases the pmd is the only mapping of the page as we
1442                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1443                  * writable private mappings in populate_vma_page_range().
1444                  *
1445                  * The only scenario when we have the page shared here is if we
1446                  * mlocking read-only mapping shared over fork(). We skip
1447                  * mlocking such pages.
1448                  */
1449                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1450                                 page->mapping && trylock_page(page)) {
1451                         lru_add_drain();
1452                         if (page->mapping)
1453                                 mlock_vma_page(page);
1454                         unlock_page(page);
1455                 }
1456         }
1457         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1458         VM_BUG_ON_PAGE(!PageCompound(page), page);
1459         if (flags & FOLL_GET)
1460                 get_page(page);
1461
1462 out:
1463         return page;
1464 }
1465
1466 /* NUMA hinting page fault entry point for trans huge pmds */
1467 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1468 {
1469         struct vm_area_struct *vma = fe->vma;
1470         struct anon_vma *anon_vma = NULL;
1471         struct page *page;
1472         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1473         int page_nid = -1, this_nid = numa_node_id();
1474         int target_nid, last_cpupid = -1;
1475         bool page_locked;
1476         bool migrated = false;
1477         bool was_writable;
1478         int flags = 0;
1479
1480         /* A PROT_NONE fault should not end up here */
1481         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1482
1483         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1484         if (unlikely(!pmd_same(pmd, *fe->pmd)))
1485                 goto out_unlock;
1486
1487         /*
1488          * If there are potential migrations, wait for completion and retry
1489          * without disrupting NUMA hinting information. Do not relock and
1490          * check_same as the page may no longer be mapped.
1491          */
1492         if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1493                 page = pmd_page(*fe->pmd);
1494                 spin_unlock(fe->ptl);
1495                 wait_on_page_locked(page);
1496                 goto out;
1497         }
1498
1499         page = pmd_page(pmd);
1500         BUG_ON(is_huge_zero_page(page));
1501         page_nid = page_to_nid(page);
1502         last_cpupid = page_cpupid_last(page);
1503         count_vm_numa_event(NUMA_HINT_FAULTS);
1504         if (page_nid == this_nid) {
1505                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1506                 flags |= TNF_FAULT_LOCAL;
1507         }
1508
1509         /* See similar comment in do_numa_page for explanation */
1510         if (!(vma->vm_flags & VM_WRITE))
1511                 flags |= TNF_NO_GROUP;
1512
1513         /*
1514          * Acquire the page lock to serialise THP migrations but avoid dropping
1515          * page_table_lock if at all possible
1516          */
1517         page_locked = trylock_page(page);
1518         target_nid = mpol_misplaced(page, vma, haddr);
1519         if (target_nid == -1) {
1520                 /* If the page was locked, there are no parallel migrations */
1521                 if (page_locked)
1522                         goto clear_pmdnuma;
1523         }
1524
1525         /* Migration could have started since the pmd_trans_migrating check */
1526         if (!page_locked) {
1527                 spin_unlock(fe->ptl);
1528                 wait_on_page_locked(page);
1529                 page_nid = -1;
1530                 goto out;
1531         }
1532
1533         /*
1534          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1535          * to serialises splits
1536          */
1537         get_page(page);
1538         spin_unlock(fe->ptl);
1539         anon_vma = page_lock_anon_vma_read(page);
1540
1541         /* Confirm the PMD did not change while page_table_lock was released */
1542         spin_lock(fe->ptl);
1543         if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1544                 unlock_page(page);
1545                 put_page(page);
1546                 page_nid = -1;
1547                 goto out_unlock;
1548         }
1549
1550         /* Bail if we fail to protect against THP splits for any reason */
1551         if (unlikely(!anon_vma)) {
1552                 put_page(page);
1553                 page_nid = -1;
1554                 goto clear_pmdnuma;
1555         }
1556
1557         /*
1558          * Migrate the THP to the requested node, returns with page unlocked
1559          * and access rights restored.
1560          */
1561         spin_unlock(fe->ptl);
1562         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1563                                 fe->pmd, pmd, fe->address, page, target_nid);
1564         if (migrated) {
1565                 flags |= TNF_MIGRATED;
1566                 page_nid = target_nid;
1567         } else
1568                 flags |= TNF_MIGRATE_FAIL;
1569
1570         goto out;
1571 clear_pmdnuma:
1572         BUG_ON(!PageLocked(page));
1573         was_writable = pmd_write(pmd);
1574         pmd = pmd_modify(pmd, vma->vm_page_prot);
1575         pmd = pmd_mkyoung(pmd);
1576         if (was_writable)
1577                 pmd = pmd_mkwrite(pmd);
1578         set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1579         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1580         unlock_page(page);
1581 out_unlock:
1582         spin_unlock(fe->ptl);
1583
1584 out:
1585         if (anon_vma)
1586                 page_unlock_anon_vma_read(anon_vma);
1587
1588         if (page_nid != -1)
1589                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1590
1591         return 0;
1592 }
1593
1594 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1595                 pmd_t *pmd, unsigned long addr, unsigned long next)
1596
1597 {
1598         spinlock_t *ptl;
1599         pmd_t orig_pmd;
1600         struct page *page;
1601         struct mm_struct *mm = tlb->mm;
1602         int ret = 0;
1603
1604         ptl = pmd_trans_huge_lock(pmd, vma);
1605         if (!ptl)
1606                 goto out_unlocked;
1607
1608         orig_pmd = *pmd;
1609         if (is_huge_zero_pmd(orig_pmd)) {
1610                 ret = 1;
1611                 goto out;
1612         }
1613
1614         page = pmd_page(orig_pmd);
1615         /*
1616          * If other processes are mapping this page, we couldn't discard
1617          * the page unless they all do MADV_FREE so let's skip the page.
1618          */
1619         if (page_mapcount(page) != 1)
1620                 goto out;
1621
1622         if (!trylock_page(page))
1623                 goto out;
1624
1625         /*
1626          * If user want to discard part-pages of THP, split it so MADV_FREE
1627          * will deactivate only them.
1628          */
1629         if (next - addr != HPAGE_PMD_SIZE) {
1630                 get_page(page);
1631                 spin_unlock(ptl);
1632                 split_huge_page(page);
1633                 put_page(page);
1634                 unlock_page(page);
1635                 goto out_unlocked;
1636         }
1637
1638         if (PageDirty(page))
1639                 ClearPageDirty(page);
1640         unlock_page(page);
1641
1642         if (PageActive(page))
1643                 deactivate_page(page);
1644
1645         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1646                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1647                         tlb->fullmm);
1648                 orig_pmd = pmd_mkold(orig_pmd);
1649                 orig_pmd = pmd_mkclean(orig_pmd);
1650
1651                 set_pmd_at(mm, addr, pmd, orig_pmd);
1652                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1653         }
1654         ret = 1;
1655 out:
1656         spin_unlock(ptl);
1657 out_unlocked:
1658         return ret;
1659 }
1660
1661 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1662                  pmd_t *pmd, unsigned long addr)
1663 {
1664         pmd_t orig_pmd;
1665         spinlock_t *ptl;
1666
1667         ptl = __pmd_trans_huge_lock(pmd, vma);
1668         if (!ptl)
1669                 return 0;
1670         /*
1671          * For architectures like ppc64 we look at deposited pgtable
1672          * when calling pmdp_huge_get_and_clear. So do the
1673          * pgtable_trans_huge_withdraw after finishing pmdp related
1674          * operations.
1675          */
1676         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1677                         tlb->fullmm);
1678         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1679         if (vma_is_dax(vma)) {
1680                 spin_unlock(ptl);
1681                 if (is_huge_zero_pmd(orig_pmd))
1682                         tlb_remove_page(tlb, pmd_page(orig_pmd));
1683         } else if (is_huge_zero_pmd(orig_pmd)) {
1684                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1685                 atomic_long_dec(&tlb->mm->nr_ptes);
1686                 spin_unlock(ptl);
1687                 tlb_remove_page(tlb, pmd_page(orig_pmd));
1688         } else {
1689                 struct page *page = pmd_page(orig_pmd);
1690                 page_remove_rmap(page, true);
1691                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1692                 VM_BUG_ON_PAGE(!PageHead(page), page);
1693                 if (PageAnon(page)) {
1694                         pgtable_t pgtable;
1695                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1696                         pte_free(tlb->mm, pgtable);
1697                         atomic_long_dec(&tlb->mm->nr_ptes);
1698                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1699                 } else {
1700                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1701                 }
1702                 spin_unlock(ptl);
1703                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1704         }
1705         return 1;
1706 }
1707
1708 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1709                   unsigned long new_addr, unsigned long old_end,
1710                   pmd_t *old_pmd, pmd_t *new_pmd)
1711 {
1712         spinlock_t *old_ptl, *new_ptl;
1713         pmd_t pmd;
1714         struct mm_struct *mm = vma->vm_mm;
1715
1716         if ((old_addr & ~HPAGE_PMD_MASK) ||
1717             (new_addr & ~HPAGE_PMD_MASK) ||
1718             old_end - old_addr < HPAGE_PMD_SIZE)
1719                 return false;
1720
1721         /*
1722          * The destination pmd shouldn't be established, free_pgtables()
1723          * should have release it.
