mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/dax.h>
  20 #include <linux/kthread.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/mman.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/migrate.h>
  26 #include <linux/hashtable.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/page_idle.h>
  29
  30 #include <asm/tlb.h>
  31 #include <asm/pgalloc.h>
  32 #include "internal.h"
  33
  34 /*
  35  * By default transparent hugepage support is disabled in order that avoid
  36  * to risk increase the memory footprint of applications without a guaranteed
  37  * benefit. When transparent hugepage support is enabled, is for all mappings,
  38  * and khugepaged scans all mappings.
  39  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  40  * for all hugepage allocations.
  41  */
  42 unsigned long transparent_hugepage_flags __read_mostly =
  43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  44         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  45 #endif
  46 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  47         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  48 #endif
  49         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  50         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  51         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  52
  53 /* default scan 8*512 pte (or vmas) every 30 second */
  54 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  55 static unsigned int khugepaged_pages_collapsed;
  56 static unsigned int khugepaged_full_scans;
  57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  58 /* during fragmentation poll the hugepage allocator once every minute */
  59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  60 static struct task_struct *khugepaged_thread __read_mostly;
  61 static DEFINE_MUTEX(khugepaged_mutex);
  62 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  63 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  64 /*
  65  * default collapse hugepages if there is at least one pte mapped like
  66  * it would have happened if the vma was large enough during page
  67  * fault.
  68  */
  69 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  70
  71 static int khugepaged(void *none);
  72 static int khugepaged_slab_init(void);
  73 static void khugepaged_slab_exit(void);
  74
  75 #define MM_SLOTS_HASH_BITS 10
  76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  77
  78 static struct kmem_cache *mm_slot_cache __read_mostly;
  79
  80 /**
  81  * struct mm_slot - hash lookup from mm to mm_slot
  82  * @hash: hash collision list
  83  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  84  * @mm: the mm that this information is valid for
  85  */
  86 struct mm_slot {
  87         struct hlist_node hash;
  88         struct list_head mm_node;
  89         struct mm_struct *mm;
  90 };
  91
  92 /**
  93  * struct khugepaged_scan - cursor for scanning
  94  * @mm_head: the head of the mm list to scan
  95  * @mm_slot: the current mm_slot we are scanning
  96  * @address: the next address inside that to be scanned
  97  *
  98  * There is only the one khugepaged_scan instance of this cursor structure.
  99  */
 100 struct khugepaged_scan {
 101         struct list_head mm_head;
 102         struct mm_slot *mm_slot;
 103         unsigned long address;
 104 };
 105 static struct khugepaged_scan khugepaged_scan = {
 106         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 107 };
 108
 109
 110 static void set_recommended_min_free_kbytes(void)
 111 {
 112         struct zone *zone;
 113         int nr_zones = 0;
 114         unsigned long recommended_min;
 115
 116         for_each_populated_zone(zone)
 117                 nr_zones++;
 118
 119         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 120         recommended_min = pageblock_nr_pages * nr_zones * 2;
 121
 122         /*
 123          * Make sure that on average at least two pageblocks are almost free
 124          * of another type, one for a migratetype to fall back to and a
 125          * second to avoid subsequent fallbacks of other types There are 3
 126          * MIGRATE_TYPES we care about.
 127          */
 128         recommended_min += pageblock_nr_pages * nr_zones *
 129                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 130
 131         /* don't ever allow to reserve more than 5% of the lowmem */
 132         recommended_min = min(recommended_min,
 133                               (unsigned long) nr_free_buffer_pages() / 20);
 134         recommended_min <<= (PAGE_SHIFT-10);
 135
 136         if (recommended_min > min_free_kbytes) {
 137                 if (user_min_free_kbytes >= 0)
 138                         pr_info("raising min_free_kbytes from %d to %lu "
 139                                 "to help transparent hugepage allocations\n",
 140                                 min_free_kbytes, recommended_min);
 141
 142                 min_free_kbytes = recommended_min;
 143         }
 144         setup_per_zone_wmarks();
 145 }
 146
 147 static int start_stop_khugepaged(void)
 148 {
 149         int err = 0;
 150         if (khugepaged_enabled()) {
 151                 if (!khugepaged_thread)
 152                         khugepaged_thread = kthread_run(khugepaged, NULL,
 153                                                         "khugepaged");
 154                 if (IS_ERR(khugepaged_thread)) {
 155                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 156                         err = PTR_ERR(khugepaged_thread);
 157                         khugepaged_thread = NULL;
 158                         goto fail;
 159                 }
 160
 161                 if (!list_empty(&khugepaged_scan.mm_head))
 162                         wake_up_interruptible(&khugepaged_wait);
 163
 164                 set_recommended_min_free_kbytes();
 165         } else if (khugepaged_thread) {
 166                 kthread_stop(khugepaged_thread);
 167                 khugepaged_thread = NULL;
 168         }
 169 fail:
 170         return err;
 171 }
 172
 173 static atomic_t huge_zero_refcount;
 174 struct page *huge_zero_page __read_mostly;
 175
 176 struct page *get_huge_zero_page(void)
 177 {
 178         struct page *zero_page;
 179 retry:
 180         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 181                 return READ_ONCE(huge_zero_page);
 182
 183         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 184                         HPAGE_PMD_ORDER);
 185         if (!zero_page) {
 186                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 187                 return NULL;
 188         }
 189         count_vm_event(THP_ZERO_PAGE_ALLOC);
 190         preempt_disable();
 191         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 192                 preempt_enable();
 193                 __free_pages(zero_page, compound_order(zero_page));
 194                 goto retry;
 195         }
 196
 197         /* We take additional reference here. It will be put back by shrinker */
 198         atomic_set(&huge_zero_refcount, 2);
 199         preempt_enable();
 200         return READ_ONCE(huge_zero_page);
 201 }
 202
 203 static void put_huge_zero_page(void)
 204 {
 205         /*
 206          * Counter should never go to zero here. Only shrinker can put
 207          * last reference.
 208          */
 209         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 210 }
 211
 212 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 213                                         struct shrink_control *sc)
 214 {
 215         /* we can free zero page only if last reference remains */
 216         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 217 }
 218
 219 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 220                                        struct shrink_control *sc)
 221 {
 222         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 223                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 224                 BUG_ON(zero_page == NULL);
 225                 __free_pages(zero_page, compound_order(zero_page));
 226                 return HPAGE_PMD_NR;
 227         }
 228
 229         return 0;
 230 }
 231
 232 static struct shrinker huge_zero_page_shrinker = {
 233         .count_objects = shrink_huge_zero_page_count,
 234         .scan_objects = shrink_huge_zero_page_scan,
 235         .seeks = DEFAULT_SEEKS,
 236 };
 237
 238 #ifdef CONFIG_SYSFS
 239
 240 static ssize_t double_flag_show(struct kobject *kobj,
 241                                 struct kobj_attribute *attr, char *buf,
 242                                 enum transparent_hugepage_flag enabled,
 243                                 enum transparent_hugepage_flag req_madv)
 244 {
 245         if (test_bit(enabled, &transparent_hugepage_flags)) {
 246                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 247                 return sprintf(buf, "[always] madvise never\n");
 248         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 249                 return sprintf(buf, "always [madvise] never\n");
 250         else
 251                 return sprintf(buf, "always madvise [never]\n");
 252 }
 253 static ssize_t double_flag_store(struct kobject *kobj,
 254                                  struct kobj_attribute *attr,
 255                                  const char *buf, size_t count,
 256                                  enum transparent_hugepage_flag enabled,
 257                                  enum transparent_hugepage_flag req_madv)
 258 {
 259         if (!memcmp("always", buf,
 260                     min(sizeof("always")-1, count))) {
 261                 set_bit(enabled, &transparent_hugepage_flags);
 262                 clear_bit(req_madv, &transparent_hugepage_flags);
 263         } else if (!memcmp("madvise", buf,
 264                            min(sizeof("madvise")-1, count))) {
 265                 clear_bit(enabled, &transparent_hugepage_flags);
 266                 set_bit(req_madv, &transparent_hugepage_flags);
 267         } else if (!memcmp("never", buf,
 268                            min(sizeof("never")-1, count))) {
 269                 clear_bit(enabled, &transparent_hugepage_flags);
 270                 clear_bit(req_madv, &transparent_hugepage_flags);
 271         } else
 272                 return -EINVAL;
 273
 274         return count;
 275 }
 276
 277 static ssize_t enabled_show(struct kobject *kobj,
 278                             struct kobj_attribute *attr, char *buf)
 279 {
 280         return double_flag_show(kobj, attr, buf,
 281                                 TRANSPARENT_HUGEPAGE_FLAG,
 282                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 283 }
 284 static ssize_t enabled_store(struct kobject *kobj,
 285                              struct kobj_attribute *attr,
 286                              const char *buf, size_t count)
 287 {
 288         ssize_t ret;
 289
 290         ret = double_flag_store(kobj, attr, buf, count,
 291                                 TRANSPARENT_HUGEPAGE_FLAG,
 292                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 293
 294         if (ret > 0) {
 295                 int err;
 296
 297                 mutex_lock(&khugepaged_mutex);
 298                 err = start_stop_khugepaged();
 299                 mutex_unlock(&khugepaged_mutex);
 300
 301                 if (err)
 302                         ret = err;
 303         }
 304
 305         return ret;
 306 }
 307 static struct kobj_attribute enabled_attr =
 308         __ATTR(enabled, 0644, enabled_show, enabled_store);
 309
 310 static ssize_t single_flag_show(struct kobject *kobj,
 311                                 struct kobj_attribute *attr, char *buf,
 312                                 enum transparent_hugepage_flag flag)
 313 {
 314         return sprintf(buf, "%d\n",
 315                        !!test_bit(flag, &transparent_hugepage_flags));
 316 }
 317
 318 static ssize_t single_flag_store(struct kobject *kobj,
 319                                  struct kobj_attribute *attr,
 320                                  const char *buf, size_t count,
 321                                  enum transparent_hugepage_flag flag)
 322 {
 323         unsigned long value;
 324         int ret;
 325
 326         ret = kstrtoul(buf, 10, &value);
 327         if (ret < 0)
 328                 return ret;
 329         if (value > 1)
 330                 return -EINVAL;
 331
 332         if (value)
 333                 set_bit(flag, &transparent_hugepage_flags);
 334         else
 335                 clear_bit(flag, &transparent_hugepage_flags);
 336
 337         return count;
 338 }
 339
 340 /*
 341  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 342  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 343  * memory just to allocate one more hugepage.
