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 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <linux/pagemap.h>
  21 #include <asm/tlb.h>
  22 #include <asm/pgalloc.h>
  23 #include "internal.h"
  24
  25 /*
  26  * By default transparent hugepage support is enabled for all mappings
  27  * and khugepaged scans all mappings. Defrag is only invoked by
  28  * khugepaged hugepage allocations and by page faults inside
  29  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  30  * allocations.
  31  */
  32 unsigned long transparent_hugepage_flags __read_mostly =
  33 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  34         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  35 #endif
  36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  37         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  38 #endif
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  40         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  41
  42 /* default scan 8*512 pte (or vmas) every 30 second */
  43 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  44 static unsigned int khugepaged_pages_collapsed;
  45 static unsigned int khugepaged_full_scans;
  46 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  47 /* during fragmentation poll the hugepage allocator once every minute */
  48 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  49 static struct task_struct *khugepaged_thread __read_mostly;
  50 static DEFINE_MUTEX(khugepaged_mutex);
  51 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  52 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  53 /*
  54  * default collapse hugepages if there is at least one pte mapped like
  55  * it would have happened if the vma was large enough during page
  56  * fault.
  57  */
  58 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  59
  60 static int khugepaged(void *none);
  61 static int mm_slots_hash_init(void);
  62 static int khugepaged_slab_init(void);
  63 static void khugepaged_slab_free(void);
  64
  65 #define MM_SLOTS_HASH_HEADS 1024
  66 static struct hlist_head *mm_slots_hash __read_mostly;
  67 static struct kmem_cache *mm_slot_cache __read_mostly;
  68
  69 /**
  70  * struct mm_slot - hash lookup from mm to mm_slot
  71  * @hash: hash collision list
  72  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  73  * @mm: the mm that this information is valid for
  74  */
  75 struct mm_slot {
  76         struct hlist_node hash;
  77         struct list_head mm_node;
  78         struct mm_struct *mm;
  79 };
  80
  81 /**
  82  * struct khugepaged_scan - cursor for scanning
  83  * @mm_head: the head of the mm list to scan
  84  * @mm_slot: the current mm_slot we are scanning
  85  * @address: the next address inside that to be scanned
  86  *
  87  * There is only the one khugepaged_scan instance of this cursor structure.
  88  */
  89 struct khugepaged_scan {
  90         struct list_head mm_head;
  91         struct mm_slot *mm_slot;
  92         unsigned long address;
  93 };
  94 static struct khugepaged_scan khugepaged_scan = {
  95         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  96 };
  97
  98
  99 static int set_recommended_min_free_kbytes(void)
 100 {
 101         struct zone *zone;
 102         int nr_zones = 0;
 103         unsigned long recommended_min;
 104         extern int min_free_kbytes;
 105
 106         if (!khugepaged_enabled())
 107                 return 0;
 108
 109         for_each_populated_zone(zone)
 110                 nr_zones++;
 111
 112         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 113         recommended_min = pageblock_nr_pages * nr_zones * 2;
 114
 115         /*
 116          * Make sure that on average at least two pageblocks are almost free
 117          * of another type, one for a migratetype to fall back to and a
 118          * second to avoid subsequent fallbacks of other types There are 3
 119          * MIGRATE_TYPES we care about.