1724          */
1725         if (WARN_ON(!pmd_none(*new_pmd))) {
1726                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1727                 return false;
1728         }
1729
1730         /*
1731          * We don't have to worry about the ordering of src and dst
1732          * ptlocks because exclusive mmap_sem prevents deadlock.
1733          */
1734         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1735         if (old_ptl) {
1736                 new_ptl = pmd_lockptr(mm, new_pmd);
1737                 if (new_ptl != old_ptl)
1738                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1739                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1740                 VM_BUG_ON(!pmd_none(*new_pmd));
1741
1742                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1743                                 vma_is_anonymous(vma)) {
1744                         pgtable_t pgtable;
1745                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1746                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1747                 }
1748                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1749                 if (new_ptl != old_ptl)
1750                         spin_unlock(new_ptl);
1751                 spin_unlock(old_ptl);
1752                 return true;
1753         }
1754         return false;
1755 }
1756
1757 /*
1758  * Returns
1759  *  - 0 if PMD could not be locked
1760  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1761  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1762  */
1763 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1764                 unsigned long addr, pgprot_t newprot, int prot_numa)
1765 {
1766         struct mm_struct *mm = vma->vm_mm;
1767         spinlock_t *ptl;
1768         int ret = 0;
1769
1770         ptl = __pmd_trans_huge_lock(pmd, vma);
1771         if (ptl) {
1772                 pmd_t entry;
1773                 bool preserve_write = prot_numa && pmd_write(*pmd);
1774                 ret = 1;
1775
1776                 /*
1777                  * Avoid trapping faults against the zero page. The read-only
1778                  * data is likely to be read-cached on the local CPU and
1779                  * local/remote hits to the zero page are not interesting.
1780                  */
1781                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1782                         spin_unlock(ptl);
1783                         return ret;
1784                 }
1785
1786                 if (!prot_numa || !pmd_protnone(*pmd)) {
1787                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1788                         entry = pmd_modify(entry, newprot);
1789                         if (preserve_write)
1790                                 entry = pmd_mkwrite(entry);
1791                         ret = HPAGE_PMD_NR;
1792                         set_pmd_at(mm, addr, pmd, entry);
1793                         BUG_ON(!preserve_write && pmd_write(entry));
1794                 }
1795                 spin_unlock(ptl);
1796         }
1797
1798         return ret;
1799 }
1800
1801 /*
1802  * Returns true if a given pmd maps a thp, false otherwise.
1803  *
1804  * Note that if it returns true, this routine returns without unlocking page
1805  * table lock. So callers must unlock it.
1806  */
1807 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1808 {
1809         spinlock_t *ptl;
1810         ptl = pmd_lock(vma->vm_mm, pmd);
1811         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1812                 return ptl;
1813         spin_unlock(ptl);
1814         return NULL;
1815 }
1816
1817 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1818
1819 int hugepage_madvise(struct vm_area_struct *vma,
1820                      unsigned long *vm_flags, int advice)
1821 {
1822         switch (advice) {
1823         case MADV_HUGEPAGE:
1824 #ifdef CONFIG_S390
1825                 /*
1826                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1827                  * can't handle this properly after s390_enable_sie, so we simply
1828                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1829                  */
1830                 if (mm_has_pgste(vma->vm_mm))
1831                         return 0;
1832 #endif
1833                 /*
1834                  * Be somewhat over-protective like KSM for now!
1835                  */
1836                 if (*vm_flags & VM_NO_THP)
1837                         return -EINVAL;
1838                 *vm_flags &= ~VM_NOHUGEPAGE;
1839                 *vm_flags |= VM_HUGEPAGE;
1840                 /*
1841                  * If the vma become good for khugepaged to scan,
1842                  * register it here without waiting a page fault that
1843                  * may not happen any time soon.
1844                  */
1845                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1846                         return -ENOMEM;
1847                 break;
1848         case MADV_NOHUGEPAGE:
1849                 /*
1850                  * Be somewhat over-protective like KSM for now!
1851                  */
1852                 if (*vm_flags & VM_NO_THP)
1853                         return -EINVAL;
1854                 *vm_flags &= ~VM_HUGEPAGE;
1855                 *vm_flags |= VM_NOHUGEPAGE;
1856                 /*
1857                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1858                  * this vma even if we leave the mm registered in khugepaged if
1859                  * it got registered before VM_NOHUGEPAGE was set.
1860                  */
1861                 break;
1862         }
1863
1864         return 0;
1865 }
1866
1867 static int __init khugepaged_slab_init(void)
1868 {
1869         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1870                                           sizeof(struct mm_slot),
1871                                           __alignof__(struct mm_slot), 0, NULL);
1872         if (!mm_slot_cache)
1873                 return -ENOMEM;
1874
1875         return 0;
1876 }
1877
1878 static void __init khugepaged_slab_exit(void)
1879 {
1880         kmem_cache_destroy(mm_slot_cache);
1881 }
1882
1883 static inline struct mm_slot *alloc_mm_slot(void)
1884 {
1885         if (!mm_slot_cache)     /* initialization failed */
1886                 return NULL;
1887         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1888 }
1889
1890 static inline void free_mm_slot(struct mm_slot *mm_slot)
1891 {
1892         kmem_cache_free(mm_slot_cache, mm_slot);
1893 }
1894
1895 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1896 {
1897         struct mm_slot *mm_slot;
1898
1899         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1900                 if (mm == mm_slot->mm)
1901                         return mm_slot;
1902
1903         return NULL;
1904 }
1905
1906 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1907                                     struct mm_slot *mm_slot)
1908 {
1909         mm_slot->mm = mm;
1910         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1911 }
1912
1913 static inline int khugepaged_test_exit(struct mm_struct *mm)
1914 {
1915         return atomic_read(&mm->mm_users) == 0;
1916 }
1917
1918 int __khugepaged_enter(struct mm_struct *mm)
1919 {
1920         struct mm_slot *mm_slot;
1921         int wakeup;
1922
1923         mm_slot = alloc_mm_slot();
1924         if (!mm_slot)
1925                 return -ENOMEM;
1926
1927         /* __khugepaged_exit() must not run from under us */
1928         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1929         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1930                 free_mm_slot(mm_slot);
1931                 return 0;
1932         }
1933
1934         spin_lock(&khugepaged_mm_lock);
1935         insert_to_mm_slots_hash(mm, mm_slot);
1936         /*
1937          * Insert just behind the scanning cursor, to let the area settle
1938          * down a little.
1939          */
1940         wakeup = list_empty(&khugepaged_scan.mm_head);
1941         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1942         spin_unlock(&khugepaged_mm_lock);
1943
1944         atomic_inc(&mm->mm_count);
1945         if (wakeup)
1946                 wake_up_interruptible(&khugepaged_wait);
1947
1948         return 0;
1949 }
1950
1951 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1952                                unsigned long vm_flags)
1953 {
1954         unsigned long hstart, hend;
1955         if (!vma->anon_vma)
1956                 /*
1957                  * Not yet faulted in so we will register later in the
1958                  * page fault if needed.
1959                  */
1960                 return 0;
1961         if (vma->vm_ops || (vm_flags & VM_NO_THP))
1962                 /* khugepaged not yet working on file or special mappings */
1963                 return 0;
1964         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1965         hend = vma->vm_end & HPAGE_PMD_MASK;
1966         if (hstart < hend)
1967                 return khugepaged_enter(vma, vm_flags);
1968         return 0;
1969 }
1970
1971 void __khugepaged_exit(struct mm_struct *mm)
1972 {
1973         struct mm_slot *mm_slot;
1974         int free = 0;
1975
1976         spin_lock(&khugepaged_mm_lock);
1977         mm_slot = get_mm_slot(mm);
1978         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1979                 hash_del(&mm_slot->hash);
1980                 list_del(&mm_slot->mm_node);
1981                 free = 1;
1982         }
1983         spin_unlock(&khugepaged_mm_lock);
1984
1985         if (free) {
1986                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1987                 free_mm_slot(mm_slot);
1988                 mmdrop(mm);
1989         } else if (mm_slot) {
1990                 /*
1991                  * This is required to serialize against
1992                  * khugepaged_test_exit() (which is guaranteed to run
1993                  * under mmap sem read mode). Stop here (after we
1994                  * return all pagetables will be destroyed) until
1995                  * khugepaged has finished working on the pagetables
1996                  * under the mmap_sem.
