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