 344  */
 345 static ssize_t defrag_show(struct kobject *kobj,
 346                            struct kobj_attribute *attr, char *buf)
 347 {
 348         return double_flag_show(kobj, attr, buf,
 349                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 350                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 351 }
 352 static ssize_t defrag_store(struct kobject *kobj,
 353                             struct kobj_attribute *attr,
 354                             const char *buf, size_t count)
 355 {
 356         return double_flag_store(kobj, attr, buf, count,
 357                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 358                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 359 }
 360 static struct kobj_attribute defrag_attr =
 361         __ATTR(defrag, 0644, defrag_show, defrag_store);
 362
 363 static ssize_t use_zero_page_show(struct kobject *kobj,
 364                 struct kobj_attribute *attr, char *buf)
 365 {
 366         return single_flag_show(kobj, attr, buf,
 367                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 368 }
 369 static ssize_t use_zero_page_store(struct kobject *kobj,
 370                 struct kobj_attribute *attr, const char *buf, size_t count)
 371 {
 372         return single_flag_store(kobj, attr, buf, count,
 373                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 374 }
 375 static struct kobj_attribute use_zero_page_attr =
 376         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 377 #ifdef CONFIG_DEBUG_VM
 378 static ssize_t debug_cow_show(struct kobject *kobj,
 379                                 struct kobj_attribute *attr, char *buf)
 380 {
 381         return single_flag_show(kobj, attr, buf,
 382                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 383 }
 384 static ssize_t debug_cow_store(struct kobject *kobj,
 385                                struct kobj_attribute *attr,
 386                                const char *buf, size_t count)
 387 {
 388         return single_flag_store(kobj, attr, buf, count,
 389                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 390 }
 391 static struct kobj_attribute debug_cow_attr =
 392         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 393 #endif /* CONFIG_DEBUG_VM */
 394
 395 static struct attribute *hugepage_attr[] = {
 396         &enabled_attr.attr,
 397         &defrag_attr.attr,
 398         &use_zero_page_attr.attr,
 399 #ifdef CONFIG_DEBUG_VM
 400         &debug_cow_attr.attr,
 401 #endif
 402         NULL,
 403 };
 404
 405 static struct attribute_group hugepage_attr_group = {
 406         .attrs = hugepage_attr,
 407 };
 408
 409 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 410                                          struct kobj_attribute *attr,
 411                                          char *buf)
 412 {
 413         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 414 }
 415
 416 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 417                                           struct kobj_attribute *attr,
 418                                           const char *buf, size_t count)
 419 {
 420         unsigned long msecs;
 421         int err;
 422
 423         err = kstrtoul(buf, 10, &msecs);
 424         if (err || msecs > UINT_MAX)
 425                 return -EINVAL;
 426
 427         khugepaged_scan_sleep_millisecs = msecs;
 428         wake_up_interruptible(&khugepaged_wait);
 429
 430         return count;
 431 }
 432 static struct kobj_attribute scan_sleep_millisecs_attr =
 433         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 434                scan_sleep_millisecs_store);
 435
 436 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 437                                           struct kobj_attribute *attr,
 438                                           char *buf)
 439 {
 440         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 441 }
 442
 443 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 444                                            struct kobj_attribute *attr,
 445                                            const char *buf, size_t count)
 446 {
 447         unsigned long msecs;
 448         int err;
 449
 450         err = kstrtoul(buf, 10, &msecs);
 451         if (err || msecs > UINT_MAX)
 452                 return -EINVAL;
 453
 454         khugepaged_alloc_sleep_millisecs = msecs;
 455         wake_up_interruptible(&khugepaged_wait);
 456
 457         return count;
 458 }
 459 static struct kobj_attribute alloc_sleep_millisecs_attr =
 460         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 461                alloc_sleep_millisecs_store);
 462
 463 static ssize_t pages_to_scan_show(struct kobject *kobj,
 464                                   struct kobj_attribute *attr,
 465                                   char *buf)
 466 {
 467         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 468 }
 469 static ssize_t pages_to_scan_store(struct kobject *kobj,
 470                                    struct kobj_attribute *attr,
 471                                    const char *buf, size_t count)
 472 {
 473         int err;
 474         unsigned long pages;
 475
 476         err = kstrtoul(buf, 10, &pages);
 477         if (err || !pages || pages > UINT_MAX)
 478                 return -EINVAL;
 479
 480         khugepaged_pages_to_scan = pages;
 481
 482         return count;
 483 }
 484 static struct kobj_attribute pages_to_scan_attr =
 485         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 486                pages_to_scan_store);
 487
 488 static ssize_t pages_collapsed_show(struct kobject *kobj,
 489                                     struct kobj_attribute *attr,
 490                                     char *buf)
 491 {
 492         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 493 }
 494 static struct kobj_attribute pages_collapsed_attr =
 495         __ATTR_RO(pages_collapsed);
 496
 497 static ssize_t full_scans_show(struct kobject *kobj,
 498                                struct kobj_attribute *attr,
 499                                char *buf)
 500 {
 501         return sprintf(buf, "%u\n", khugepaged_full_scans);
 502 }
 503 static struct kobj_attribute full_scans_attr =
 504         __ATTR_RO(full_scans);
 505
 506 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 507                                       struct kobj_attribute *attr, char *buf)
 508 {
 509         return single_flag_show(kobj, attr, buf,
 510                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 511 }
 512 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 513                                        struct kobj_attribute *attr,
 514                                        const char *buf, size_t count)
 515 {
 516         return single_flag_store(kobj, attr, buf, count,
 517                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 518 }
 519 static struct kobj_attribute khugepaged_defrag_attr =
 520         __ATTR(defrag, 0644, khugepaged_defrag_show,
 521                khugepaged_defrag_store);
 522
 523 /*
 524  * max_ptes_none controls if khugepaged should collapse hugepages over
 525  * any unmapped ptes in turn potentially increasing the memory
 526  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 527  * reduce the available free memory in the system as it
 528  * runs. Increasing max_ptes_none will instead potentially reduce the
 529  * free memory in the system during the khugepaged scan.
 530  */
 531 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 532                                              struct kobj_attribute *attr,
 533                                              char *buf)
 534 {
 535         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 536 }
 537 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 538                                               struct kobj_attribute *attr,
 539                                               const char *buf, size_t count)
 540 {
 541         int err;
 542         unsigned long max_ptes_none;
 543
 544         err = kstrtoul(buf, 10, &max_ptes_none);
 545         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 546                 return -EINVAL;
 547
 548         khugepaged_max_ptes_none = max_ptes_none;
 549
 550         return count;
 551 }
 552 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 553         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 554                khugepaged_max_ptes_none_store);
 555
 556 static struct attribute *khugepaged_attr[] = {
 557         &khugepaged_defrag_attr.attr,
 558         &khugepaged_max_ptes_none_attr.attr,
 559         &pages_to_scan_attr.attr,
 560         &pages_collapsed_attr.attr,
 561         &full_scans_attr.attr,
 562         &scan_sleep_millisecs_attr.attr,
 563         &alloc_sleep_millisecs_attr.attr,
 564         NULL,
 565 };
 566
 567 static struct attribute_group khugepaged_attr_group = {
 568         .attrs = khugepaged_attr,
 569         .name = "khugepaged",
 570 };
 571
 572 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 573 {
 574         int err;
 575
 576         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 577         if (unlikely(!*hugepage_kobj)) {
 578                 pr_err("failed to create transparent hugepage kobject\n");
 579                 return -ENOMEM;
 580         }
 581
 582         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 583         if (err) {
 584                 pr_err("failed to register transparent hugepage group\n");
 585                 goto delete_obj;
 586         }
 587
 588         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 589         if (err) {
 590                 pr_err("failed to register transparent hugepage group\n");
 591                 goto remove_hp_group;
 592         }
 593
 594         return 0;
 595
 596 remove_hp_group:
 597         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 598 delete_obj:
 599         kobject_put(*hugepage_kobj);
 600         return err;
 601 }
 602
 603 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 604 {
 605         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 606         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 607         kobject_put(hugepage_kobj);
 608 }
 609 #else
 610 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 611 {
 612         return 0;
 613 }
 614
 615 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 616 {
 617 }
 618 #endif /* CONFIG_SYSFS */
 619
 620 static int __init hugepage_init(void)
 621 {
 622         int err;
 623         struct kobject *hugepage_kobj;
 624
 625         if (!has_transparent_hugepage()) {
 626                 transparent_hugepage_flags = 0;
 627                 return -EINVAL;
 628         }
 629
 630         err = hugepage_init_sysfs(&hugepage_kobj);
 631         if (err)
 632                 goto err_sysfs;
 633
 634         err = khugepaged_slab_init();
 635         if (err)
 636                 goto err_slab;
 637
 638         err = register_shrinker(&huge_zero_page_shrinker);
 639         if (err)
 640                 goto err_hzp_shrinker;
 641
 642         /*
 643          * By default disable transparent hugepages on smaller systems,
 644          * where the extra memory used could hurt more than TLB overhead
 645          * is likely to save.  The admin can still enable it through /sys.
 646          */
 647         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 648                 transparent_hugepage_flags = 0;
 649                 return 0;
 650         }
 651
 652         err = start_stop_khugepaged();
 653         if (err)
 654                 goto err_khugepaged;
 655
 656         return 0;
 657 err_khugepaged:
 658         unregister_shrinker(&huge_zero_page_shrinker);
 659 err_hzp_shrinker:
 660         khugepaged_slab_exit();
 661 err_slab:
 662         hugepage_exit_sysfs(hugepage_kobj);
 663 err_sysfs:
 664         return err;
 665 }
 666 subsys_initcall(hugepage_init);
 667
 668 static int __init setup_transparent_hugepage(char *str)
 669 {
 670         int ret = 0;
 671         if (!str)
 672                 goto out;
 673         if (!strcmp(str, "always")) {
 674                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 675                         &transparent_hugepage_flags);
 676                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 677                           &transparent_hugepage_flags);
 678                 ret = 1;
 679         } else if (!strcmp(str, "madvise")) {
 680                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 681                           &transparent_hugepage_flags);
 682                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 683                         &transparent_hugepage_flags);
 684                 ret = 1;
 685         } else if (!strcmp(str, "never")) {
 686                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 687                           &transparent_hugepage_flags);
 688                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 689                           &transparent_hugepage_flags);
 690                 ret = 1;
 691         }
 692 out:
 693         if (!ret)
 694                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 695         return ret;
 696 }
 697 __setup("transparent_hugepage=", setup_transparent_hugepage);
 698
 699 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 700 {
 701         if (likely(vma->vm_flags & VM_WRITE))
 702                 pmd = pmd_mkwrite(pmd);
 703         return pmd;
 704 }
 705
 706 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 707 {
 708         pmd_t entry;
 709         entry = mk_pmd(page, prot);
 710         entry = pmd_mkhuge(entry);
 711         return entry;
 712 }
 713
 714 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 715                                         struct vm_area_struct *vma,
 716                                         unsigned long address, pmd_t *pmd,
 717                                         struct page *page, gfp_t gfp,
 718                                         unsigned int flags)
 719 {
 720         struct mem_cgroup *memcg;
 721         pgtable_t pgtable;
 722         spinlock_t *ptl;
 723         unsigned long haddr = address & HPAGE_PMD_MASK;
 724
 725         VM_BUG_ON_PAGE(!PageCompound(page), page);
 726
 727         if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
 728                 put_page(page);
 729                 count_vm_event(THP_FAULT_FALLBACK);
 730                 return VM_FAULT_FALLBACK;
 731         }
 732
 733         pgtable = pte_alloc_one(mm, haddr);
 734         if (unlikely(!pgtable)) {
 735                 mem_cgroup_cancel_charge(page, memcg);
 736                 put_page(page);
 737                 return VM_FAULT_OOM;
 738         }
 739
 740         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 741         /*
 742          * The memory barrier inside __SetPageUptodate makes sure that
 743          * clear_huge_page writes become visible before the set_pmd_at()
 744          * write.
 745          */
 746         __SetPageUptodate(page);
 747
 748         ptl = pmd_lock(mm, pmd);
 749         if (unlikely(!pmd_none(*pmd))) {
 750                 spin_unlock(ptl);
 751                 mem_cgroup_cancel_charge(page, memcg);
 752                 put_page(page);
 753                 pte_free(mm, pgtable);
 754         } else {
 755                 pmd_t entry;
 756
 757                 /* Deliver the page fault to userland */
 758                 if (userfaultfd_missing(vma)) {
 759                         int ret;
 760
 761                         spin_unlock(ptl);
 762                         mem_cgroup_cancel_charge(page, memcg);
 763                         put_page(page);
 764                         pte_free(mm, pgtable);
 765                         ret = handle_userfault(vma, address, flags,
 766                                                VM_UFFD_MISSING);
 767                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 768                         return ret;
 769                 }
 770
 771                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 772                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 773                 page_add_new_anon_rmap(page, vma, haddr);
 774                 mem_cgroup_commit_charge(page, memcg, false);
 775                 lru_cache_add_active_or_unevictable(page, vma);
 776                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 777                 set_pmd_at(mm, haddr, pmd, entry);
 778                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 779                 atomic_long_inc(&mm->nr_ptes);
 780                 spin_unlock(ptl);
 781                 count_vm_event(THP_FAULT_ALLOC);
 782         }
 783
 784         return 0;
 785 }
 786
 787 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 788 {
 789         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 790 }
 791
 792 /* Caller must hold page table lock. */
 793 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 794                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 795                 struct page *zero_page)
 796 {
 797         pmd_t entry;
 798         if (!pmd_none(*pmd))
 799                 return false;
 800         entry = mk_pmd(zero_page, vma->vm_page_prot);
 801         entry = pmd_mkhuge(entry);
 802         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 803         set_pmd_at(mm, haddr, pmd, entry);
 804         atomic_long_inc(&mm->nr_ptes);
 805         return true;
 806 }
 807
 808 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 809                                unsigned long address, pmd_t *pmd,
 810                                unsigned int flags)
 811 {
 812         gfp_t gfp;
 813         struct page *page;
 814         unsigned long haddr = address & HPAGE_PMD_MASK;
 815
 816         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 817                 return VM_FAULT_FALLBACK;
 818         if (unlikely(anon_vma_prepare(vma)))
 819                 return VM_FAULT_OOM;
 820         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 821                 return VM_FAULT_OOM;
 822         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 823                         transparent_hugepage_use_zero_page()) {
 824                 spinlock_t *ptl;
 825                 pgtable_t pgtable;
 826                 struct page *zero_page;
 827                 bool set;
 828                 int ret;
 829                 pgtable = pte_alloc_one(mm, haddr);
 830                 if (unlikely(!pgtable))
 831                         return VM_FAULT_OOM;
 832                 zero_page = get_huge_zero_page();
 833                 if (unlikely(!zero_page)) {
 834                         pte_free(mm, pgtable);
 835                         count_vm_event(THP_FAULT_FALLBACK);
 836                         return VM_FAULT_FALLBACK;
 837                 }
 838                 ptl = pmd_lock(mm, pmd);
 839                 ret = 0;
 840                 set = false;
 841                 if (pmd_none(*pmd)) {
 842                         if (userfaultfd_missing(vma)) {
 843                                 spin_unlock(ptl);
 844                                 ret = handle_userfault(vma, address, flags,
 845                                                        VM_UFFD_MISSING);
 846                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 847                         } else {
 848                                 set_huge_zero_page(pgtable, mm, vma,
 849                                                    haddr, pmd,
 850                                                    zero_page);
 851                                 spin_unlock(ptl);
 852                                 set = true;
 853                         }
 854                 } else
 855                         spin_unlock(ptl);
 856                 if (!set) {
 857                         pte_free(mm, pgtable);
 858                         put_huge_zero_page();
 859                 }
 860                 return ret;
 861         }
 862         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 863         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 864         if (unlikely(!page)) {
 865                 count_vm_event(THP_FAULT_FALLBACK);
 866                 return VM_FAULT_FALLBACK;
 867         }
 868         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 869                                             flags);
 870 }
 871
 872 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 873                 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
 874 {
 875         struct mm_struct *mm = vma->vm_mm;
 876         pmd_t entry;
 877         spinlock_t *ptl;
 878
 879         ptl = pmd_lock(mm, pmd);
 880         if (pmd_none(*pmd)) {
 881                 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
 882                 if (write) {
 883                         entry = pmd_mkyoung(pmd_mkdirty(entry));
 884                         entry = maybe_pmd_mkwrite(entry, vma);
 885                 }
 886                 set_pmd_at(mm, addr, pmd, entry);
 887                 update_mmu_cache_pmd(vma, addr, pmd);
 888         }
 889         spin_unlock(ptl);
 890 }
 891
 892 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 893                         pmd_t *pmd, unsigned long pfn, bool write)
 894 {
 895         pgprot_t pgprot = vma->vm_page_prot;
 896         /*
 897          * If we had pmd_special, we could avoid all these restrictions,
 898          * but we need to be consistent with PTEs and architectures that
 899          * can't support a 'special' bit.