 120          */
 121         recommended_min += pageblock_nr_pages * nr_zones *
 122                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 123
 124         /* don't ever allow to reserve more than 5% of the lowmem */
 125         recommended_min = min(recommended_min,
 126                               (unsigned long) nr_free_buffer_pages() / 20);
 127         recommended_min <<= (PAGE_SHIFT-10);
 128
 129         if (recommended_min > min_free_kbytes)
 130                 min_free_kbytes = recommended_min;
 131         setup_per_zone_wmarks();
 132         return 0;
 133 }
 134 late_initcall(set_recommended_min_free_kbytes);
 135
 136 static int start_khugepaged(void)
 137 {
 138         int err = 0;
 139         if (khugepaged_enabled()) {
 140                 if (!khugepaged_thread)
 141                         khugepaged_thread = kthread_run(khugepaged, NULL,
 142                                                         "khugepaged");
 143                 if (unlikely(IS_ERR(khugepaged_thread))) {
 144                         printk(KERN_ERR
 145                                "khugepaged: kthread_run(khugepaged) failed\n");
 146                         err = PTR_ERR(khugepaged_thread);
 147                         khugepaged_thread = NULL;
 148                 }
 149
 150                 if (!list_empty(&khugepaged_scan.mm_head))
 151                         wake_up_interruptible(&khugepaged_wait);
 152
 153                 set_recommended_min_free_kbytes();
 154         } else if (khugepaged_thread) {
 155                 kthread_stop(khugepaged_thread);
 156                 khugepaged_thread = NULL;
 157         }
 158
 159         return err;
 160 }
 161
 162 #ifdef CONFIG_SYSFS
 163
 164 static ssize_t double_flag_show(struct kobject *kobj,
 165                                 struct kobj_attribute *attr, char *buf,
 166                                 enum transparent_hugepage_flag enabled,
 167                                 enum transparent_hugepage_flag req_madv)
 168 {
 169         if (test_bit(enabled, &transparent_hugepage_flags)) {
 170                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 171                 return sprintf(buf, "[always] madvise never\n");
 172         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 173                 return sprintf(buf, "always [madvise] never\n");
 174         else
 175                 return sprintf(buf, "always madvise [never]\n");
 176 }
 177 static ssize_t double_flag_store(struct kobject *kobj,
 178                                  struct kobj_attribute *attr,
 179                                  const char *buf, size_t count,
 180                                  enum transparent_hugepage_flag enabled,
 181                                  enum transparent_hugepage_flag req_madv)
 182 {
 183         if (!memcmp("always", buf,
 184                     min(sizeof("always")-1, count))) {
 185                 set_bit(enabled, &transparent_hugepage_flags);
 186                 clear_bit(req_madv, &transparent_hugepage_flags);
 187         } else if (!memcmp("madvise", buf,
 188                            min(sizeof("madvise")-1, count))) {
 189                 clear_bit(enabled, &transparent_hugepage_flags);
 190                 set_bit(req_madv, &transparent_hugepage_flags);
 191         } else if (!memcmp("never", buf,
 192                            min(sizeof("never")-1, count))) {
 193                 clear_bit(enabled, &transparent_hugepage_flags);
 194                 clear_bit(req_madv, &transparent_hugepage_flags);
 195         } else
 196                 return -EINVAL;
 197
 198         return count;
 199 }
 200
 201 static ssize_t enabled_show(struct kobject *kobj,
 202                             struct kobj_attribute *attr, char *buf)
 203 {
 204         return double_flag_show(kobj, attr, buf,
 205                                 TRANSPARENT_HUGEPAGE_FLAG,
 206                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 207 }
 208 static ssize_t enabled_store(struct kobject *kobj,
 209                              struct kobj_attribute *attr,
 210                              const char *buf, size_t count)
 211 {
 212         ssize_t ret;
 213
 214         ret = double_flag_store(kobj, attr, buf, count,
 215                                 TRANSPARENT_HUGEPAGE_FLAG,
 216                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 217
 218         if (ret > 0) {
 219                 int err;
 220
 221                 mutex_lock(&khugepaged_mutex);
 222                 err = start_khugepaged();
 223                 mutex_unlock(&khugepaged_mutex);
 224
 225                 if (err)
 226                         ret = err;
 227         }
 228
 229         return ret;
 230 }
 231 static struct kobj_attribute enabled_attr =
 232         __ATTR(enabled, 0644, enabled_show, enabled_store);
 233
 234 static ssize_t single_flag_show(struct kobject *kobj,
 235                                 struct kobj_attribute *attr, char *buf,
 236                                 enum transparent_hugepage_flag flag)
 237 {
 238         return sprintf(buf, "%d\n",
 239                        !!test_bit(flag, &transparent_hugepage_flags));
 240 }
 241
 242 static ssize_t single_flag_store(struct kobject *kobj,
 243                                  struct kobj_attribute *attr,
 244                                  const char *buf, size_t count,
 245                                  enum transparent_hugepage_flag flag)
 246 {
 247         unsigned long value;
 248         int ret;
 249
 250         ret = kstrtoul(buf, 10, &value);
 251         if (ret < 0)
 252                 return ret;
 253         if (value > 1)
 254                 return -EINVAL;
 255
 256         if (value)
 257                 set_bit(flag, &transparent_hugepage_flags);
 258         else
 259                 clear_bit(flag, &transparent_hugepage_flags);
 260
 261         return count;
 262 }
 263
 264 /*
 265  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 266  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 267  * memory just to allocate one more hugepage.