1997                  */
1998                 down_write(&mm->mmap_sem);
1999                 up_write(&mm->mmap_sem);
2000         }
2001 }
2002
2003 static void release_pte_page(struct page *page)
2004 {
2005         /* 0 stands for page_is_file_cache(page) == false */
2006         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2007         unlock_page(page);
2008         putback_lru_page(page);
2009 }
2010
2011 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2012 {
2013         while (--_pte >= pte) {
2014                 pte_t pteval = *_pte;
2015                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2016                         release_pte_page(pte_page(pteval));
2017         }
2018 }
2019
2020 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2021                                         unsigned long address,
2022                                         pte_t *pte)
2023 {
2024         struct page *page = NULL;
2025         pte_t *_pte;
2026         int none_or_zero = 0, result = 0;
2027         bool referenced = false, writable = false;
2028
2029         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2030              _pte++, address += PAGE_SIZE) {
2031                 pte_t pteval = *_pte;
2032                 if (pte_none(pteval) || (pte_present(pteval) &&
2033                                 is_zero_pfn(pte_pfn(pteval)))) {
2034                         if (!userfaultfd_armed(vma) &&
2035                             ++none_or_zero <= khugepaged_max_ptes_none) {
2036                                 continue;
2037                         } else {
2038                                 result = SCAN_EXCEED_NONE_PTE;
2039                                 goto out;
2040                         }
2041                 }
2042                 if (!pte_present(pteval)) {
2043                         result = SCAN_PTE_NON_PRESENT;
2044                         goto out;
2045                 }
2046                 page = vm_normal_page(vma, address, pteval);
2047                 if (unlikely(!page)) {
2048                         result = SCAN_PAGE_NULL;
2049                         goto out;
2050                 }
2051
2052                 VM_BUG_ON_PAGE(PageCompound(page), page);
2053                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2054                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2055
2056                 /*
2057                  * We can do it before isolate_lru_page because the
2058                  * page can't be freed from under us. NOTE: PG_lock
2059                  * is needed to serialize against split_huge_page
2060                  * when invoked from the VM.
2061                  */
2062                 if (!trylock_page(page)) {
2063                         result = SCAN_PAGE_LOCK;
2064                         goto out;
2065                 }
2066
2067                 /*
2068                  * cannot use mapcount: can't collapse if there's a gup pin.
2069                  * The page must only be referenced by the scanned process
2070                  * and page swap cache.
2071                  */
2072                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2073                         unlock_page(page);
2074                         result = SCAN_PAGE_COUNT;
2075                         goto out;
2076                 }
2077                 if (pte_write(pteval)) {
2078                         writable = true;
2079                 } else {
2080                         if (PageSwapCache(page) &&
2081                             !reuse_swap_page(page, NULL)) {
2082                                 unlock_page(page);
2083                                 result = SCAN_SWAP_CACHE_PAGE;
2084                                 goto out;
2085                         }
2086                         /*
2087                          * Page is not in the swap cache. It can be collapsed
2088                          * into a THP.
2089                          */
2090                 }
2091
2092                 /*
2093                  * Isolate the page to avoid collapsing an hugepage
2094                  * currently in use by the VM.
2095                  */
2096                 if (isolate_lru_page(page)) {
2097                         unlock_page(page);
2098                         result = SCAN_DEL_PAGE_LRU;
2099                         goto out;
2100                 }
2101                 /* 0 stands for page_is_file_cache(page) == false */
2102                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2103                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2104                 VM_BUG_ON_PAGE(PageLRU(page), page);
2105
2106                 /* If there is no mapped pte young don't collapse the page */
2107                 if (pte_young(pteval) ||
2108                     page_is_young(page) || PageReferenced(page) ||
2109                     mmu_notifier_test_young(vma->vm_mm, address))
2110                         referenced = true;
2111         }
2112         if (likely(writable)) {
2113                 if (likely(referenced)) {
2114                         result = SCAN_SUCCEED;
2115                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2116                                                             referenced, writable, result);
2117                         return 1;
2118                 }
2119         } else {
2120                 result = SCAN_PAGE_RO;
2121         }
2122
2123 out:
2124         release_pte_pages(pte, _pte);
2125         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2126                                             referenced, writable, result);
2127         return 0;
2128 }
2129
2130 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2131                                       struct vm_area_struct *vma,
2132                                       unsigned long address,
2133                                       spinlock_t *ptl)
2134 {
2135         pte_t *_pte;
2136         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2137                 pte_t pteval = *_pte;
2138                 struct page *src_page;
2139
2140                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2141                         clear_user_highpage(page, address);
2142                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2143                         if (is_zero_pfn(pte_pfn(pteval))) {
2144                                 /*
2145                                  * ptl mostly unnecessary.
2146                                  */
2147                                 spin_lock(ptl);
2148                                 /*
2149                                  * paravirt calls inside pte_clear here are
2150                                  * superfluous.
2151                                  */
2152                                 pte_clear(vma->vm_mm, address, _pte);
2153                                 spin_unlock(ptl);
2154                         }
2155                 } else {
2156                         src_page = pte_page(pteval);
2157                         copy_user_highpage(page, src_page, address, vma);
2158                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2159                         release_pte_page(src_page);
2160                         /*
2161                          * ptl mostly unnecessary, but preempt has to
2162                          * be disabled to update the per-cpu stats
2163                          * inside page_remove_rmap().
2164                          */
2165                         spin_lock(ptl);
2166                         /*
2167                          * paravirt calls inside pte_clear here are
2168                          * superfluous.
2169                          */
2170                         pte_clear(vma->vm_mm, address, _pte);
2171                         page_remove_rmap(src_page, false);
2172                         spin_unlock(ptl);
2173                         free_page_and_swap_cache(src_page);
2174                 }
2175
2176                 address += PAGE_SIZE;
2177                 page++;
2178         }
2179 }
2180
2181 static void khugepaged_alloc_sleep(void)
2182 {
2183         DEFINE_WAIT(wait);
2184
2185         add_wait_queue(&khugepaged_wait, &wait);
2186         freezable_schedule_timeout_interruptible(
2187                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2188         remove_wait_queue(&khugepaged_wait, &wait);
2189 }
2190
2191 static int khugepaged_node_load[MAX_NUMNODES];
2192
2193 static bool khugepaged_scan_abort(int nid)
2194 {
2195         int i;
2196
2197         /*
2198          * If zone_reclaim_mode is disabled, then no extra effort is made to
2199          * allocate memory locally.
2200          */
2201         if (!zone_reclaim_mode)
2202                 return false;
2203
2204         /* If there is a count for this node already, it must be acceptable */
2205         if (khugepaged_node_load[nid])
2206                 return false;
2207
2208         for (i = 0; i < MAX_NUMNODES; i++) {
2209                 if (!khugepaged_node_load[i])
2210                         continue;
2211                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2212                         return true;
2213         }
2214         return false;
2215 }
2216
2217 #ifdef CONFIG_NUMA
2218 static int khugepaged_find_target_node(void)
2219 {
2220         static int last_khugepaged_target_node = NUMA_NO_NODE;
2221         int nid, target_node = 0, max_value = 0;
2222
2223         /* find first node with max normal pages hit */
2224         for (nid = 0; nid < MAX_NUMNODES; nid++)
2225                 if (khugepaged_node_load[nid] > max_value) {
2226                         max_value = khugepaged_node_load[nid];
2227                         target_node = nid;
2228                 }
2229
2230         /* do some balance if several nodes have the same hit record */
2231         if (target_node <= last_khugepaged_target_node)
2232                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2233                                 nid++)
2234                         if (max_value == khugepaged_node_load[nid]) {
2235                                 target_node = nid;
2236                                 break;
2237                         }
2238
2239         last_khugepaged_target_node = target_node;
2240         return target_node;
2241 }
2242
2243 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2244 {
2245         if (IS_ERR(*hpage)) {
2246                 if (!*wait)
2247                         return false;
2248
2249                 *wait = false;
2250                 *hpage = NULL;
2251                 khugepaged_alloc_sleep();
2252         } else if (*hpage) {
2253                 put_page(*hpage);
2254                 *hpage = NULL;
2255         }
2256
2257         return true;
2258 }
2259
2260 static struct page *
2261 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2262                        unsigned long address, int node)
2263 {
2264         VM_BUG_ON_PAGE(*hpage, *hpage);
2265
2266         /*
2267          * Before allocating the hugepage, release the mmap_sem read lock.
2268          * The allocation can take potentially a long time if it involves
2269          * sync compaction, and we do not need to hold the mmap_sem during
2270          * that. We will recheck the vma after taking it again in write mode.