 900          */
 901         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 902         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 903                                                 (VM_PFNMAP|VM_MIXEDMAP));
 904         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 905         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
 906
 907         if (addr < vma->vm_start || addr >= vma->vm_end)
 908                 return VM_FAULT_SIGBUS;
 909         if (track_pfn_insert(vma, &pgprot, pfn))
 910                 return VM_FAULT_SIGBUS;
 911         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 912         return VM_FAULT_NOPAGE;
 913 }
 914
 915 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 916                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 917                   struct vm_area_struct *vma)
 918 {
 919         spinlock_t *dst_ptl, *src_ptl;
 920         struct page *src_page;
 921         pmd_t pmd;
 922         pgtable_t pgtable;
 923         int ret;
 924
 925         ret = -ENOMEM;
 926         pgtable = pte_alloc_one(dst_mm, addr);
 927         if (unlikely(!pgtable))
 928                 goto out;
 929
 930         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 931         src_ptl = pmd_lockptr(src_mm, src_pmd);
 932         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 933
 934         ret = -EAGAIN;
 935         pmd = *src_pmd;
 936         if (unlikely(!pmd_trans_huge(pmd))) {
 937                 pte_free(dst_mm, pgtable);
 938                 goto out_unlock;
 939         }
 940         /*
 941          * When page table lock is held, the huge zero pmd should not be
 942          * under splitting since we don't split the page itself, only pmd to
 943          * a page table.
 944          */
 945         if (is_huge_zero_pmd(pmd)) {
 946                 struct page *zero_page;
 947                 /*
 948                  * get_huge_zero_page() will never allocate a new page here,
 949                  * since we already have a zero page to copy. It just takes a
 950                  * reference.
 951                  */
 952                 zero_page = get_huge_zero_page();
 953                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 954                                 zero_page);
 955                 ret = 0;
 956                 goto out_unlock;
 957         }
 958
 959         if (unlikely(pmd_trans_splitting(pmd))) {
 960                 /* split huge page running from under us */
 961                 spin_unlock(src_ptl);
 962                 spin_unlock(dst_ptl);
 963                 pte_free(dst_mm, pgtable);
 964
 965                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 966                 goto out;
 967         }
 968         src_page = pmd_page(pmd);
 969         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 970         get_page(src_page);
 971         page_dup_rmap(src_page);
 972         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 973
 974         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 975         pmd = pmd_mkold(pmd_wrprotect(pmd));
 976         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 977         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 978         atomic_long_inc(&dst_mm->nr_ptes);
 979
 980         ret = 0;
 981 out_unlock:
 982         spin_unlock(src_ptl);
 983         spin_unlock(dst_ptl);
 984 out:
 985         return ret;
 986 }
 987
 988 void huge_pmd_set_accessed(struct mm_struct *mm,
 989                            struct vm_area_struct *vma,
 990                            unsigned long address,
 991                            pmd_t *pmd, pmd_t orig_pmd,
 992                            int dirty)
 993 {
 994         spinlock_t *ptl;
 995         pmd_t entry;
 996         unsigned long haddr;
 997
 998         ptl = pmd_lock(mm, pmd);
 999         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1000                 goto unlock;
1001
1002         entry = pmd_mkyoung(orig_pmd);
1003         haddr = address & HPAGE_PMD_MASK;
1004         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1005                 update_mmu_cache_pmd(vma, address, pmd);
1006
1007 unlock:
1008         spin_unlock(ptl);
1009 }
1010
1011 /*
1012  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1013  * during copy_user_huge_page()'s copy_page_rep(): in the case when
1014  * the source page gets split and a tail freed before copy completes.
1015  * Called under pmd_lock of checked pmd, so safe from splitting itself.
1016  */
1017 static void get_user_huge_page(struct page *page)
1018 {
1019         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1020                 struct page *endpage = page + HPAGE_PMD_NR;
1021
1022                 atomic_add(HPAGE_PMD_NR, &page->_count);
1023                 while (++page < endpage)
1024                         get_huge_page_tail(page);
1025         } else {
1026                 get_page(page);
1027         }
1028 }
1029
1030 static void put_user_huge_page(struct page *page)
1031 {
1032         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1033                 struct page *endpage = page + HPAGE_PMD_NR;
1034
1035                 while (page < endpage)
1036                         put_page(page++);
1037         } else {
1038                 put_page(page);
1039         }
1040 }
1041
1042 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1043                                         struct vm_area_struct *vma,
1044                                         unsigned long address,
1045                                         pmd_t *pmd, pmd_t orig_pmd,
1046                                         struct page *page,
1047                                         unsigned long haddr)
1048 {
1049         struct mem_cgroup *memcg;
1050         spinlock_t *ptl;
1051         pgtable_t pgtable;
1052         pmd_t _pmd;
1053         int ret = 0, i;
1054         struct page **pages;
1055         unsigned long mmun_start;       /* For mmu_notifiers */
1056         unsigned long mmun_end;         /* For mmu_notifiers */
1057
1058         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1059                         GFP_KERNEL);
1060         if (unlikely(!pages)) {
1061                 ret |= VM_FAULT_OOM;
1062                 goto out;
1063         }
1064
1065         for (i = 0; i < HPAGE_PMD_NR; i++) {
1066                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1067                                                __GFP_OTHER_NODE,
1068                                                vma, address, page_to_nid(page));
1069                 if (unlikely(!pages[i] ||
1070                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1071                                                    &memcg))) {
1072                         if (pages[i])
1073                                 put_page(pages[i]);
1074                         while (--i >= 0) {
1075                                 memcg = (void *)page_private(pages[i]);
1076                                 set_page_private(pages[i], 0);
1077                                 mem_cgroup_cancel_charge(pages[i], memcg);
1078                                 put_page(pages[i]);
1079                         }
1080                         kfree(pages);
1081                         ret |= VM_FAULT_OOM;
1082                         goto out;
1083                 }
1084                 set_page_private(pages[i], (unsigned long)memcg);
1085         }
1086
1087         for (i = 0; i < HPAGE_PMD_NR; i++) {
1088                 copy_user_highpage(pages[i], page + i,
1089                                    haddr + PAGE_SIZE * i, vma);
1090                 __SetPageUptodate(pages[i]);
1091                 cond_resched();
1092         }
1093
1094         mmun_start = haddr;
1095         mmun_end   = haddr + HPAGE_PMD_SIZE;
1096         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1097
1098         ptl = pmd_lock(mm, pmd);
1099         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1100                 goto out_free_pages;
1101         VM_BUG_ON_PAGE(!PageHead(page), page);
1102
1103         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1104         /* leave pmd empty until pte is filled */
1105
1106         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1107         pmd_populate(mm, &_pmd, pgtable);
1108
1109         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1110                 pte_t *pte, entry;
1111                 entry = mk_pte(pages[i], vma->vm_page_prot);
1112                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1113                 memcg = (void *)page_private(pages[i]);
1114                 set_page_private(pages[i], 0);
1115                 page_add_new_anon_rmap(pages[i], vma, haddr);
1116                 mem_cgroup_commit_charge(pages[i], memcg, false);
1117                 lru_cache_add_active_or_unevictable(pages[i], vma);
1118                 pte = pte_offset_map(&_pmd, haddr);
1119                 VM_BUG_ON(!pte_none(*pte));
1120                 set_pte_at(mm, haddr, pte, entry);
1121                 pte_unmap(pte);
1122         }
1123         kfree(pages);
1124
1125         smp_wmb(); /* make pte visible before pmd */
1126         pmd_populate(mm, pmd, pgtable);
1127         page_remove_rmap(page);
1128         spin_unlock(ptl);
1129
1130         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1131
1132         ret |= VM_FAULT_WRITE;
1133         put_page(page);
1134
1135 out:
1136         return ret;
1137
1138 out_free_pages:
1139         spin_unlock(ptl);
1140         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1141         for (i = 0; i < HPAGE_PMD_NR; i++) {
1142                 memcg = (void *)page_private(pages[i]);
1143                 set_page_private(pages[i], 0);
1144                 mem_cgroup_cancel_charge(pages[i], memcg);
1145                 put_page(pages[i]);
1146         }
1147         kfree(pages);
1148         goto out;
1149 }
1150
1151 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1152                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1153 {
1154         spinlock_t *ptl;
1155         int ret = 0;
1156         struct page *page = NULL, *new_page;
1157         struct mem_cgroup *memcg;
1158         unsigned long haddr;
1159         unsigned long mmun_start;       /* For mmu_notifiers */
1160         unsigned long mmun_end;         /* For mmu_notifiers */
1161         gfp_t huge_gfp;                 /* for allocation and charge */
1162
1163         ptl = pmd_lockptr(mm, pmd);
1164         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1165         haddr = address & HPAGE_PMD_MASK;
1166         if (is_huge_zero_pmd(orig_pmd))
1167                 goto alloc;
1168         spin_lock(ptl);
1169         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1170                 goto out_unlock;
1171
1172         page = pmd_page(orig_pmd);
1173         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1174         if (page_mapcount(page) == 1) {
1175                 pmd_t entry;
1176                 entry = pmd_mkyoung(orig_pmd);
1177                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1178                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1179                         update_mmu_cache_pmd(vma, address, pmd);
1180                 ret |= VM_FAULT_WRITE;
1181                 goto out_unlock;
1182         }
1183         get_user_huge_page(page);
1184         spin_unlock(ptl);
1185 alloc:
1186         if (transparent_hugepage_enabled(vma) &&
1187             !transparent_hugepage_debug_cow()) {
1188                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1189                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1190         } else
1191                 new_page = NULL;
1192
1193         if (unlikely(!new_page)) {
1194                 if (!page) {
1195                         split_huge_page_pmd(vma, address, pmd);
1196                         ret |= VM_FAULT_FALLBACK;
1197                 } else {
1198                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1199                                         pmd, orig_pmd, page, haddr);
1200                         if (ret & VM_FAULT_OOM) {
1201                                 split_huge_page(page);
1202                                 ret |= VM_FAULT_FALLBACK;
1203                         }
1204                         put_user_huge_page(page);
1205                 }
1206                 count_vm_event(THP_FAULT_FALLBACK);
1207                 goto out;
1208         }
1209
1210         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1211                 put_page(new_page);
1212                 if (page) {
1213                         split_huge_page(page);
1214                         put_user_huge_page(page);
1215                 } else
1216                         split_huge_page_pmd(vma, address, pmd);
1217                 ret |= VM_FAULT_FALLBACK;
1218                 count_vm_event(THP_FAULT_FALLBACK);
1219                 goto out;
1220         }
1221
1222         count_vm_event(THP_FAULT_ALLOC);
1223
1224         if (!page)
1225                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1226         else
1227                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1228         __SetPageUptodate(new_page);
1229
1230         mmun_start = haddr;
1231         mmun_end   = haddr + HPAGE_PMD_SIZE;
1232         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1233
1234         spin_lock(ptl);
1235         if (page)
1236                 put_user_huge_page(page);
1237         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1238                 spin_unlock(ptl);
1239                 mem_cgroup_cancel_charge(new_page, memcg);
1240                 put_page(new_page);
1241                 goto out_mn;
1242         } else {
1243                 pmd_t entry;
1244                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1245                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1246                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1247                 page_add_new_anon_rmap(new_page, vma, haddr);
1248                 mem_cgroup_commit_charge(new_page, memcg, false);
1249                 lru_cache_add_active_or_unevictable(new_page, vma);
1250                 set_pmd_at(mm, haddr, pmd, entry);
1251                 update_mmu_cache_pmd(vma, address, pmd);
1252                 if (!page) {
1253                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1254                         put_huge_zero_page();
1255                 } else {
1256                         VM_BUG_ON_PAGE(!PageHead(page), page);
1257                         page_remove_rmap(page);
1258                         put_page(page);
1259                 }
1260                 ret |= VM_FAULT_WRITE;
1261         }
1262         spin_unlock(ptl);
1263 out_mn:
1264         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1265 out:
1266         return ret;
1267 out_unlock:
1268         spin_unlock(ptl);
1269         return ret;
1270 }
1271
1272 /*
1273  * FOLL_FORCE can write to even unwritable pmd's, but only
1274  * after we've gone through a COW cycle and they are dirty.