 268  */
 269 static ssize_t defrag_show(struct kobject *kobj,
 270                            struct kobj_attribute *attr, char *buf)
 271 {
 272         return double_flag_show(kobj, attr, buf,
 273                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 274                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 275 }
 276 static ssize_t defrag_store(struct kobject *kobj,
 277                             struct kobj_attribute *attr,
 278                             const char *buf, size_t count)
 279 {
 280         return double_flag_store(kobj, attr, buf, count,
 281                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 282                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 283 }
 284 static struct kobj_attribute defrag_attr =
 285         __ATTR(defrag, 0644, defrag_show, defrag_store);
 286
 287 #ifdef CONFIG_DEBUG_VM
 288 static ssize_t debug_cow_show(struct kobject *kobj,
 289                                 struct kobj_attribute *attr, char *buf)
 290 {
 291         return single_flag_show(kobj, attr, buf,
 292                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 293 }
 294 static ssize_t debug_cow_store(struct kobject *kobj,
 295                                struct kobj_attribute *attr,
 296                                const char *buf, size_t count)
 297 {
 298         return single_flag_store(kobj, attr, buf, count,
 299                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 300 }
 301 static struct kobj_attribute debug_cow_attr =
 302         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 303 #endif /* CONFIG_DEBUG_VM */
 304
 305 static struct attribute *hugepage_attr[] = {
 306         &enabled_attr.attr,
 307         &defrag_attr.attr,
 308 #ifdef CONFIG_DEBUG_VM
 309         &debug_cow_attr.attr,
 310 #endif
 311         NULL,
 312 };
 313
 314 static struct attribute_group hugepage_attr_group = {
 315         .attrs = hugepage_attr,
 316 };
 317
 318 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 319                                          struct kobj_attribute *attr,
 320                                          char *buf)
 321 {
 322         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 323 }
 324
 325 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 326                                           struct kobj_attribute *attr,
 327                                           const char *buf, size_t count)
 328 {
 329         unsigned long msecs;
 330         int err;
 331
 332         err = strict_strtoul(buf, 10, &msecs);
 333         if (err || msecs > UINT_MAX)
 334                 return -EINVAL;
 335
 336         khugepaged_scan_sleep_millisecs = msecs;
 337         wake_up_interruptible(&khugepaged_wait);
 338
 339         return count;
 340 }
 341 static struct kobj_attribute scan_sleep_millisecs_attr =
 342         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 343                scan_sleep_millisecs_store);
 344
 345 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 346                                           struct kobj_attribute *attr,
 347                                           char *buf)
 348 {
 349         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 350 }
 351
 352 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 353                                            struct kobj_attribute *attr,
 354                                            const char *buf, size_t count)
 355 {
 356         unsigned long msecs;
 357         int err;
 358
 359         err = strict_strtoul(buf, 10, &msecs);
 360         if (err || msecs > UINT_MAX)
 361                 return -EINVAL;
 362
 363         khugepaged_alloc_sleep_millisecs = msecs;
 364         wake_up_interruptible(&khugepaged_wait);
 365
 366         return count;
 367 }
 368 static struct kobj_attribute alloc_sleep_millisecs_attr =
 369         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 370                alloc_sleep_millisecs_store);
 371
 372 static ssize_t pages_to_scan_show(struct kobject *kobj,
 373                                   struct kobj_attribute *attr,
 374                                   char *buf)
 375 {
 376         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 377 }
 378 static ssize_t pages_to_scan_store(struct kobject *kobj,
 379                                    struct kobj_attribute *attr,
 380                                    const char *buf, size_t count)
 381 {
 382         int err;
 383         unsigned long pages;
 384
 385         err = strict_strtoul(buf, 10, &pages);
 386         if (err || !pages || pages > UINT_MAX)
 387                 return -EINVAL;
 388
 389         khugepaged_pages_to_scan = pages;
 390
 391         return count;
 392 }
 393 static struct kobj_attribute pages_to_scan_attr =
 394         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 395                pages_to_scan_store);
 396
 397 static ssize_t pages_collapsed_show(struct kobject *kobj,
 398                                     struct kobj_attribute *attr,
 399                                     char *buf)
 400 {
 401         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 402 }
 403 static struct kobj_attribute pages_collapsed_attr =
 404         __ATTR_RO(pages_collapsed);
 405
 406 static ssize_t full_scans_show(struct kobject *kobj,