2271          */
2272         up_read(&mm->mmap_sem);
2273
2274         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2275         if (unlikely(!*hpage)) {
2276                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2277                 *hpage = ERR_PTR(-ENOMEM);
2278                 return NULL;
2279         }
2280
2281         prep_transhuge_page(*hpage);
2282         count_vm_event(THP_COLLAPSE_ALLOC);
2283         return *hpage;
2284 }
2285 #else
2286 static int khugepaged_find_target_node(void)
2287 {
2288         return 0;
2289 }
2290
2291 static inline struct page *alloc_khugepaged_hugepage(void)
2292 {
2293         struct page *page;
2294
2295         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2296                            HPAGE_PMD_ORDER);
2297         if (page)
2298                 prep_transhuge_page(page);
2299         return page;
2300 }
2301
2302 static struct page *khugepaged_alloc_hugepage(bool *wait)
2303 {
2304         struct page *hpage;
2305
2306         do {
2307                 hpage = alloc_khugepaged_hugepage();
2308                 if (!hpage) {
2309                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2310                         if (!*wait)
2311                                 return NULL;
2312
2313                         *wait = false;
2314                         khugepaged_alloc_sleep();
2315                 } else
2316                         count_vm_event(THP_COLLAPSE_ALLOC);
2317         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2318
2319         return hpage;
2320 }
2321
2322 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2323 {
2324         if (!*hpage)
2325                 *hpage = khugepaged_alloc_hugepage(wait);
2326
2327         if (unlikely(!*hpage))
2328                 return false;
2329
2330         return true;
2331 }
2332
2333 static struct page *
2334 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2335                        unsigned long address, int node)
2336 {
2337         up_read(&mm->mmap_sem);
2338         VM_BUG_ON(!*hpage);
2339
2340         return  *hpage;
2341 }
2342 #endif
2343
2344 static bool hugepage_vma_check(struct vm_area_struct *vma)
2345 {
2346         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2347             (vma->vm_flags & VM_NOHUGEPAGE))
2348                 return false;
2349         if (!vma->anon_vma || vma->vm_ops)
2350                 return false;
2351         if (is_vma_temporary_stack(vma))
2352                 return false;
2353         return !(vma->vm_flags & VM_NO_THP);
2354 }
2355
2356 /*
2357  * If mmap_sem temporarily dropped, revalidate vma
2358  * before taking mmap_sem.
2359  * Return 0 if succeeds, otherwise return none-zero
2360  * value (scan code).
2361  */
2362
2363 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address)
2364 {
2365         struct vm_area_struct *vma;
2366         unsigned long hstart, hend;
2367
2368         if (unlikely(khugepaged_test_exit(mm)))
2369                 return SCAN_ANY_PROCESS;
2370
2371         vma = find_vma(mm, address);
2372         if (!vma)
2373                 return SCAN_VMA_NULL;
2374
2375         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2376         hend = vma->vm_end & HPAGE_PMD_MASK;
2377         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2378                 return SCAN_ADDRESS_RANGE;
2379         if (!hugepage_vma_check(vma))
2380                 return SCAN_VMA_CHECK;
2381         return 0;
2382 }
2383
2384 /*
2385  * Bring missing pages in from swap, to complete THP collapse.
2386  * Only done if khugepaged_scan_pmd believes it is worthwhile.
2387  *
2388  * Called and returns without pte mapped or spinlocks held,
2389  * but with mmap_sem held to protect against vma changes.
2390  */
2391
2392 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
2393                                         struct vm_area_struct *vma,
2394                                         unsigned long address, pmd_t *pmd)
2395 {
2396         pte_t pteval;
2397         int swapped_in = 0, ret = 0;
2398         struct fault_env fe = {
2399                 .vma = vma,
2400                 .address = address,
2401                 .flags = FAULT_FLAG_ALLOW_RETRY,
2402                 .pmd = pmd,
2403         };
2404
2405         fe.pte = pte_offset_map(pmd, address);
2406         for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
2407                         fe.pte++, fe.address += PAGE_SIZE) {
2408                 pteval = *fe.pte;
2409                 if (!is_swap_pte(pteval))
2410                         continue;
2411                 swapped_in++;
2412                 ret = do_swap_page(&fe, pteval);
2413                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
2414                 if (ret & VM_FAULT_RETRY) {
2415                         down_read(&mm->mmap_sem);
2416                         /* vma is no longer available, don't continue to swapin */
2417                         if (hugepage_vma_revalidate(mm, address))
2418                                 return false;
2419                         /* check if the pmd is still valid */
2420                         if (mm_find_pmd(mm, address) != pmd)
2421                                 return false;
2422                 }
2423                 if (ret & VM_FAULT_ERROR) {
2424                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, 0);
2425                         return false;
2426                 }
2427                 /* pte is unmapped now, we need to map it */
2428                 fe.pte = pte_offset_map(pmd, fe.address);
2429         }
2430         fe.pte--;
2431         pte_unmap(fe.pte);
2432         trace_mm_collapse_huge_page_swapin(mm, swapped_in, 1);
2433         return true;
2434 }
2435
2436 static void collapse_huge_page(struct mm_struct *mm,
2437                                    unsigned long address,
2438                                    struct page **hpage,
2439                                    struct vm_area_struct *vma,
2440                                    int node)
2441 {
2442         pmd_t *pmd, _pmd;
2443         pte_t *pte;
2444         pgtable_t pgtable;
2445         struct page *new_page;
2446         spinlock_t *pmd_ptl, *pte_ptl;
2447         int isolated = 0, result = 0;
2448         struct mem_cgroup *memcg;
2449         unsigned long mmun_start;       /* For mmu_notifiers */
2450         unsigned long mmun_end;         /* For mmu_notifiers */
2451         gfp_t gfp;
2452
2453         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2454
2455         /* Only allocate from the target node */
2456         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
2457
2458         /* release the mmap_sem read lock. */
2459         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2460         if (!new_page) {
2461                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2462                 goto out_nolock;
2463         }
2464
2465         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2466                 result = SCAN_CGROUP_CHARGE_FAIL;
2467                 goto out_nolock;
2468         }
2469
2470         down_read(&mm->mmap_sem);
2471         result = hugepage_vma_revalidate(mm, address);
2472         if (result) {
2473                 mem_cgroup_cancel_charge(new_page, memcg, true);
2474                 up_read(&mm->mmap_sem);
2475                 goto out_nolock;
2476         }
2477
2478         pmd = mm_find_pmd(mm, address);
2479         if (!pmd) {
2480                 result = SCAN_PMD_NULL;
2481                 mem_cgroup_cancel_charge(new_page, memcg, true);
2482                 up_read(&mm->mmap_sem);
2483                 goto out_nolock;
2484         }
2485
2486         /*
2487          * __collapse_huge_page_swapin always returns with mmap_sem locked.
2488          * If it fails, release mmap_sem and jump directly out.
2489          * Continuing to collapse causes inconsistency.
2490          */
2491         if (!__collapse_huge_page_swapin(mm, vma, address, pmd)) {
2492                 mem_cgroup_cancel_charge(new_page, memcg, true);
2493                 up_read(&mm->mmap_sem);
2494                 goto out_nolock;
2495         }
2496
2497         up_read(&mm->mmap_sem);
2498         /*
2499          * Prevent all access to pagetables with the exception of
2500          * gup_fast later handled by the ptep_clear_flush and the VM
2501          * handled by the anon_vma lock + PG_lock.
2502          */
2503         down_write(&mm->mmap_sem);
2504         result = hugepage_vma_revalidate(mm, address);
2505         if (result)
2506                 goto out;
2507         /* check if the pmd is still valid */
2508         if (mm_find_pmd(mm, address) != pmd)
2509                 goto out;
2510
2511         anon_vma_lock_write(vma->anon_vma);
2512
2513         pte = pte_offset_map(pmd, address);
2514         pte_ptl = pte_lockptr(mm, pmd);
2515
2516         mmun_start = address;
2517         mmun_end   = address + HPAGE_PMD_SIZE;
2518         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2519         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2520         /*
2521          * After this gup_fast can't run anymore. This also removes
2522          * any huge TLB entry from the CPU so we won't allow
2523          * huge and small TLB entries for the same virtual address
2524          * to avoid the risk of CPU bugs in that area.
2525          */
2526         _pmd = pmdp_collapse_flush(vma, address, pmd);
2527         spin_unlock(pmd_ptl);
2528         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2529
2530         spin_lock(pte_ptl);
2531         isolated = __collapse_huge_page_isolate(vma, address, pte);
2532         spin_unlock(pte_ptl);
2533
2534         if (unlikely(!isolated)) {
2535                 pte_unmap(pte);
2536                 spin_lock(pmd_ptl);
2537                 BUG_ON(!pmd_none(*pmd));
2538                 /*
2539                  * We can only use set_pmd_at when establishing
2540                  * hugepmds and never for establishing regular pmds that
2541                  * points to regular pagetables. Use pmd_populate for that
2542                  */
2543                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2544                 spin_unlock(pmd_ptl);
2545                 anon_vma_unlock_write(vma->anon_vma);
2546                 result = SCAN_FAIL;
2547                 goto out;
2548         }
2549
2550         /*
2551          * All pages are isolated and locked so anon_vma rmap
2552          * can't run anymore.
2553          */
2554         anon_vma_unlock_write(vma->anon_vma);
2555
2556         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2557         pte_unmap(pte);
2558         __SetPageUptodate(new_page);
2559         pgtable = pmd_pgtable(_pmd);
2560
2561         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2562         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2563
2564         /*
2565          * spin_lock() below is not the equivalent of smp_wmb(), so
2566          * this is needed to avoid the copy_huge_page writes to become
2567          * visible after the set_pmd_at() write.