1275  */
1276 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1277 {
1278         return pmd_write(pmd) ||
1279                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1280 }
1281
1282 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1283                                    unsigned long addr,
1284                                    pmd_t *pmd,
1285                                    unsigned int flags)
1286 {
1287         struct mm_struct *mm = vma->vm_mm;
1288         struct page *page = NULL;
1289
1290         assert_spin_locked(pmd_lockptr(mm, pmd));
1291
1292         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1293                 goto out;
1294
1295         /* Avoid dumping huge zero page */
1296         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1297                 return ERR_PTR(-EFAULT);
1298
1299         /* Full NUMA hinting faults to serialise migration in fault paths */
1300         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1301                 goto out;
1302
1303         page = pmd_page(*pmd);
1304         VM_BUG_ON_PAGE(!PageHead(page), page);
1305         if (flags & FOLL_TOUCH) {
1306                 pmd_t _pmd;
1307                 _pmd = pmd_mkyoung(*pmd);
1308                 if (flags & FOLL_WRITE)
1309                         _pmd = pmd_mkdirty(_pmd);
1310                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1311                                           pmd, _pmd, flags & FOLL_WRITE))
1312                         update_mmu_cache_pmd(vma, addr, pmd);
1313         }
1314         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1315                 if (page->mapping && trylock_page(page)) {
1316                         lru_add_drain();
1317                         if (page->mapping)
1318                                 mlock_vma_page(page);
1319                         unlock_page(page);
1320                 }
1321         }
1322         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1323         VM_BUG_ON_PAGE(!PageCompound(page), page);
1324         if (flags & FOLL_GET)
1325                 get_page_foll(page);
1326
1327 out:
1328         return page;
1329 }
1330
1331 /* NUMA hinting page fault entry point for trans huge pmds */
1332 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1333                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1334 {
1335         spinlock_t *ptl;
1336         struct anon_vma *anon_vma = NULL;
1337         struct page *page;
1338         unsigned long haddr = addr & HPAGE_PMD_MASK;
1339         int page_nid = -1, this_nid = numa_node_id();
1340         int target_nid, last_cpupid = -1;
1341         bool page_locked;
1342         bool migrated = false;
1343         bool was_writable;
1344         int flags = 0;
1345
1346         /* A PROT_NONE fault should not end up here */
1347         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1348
1349         ptl = pmd_lock(mm, pmdp);
1350         if (unlikely(!pmd_same(pmd, *pmdp)))
1351                 goto out_unlock;
1352
1353         /*
1354          * If there are potential migrations, wait for completion and retry
1355          * without disrupting NUMA hinting information. Do not relock and
1356          * check_same as the page may no longer be mapped.
1357          */
1358         if (unlikely(pmd_trans_migrating(*pmdp))) {
1359                 page = pmd_page(*pmdp);
1360                 if (!get_page_unless_zero(page))
1361                         goto out_unlock;
1362                 spin_unlock(ptl);
1363                 wait_on_page_locked(page);
1364                 put_page(page);
1365                 goto out;
1366         }
1367
1368         page = pmd_page(pmd);
1369         BUG_ON(is_huge_zero_page(page));
1370         page_nid = page_to_nid(page);
1371         last_cpupid = page_cpupid_last(page);
1372         count_vm_numa_event(NUMA_HINT_FAULTS);
1373         if (page_nid == this_nid) {
1374                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1375                 flags |= TNF_FAULT_LOCAL;
1376         }
1377
1378         /* See similar comment in do_numa_page for explanation */
1379         if (!(vma->vm_flags & VM_WRITE))
1380                 flags |= TNF_NO_GROUP;
1381
1382         /*
1383          * Acquire the page lock to serialise THP migrations but avoid dropping
1384          * page_table_lock if at all possible
1385          */
1386         page_locked = trylock_page(page);
1387         target_nid = mpol_misplaced(page, vma, haddr);
1388         if (target_nid == -1) {
1389                 /* If the page was locked, there are no parallel migrations */
1390                 if (page_locked)
1391                         goto clear_pmdnuma;
1392         }
1393
1394         /* Migration could have started since the pmd_trans_migrating check */
1395         if (!page_locked) {
1396                 page_nid = -1;
1397                 if (!get_page_unless_zero(page))
1398                         goto out_unlock;
1399                 spin_unlock(ptl);
1400                 wait_on_page_locked(page);
1401                 put_page(page);
1402                 goto out;
1403         }
1404
1405         /*
1406          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1407          * to serialises splits
1408          */
1409         get_page(page);
1410         spin_unlock(ptl);
1411         anon_vma = page_lock_anon_vma_read(page);
1412
1413         /* Confirm the PMD did not change while page_table_lock was released */
1414         spin_lock(ptl);
1415         if (unlikely(!pmd_same(pmd, *pmdp))) {
1416                 unlock_page(page);
1417                 put_page(page);
1418                 page_nid = -1;
1419                 goto out_unlock;
1420         }
1421
1422         /* Bail if we fail to protect against THP splits for any reason */
1423         if (unlikely(!anon_vma)) {
1424                 put_page(page);
1425                 page_nid = -1;
1426                 goto clear_pmdnuma;
1427         }
1428
1429         /*
1430          * Migrate the THP to the requested node, returns with page unlocked
1431          * and access rights restored.
1432          */
1433         spin_unlock(ptl);
1434         migrated = migrate_misplaced_transhuge_page(mm, vma,
1435                                 pmdp, pmd, addr, page, target_nid);
1436         if (migrated) {
1437                 flags |= TNF_MIGRATED;
1438                 page_nid = target_nid;
1439         } else
1440                 flags |= TNF_MIGRATE_FAIL;
1441
1442         goto out;
1443 clear_pmdnuma:
1444         BUG_ON(!PageLocked(page));
1445         was_writable = pmd_write(pmd);
1446         pmd = pmd_modify(pmd, vma->vm_page_prot);
1447         pmd = pmd_mkyoung(pmd);
1448         if (was_writable)
1449                 pmd = pmd_mkwrite(pmd);
1450         set_pmd_at(mm, haddr, pmdp, pmd);
1451         update_mmu_cache_pmd(vma, addr, pmdp);
1452         unlock_page(page);
1453 out_unlock:
1454         spin_unlock(ptl);
1455
1456 out:
1457         if (anon_vma)
1458                 page_unlock_anon_vma_read(anon_vma);
1459
1460         if (page_nid != -1)
1461                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1462
1463         return 0;
1464 }
1465
1466 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1467                  pmd_t *pmd, unsigned long addr)
1468 {
1469         pmd_t orig_pmd;
1470         spinlock_t *ptl;
1471
1472         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1473                 return 0;
1474         /*
1475          * For architectures like ppc64 we look at deposited pgtable
1476          * when calling pmdp_huge_get_and_clear. So do the
1477          * pgtable_trans_huge_withdraw after finishing pmdp related
1478          * operations.
1479          */
1480         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1481                         tlb->fullmm);
1482         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1483         if (vma_is_dax(vma)) {
1484                 spin_unlock(ptl);
1485                 if (is_huge_zero_pmd(orig_pmd))
1486                         put_huge_zero_page();
1487         } else if (is_huge_zero_pmd(orig_pmd)) {
1488                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1489                 atomic_long_dec(&tlb->mm->nr_ptes);
1490                 spin_unlock(ptl);
1491                 put_huge_zero_page();
1492         } else {
1493                 struct page *page = pmd_page(orig_pmd);
1494                 page_remove_rmap(page);
1495                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1496                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1497                 VM_BUG_ON_PAGE(!PageHead(page), page);
1498                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1499                 atomic_long_dec(&tlb->mm->nr_ptes);
1500                 spin_unlock(ptl);
1501                 tlb_remove_page(tlb, page);
1502         }
1503         return 1;
1504 }
1505
1506 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1507                   unsigned long old_addr,
1508                   unsigned long new_addr, unsigned long old_end,
1509                   pmd_t *old_pmd, pmd_t *new_pmd)
1510 {
1511         spinlock_t *old_ptl, *new_ptl;
1512         int ret = 0;
1513         pmd_t pmd;
1514         bool force_flush = false;
1515         struct mm_struct *mm = vma->vm_mm;
1516
1517         if ((old_addr & ~HPAGE_PMD_MASK) ||
1518             (new_addr & ~HPAGE_PMD_MASK) ||
1519             old_end - old_addr < HPAGE_PMD_SIZE ||
1520             (new_vma->vm_flags & VM_NOHUGEPAGE))
1521                 goto out;
1522
1523         /*
1524          * The destination pmd shouldn't be established, free_pgtables()
1525          * should have release it.
1526          */
1527         if (WARN_ON(!pmd_none(*new_pmd))) {
1528                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1529                 goto out;
1530         }
1531
1532         /*
1533          * We don't have to worry about the ordering of src and dst
1534          * ptlocks because exclusive mmap_sem prevents deadlock.
1535          */
1536         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1537         if (ret == 1) {
1538                 new_ptl = pmd_lockptr(mm, new_pmd);
1539                 if (new_ptl != old_ptl)
1540                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1541                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1542                 if (pmd_present(pmd))
1543                         force_flush = true;
1544                 VM_BUG_ON(!pmd_none(*new_pmd));
1545
1546                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1547                         pgtable_t pgtable;
1548                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1549                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1550                 }
1551                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1552                 if (force_flush)
1553                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1554                 if (new_ptl != old_ptl)
1555                         spin_unlock(new_ptl);
1556                 spin_unlock(old_ptl);
1557         }
1558 out:
1559         return ret;
1560 }
1561
1562 /*
1563  * Returns
1564  *  - 0 if PMD could not be locked
1565  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1566  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1567  */
1568 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1569                 unsigned long addr, pgprot_t newprot, int prot_numa)
1570 {
1571         struct mm_struct *mm = vma->vm_mm;
1572         spinlock_t *ptl;
1573         pmd_t entry;
1574         bool preserve_write;
1575
1576         int ret = 0;
1577
1578         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1579                 return 0;
1580
1581         preserve_write = prot_numa && pmd_write(*pmd);
1582         ret = 1;
1583
1584         /*
1585          * Avoid trapping faults against the zero page. The read-only
1586          * data is likely to be read-cached on the local CPU and
1587          * local/remote hits to the zero page are not interesting.
1588          */
1589         if (prot_numa && is_huge_zero_pmd(*pmd))
1590                 goto unlock;
1591
1592         if (prot_numa && pmd_protnone(*pmd))
1593                 goto unlock;
1594
1595         /*
1596          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1597          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1598          * which is also under down_read(mmap_sem):
1599          *
1600          *      CPU0:                           CPU1:
1601          *                              change_huge_pmd(prot_numa=1)
1602          *                               pmdp_huge_get_and_clear_notify()
1603          * madvise_dontneed()
1604          *  zap_pmd_range()
1605          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1606          *   // skip the pmd
1607          *                               set_pmd_at();
1608          *                               // pmd is re-established
1609          *
1610          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1611          * which may break userspace.