 407                                struct kobj_attribute *attr,
 408                                char *buf)
 409 {
 410         return sprintf(buf, "%u\n", khugepaged_full_scans);
 411 }
 412 static struct kobj_attribute full_scans_attr =
 413         __ATTR_RO(full_scans);
 414
 415 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 416                                       struct kobj_attribute *attr, char *buf)
 417 {
 418         return single_flag_show(kobj, attr, buf,
 419                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 420 }
 421 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 422                                        struct kobj_attribute *attr,
 423                                        const char *buf, size_t count)
 424 {
 425         return single_flag_store(kobj, attr, buf, count,
 426                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 427 }
 428 static struct kobj_attribute khugepaged_defrag_attr =
 429         __ATTR(defrag, 0644, khugepaged_defrag_show,
 430                khugepaged_defrag_store);
 431
 432 /*
 433  * max_ptes_none controls if khugepaged should collapse hugepages over
 434  * any unmapped ptes in turn potentially increasing the memory
 435  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 436  * reduce the available free memory in the system as it
 437  * runs. Increasing max_ptes_none will instead potentially reduce the
 438  * free memory in the system during the khugepaged scan.
 439  */
 440 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 441                                              struct kobj_attribute *attr,
 442                                              char *buf)
 443 {
 444         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 445 }
 446 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 447                                               struct kobj_attribute *attr,
 448                                               const char *buf, size_t count)
 449 {
 450         int err;
 451         unsigned long max_ptes_none;
 452
 453         err = strict_strtoul(buf, 10, &max_ptes_none);
 454         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 455                 return -EINVAL;
 456
 457         khugepaged_max_ptes_none = max_ptes_none;
 458
 459         return count;
 460 }
 461 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 462         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 463                khugepaged_max_ptes_none_store);
 464
 465 static struct attribute *khugepaged_attr[] = {
 466         &khugepaged_defrag_attr.attr,
 467         &khugepaged_max_ptes_none_attr.attr,
 468         &pages_to_scan_attr.attr,
 469         &pages_collapsed_attr.attr,
 470         &full_scans_attr.attr,
 471         &scan_sleep_millisecs_attr.attr,
 472         &alloc_sleep_millisecs_attr.attr,
 473         NULL,
 474 };
 475
 476 static struct attribute_group khugepaged_attr_group = {
 477         .attrs = khugepaged_attr,
 478         .name = "khugepaged",
 479 };
 480
 481 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 482 {
 483         int err;
 484
 485         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 486         if (unlikely(!*hugepage_kobj)) {
 487                 printk(KERN_ERR "hugepage: failed kobject create\n");
 488                 return -ENOMEM;
 489         }
 490
 491         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 492         if (err) {
 493                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 494                 goto delete_obj;
 495         }
 496
 497         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 498         if (err) {
 499                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 500                 goto remove_hp_group;
 501         }
 502
 503         return 0;
 504
 505 remove_hp_group:
 506         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 507 delete_obj:
 508         kobject_put(*hugepage_kobj);
 509         return err;
 510 }
 511
 512 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 513 {
 514         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 515         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 516         kobject_put(hugepage_kobj);
 517 }
 518 #else
 519 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 520 {
 521         return 0;
 522 }
 523
 524 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 525 {
 526 }
 527 #endif /* CONFIG_SYSFS */
 528
 529 static int __init hugepage_init(void)
 530 {
 531         int err;
 532         struct kobject *hugepage_kobj;
 533
 534         if (!has_transparent_hugepage()) {
 535                 transparent_hugepage_flags = 0;
 536                 return -EINVAL;
 537         }
 538
 539         err = hugepage_init_sysfs(&hugepage_kobj);
 540         if (err)
 541                 return err;
 542
 543         err = khugepaged_slab_init();
 544         if (err)
 545                 goto out;
 546
 547         err = mm_slots_hash_init();
 548         if (err) {
 549                 khugepaged_slab_free();
 550                 goto out;
 551         }
 552
 553         /*
 554          * By default disable transparent hugepages on smaller systems,
 555          * where the extra memory used could hurt more than TLB overhead
 556          * is likely to save.  The admin can still enable it through /sys.