2568          */
2569         smp_wmb();
2570
2571         spin_lock(pmd_ptl);
2572         BUG_ON(!pmd_none(*pmd));
2573         page_add_new_anon_rmap(new_page, vma, address, true);
2574         mem_cgroup_commit_charge(new_page, memcg, false, true);
2575         lru_cache_add_active_or_unevictable(new_page, vma);
2576         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2577         set_pmd_at(mm, address, pmd, _pmd);
2578         update_mmu_cache_pmd(vma, address, pmd);
2579         spin_unlock(pmd_ptl);
2580
2581         *hpage = NULL;
2582
2583         khugepaged_pages_collapsed++;
2584         result = SCAN_SUCCEED;
2585 out_up_write:
2586         up_write(&mm->mmap_sem);
2587 out_nolock:
2588         trace_mm_collapse_huge_page(mm, isolated, result);
2589         return;
2590 out:
2591         mem_cgroup_cancel_charge(new_page, memcg, true);
2592         goto out_up_write;
2593 }
2594
2595 static int khugepaged_scan_pmd(struct mm_struct *mm,
2596                                struct vm_area_struct *vma,
2597                                unsigned long address,
2598                                struct page **hpage)
2599 {
2600         pmd_t *pmd;
2601         pte_t *pte, *_pte;
2602         int ret = 0, none_or_zero = 0, result = 0;
2603         struct page *page = NULL;
2604         unsigned long _address;
2605         spinlock_t *ptl;
2606         int node = NUMA_NO_NODE, unmapped = 0;
2607         bool writable = false, referenced = false;
2608
2609         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2610
2611         pmd = mm_find_pmd(mm, address);
2612         if (!pmd) {
2613                 result = SCAN_PMD_NULL;
2614                 goto out;
2615         }
2616
2617         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2618         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2619         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2620              _pte++, _address += PAGE_SIZE) {
2621                 pte_t pteval = *_pte;
2622                 if (is_swap_pte(pteval)) {
2623                         if (++unmapped <= khugepaged_max_ptes_swap) {
2624                                 continue;
2625                         } else {
2626                                 result = SCAN_EXCEED_SWAP_PTE;
2627                                 goto out_unmap;
2628                         }
2629                 }
2630                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2631                         if (!userfaultfd_armed(vma) &&
2632                             ++none_or_zero <= khugepaged_max_ptes_none) {
2633                                 continue;
2634                         } else {
2635                                 result = SCAN_EXCEED_NONE_PTE;
2636                                 goto out_unmap;
2637                         }
2638                 }
2639                 if (!pte_present(pteval)) {
2640                         result = SCAN_PTE_NON_PRESENT;
2641                         goto out_unmap;
2642                 }
2643                 if (pte_write(pteval))
2644                         writable = true;
2645
2646                 page = vm_normal_page(vma, _address, pteval);
2647                 if (unlikely(!page)) {
2648                         result = SCAN_PAGE_NULL;
2649                         goto out_unmap;
2650                 }
2651
2652                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2653                 if (PageCompound(page)) {
2654                         result = SCAN_PAGE_COMPOUND;
2655                         goto out_unmap;
2656                 }
2657
2658                 /*
2659                  * Record which node the original page is from and save this
2660                  * information to khugepaged_node_load[].
2661                  * Khupaged will allocate hugepage from the node has the max
2662                  * hit record.
2663                  */
2664                 node = page_to_nid(page);
2665                 if (khugepaged_scan_abort(node)) {
2666                         result = SCAN_SCAN_ABORT;
2667                         goto out_unmap;
2668                 }
2669                 khugepaged_node_load[node]++;
2670                 if (!PageLRU(page)) {
2671                         result = SCAN_PAGE_LRU;
2672                         goto out_unmap;
2673                 }
2674                 if (PageLocked(page)) {
2675                         result = SCAN_PAGE_LOCK;
2676                         goto out_unmap;
2677                 }
2678                 if (!PageAnon(page)) {
2679                         result = SCAN_PAGE_ANON;
2680                         goto out_unmap;
2681                 }
2682
2683                 /*
2684                  * cannot use mapcount: can't collapse if there's a gup pin.
2685                  * The page must only be referenced by the scanned process
2686                  * and page swap cache.
2687                  */
2688                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2689                         result = SCAN_PAGE_COUNT;
2690                         goto out_unmap;
2691                 }
2692                 if (pte_young(pteval) ||
2693                     page_is_young(page) || PageReferenced(page) ||
2694                     mmu_notifier_test_young(vma->vm_mm, address))
2695                         referenced = true;
2696         }
2697         if (writable) {
2698                 if (referenced) {
2699                         result = SCAN_SUCCEED;
2700                         ret = 1;
2701                 } else {
2702                         result = SCAN_NO_REFERENCED_PAGE;
2703                 }
2704         } else {
2705                 result = SCAN_PAGE_RO;
2706         }
2707 out_unmap:
2708         pte_unmap_unlock(pte, ptl);
2709         if (ret) {
2710                 node = khugepaged_find_target_node();
2711                 /* collapse_huge_page will return with the mmap_sem released */
2712                 collapse_huge_page(mm, address, hpage, vma, node);
2713         }
2714 out:
2715         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2716                                      none_or_zero, result, unmapped);
2717         return ret;
2718 }
2719
2720 static void collect_mm_slot(struct mm_slot *mm_slot)
2721 {
2722         struct mm_struct *mm = mm_slot->mm;
2723
2724         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2725
2726         if (khugepaged_test_exit(mm)) {
2727                 /* free mm_slot */
2728                 hash_del(&mm_slot->hash);
2729                 list_del(&mm_slot->mm_node);
2730
2731                 /*
2732                  * Not strictly needed because the mm exited already.
2733                  *
2734                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2735                  */
2736
2737                 /* khugepaged_mm_lock actually not necessary for the below */
2738                 free_mm_slot(mm_slot);
2739                 mmdrop(mm);
2740         }
2741 }
2742
2743 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2744                                             struct page **hpage)
2745         __releases(&khugepaged_mm_lock)
2746         __acquires(&khugepaged_mm_lock)
2747 {
2748         struct mm_slot *mm_slot;
2749         struct mm_struct *mm;
2750         struct vm_area_struct *vma;
2751         int progress = 0;
2752
2753         VM_BUG_ON(!pages);
2754         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2755
2756         if (khugepaged_scan.mm_slot)
2757                 mm_slot = khugepaged_scan.mm_slot;
2758         else {
2759                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2760                                      struct mm_slot, mm_node);
2761                 khugepaged_scan.address = 0;
2762                 khugepaged_scan.mm_slot = mm_slot;
2763         }
2764         spin_unlock(&khugepaged_mm_lock);
2765
2766         mm = mm_slot->mm;
2767         down_read(&mm->mmap_sem);
2768         if (unlikely(khugepaged_test_exit(mm)))
2769                 vma = NULL;
2770         else
2771                 vma = find_vma(mm, khugepaged_scan.address);
2772
2773         progress++;
2774         for (; vma; vma = vma->vm_next) {
2775                 unsigned long hstart, hend;
2776
2777                 cond_resched();
2778                 if (unlikely(khugepaged_test_exit(mm))) {
2779                         progress++;
2780                         break;
2781                 }
2782                 if (!hugepage_vma_check(vma)) {
2783 skip:
2784                         progress++;
2785                         continue;
2786                 }
2787                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2788                 hend = vma->vm_end & HPAGE_PMD_MASK;
2789                 if (hstart >= hend)
2790                         goto skip;
2791                 if (khugepaged_scan.address > hend)
2792                         goto skip;
2793                 if (khugepaged_scan.address < hstart)
2794                         khugepaged_scan.address = hstart;
2795                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2796
2797                 while (khugepaged_scan.address < hend) {
2798                         int ret;
2799                         cond_resched();
2800                         if (unlikely(khugepaged_test_exit(mm)))
2801                                 goto breakouterloop;
2802
2803                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2804                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2805                                   hend);
2806                         ret = khugepaged_scan_pmd(mm, vma,
2807                                                   khugepaged_scan.address,
2808                                                   hpage);
2809                         /* move to next address */
2810                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2811                         progress += HPAGE_PMD_NR;
2812                         if (ret)
2813                                 /* we released mmap_sem so break loop */
2814                                 goto breakouterloop_mmap_sem;
2815                         if (progress >= pages)
2816                                 goto breakouterloop;
2817                 }
2818         }
2819 breakouterloop:
2820         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2821 breakouterloop_mmap_sem:
2822
2823         spin_lock(&khugepaged_mm_lock);
2824         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2825         /*
2826          * Release the current mm_slot if this mm is about to die, or
2827          * if we scanned all vmas of this mm.
2828          */
2829         if (khugepaged_test_exit(mm) || !vma) {
2830                 /*
2831                  * Make sure that if mm_users is reaching zero while
2832                  * khugepaged runs here, khugepaged_exit will find
2833                  * mm_slot not pointing to the exiting mm.