1612          *
1613          * pmdp_invalidate() is required to make sure we don't miss
1614          * dirty/young flags set by hardware.
1615          */
1616         entry = *pmd;
1617         pmdp_invalidate(vma, addr, pmd);
1618
1619         /*
1620          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1621          * corrupt them.
1622          */
1623         if (pmd_dirty(*pmd))
1624                 entry = pmd_mkdirty(entry);
1625         if (pmd_young(*pmd))
1626                 entry = pmd_mkyoung(entry);
1627
1628         entry = pmd_modify(entry, newprot);
1629         if (preserve_write)
1630                 entry = pmd_mkwrite(entry);
1631         ret = HPAGE_PMD_NR;
1632         set_pmd_at(mm, addr, pmd, entry);
1633         BUG_ON(!preserve_write && pmd_write(entry));
1634 unlock:
1635         spin_unlock(ptl);
1636         return ret;
1637 }
1638
1639 /*
1640  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1641  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1642  *
1643  * Note that if it returns 1, this routine returns without unlocking page
1644  * table locks. So callers must unlock them.
1645  */
1646 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1647                 spinlock_t **ptl)
1648 {
1649         *ptl = pmd_lock(vma->vm_mm, pmd);
1650         if (likely(pmd_trans_huge(*pmd))) {
1651                 if (unlikely(pmd_trans_splitting(*pmd))) {
1652                         spin_unlock(*ptl);
1653                         wait_split_huge_page(vma->anon_vma, pmd);
1654                         return -1;
1655                 } else {
1656                         /* Thp mapped by 'pmd' is stable, so we can
1657                          * handle it as it is. */
1658                         return 1;
1659                 }
1660         }
1661         spin_unlock(*ptl);
1662         return 0;
1663 }
1664
1665 /*
1666  * This function returns whether a given @page is mapped onto the @address
1667  * in the virtual space of @mm.
1668  *
1669  * When it's true, this function returns *pmd with holding the page table lock
1670  * and passing it back to the caller via @ptl.
1671  * If it's false, returns NULL without holding the page table lock.
1672  */
1673 pmd_t *page_check_address_pmd(struct page *page,
1674                               struct mm_struct *mm,
1675                               unsigned long address,
1676                               enum page_check_address_pmd_flag flag,
1677                               spinlock_t **ptl)
1678 {
1679         pgd_t *pgd;
1680         pud_t *pud;
1681         pmd_t *pmd;
1682
1683         if (address & ~HPAGE_PMD_MASK)
1684                 return NULL;
1685
1686         pgd = pgd_offset(mm, address);
1687         if (!pgd_present(*pgd))
1688                 return NULL;
1689         pud = pud_offset(pgd, address);
1690         if (!pud_present(*pud))
1691                 return NULL;
1692         pmd = pmd_offset(pud, address);
1693
1694         *ptl = pmd_lock(mm, pmd);
1695         if (!pmd_present(*pmd))
1696                 goto unlock;
1697         if (pmd_page(*pmd) != page)
1698                 goto unlock;
1699         /*
1700          * split_vma() may create temporary aliased mappings. There is
1701          * no risk as long as all huge pmd are found and have their
1702          * splitting bit set before __split_huge_page_refcount
1703          * runs. Finding the same huge pmd more than once during the
1704          * same rmap walk is not a problem.
1705          */
1706         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1707             pmd_trans_splitting(*pmd))
1708                 goto unlock;
1709         if (pmd_trans_huge(*pmd)) {
1710                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1711                           !pmd_trans_splitting(*pmd));
1712                 return pmd;
1713         }
1714 unlock:
1715         spin_unlock(*ptl);
1716         return NULL;
1717 }
1718
1719 static int __split_huge_page_splitting(struct page *page,
1720                                        struct vm_area_struct *vma,
1721                                        unsigned long address)
1722 {
1723         struct mm_struct *mm = vma->vm_mm;
1724         spinlock_t *ptl;
1725         pmd_t *pmd;
1726         int ret = 0;
1727         /* For mmu_notifiers */
1728         const unsigned long mmun_start = address;
1729         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1730
1731         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1732         pmd = page_check_address_pmd(page, mm, address,
1733                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1734         if (pmd) {
1735                 /*
1736                  * We can't temporarily set the pmd to null in order
1737                  * to split it, the pmd must remain marked huge at all
1738                  * times or the VM won't take the pmd_trans_huge paths
1739                  * and it won't wait on the anon_vma->root->rwsem to
1740                  * serialize against split_huge_page*.
1741                  */
1742                 pmdp_splitting_flush(vma, address, pmd);
1743
1744                 ret = 1;
1745                 spin_unlock(ptl);
1746         }
1747         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1748
1749         return ret;
1750 }
1751
1752 static void __split_huge_page_refcount(struct page *page,
1753                                        struct list_head *list)
1754 {
1755         int i;
1756         struct zone *zone = page_zone(page);
1757         struct lruvec *lruvec;
1758         int tail_count = 0;
1759
1760         /* prevent PageLRU to go away from under us, and freeze lru stats */
1761         spin_lock_irq(&zone->lru_lock);
1762         lruvec = mem_cgroup_page_lruvec(page, zone);
1763
1764         compound_lock(page);
1765         /* complete memcg works before add pages to LRU */
1766         mem_cgroup_split_huge_fixup(page);
1767
1768         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1769                 struct page *page_tail = page + i;
1770
1771                 /* tail_page->_mapcount cannot change */
1772                 BUG_ON(page_mapcount(page_tail) < 0);
1773                 tail_count += page_mapcount(page_tail);
1774                 /* check for overflow */
1775                 BUG_ON(tail_count < 0);
1776                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1777                 /*
1778                  * tail_page->_count is zero and not changing from
1779                  * under us. But get_page_unless_zero() may be running
1780                  * from under us on the tail_page. If we used
1781                  * atomic_set() below instead of atomic_add(), we
1782                  * would then run atomic_set() concurrently with
1783                  * get_page_unless_zero(), and atomic_set() is
1784                  * implemented in C not using locked ops. spin_unlock
1785                  * on x86 sometime uses locked ops because of PPro
1786                  * errata 66, 92, so unless somebody can guarantee
1787                  * atomic_set() here would be safe on all archs (and
1788                  * not only on x86), it's safer to use atomic_add().
1789                  */
1790                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1791                            &page_tail->_count);
1792
1793                 /* after clearing PageTail the gup refcount can be released */
1794                 smp_mb__after_atomic();
1795
1796                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1797                 page_tail->flags |= (page->flags &
1798                                      ((1L << PG_referenced) |
1799                                       (1L << PG_swapbacked) |
1800                                       (1L << PG_mlocked) |
1801                                       (1L << PG_uptodate) |
1802                                       (1L << PG_active) |
1803                                       (1L << PG_unevictable)));
1804                 page_tail->flags |= (1L << PG_dirty);
1805
1806                 clear_compound_head(page_tail);
1807
1808                 if (page_is_young(page))
1809                         set_page_young(page_tail);
1810                 if (page_is_idle(page))
1811                         set_page_idle(page_tail);
1812
1813                 /*
1814                  * __split_huge_page_splitting() already set the
1815                  * splitting bit in all pmd that could map this
1816                  * hugepage, that will ensure no CPU can alter the
1817                  * mapcount on the head page. The mapcount is only
1818                  * accounted in the head page and it has to be
1819                  * transferred to all tail pages in the below code. So
1820                  * for this code to be safe, the split the mapcount
1821                  * can't change. But that doesn't mean userland can't
1822                  * keep changing and reading the page contents while
1823                  * we transfer the mapcount, so the pmd splitting
1824                  * status is achieved setting a reserved bit in the
1825                  * pmd, not by clearing the present bit.
1826                 */
1827                 page_tail->_mapcount = page->_mapcount;
1828
1829                 BUG_ON(page_tail->mapping);
1830                 page_tail->mapping = page->mapping;
1831
1832                 page_tail->index = page->index + i;
1833                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1834
1835                 BUG_ON(!PageAnon(page_tail));
1836                 BUG_ON(!PageUptodate(page_tail));
1837                 BUG_ON(!PageDirty(page_tail));
1838                 BUG_ON(!PageSwapBacked(page_tail));
1839
1840                 lru_add_page_tail(page, page_tail, lruvec, list);
1841         }
1842         atomic_sub(tail_count, &page->_count);
1843         BUG_ON(atomic_read(&page->_count) <= 0);
1844
1845         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1846
1847         ClearPageCompound(page);
1848         compound_unlock(page);
1849         spin_unlock_irq(&zone->lru_lock);
1850
1851         for (i = 1; i < HPAGE_PMD_NR; i++) {
1852                 struct page *page_tail = page + i;
1853                 BUG_ON(page_count(page_tail) <= 0);
1854                 /*
1855                  * Tail pages may be freed if there wasn't any mapping
1856                  * like if add_to_swap() is running on a lru page that
1857                  * had its mapping zapped. And freeing these pages
1858                  * requires taking the lru_lock so we do the put_page
1859                  * of the tail pages after the split is complete.
1860                  */
1861                 put_page(page_tail);
1862         }
1863
1864         /*
1865          * Only the head page (now become a regular page) is required
1866          * to be pinned by the caller.
1867          */
1868         BUG_ON(page_count(page) <= 0);
1869 }
1870
1871 static int __split_huge_page_map(struct page *page,
1872                                  struct vm_area_struct *vma,
1873                                  unsigned long address)
1874 {
1875         struct mm_struct *mm = vma->vm_mm;
1876         spinlock_t *ptl;
1877         pmd_t *pmd, _pmd;
1878         int ret = 0, i;
1879         pgtable_t pgtable;
1880         unsigned long haddr;
1881
1882         pmd = page_check_address_pmd(page, mm, address,
1883                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1884         if (pmd) {
1885                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1886                 pmd_populate(mm, &_pmd, pgtable);
1887                 if (pmd_write(*pmd))
1888                         BUG_ON(page_mapcount(page) != 1);
1889
1890                 haddr = address;
1891                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1892                         pte_t *pte, entry;
1893                         BUG_ON(PageCompound(page+i));
1894                         /*
1895                          * Note that NUMA hinting access restrictions are not
1896                          * transferred to avoid any possibility of altering
1897                          * permissions across VMAs.
1898                          */
1899                         entry = mk_pte(page + i, vma->vm_page_prot);
1900                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1901                         if (!pmd_write(*pmd))
1902                                 entry = pte_wrprotect(entry);
1903                         if (!pmd_young(*pmd))
1904                                 entry = pte_mkold(entry);
1905                         pte = pte_offset_map(&_pmd, haddr);
1906                         BUG_ON(!pte_none(*pte));
1907                         set_pte_at(mm, haddr, pte, entry);
1908                         pte_unmap(pte);
1909                 }
1910
1911                 smp_wmb(); /* make pte visible before pmd */
1912                 /*
1913                  * Up to this point the pmd is present and huge and
1914                  * userland has the whole access to the hugepage
1915                  * during the split (which happens in place). If we
1916                  * overwrite the pmd with the not-huge version
1917                  * pointing to the pte here (which of course we could
1918                  * if all CPUs were bug free), userland could trigger
1919                  * a small page size TLB miss on the small sized TLB
1920                  * while the hugepage TLB entry is still established
1921                  * in the huge TLB. Some CPU doesn't like that. See
1922                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1923                  * Erratum 383 on page 93. Intel should be safe but is
1924                  * also warns that it's only safe if the permission
1925                  * and cache attributes of the two entries loaded in
1926                  * the two TLB is identical (which should be the case
1927                  * here). But it is generally safer to never allow
1928                  * small and huge TLB entries for the same virtual
1929                  * address to be loaded simultaneously. So instead of
1930                  * doing "pmd_populate(); flush_pmd_tlb_range();" we first
1931                  * mark the current pmd notpresent (atomically because
1932                  * here the pmd_trans_huge and pmd_trans_splitting
1933                  * must remain set at all times on the pmd until the
1934                  * split is complete for this pmd), then we flush the
1935                  * SMP TLB and finally we write the non-huge version
1936                  * of the pmd entry with pmd_populate.