2834                  */
2835                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2836                         khugepaged_scan.mm_slot = list_entry(
2837                                 mm_slot->mm_node.next,
2838                                 struct mm_slot, mm_node);
2839                         khugepaged_scan.address = 0;
2840                 } else {
2841                         khugepaged_scan.mm_slot = NULL;
2842                         khugepaged_full_scans++;
2843                 }
2844
2845                 collect_mm_slot(mm_slot);
2846         }
2847
2848         return progress;
2849 }
2850
2851 static int khugepaged_has_work(void)
2852 {
2853         return !list_empty(&khugepaged_scan.mm_head) &&
2854                 khugepaged_enabled();
2855 }
2856
2857 static int khugepaged_wait_event(void)
2858 {
2859         return !list_empty(&khugepaged_scan.mm_head) ||
2860                 kthread_should_stop();
2861 }
2862
2863 static void khugepaged_do_scan(void)
2864 {
2865         struct page *hpage = NULL;
2866         unsigned int progress = 0, pass_through_head = 0;
2867         unsigned int pages = khugepaged_pages_to_scan;
2868         bool wait = true;
2869
2870         barrier(); /* write khugepaged_pages_to_scan to local stack */
2871
2872         while (progress < pages) {
2873                 if (!khugepaged_prealloc_page(&hpage, &wait))
2874                         break;
2875
2876                 cond_resched();
2877
2878                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2879                         break;
2880
2881                 spin_lock(&khugepaged_mm_lock);
2882                 if (!khugepaged_scan.mm_slot)
2883                         pass_through_head++;
2884                 if (khugepaged_has_work() &&
2885                     pass_through_head < 2)
2886                         progress += khugepaged_scan_mm_slot(pages - progress,
2887                                                             &hpage);
2888                 else
2889                         progress = pages;
2890                 spin_unlock(&khugepaged_mm_lock);
2891         }
2892
2893         if (!IS_ERR_OR_NULL(hpage))
2894                 put_page(hpage);
2895 }
2896
2897 static bool khugepaged_should_wakeup(void)
2898 {
2899         return kthread_should_stop() ||
2900                time_after_eq(jiffies, khugepaged_sleep_expire);
2901 }
2902
2903 static void khugepaged_wait_work(void)
2904 {
2905         if (khugepaged_has_work()) {
2906                 const unsigned long scan_sleep_jiffies =
2907                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2908
2909                 if (!scan_sleep_jiffies)
2910                         return;
2911
2912                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2913                 wait_event_freezable_timeout(khugepaged_wait,
2914                                              khugepaged_should_wakeup(),
2915                                              scan_sleep_jiffies);
2916                 return;
2917         }
2918
2919         if (khugepaged_enabled())
2920                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2921 }
2922
2923 static int khugepaged(void *none)
2924 {
2925         struct mm_slot *mm_slot;
2926
2927         set_freezable();
2928         set_user_nice(current, MAX_NICE);
2929
2930         while (!kthread_should_stop()) {
2931                 khugepaged_do_scan();
2932                 khugepaged_wait_work();
2933         }
2934
2935         spin_lock(&khugepaged_mm_lock);
2936         mm_slot = khugepaged_scan.mm_slot;
2937         khugepaged_scan.mm_slot = NULL;
2938         if (mm_slot)
2939                 collect_mm_slot(mm_slot);
2940         spin_unlock(&khugepaged_mm_lock);
2941         return 0;
2942 }
2943
2944 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2945                 unsigned long haddr, pmd_t *pmd)
2946 {
2947         struct mm_struct *mm = vma->vm_mm;
2948         pgtable_t pgtable;
2949         pmd_t _pmd;
2950         int i;
2951
2952         /* leave pmd empty until pte is filled */
2953         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2954
2955         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2956         pmd_populate(mm, &_pmd, pgtable);
2957
2958         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2959                 pte_t *pte, entry;
2960                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2961                 entry = pte_mkspecial(entry);
2962                 pte = pte_offset_map(&_pmd, haddr);
2963                 VM_BUG_ON(!pte_none(*pte));
2964                 set_pte_at(mm, haddr, pte, entry);
2965                 pte_unmap(pte);
2966         }
2967         smp_wmb(); /* make pte visible before pmd */
2968         pmd_populate(mm, pmd, pgtable);
2969         put_huge_zero_page();
2970 }
2971
2972 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2973                 unsigned long haddr, bool freeze)
2974 {
2975         struct mm_struct *mm = vma->vm_mm;
2976         struct page *page;
2977         pgtable_t pgtable;
2978         pmd_t _pmd;
2979         bool young, write, dirty;
2980         unsigned long addr;
2981         int i;
2982
2983         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2984         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2985         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2986         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2987
2988         count_vm_event(THP_SPLIT_PMD);
2989
2990         if (!vma_is_anonymous(vma)) {
2991                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2992                 if (is_huge_zero_pmd(_pmd))
2993                         put_huge_zero_page();
2994                 if (vma_is_dax(vma))
2995                         return;
2996                 page = pmd_page(_pmd);
2997                 if (!PageReferenced(page) && pmd_young(_pmd))
2998                         SetPageReferenced(page);
2999                 page_remove_rmap(page, true);
3000                 put_page(page);
3001                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
3002                 return;
3003         } else if (is_huge_zero_pmd(*pmd)) {
3004                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
3005         }
3006
3007         page = pmd_page(*pmd);
3008         VM_BUG_ON_PAGE(!page_count(page), page);
3009         page_ref_add(page, HPAGE_PMD_NR - 1);
3010         write = pmd_write(*pmd);
3011         young = pmd_young(*pmd);
3012         dirty = pmd_dirty(*pmd);
3013
3014         pmdp_huge_split_prepare(vma, haddr, pmd);
3015         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
3016         pmd_populate(mm, &_pmd, pgtable);
3017
3018         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
3019                 pte_t entry, *pte;
3020                 /*
3021                  * Note that NUMA hinting access restrictions are not
3022                  * transferred to avoid any possibility of altering
3023                  * permissions across VMAs.
3024                  */
3025                 if (freeze) {
3026                         swp_entry_t swp_entry;
3027                         swp_entry = make_migration_entry(page + i, write);
3028                         entry = swp_entry_to_pte(swp_entry);
3029                 } else {
3030                         entry = mk_pte(page + i, vma->vm_page_prot);
3031                         entry = maybe_mkwrite(entry, vma);
3032                         if (!write)
3033                                 entry = pte_wrprotect(entry);
3034                         if (!young)
3035                                 entry = pte_mkold(entry);
3036                 }
3037                 if (dirty)
3038                         SetPageDirty(page + i);
3039                 pte = pte_offset_map(&_pmd, addr);
3040                 BUG_ON(!pte_none(*pte));
3041                 set_pte_at(mm, addr, pte, entry);
3042                 atomic_inc(&page[i]._mapcount);
3043                 pte_unmap(pte);
3044         }
3045
3046         /*
3047          * Set PG_double_map before dropping compound_mapcount to avoid
3048          * false-negative page_mapped().
3049          */
3050         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
3051                 for (i = 0; i < HPAGE_PMD_NR; i++)
3052                         atomic_inc(&page[i]._mapcount);
3053         }
3054
3055         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
3056                 /* Last compound_mapcount is gone. */
3057                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
3058                 if (TestClearPageDoubleMap(page)) {
3059                         /* No need in mapcount reference anymore */
3060                         for (i = 0; i < HPAGE_PMD_NR; i++)
3061                                 atomic_dec(&page[i]._mapcount);
3062                 }
3063         }
3064
3065         smp_wmb(); /* make pte visible before pmd */
3066         /*
3067          * Up to this point the pmd is present and huge and userland has the
3068          * whole access to the hugepage during the split (which happens in
3069          * place). If we overwrite the pmd with the not-huge version pointing
3070          * to the pte here (which of course we could if all CPUs were bug
3071          * free), userland could trigger a small page size TLB miss on the
3072          * small sized TLB while the hugepage TLB entry is still established in
3073          * the huge TLB. Some CPU doesn't like that.
3074          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
3075          * 383 on page 93. Intel should be safe but is also warns that it's
3076          * only safe if the permission and cache attributes of the two entries
3077          * loaded in the two TLB is identical (which should be the case here).
3078          * But it is generally safer to never allow small and huge TLB entries
3079          * for the same virtual address to be loaded simultaneously. So instead
3080          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
3081          * current pmd notpresent (atomically because here the pmd_trans_huge
3082          * and pmd_trans_splitting must remain set at all times on the pmd
3083          * until the split is complete for this pmd), then we flush the SMP TLB
3084          * and finally we write the non-huge version of the pmd entry with
3085          * pmd_populate.