1937                  */
1938                 pmdp_invalidate(vma, address, pmd);
1939                 pmd_populate(mm, pmd, pgtable);
1940                 ret = 1;
1941                 spin_unlock(ptl);
1942         }
1943
1944         return ret;
1945 }
1946
1947 /* must be called with anon_vma->root->rwsem held */
1948 static void __split_huge_page(struct page *page,
1949                               struct anon_vma *anon_vma,
1950                               struct list_head *list)
1951 {
1952         int mapcount, mapcount2;
1953         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1954         struct anon_vma_chain *avc;
1955
1956         BUG_ON(!PageHead(page));
1957         BUG_ON(PageTail(page));
1958
1959         mapcount = 0;
1960         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1961                 struct vm_area_struct *vma = avc->vma;
1962                 unsigned long addr = vma_address(page, vma);
1963                 BUG_ON(is_vma_temporary_stack(vma));
1964                 mapcount += __split_huge_page_splitting(page, vma, addr);
1965         }
1966         /*
1967          * It is critical that new vmas are added to the tail of the
1968          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1969          * and establishes a child pmd before
1970          * __split_huge_page_splitting() freezes the parent pmd (so if
1971          * we fail to prevent copy_huge_pmd() from running until the
1972          * whole __split_huge_page() is complete), we will still see
1973          * the newly established pmd of the child later during the
1974          * walk, to be able to set it as pmd_trans_splitting too.
1975          */
1976         if (mapcount != page_mapcount(page)) {
1977                 pr_err("mapcount %d page_mapcount %d\n",
1978                         mapcount, page_mapcount(page));
1979                 BUG();
1980         }
1981
1982         __split_huge_page_refcount(page, list);
1983
1984         mapcount2 = 0;
1985         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1986                 struct vm_area_struct *vma = avc->vma;
1987                 unsigned long addr = vma_address(page, vma);
1988                 BUG_ON(is_vma_temporary_stack(vma));
1989                 mapcount2 += __split_huge_page_map(page, vma, addr);
1990         }
1991         if (mapcount != mapcount2) {
1992                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1993                         mapcount, mapcount2, page_mapcount(page));
1994                 BUG();
1995         }
1996 }
1997
1998 /*
1999  * Split a hugepage into normal pages. This doesn't change the position of head
2000  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
2001  * @list. Both head page and tail pages will inherit mapping, flags, and so on
2002  * from the hugepage.
2003  * Return 0 if the hugepage is split successfully otherwise return 1.
2004  */
2005 int split_huge_page_to_list(struct page *page, struct list_head *list)
2006 {
2007         struct anon_vma *anon_vma;
2008         int ret = 1;
2009
2010         BUG_ON(is_huge_zero_page(page));
2011         BUG_ON(!PageAnon(page));
2012
2013         /*
2014          * The caller does not necessarily hold an mmap_sem that would prevent
2015          * the anon_vma disappearing so we first we take a reference to it
2016          * and then lock the anon_vma for write. This is similar to
2017          * page_lock_anon_vma_read except the write lock is taken to serialise
2018          * against parallel split or collapse operations.
2019          */
2020         anon_vma = page_get_anon_vma(page);
2021         if (!anon_vma)
2022                 goto out;
2023         anon_vma_lock_write(anon_vma);
2024
2025         ret = 0;
2026         if (!PageCompound(page))
2027                 goto out_unlock;
2028
2029         BUG_ON(!PageSwapBacked(page));
2030         __split_huge_page(page, anon_vma, list);
2031         count_vm_event(THP_SPLIT);
2032
2033         BUG_ON(PageCompound(page));
2034 out_unlock:
2035         anon_vma_unlock_write(anon_vma);
2036         put_anon_vma(anon_vma);
2037 out:
2038         return ret;
2039 }
2040
2041 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
2042
2043 int hugepage_madvise(struct vm_area_struct *vma,
2044                      unsigned long *vm_flags, int advice)
2045 {
2046         switch (advice) {
2047         case MADV_HUGEPAGE:
2048 #ifdef CONFIG_S390
2049                 /*
2050                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2051                  * can't handle this properly after s390_enable_sie, so we simply
2052                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
2053                  */
2054                 if (mm_has_pgste(vma->vm_mm))
2055                         return 0;
2056 #endif
2057                 /*
2058                  * Be somewhat over-protective like KSM for now!
2059                  */
2060                 if (*vm_flags & VM_NO_THP)
2061                         return -EINVAL;
2062                 *vm_flags &= ~VM_NOHUGEPAGE;
2063                 *vm_flags |= VM_HUGEPAGE;
2064                 /*
2065                  * If the vma become good for khugepaged to scan,
2066                  * register it here without waiting a page fault that
2067                  * may not happen any time soon.
2068                  */
2069                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2070                         return -ENOMEM;
2071                 break;
2072         case MADV_NOHUGEPAGE:
2073                 /*
2074                  * Be somewhat over-protective like KSM for now!
2075                  */
2076                 if (*vm_flags & VM_NO_THP)
2077                         return -EINVAL;
2078                 *vm_flags &= ~VM_HUGEPAGE;
2079                 *vm_flags |= VM_NOHUGEPAGE;
2080                 /*
2081                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2082                  * this vma even if we leave the mm registered in khugepaged if
2083                  * it got registered before VM_NOHUGEPAGE was set.
2084                  */
2085                 break;
2086         }
2087
2088         return 0;
2089 }
2090
2091 static int __init khugepaged_slab_init(void)
2092 {
2093         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2094                                           sizeof(struct mm_slot),
2095                                           __alignof__(struct mm_slot), 0, NULL);
2096         if (!mm_slot_cache)
2097                 return -ENOMEM;
2098
2099         return 0;
2100 }
2101
2102 static void __init khugepaged_slab_exit(void)
2103 {
2104         kmem_cache_destroy(mm_slot_cache);
2105 }
2106
2107 static inline struct mm_slot *alloc_mm_slot(void)
2108 {
2109         if (!mm_slot_cache)     /* initialization failed */
2110                 return NULL;
2111         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2112 }
2113
2114 static inline void free_mm_slot(struct mm_slot *mm_slot)
2115 {
2116         kmem_cache_free(mm_slot_cache, mm_slot);
2117 }
2118
2119 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2120 {
2121         struct mm_slot *mm_slot;
2122
2123         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2124                 if (mm == mm_slot->mm)
2125                         return mm_slot;
2126
2127         return NULL;
2128 }
2129
2130 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2131                                     struct mm_slot *mm_slot)
2132 {
2133         mm_slot->mm = mm;
2134         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2135 }
2136
2137 static inline int khugepaged_test_exit(struct mm_struct *mm)
2138 {
2139         return atomic_read(&mm->mm_users) == 0;
2140 }
2141
2142 int __khugepaged_enter(struct mm_struct *mm)
2143 {
2144         struct mm_slot *mm_slot;
2145         int wakeup;
2146
2147         mm_slot = alloc_mm_slot();
2148         if (!mm_slot)
2149                 return -ENOMEM;
2150
2151         /* __khugepaged_exit() must not run from under us */
2152         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2153         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2154                 free_mm_slot(mm_slot);
2155                 return 0;
2156         }
2157
2158         spin_lock(&khugepaged_mm_lock);
2159         insert_to_mm_slots_hash(mm, mm_slot);
2160         /*
2161          * Insert just behind the scanning cursor, to let the area settle
2162          * down a little.
2163          */
2164         wakeup = list_empty(&khugepaged_scan.mm_head);
2165         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2166         spin_unlock(&khugepaged_mm_lock);
2167
2168         atomic_inc(&mm->mm_count);
2169         if (wakeup)
2170                 wake_up_interruptible(&khugepaged_wait);
2171
2172         return 0;
2173 }
2174
2175 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2176                                unsigned long vm_flags)
2177 {
2178         unsigned long hstart, hend;
2179         if (!vma->anon_vma)
2180                 /*
2181                  * Not yet faulted in so we will register later in the
2182                  * page fault if needed.
2183                  */
2184                 return 0;
2185         if (vma->vm_ops || (vm_flags & VM_NO_THP))
2186                 /* khugepaged not yet working on file or special mappings */
2187                 return 0;
2188         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2189         hend = vma->vm_end & HPAGE_PMD_MASK;
2190         if (hstart < hend)
2191                 return khugepaged_enter(vma, vm_flags);
2192         return 0;
2193 }
2194
2195 void __khugepaged_exit(struct mm_struct *mm)
2196 {
2197         struct mm_slot *mm_slot;
2198         int free = 0;
2199
2200         spin_lock(&khugepaged_mm_lock);
2201         mm_slot = get_mm_slot(mm);
2202         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2203                 hash_del(&mm_slot->hash);
2204                 list_del(&mm_slot->mm_node);
2205                 free = 1;
2206         }
2207         spin_unlock(&khugepaged_mm_lock);
2208
2209         if (free) {
2210                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2211                 free_mm_slot(mm_slot);
2212                 mmdrop(mm);
2213         } else if (mm_slot) {
2214                 /*
2215                  * This is required to serialize against
2216                  * khugepaged_test_exit() (which is guaranteed to run
2217                  * under mmap sem read mode). Stop here (after we
2218                  * return all pagetables will be destroyed) until
2219                  * khugepaged has finished working on the pagetables
2220                  * under the mmap_sem.
2221                  */
2222                 down_write(&mm->mmap_sem);
2223                 up_write(&mm->mmap_sem);
2224         }
2225 }
2226
2227 static void release_pte_page(struct page *page)
2228 {
2229         /* 0 stands for page_is_file_cache(page) == false */
2230         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2231         unlock_page(page);
2232         putback_lru_page(page);
2233 }
2234
2235 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2236 {
2237         while (--_pte >= pte) {
2238                 pte_t pteval = *_pte;
2239                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2240                         release_pte_page(pte_page(pteval));
2241         }
2242 }
2243
2244 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2245                                         unsigned long address,
2246                                         pte_t *pte)
2247 {
2248         struct page *page;
2249         pte_t *_pte;
2250         int none_or_zero = 0;
2251         bool referenced = false, writable = false;
2252         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2253              _pte++, address += PAGE_SIZE) {
2254                 pte_t pteval = *_pte;
2255                 if (pte_none(pteval) || (pte_present(pteval) &&
2256                                 is_zero_pfn(pte_pfn(pteval)))) {
2257                         if (!userfaultfd_armed(vma) &&
2258                             ++none_or_zero <= khugepaged_max_ptes_none)
2259                                 continue;
2260                         else
2261                                 goto out;
2262                 }
2263                 if (!pte_present(pteval))
2264                         goto out;
2265                 page = vm_normal_page(vma, address, pteval);
2266                 if (unlikely(!page))
2267                         goto out;
2268
2269                 VM_BUG_ON_PAGE(PageCompound(page), page);
2270                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2271                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2272
2273                 /*
2274                  * We can do it before isolate_lru_page because the
2275                  * page can't be freed from under us. NOTE: PG_lock
2276                  * is needed to serialize against split_huge_page
2277                  * when invoked from the VM.
2278                  */
2279                 if (!trylock_page(page))
2280                         goto out;
2281
2282                 /*
2283                  * cannot use mapcount: can't collapse if there's a gup pin.
2284                  * The page must only be referenced by the scanned process
2285                  * and page swap cache.
2286                  */
2287                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2288                         unlock_page(page);
2289                         goto out;
2290                 }
2291                 if (pte_write(pteval)) {
2292                         writable = true;
2293                 } else {
2294                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2295                                 unlock_page(page);
2296                                 goto out;
2297                         }
2298                         /*
2299                          * Page is not in the swap cache. It can be collapsed
2300                          * into a THP.
2301                          */
2302                 }
2303
2304                 /*
2305                  * Isolate the page to avoid collapsing an hugepage
2306                  * currently in use by the VM.
2307                  */
2308                 if (isolate_lru_page(page)) {
2309                         unlock_page(page);
2310                         goto out;
2311                 }
2312                 /* 0 stands for page_is_file_cache(page) == false */
2313                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2314                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2315                 VM_BUG_ON_PAGE(PageLRU(page), page);
2316
2317                 /* If there is no mapped pte young don't collapse the page */
2318                 if (pte_young(pteval) ||
2319                     page_is_young(page) || PageReferenced(page) ||
2320                     mmu_notifier_test_young(vma->vm_mm, address))
2321                         referenced = true;
2322         }
2323         if (likely(referenced && writable))
2324                 return 1;
2325 out:
2326         release_pte_pages(pte, _pte);
2327         return 0;
2328 }
2329
2330 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2331                                       struct vm_area_struct *vma,
2332                                       unsigned long address,
2333                                       spinlock_t *ptl)
2334 {
2335         pte_t *_pte;
2336         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2337                 pte_t pteval = *_pte;
2338                 struct page *src_page;
2339
2340                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2341                         clear_user_highpage(page, address);
2342                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2343                         if (is_zero_pfn(pte_pfn(pteval))) {
2344                                 /*
2345                                  * ptl mostly unnecessary.
2346                                  */
2347                                 spin_lock(ptl);
2348                                 /*
2349                                  * paravirt calls inside pte_clear here are
2350                                  * superfluous.
2351                                  */
2352                                 pte_clear(vma->vm_mm, address, _pte);
2353                                 spin_unlock(ptl);
2354                         }
2355                 } else {
2356                         src_page = pte_page(pteval);
2357                         copy_user_highpage(page, src_page, address, vma);
2358                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2359                         release_pte_page(src_page);
2360                         /*
2361                          * ptl mostly unnecessary, but preempt has to
2362                          * be disabled to update the per-cpu stats
2363                          * inside page_remove_rmap().
2364                          */
2365                         spin_lock(ptl);
2366                         /*
2367                          * paravirt calls inside pte_clear here are
2368                          * superfluous.
2369                          */
2370                         pte_clear(vma->vm_mm, address, _pte);
2371                         page_remove_rmap(src_page);
2372                         spin_unlock(ptl);
2373                         free_page_and_swap_cache(src_page);
2374                 }
2375
2376                 address += PAGE_SIZE;
2377                 page++;
2378         }
2379 }
2380
2381 static void khugepaged_alloc_sleep(void)
2382 {
2383         DEFINE_WAIT(wait);
2384
2385         add_wait_queue(&khugepaged_wait, &wait);
2386         freezable_schedule_timeout_interruptible(
2387                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2388         remove_wait_queue(&khugepaged_wait, &wait);
2389 }
2390
2391 static int khugepaged_node_load[MAX_NUMNODES];
2392
2393 static bool khugepaged_scan_abort(int nid)
2394 {
2395         int i;
2396
2397         /*
2398          * If zone_reclaim_mode is disabled, then no extra effort is made to
2399          * allocate memory locally.
2400          */
2401         if (!zone_reclaim_mode)
2402                 return false;
2403
2404         /* If there is a count for this node already, it must be acceptable */
2405         if (khugepaged_node_load[nid])
2406                 return false;
2407
2408         for (i = 0; i < MAX_NUMNODES; i++) {
2409                 if (!khugepaged_node_load[i])
2410                         continue;
2411                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2412                         return true;
2413         }
2414         return false;
2415 }
2416
2417 #ifdef CONFIG_NUMA
2418 static int khugepaged_find_target_node(void)
2419 {
2420         static int last_khugepaged_target_node = NUMA_NO_NODE;
2421         int nid, target_node = 0, max_value = 0;
2422
2423         /* find first node with max normal pages hit */
2424         for (nid = 0; nid < MAX_NUMNODES; nid++)
2425                 if (khugepaged_node_load[nid] > max_value) {
2426                         max_value = khugepaged_node_load[nid];
2427                         target_node = nid;
2428                 }
2429
2430         /* do some balance if several nodes have the same hit record */
2431         if (target_node <= last_khugepaged_target_node)
2432                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2433                                 nid++)
2434                         if (max_value == khugepaged_node_load[nid]) {
2435                                 target_node = nid;
2436                                 break;
2437                         }
2438
2439         last_khugepaged_target_node = target_node;
2440         return target_node;
2441 }
2442
2443 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2444 {
2445         if (IS_ERR(*hpage)) {
2446                 if (!*wait)
2447                         return false;
2448
2449                 *wait = false;
2450                 *hpage = NULL;
2451                 khugepaged_alloc_sleep();
2452         } else if (*hpage) {
2453                 put_page(*hpage);
2454                 *hpage = NULL;
2455         }
2456
2457         return true;
2458 }
2459
2460 static struct page *
2461 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2462                        unsigned long address, int node)
2463 {
2464         VM_BUG_ON_PAGE(*hpage, *hpage);
2465
2466         /*
2467          * Before allocating the hugepage, release the mmap_sem read lock.
2468          * The allocation can take potentially a long time if it involves
2469          * sync compaction, and we do not need to hold the mmap_sem during
2470          * that. We will recheck the vma after taking it again in write mode.
2471          */
2472         up_read(&mm->mmap_sem);
2473
2474         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2475         if (unlikely(!*hpage)) {
2476                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2477                 *hpage = ERR_PTR(-ENOMEM);
2478                 return NULL;
2479         }
2480
2481         count_vm_event(THP_COLLAPSE_ALLOC);
2482         return *hpage;
2483 }
2484 #else
2485 static int khugepaged_find_target_node(void)
2486 {
2487         return 0;
2488 }
2489
2490 static inline struct page *alloc_hugepage(int defrag)
2491 {
2492         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2493                            HPAGE_PMD_ORDER);
2494 }
2495
2496 static struct page *khugepaged_alloc_hugepage(bool *wait)
2497 {
2498         struct page *hpage;
2499
2500         do {
2501                 hpage = alloc_hugepage(khugepaged_defrag());
2502                 if (!hpage) {
2503                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2504                         if (!*wait)
2505                                 return NULL;
2506
2507                         *wait = false;
2508                         khugepaged_alloc_sleep();
2509                 } else
2510                         count_vm_event(THP_COLLAPSE_ALLOC);
2511         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2512
2513         return hpage;
2514 }
2515
2516 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2517 {
2518         if (!*hpage)
2519                 *hpage = khugepaged_alloc_hugepage(wait);
2520
2521         if (unlikely(!*hpage))
2522                 return false;
2523
2524         return true;
2525 }
2526
2527 static struct page *
2528 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2529                        unsigned long address, int node)
2530 {
2531         up_read(&mm->mmap_sem);
2532         VM_BUG_ON(!*hpage);
2533
2534         return  *hpage;
2535 }
2536 #endif
2537
2538 static bool hugepage_vma_check(struct vm_area_struct *vma)
2539 {
2540         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2541             (vma->vm_flags & VM_NOHUGEPAGE))
2542                 return false;
2543
2544         if (!vma->anon_vma || vma->vm_ops)
2545                 return false;
2546         if (is_vma_temporary_stack(vma))
2547                 return false;
2548         return !(vma->vm_flags & VM_NO_THP);
2549 }
2550
2551 static void collapse_huge_page(struct mm_struct *mm,
2552                                    unsigned long address,
2553                                    struct page **hpage,
2554                                    struct vm_area_struct *vma,
2555                                    int node)
2556 {
2557         pmd_t *pmd, _pmd;
2558         pte_t *pte;
2559         pgtable_t pgtable;
2560         struct page *new_page;
2561         spinlock_t *pmd_ptl, *pte_ptl;
2562         int isolated;
2563         unsigned long hstart, hend;
2564         struct mem_cgroup *memcg;
2565         unsigned long mmun_start;       /* For mmu_notifiers */
2566         unsigned long mmun_end;         /* For mmu_notifiers */
2567         gfp_t gfp;
2568
2569         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2570
2571         /* Only allocate from the target node */
2572         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2573                 __GFP_THISNODE;
2574
2575         /* release the mmap_sem read lock. */
2576         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2577         if (!new_page)
2578                 return;
2579
2580         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2581                                            gfp, &memcg)))
2582                 return;
2583
2584         /*
2585          * Prevent all access to pagetables with the exception of
2586          * gup_fast later hanlded by the ptep_clear_flush and the VM
2587          * handled by the anon_vma lock + PG_lock.
2588          */
2589         down_write(&mm->mmap_sem);
2590         if (unlikely(khugepaged_test_exit(mm)))
2591                 goto out;
2592
2593         vma = find_vma(mm, address);
2594         if (!vma)
2595                 goto out;
2596         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2597         hend = vma->vm_end & HPAGE_PMD_MASK;
2598         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2599                 goto out;
2600         if (!hugepage_vma_check(vma))
2601                 goto out;
2602         pmd = mm_find_pmd(mm, address);
2603         if (!pmd)
2604                 goto out;
2605
2606         anon_vma_lock_write(vma->anon_vma);
2607
2608         pte = pte_offset_map(pmd, address);
2609         pte_ptl = pte_lockptr(mm, pmd);
2610
2611         mmun_start = address;
2612         mmun_end   = address + HPAGE_PMD_SIZE;
2613         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2614         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2615         /*
2616          * After this gup_fast can't run anymore. This also removes
2617          * any huge TLB entry from the CPU so we won't allow
2618          * huge and small TLB entries for the same virtual address
2619          * to avoid the risk of CPU bugs in that area.
2620          */
2621         _pmd = pmdp_collapse_flush(vma, address, pmd);
2622         spin_unlock(pmd_ptl);
2623         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2624
2625         spin_lock(pte_ptl);
2626         isolated = __collapse_huge_page_isolate(vma, address, pte);
2627         spin_unlock(pte_ptl);
2628
2629         if (unlikely(!isolated)) {
2630                 pte_unmap(pte);
2631                 spin_lock(pmd_ptl);
2632                 BUG_ON(!pmd_none(*pmd));
2633                 /*
2634                  * We can only use set_pmd_at when establishing
2635                  * hugepmds and never for establishing regular pmds that
2636                  * points to regular pagetables. Use pmd_populate for that
2637                  */
2638                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2639                 spin_unlock(pmd_ptl);
2640                 anon_vma_unlock_write(vma->anon_vma);
2641                 goto out;
2642         }
2643
2644         /*
2645          * All pages are isolated and locked so anon_vma rmap
2646          * can't run anymore.
2647          */
2648         anon_vma_unlock_write(vma->anon_vma);
2649
2650         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2651         pte_unmap(pte);
2652         __SetPageUptodate(new_page);
2653         pgtable = pmd_pgtable(_pmd);
2654
2655         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2656         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2657
2658         /*
2659          * spin_lock() below is not the equivalent of smp_wmb(), so
2660          * this is needed to avoid the copy_huge_page writes to become
2661          * visible after the set_pmd_at() write.
2662          */
2663         smp_wmb();
2664
2665         spin_lock(pmd_ptl);
2666         BUG_ON(!pmd_none(*pmd));
2667         page_add_new_anon_rmap(new_page, vma, address);
2668         mem_cgroup_commit_charge(new_page, memcg, false);
2669         lru_cache_add_active_or_unevictable(new_page, vma);
2670         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2671         set_pmd_at(mm, address, pmd, _pmd);
2672         update_mmu_cache_pmd(vma, address, pmd);
2673         spin_unlock(pmd_ptl);
2674
2675         *hpage = NULL;
2676
2677         khugepaged_pages_collapsed++;
2678 out_up_write:
2679         up_write(&mm->mmap_sem);
2680         return;
2681
2682 out:
2683         mem_cgroup_cancel_charge(new_page, memcg);
2684         goto out_up_write;
2685 }
2686
2687 static int khugepaged_scan_pmd(struct mm_struct *mm,
2688                                struct vm_area_struct *vma,
2689                                unsigned long address,
2690                                struct page **hpage)
2691 {
2692         pmd_t *pmd;
2693         pte_t *pte, *_pte;
2694         int ret = 0, none_or_zero = 0;
2695         struct page *page;
2696         unsigned long _address;
2697         spinlock_t *ptl;
2698         int node = NUMA_NO_NODE;
2699         bool writable = false, referenced = false;
2700
2701         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2702
2703         pmd = mm_find_pmd(mm, address);
2704         if (!pmd)
2705                 goto out;
2706
2707         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2708         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2709         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2710              _pte++, _address += PAGE_SIZE) {
2711                 pte_t pteval = *_pte;
2712                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2713                         if (!userfaultfd_armed(vma) &&
2714                             ++none_or_zero <= khugepaged_max_ptes_none)
2715                                 continue;
2716                         else
2717                                 goto out_unmap;
2718                 }
2719                 if (!pte_present(pteval))
2720                         goto out_unmap;
2721                 if (pte_write(pteval))
2722                         writable = true;
2723
2724                 page = vm_normal_page(vma, _address, pteval);
2725                 if (unlikely(!page))
2726                         goto out_unmap;
2727                 /*
2728                  * Record which node the original page is from and save this
2729                  * information to khugepaged_node_load[].
2730                  * Khupaged will allocate hugepage from the node has the max
2731                  * hit record.
2732                  */
2733                 node = page_to_nid(page);
2734                 if (khugepaged_scan_abort(node))
2735                         goto out_unmap;
2736                 khugepaged_node_load[node]++;
2737                 VM_BUG_ON_PAGE(PageCompound(page), page);
2738                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2739                         goto out_unmap;
2740                 /*
2741                  * cannot use mapcount: can't collapse if there's a gup pin.