3086          */
3087         pmdp_invalidate(vma, haddr, pmd);
3088         pmd_populate(mm, pmd, pgtable);
3089
3090         if (freeze) {
3091                 for (i = 0; i < HPAGE_PMD_NR; i++) {
3092                         page_remove_rmap(page + i, false);
3093                         put_page(page + i);
3094                 }
3095         }
3096 }
3097
3098 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
3099                 unsigned long address, bool freeze, struct page *page)
3100 {
3101         spinlock_t *ptl;
3102         struct mm_struct *mm = vma->vm_mm;
3103         unsigned long haddr = address & HPAGE_PMD_MASK;
3104
3105         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
3106         ptl = pmd_lock(mm, pmd);
3107
3108         /*
3109          * If caller asks to setup a migration entries, we need a page to check
3110          * pmd against. Otherwise we can end up replacing wrong page.
3111          */
3112         VM_BUG_ON(freeze && !page);
3113         if (page && page != pmd_page(*pmd))
3114                 goto out;
3115
3116         if (pmd_trans_huge(*pmd)) {
3117                 page = pmd_page(*pmd);
3118                 if (PageMlocked(page))
3119                         clear_page_mlock(page);
3120         } else if (!pmd_devmap(*pmd))
3121                 goto out;
3122         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
3123 out:
3124         spin_unlock(ptl);
3125         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
3126 }
3127
3128 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3129                 bool freeze, struct page *page)
3130 {
3131         pgd_t *pgd;
3132         pud_t *pud;
3133         pmd_t *pmd;
3134
3135         pgd = pgd_offset(vma->vm_mm, address);
3136         if (!pgd_present(*pgd))
3137                 return;
3138
3139         pud = pud_offset(pgd, address);
3140         if (!pud_present(*pud))
3141                 return;
3142
3143         pmd = pmd_offset(pud, address);
3144
3145         __split_huge_pmd(vma, pmd, address, freeze, page);
3146 }
3147
3148 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3149                              unsigned long start,
3150                              unsigned long end,
3151                              long adjust_next)
3152 {
3153         /*
3154          * If the new start address isn't hpage aligned and it could
3155          * previously contain an hugepage: check if we need to split
3156          * an huge pmd.
3157          */
3158         if (start & ~HPAGE_PMD_MASK &&
3159             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3160             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3161                 split_huge_pmd_address(vma, start, false, NULL);
3162
3163         /*
3164          * If the new end address isn't hpage aligned and it could
3165          * previously contain an hugepage: check if we need to split
3166          * an huge pmd.
3167          */
3168         if (end & ~HPAGE_PMD_MASK &&
3169             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3170             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3171                 split_huge_pmd_address(vma, end, false, NULL);
3172
3173         /*
3174          * If we're also updating the vma->vm_next->vm_start, if the new
3175          * vm_next->vm_start isn't page aligned and it could previously
3176          * contain an hugepage: check if we need to split an huge pmd.
3177          */
3178         if (adjust_next > 0) {
3179                 struct vm_area_struct *next = vma->vm_next;
3180                 unsigned long nstart = next->vm_start;
3181                 nstart += adjust_next << PAGE_SHIFT;
3182                 if (nstart & ~HPAGE_PMD_MASK &&
3183                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3184                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3185                         split_huge_pmd_address(next, nstart, false, NULL);
3186         }
3187 }
3188
3189 static void freeze_page(struct page *page)
3190 {
3191         enum ttu_flags ttu_flags = TTU_MIGRATION | TTU_IGNORE_MLOCK |
3192                 TTU_IGNORE_ACCESS | TTU_RMAP_LOCKED;
3193         int i, ret;
3194
3195         VM_BUG_ON_PAGE(!PageHead(page), page);
3196
3197         /* We only need TTU_SPLIT_HUGE_PMD once */
3198         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
3199         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
3200                 /* Cut short if the page is unmapped */
3201                 if (page_count(page) == 1)
3202                         return;
3203
3204                 ret = try_to_unmap(page + i, ttu_flags);
3205         }
3206         VM_BUG_ON(ret);
3207 }
3208
3209 static void unfreeze_page(struct page *page)
3210 {
3211         int i;
3212
3213         for (i = 0; i < HPAGE_PMD_NR; i++)
3214                 remove_migration_ptes(page + i, page + i, true);
3215 }
3216
3217 static void __split_huge_page_tail(struct page *head, int tail,
3218                 struct lruvec *lruvec, struct list_head *list)
3219 {
3220         struct page *page_tail = head + tail;
3221
3222         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
3223         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
3224
3225         /*
3226          * tail_page->_refcount is zero and not changing from under us. But
3227          * get_page_unless_zero() may be running from under us on the
3228          * tail_page. If we used atomic_set() below instead of atomic_inc(), we
3229          * would then run atomic_set() concurrently with
3230          * get_page_unless_zero(), and atomic_set() is implemented in C not
3231          * using locked ops. spin_unlock on x86 sometime uses locked ops
3232          * because of PPro errata 66, 92, so unless somebody can guarantee
3233          * atomic_set() here would be safe on all archs (and not only on x86),
3234          * it's safer to use atomic_inc().
3235          */
3236         page_ref_inc(page_tail);
3237
3238         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3239         page_tail->flags |= (head->flags &
3240                         ((1L << PG_referenced) |
3241                          (1L << PG_swapbacked) |
3242                          (1L << PG_mlocked) |
3243                          (1L << PG_uptodate) |
3244                          (1L << PG_active) |
3245                          (1L << PG_locked) |
3246                          (1L << PG_unevictable) |
3247                          (1L << PG_dirty)));
3248
3249         /*
3250          * After clearing PageTail the gup refcount can be released.
3251          * Page flags also must be visible before we make the page non-compound.
3252          */
3253         smp_wmb();
3254
3255         clear_compound_head(page_tail);
3256
3257         if (page_is_young(head))
3258                 set_page_young(page_tail);
3259         if (page_is_idle(head))
3260                 set_page_idle(page_tail);
3261
3262         /* ->mapping in first tail page is compound_mapcount */
3263         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3264                         page_tail);
3265         page_tail->mapping = head->mapping;
3266
3267         page_tail->index = head->index + tail;
3268         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3269         lru_add_page_tail(head, page_tail, lruvec, list);
3270 }
3271
3272 static void __split_huge_page(struct page *page, struct list_head *list)
3273 {
3274         struct page *head = compound_head(page);
3275         struct zone *zone = page_zone(head);
3276         struct lruvec *lruvec;
3277         int i;
3278
3279         /* prevent PageLRU to go away from under us, and freeze lru stats */
3280         spin_lock_irq(&zone->lru_lock);
3281         lruvec = mem_cgroup_page_lruvec(head, zone);
3282
3283         /* complete memcg works before add pages to LRU */
3284         mem_cgroup_split_huge_fixup(head);
3285
3286         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3287                 __split_huge_page_tail(head, i, lruvec, list);
3288
3289         ClearPageCompound(head);
3290         spin_unlock_irq(&zone->lru_lock);
3291
3292         unfreeze_page(head);
3293
3294         for (i = 0; i < HPAGE_PMD_NR; i++) {
3295                 struct page *subpage = head + i;
3296                 if (subpage == page)
3297                         continue;
3298                 unlock_page(subpage);
3299
3300                 /*
3301                  * Subpages may be freed if there wasn't any mapping
3302                  * like if add_to_swap() is running on a lru page that
3303                  * had its mapping zapped. And freeing these pages
3304                  * requires taking the lru_lock so we do the put_page
3305                  * of the tail pages after the split is complete.
3306                  */
3307                 put_page(subpage);
3308         }
3309 }
3310
3311 int total_mapcount(struct page *page)
3312 {
3313         int i, compound, ret;
3314
3315         VM_BUG_ON_PAGE(PageTail(page), page);
3316
3317         if (likely(!PageCompound(page)))
3318                 return atomic_read(&page->_mapcount) + 1;
3319
3320         compound = compound_mapcount(page);
3321         if (PageHuge(page))
3322                 return compound;
3323         ret = compound;
3324         for (i = 0; i < HPAGE_PMD_NR; i++)
3325                 ret += atomic_read(&page[i]._mapcount) + 1;
3326         /* File pages has compound_mapcount included in _mapcount */
3327         if (!PageAnon(page))
3328                 return ret - compound * HPAGE_PMD_NR;
3329         if (PageDoubleMap(page))
3330                 ret -= HPAGE_PMD_NR;
3331         return ret;
3332 }
3333
3334 /*
3335  * This calculates accurately how many mappings a transparent hugepage
3336  * has (unlike page_mapcount() which isn't fully accurate). This full
3337  * accuracy is primarily needed to know if copy-on-write faults can
3338  * reuse the page and change the mapping to read-write instead of
3339  * copying them. At the same time this returns the total_mapcount too.
3340  *
3341  * The function returns the highest mapcount any one of the subpages
3342  * has. If the return value is one, even if different processes are
3343  * mapping different subpages of the transparent hugepage, they can
3344  * all reuse it, because each process is reusing a different subpage.
3345  *
3346  * The total_mapcount is instead counting all virtual mappings of the
3347  * subpages. If the total_mapcount is equal to "one", it tells the
3348  * caller all mappings belong to the same "mm" and in turn the
3349  * anon_vma of the transparent hugepage can become the vma->anon_vma
3350  * local one as no other process may be mapping any of the subpages.