2742                  * The page must only be referenced by the scanned process
2743                  * and page swap cache.
2744                  */
2745                 if (page_count(page) != 1 + !!PageSwapCache(page))
2746                         goto out_unmap;
2747                 if (pte_young(pteval) ||
2748                     page_is_young(page) || PageReferenced(page) ||
2749                     mmu_notifier_test_young(vma->vm_mm, address))
2750                         referenced = true;
2751         }
2752         if (referenced && writable)
2753                 ret = 1;
2754 out_unmap:
2755         pte_unmap_unlock(pte, ptl);
2756         if (ret) {
2757                 node = khugepaged_find_target_node();
2758                 /* collapse_huge_page will return with the mmap_sem released */
2759                 collapse_huge_page(mm, address, hpage, vma, node);
2760         }
2761 out:
2762         return ret;
2763 }
2764
2765 static void collect_mm_slot(struct mm_slot *mm_slot)
2766 {
2767         struct mm_struct *mm = mm_slot->mm;
2768
2769         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2770
2771         if (khugepaged_test_exit(mm)) {
2772                 /* free mm_slot */
2773                 hash_del(&mm_slot->hash);
2774                 list_del(&mm_slot->mm_node);
2775
2776                 /*
2777                  * Not strictly needed because the mm exited already.
2778                  *
2779                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2780                  */
2781
2782                 /* khugepaged_mm_lock actually not necessary for the below */
2783                 free_mm_slot(mm_slot);
2784                 mmdrop(mm);
2785         }
2786 }
2787
2788 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2789                                             struct page **hpage)
2790         __releases(&khugepaged_mm_lock)
2791         __acquires(&khugepaged_mm_lock)
2792 {
2793         struct mm_slot *mm_slot;
2794         struct mm_struct *mm;
2795         struct vm_area_struct *vma;
2796         int progress = 0;
2797
2798         VM_BUG_ON(!pages);
2799         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2800
2801         if (khugepaged_scan.mm_slot)
2802                 mm_slot = khugepaged_scan.mm_slot;
2803         else {
2804                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2805                                      struct mm_slot, mm_node);
2806                 khugepaged_scan.address = 0;
2807                 khugepaged_scan.mm_slot = mm_slot;
2808         }
2809         spin_unlock(&khugepaged_mm_lock);
2810
2811         mm = mm_slot->mm;
2812         down_read(&mm->mmap_sem);
2813         if (unlikely(khugepaged_test_exit(mm)))
2814                 vma = NULL;
2815         else
2816                 vma = find_vma(mm, khugepaged_scan.address);
2817
2818         progress++;
2819         for (; vma; vma = vma->vm_next) {
2820                 unsigned long hstart, hend;
2821
2822                 cond_resched();
2823                 if (unlikely(khugepaged_test_exit(mm))) {
2824                         progress++;
2825                         break;
2826                 }
2827                 if (!hugepage_vma_check(vma)) {
2828 skip:
2829                         progress++;
2830                         continue;
2831                 }
2832                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2833                 hend = vma->vm_end & HPAGE_PMD_MASK;
2834                 if (hstart >= hend)
2835                         goto skip;
2836                 if (khugepaged_scan.address > hend)
2837                         goto skip;
2838                 if (khugepaged_scan.address < hstart)
2839                         khugepaged_scan.address = hstart;
2840                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2841
2842                 while (khugepaged_scan.address < hend) {
2843                         int ret;
2844                         cond_resched();
2845                         if (unlikely(khugepaged_test_exit(mm)))
2846                                 goto breakouterloop;
2847
2848                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2849                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2850                                   hend);
2851                         ret = khugepaged_scan_pmd(mm, vma,
2852                                                   khugepaged_scan.address,
2853                                                   hpage);
2854                         /* move to next address */
2855                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2856                         progress += HPAGE_PMD_NR;
2857                         if (ret)
2858                                 /* we released mmap_sem so break loop */
2859                                 goto breakouterloop_mmap_sem;
2860                         if (progress >= pages)
2861                                 goto breakouterloop;
2862                 }
2863         }
2864 breakouterloop:
2865         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2866 breakouterloop_mmap_sem:
2867
2868         spin_lock(&khugepaged_mm_lock);
2869         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2870         /*
2871          * Release the current mm_slot if this mm is about to die, or
2872          * if we scanned all vmas of this mm.
2873          */
2874         if (khugepaged_test_exit(mm) || !vma) {
2875                 /*
2876                  * Make sure that if mm_users is reaching zero while
2877                  * khugepaged runs here, khugepaged_exit will find
2878                  * mm_slot not pointing to the exiting mm.
2879                  */
2880                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2881                         khugepaged_scan.mm_slot = list_entry(
2882                                 mm_slot->mm_node.next,
2883                                 struct mm_slot, mm_node);
2884                         khugepaged_scan.address = 0;
2885                 } else {
2886                         khugepaged_scan.mm_slot = NULL;
2887                         khugepaged_full_scans++;
2888                 }
2889
2890                 collect_mm_slot(mm_slot);
2891         }
2892
2893         return progress;
2894 }
2895
2896 static int khugepaged_has_work(void)
2897 {
2898         return !list_empty(&khugepaged_scan.mm_head) &&
2899                 khugepaged_enabled();
2900 }
2901
2902 static int khugepaged_wait_event(void)
2903 {
2904         return !list_empty(&khugepaged_scan.mm_head) ||
2905                 kthread_should_stop();
2906 }
2907
2908 static void khugepaged_do_scan(void)
2909 {
2910         struct page *hpage = NULL;
2911         unsigned int progress = 0, pass_through_head = 0;
2912         unsigned int pages = khugepaged_pages_to_scan;
2913         bool wait = true;
2914
2915         barrier(); /* write khugepaged_pages_to_scan to local stack */
2916
2917         while (progress < pages) {
2918                 if (!khugepaged_prealloc_page(&hpage, &wait))
2919                         break;
2920
2921                 cond_resched();
2922
2923                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2924                         break;
2925
2926                 spin_lock(&khugepaged_mm_lock);
2927                 if (!khugepaged_scan.mm_slot)
2928                         pass_through_head++;
2929                 if (khugepaged_has_work() &&
2930                     pass_through_head < 2)
2931                         progress += khugepaged_scan_mm_slot(pages - progress,
2932                                                             &hpage);
2933                 else
2934                         progress = pages;
2935                 spin_unlock(&khugepaged_mm_lock);
2936         }
2937
2938         if (!IS_ERR_OR_NULL(hpage))
2939                 put_page(hpage);
2940 }
2941
2942 static void khugepaged_wait_work(void)
2943 {
2944         if (khugepaged_has_work()) {
2945                 if (!khugepaged_scan_sleep_millisecs)
2946                         return;
2947
2948                 wait_event_freezable_timeout(khugepaged_wait,
2949                                              kthread_should_stop(),
2950                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2951                 return;
2952         }
2953
2954         if (khugepaged_enabled())
2955                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2956 }
2957
2958 static int khugepaged(void *none)
2959 {
2960         struct mm_slot *mm_slot;
2961
2962         set_freezable();
2963         set_user_nice(current, MAX_NICE);
2964
2965         while (!kthread_should_stop()) {
2966                 khugepaged_do_scan();
2967                 khugepaged_wait_work();
2968         }
2969
2970         spin_lock(&khugepaged_mm_lock);
2971         mm_slot = khugepaged_scan.mm_slot;
2972         khugepaged_scan.mm_slot = NULL;
2973         if (mm_slot)
2974                 collect_mm_slot(mm_slot);
2975         spin_unlock(&khugepaged_mm_lock);
2976         return 0;
2977 }
2978
2979 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2980                 unsigned long haddr, pmd_t *pmd)
2981 {
2982         struct mm_struct *mm = vma->vm_mm;
2983         pgtable_t pgtable;
2984         pmd_t _pmd;
2985         int i;
2986
2987         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2988         /* leave pmd empty until pte is filled */
2989
2990         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2991         pmd_populate(mm, &_pmd, pgtable);
2992
2993         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2994                 pte_t *pte, entry;
2995                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2996                 entry = pte_mkspecial(entry);
2997                 pte = pte_offset_map(&_pmd, haddr);
2998                 VM_BUG_ON(!pte_none(*pte));
2999                 set_pte_at(mm, haddr, pte, entry);
3000                 pte_unmap(pte);
3001         }
3002         smp_wmb(); /* make pte visible before pmd */
3003         pmd_populate(mm, pmd, pgtable);
3004         put_huge_zero_page();
3005 }
3006
3007 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
3008                 pmd_t *pmd)
3009 {
3010         spinlock_t *ptl;
3011         struct page *page = NULL;
3012         struct mm_struct *mm = vma->vm_mm;
3013         unsigned long haddr = address & HPAGE_PMD_MASK;
3014         unsigned long mmun_start;       /* For mmu_notifiers */
3015         unsigned long mmun_end;         /* For mmu_notifiers */
3016
3017         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
3018
3019         mmun_start = haddr;
3020         mmun_end   = haddr + HPAGE_PMD_SIZE;
3021 again:
3022         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3023         ptl = pmd_lock(mm, pmd);
3024         if (unlikely(!pmd_trans_huge(*pmd)))
3025                 goto unlock;
3026         if (vma_is_dax(vma)) {
3027                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
3028                 if (is_huge_zero_pmd(_pmd))
3029                         put_huge_zero_page();
3030         } else if (is_huge_zero_pmd(*pmd)) {
3031                 __split_huge_zero_page_pmd(vma, haddr, pmd);
3032         } else {
3033                 page = pmd_page(*pmd);
3034                 VM_BUG_ON_PAGE(!page_count(page), page);
3035                 get_page(page);
3036         }
3037  unlock:
3038         spin_unlock(ptl);
3039         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3040
3041         if (!page)
3042                 return;
3043
3044         split_huge_page(page);
3045         put_page(page);
3046
3047         /*
3048          * We don't always have down_write of mmap_sem here: a racing
3049          * do_huge_pmd_wp_page() might have copied-on-write to another
3050          * huge page before our split_huge_page() got the anon_vma lock.
3051          */
3052         if (unlikely(pmd_trans_huge(*pmd)))
3053                 goto again;
3054 }
3055
3056 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3057                 pmd_t *pmd)
3058 {
3059         struct vm_area_struct *vma;
3060
3061         vma = find_vma(mm, address);
3062         BUG_ON(vma == NULL);
3063         split_huge_page_pmd(vma, address, pmd);
3064 }
3065
3066 static void split_huge_page_address(struct mm_struct *mm,
3067                                     unsigned long address)
3068 {
3069         pgd_t *pgd;
3070         pud_t *pud;
3071         pmd_t *pmd;
3072
3073         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3074
3075         pgd = pgd_offset(mm, address);
3076         if (!pgd_present(*pgd))
3077                 return;
3078
3079         pud = pud_offset(pgd, address);
3080         if (!pud_present(*pud))
3081                 return;
3082
3083         pmd = pmd_offset(pud, address);
3084         if (!pmd_present(*pmd))
3085                 return;
3086         /*
3087          * Caller holds the mmap_sem write mode, so a huge pmd cannot
3088          * materialize from under us.
3089          */
3090         split_huge_page_pmd_mm(mm, address, pmd);
3091 }
3092
3093 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3094                              unsigned long start,
3095                              unsigned long end,
3096                              long adjust_next)
3097 {
3098         /*
3099          * If the new start address isn't hpage aligned and it could
3100          * previously contain an hugepage: check if we need to split
3101          * an huge pmd.
3102          */
3103         if (start & ~HPAGE_PMD_MASK &&
3104             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3105             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3106                 split_huge_page_address(vma->vm_mm, start);
3107
3108         /*
3109          * If the new end address isn't hpage aligned and it could
3110          * previously contain an hugepage: check if we need to split
3111          * an huge pmd.
3112          */
3113         if (end & ~HPAGE_PMD_MASK &&
3114             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3115             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3116                 split_huge_page_address(vma->vm_mm, end);
3117
3118         /*
3119          * If we're also updating the vma->vm_next->vm_start, if the new
3120          * vm_next->vm_start isn't page aligned and it could previously
3121          * contain an hugepage: check if we need to split an huge pmd.
3122          */
3123         if (adjust_next > 0) {
3124                 struct vm_area_struct *next = vma->vm_next;
3125                 unsigned long nstart = next->vm_start;
3126                 nstart += adjust_next << PAGE_SHIFT;
3127                 if (nstart & ~HPAGE_PMD_MASK &&
3128                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3129                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3130                         split_huge_page_address(next->vm_mm, nstart);
3131         }
3132 }