3351  *
3352  * It would be more accurate to replace page_mapcount() with
3353  * page_trans_huge_mapcount(), however we only use
3354  * page_trans_huge_mapcount() in the copy-on-write faults where we
3355  * need full accuracy to avoid breaking page pinning, because
3356  * page_trans_huge_mapcount() is slower than page_mapcount().
3357  */
3358 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
3359 {
3360         int i, ret, _total_mapcount, mapcount;
3361
3362         /* hugetlbfs shouldn't call it */
3363         VM_BUG_ON_PAGE(PageHuge(page), page);
3364
3365         if (likely(!PageTransCompound(page))) {
3366                 mapcount = atomic_read(&page->_mapcount) + 1;
3367                 if (total_mapcount)
3368                         *total_mapcount = mapcount;
3369                 return mapcount;
3370         }
3371
3372         page = compound_head(page);
3373
3374         _total_mapcount = ret = 0;
3375         for (i = 0; i < HPAGE_PMD_NR; i++) {
3376                 mapcount = atomic_read(&page[i]._mapcount) + 1;
3377                 ret = max(ret, mapcount);
3378                 _total_mapcount += mapcount;
3379         }
3380         if (PageDoubleMap(page)) {
3381                 ret -= 1;
3382                 _total_mapcount -= HPAGE_PMD_NR;
3383         }
3384         mapcount = compound_mapcount(page);
3385         ret += mapcount;
3386         _total_mapcount += mapcount;
3387         if (total_mapcount)
3388                 *total_mapcount = _total_mapcount;
3389         return ret;
3390 }
3391
3392 /*
3393  * This function splits huge page into normal pages. @page can point to any
3394  * subpage of huge page to split. Split doesn't change the position of @page.
3395  *
3396  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3397  * The huge page must be locked.
3398  *
3399  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3400  *
3401  * Both head page and tail pages will inherit mapping, flags, and so on from
3402  * the hugepage.
3403  *
3404  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3405  * they are not mapped.
3406  *
3407  * Returns 0 if the hugepage is split successfully.
3408  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3409  * us.
3410  */
3411 int split_huge_page_to_list(struct page *page, struct list_head *list)
3412 {
3413         struct page *head = compound_head(page);
3414         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3415         struct anon_vma *anon_vma;
3416         int count, mapcount, ret;
3417         bool mlocked;
3418         unsigned long flags;
3419
3420         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3421         VM_BUG_ON_PAGE(!PageAnon(page), page);
3422         VM_BUG_ON_PAGE(!PageLocked(page), page);
3423         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3424         VM_BUG_ON_PAGE(!PageCompound(page), page);
3425
3426         /*
3427          * The caller does not necessarily hold an mmap_sem that would prevent
3428          * the anon_vma disappearing so we first we take a reference to it
3429          * and then lock the anon_vma for write. This is similar to
3430          * page_lock_anon_vma_read except the write lock is taken to serialise
3431          * against parallel split or collapse operations.
3432          */
3433         anon_vma = page_get_anon_vma(head);
3434         if (!anon_vma) {
3435                 ret = -EBUSY;
3436                 goto out;
3437         }
3438         anon_vma_lock_write(anon_vma);
3439
3440         /*
3441          * Racy check if we can split the page, before freeze_page() will
3442          * split PMDs
3443          */
3444         if (total_mapcount(head) != page_count(head) - 1) {
3445                 ret = -EBUSY;
3446                 goto out_unlock;
3447         }
3448
3449         mlocked = PageMlocked(page);
3450         freeze_page(head);
3451         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3452
3453         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3454         if (mlocked)
3455                 lru_add_drain();
3456
3457         /* Prevent deferred_split_scan() touching ->_refcount */
3458         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3459         count = page_count(head);
3460         mapcount = total_mapcount(head);
3461         if (!mapcount && count == 1) {
3462                 if (!list_empty(page_deferred_list(head))) {
3463                         pgdata->split_queue_len--;
3464                         list_del(page_deferred_list(head));
3465                 }
3466                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3467                 __split_huge_page(page, list);
3468                 ret = 0;
3469         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3470                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3471                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3472                                 mapcount, count);
3473                 if (PageTail(page))
3474                         dump_page(head, NULL);
3475                 dump_page(page, "total_mapcount(head) > 0");
3476                 BUG();
3477         } else {
3478                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3479                 unfreeze_page(head);
3480                 ret = -EBUSY;
3481         }
3482
3483 out_unlock:
3484         anon_vma_unlock_write(anon_vma);
3485         put_anon_vma(anon_vma);
3486 out:
3487         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3488         return ret;
3489 }
3490
3491 void free_transhuge_page(struct page *page)
3492 {
3493         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3494         unsigned long flags;
3495
3496         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3497         if (!list_empty(page_deferred_list(page))) {
3498                 pgdata->split_queue_len--;
3499                 list_del(page_deferred_list(page));
3500         }
3501         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3502         free_compound_page(page);
3503 }
3504
3505 void deferred_split_huge_page(struct page *page)
3506 {
3507         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3508         unsigned long flags;
3509
3510         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3511
3512         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3513         if (list_empty(page_deferred_list(page))) {
3514                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3515                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3516                 pgdata->split_queue_len++;
3517         }
3518         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3519 }
3520
3521 static unsigned long deferred_split_count(struct shrinker *shrink,
3522                 struct shrink_control *sc)
3523 {
3524         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3525         return ACCESS_ONCE(pgdata->split_queue_len);
3526 }
3527
3528 static unsigned long deferred_split_scan(struct shrinker *shrink,
3529                 struct shrink_control *sc)
3530 {
3531         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3532         unsigned long flags;
3533         LIST_HEAD(list), *pos, *next;
3534         struct page *page;
3535         int split = 0;
3536
3537         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3538         /* Take pin on all head pages to avoid freeing them under us */
3539         list_for_each_safe(pos, next, &pgdata->split_queue) {
3540                 page = list_entry((void *)pos, struct page, mapping);
3541                 page = compound_head(page);
3542                 if (get_page_unless_zero(page)) {
3543                         list_move(page_deferred_list(page), &list);
3544                 } else {
3545                         /* We lost race with put_compound_page() */
3546                         list_del_init(page_deferred_list(page));
3547                         pgdata->split_queue_len--;
3548                 }
3549                 if (!--sc->nr_to_scan)
3550                         break;
3551         }
3552         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3553
3554         list_for_each_safe(pos, next, &list) {
3555                 page = list_entry((void *)pos, struct page, mapping);
3556                 lock_page(page);
3557                 /* split_huge_page() removes page from list on success */
3558                 if (!split_huge_page(page))
3559                         split++;
3560                 unlock_page(page);
3561                 put_page(page);
3562         }
3563
3564         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3565         list_splice_tail(&list, &pgdata->split_queue);
3566         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3567
3568         /*
3569          * Stop shrinker if we didn't split any page, but the queue is empty.
3570          * This can happen if pages were freed under us.
3571          */
3572         if (!split && list_empty(&pgdata->split_queue))
3573                 return SHRINK_STOP;
3574         return split;
3575 }
3576
3577 static struct shrinker deferred_split_shrinker = {
3578         .count_objects = deferred_split_count,
3579         .scan_objects = deferred_split_scan,
3580         .seeks = DEFAULT_SEEKS,
3581         .flags = SHRINKER_NUMA_AWARE,
3582 };
3583
3584 #ifdef CONFIG_DEBUG_FS
3585 static int split_huge_pages_set(void *data, u64 val)
3586 {
3587         struct zone *zone;
3588         struct page *page;
3589         unsigned long pfn, max_zone_pfn;
3590         unsigned long total = 0, split = 0;
3591
3592         if (val != 1)
3593                 return -EINVAL;
3594
3595         for_each_populated_zone(zone) {
3596                 max_zone_pfn = zone_end_pfn(zone);
3597                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3598                         if (!pfn_valid(pfn))
3599                                 continue;
3600
3601                         page = pfn_to_page(pfn);
3602                         if (!get_page_unless_zero(page))
3603                                 continue;
3604
3605                         if (zone != page_zone(page))
3606                                 goto next;
3607
3608                         if (!PageHead(page) || !PageAnon(page) ||
3609                                         PageHuge(page))
3610                                 goto next;
3611
3612                         total++;
3613                         lock_page(page);
3614                         if (!split_huge_page(page))
3615                                 split++;
3616                         unlock_page(page);
3617 next:
3618                         put_page(page);
3619                 }
3620         }
3621
3622         pr_info("%lu of %lu THP split\n", split, total);
3623
3624         return 0;
3625 }
3626 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3627                 "%llu\n");
3628
3629 static int __init split_huge_pages_debugfs(void)
3630 {
3631         void *ret;
3632
3633         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3634                         &split_huge_pages_fops);
3635         if (!ret)
3636                 pr_warn("Failed to create split_huge_pages in debugfs");
3637         return 0;
3638 }
3639 late_initcall(split_huge_pages_debugfs);
3640 #endif