 557          */
 558         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 559                 transparent_hugepage_flags = 0;
 560
 561         start_khugepaged();
 562
 563         return 0;
 564 out:
 565         hugepage_exit_sysfs(hugepage_kobj);
 566         return err;
 567 }
 568 module_init(hugepage_init)
 569
 570 static int __init setup_transparent_hugepage(char *str)
 571 {
 572         int ret = 0;
 573         if (!str)
 574                 goto out;
 575         if (!strcmp(str, "always")) {
 576                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 577                         &transparent_hugepage_flags);
 578                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 579                           &transparent_hugepage_flags);
 580                 ret = 1;
 581         } else if (!strcmp(str, "madvise")) {
 582                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 583                           &transparent_hugepage_flags);
 584                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 585                         &transparent_hugepage_flags);
 586                 ret = 1;
 587         } else if (!strcmp(str, "never")) {
 588                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 589                           &transparent_hugepage_flags);
 590                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 591                           &transparent_hugepage_flags);
 592                 ret = 1;
 593         }
 594 out:
 595         if (!ret)
 596                 printk(KERN_WARNING
 597                        "transparent_hugepage= cannot parse, ignored\n");
 598         return ret;
 599 }
 600 __setup("transparent_hugepage=", setup_transparent_hugepage);
 601
 602 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 603 {
 604         if (likely(vma->vm_flags & VM_WRITE))
 605                 pmd = pmd_mkwrite(pmd);
 606         return pmd;
 607 }
 608
 609 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 610                                         struct vm_area_struct *vma,
 611                                         unsigned long haddr, pmd_t *pmd,
 612                                         struct page *page)
 613 {
 614         pgtable_t pgtable;
 615
 616         VM_BUG_ON(!PageCompound(page));
 617         pgtable = pte_alloc_one(mm, haddr);
 618         if (unlikely(!pgtable))
 619                 return VM_FAULT_OOM;
 620
 621         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 622         __SetPageUptodate(page);
 623
 624         spin_lock(&mm->page_table_lock);
 625         if (unlikely(!pmd_none(*pmd))) {
 626                 spin_unlock(&mm->page_table_lock);
 627                 mem_cgroup_uncharge_page(page);
 628                 put_page(page);
 629                 pte_free(mm, pgtable);
 630         } else {
 631                 pmd_t entry;
 632                 entry = mk_pmd(page, vma->vm_page_prot);
 633                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 634                 entry = pmd_mkhuge(entry);
 635                 /*
 636                  * The spinlocking to take the lru_lock inside
 637                  * page_add_new_anon_rmap() acts as a full memory
 638                  * barrier to be sure clear_huge_page writes become
 639                  * visible after the set_pmd_at() write.
 640                  */
 641                 page_add_new_anon_rmap(page, vma, haddr);
 642                 set_pmd_at(mm, haddr, pmd, entry);
 643                 pgtable_trans_huge_deposit(mm, pgtable);
 644                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 645                 mm->nr_ptes++;
 646                 spin_unlock(&mm->page_table_lock);
 647         }
 648
 649         return 0;
 650 }
 651
 652 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 653 {
 654         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 655 }
 656
 657 static inline struct page *alloc_hugepage_vma(int defrag,
 658                                               struct vm_area_struct *vma,
 659                                               unsigned long haddr, int nd,
 660                                               gfp_t extra_gfp)
 661 {
 662         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 663                                HPAGE_PMD_ORDER, vma, haddr, nd);
 664 }
 665
 666 #ifndef CONFIG_NUMA
 667 static inline struct page *alloc_hugepage(int defrag)
 668 {
 669         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 670                            HPAGE_PMD_ORDER);
 671 }
 672 #endif
 673
 674 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 675                                unsigned long address, pmd_t *pmd,
 676                                unsigned int flags)
 677 {
 678         struct page *page;
 679         unsigned long haddr = address & HPAGE_PMD_MASK;
 680         pte_t *pte;
 681
 682         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 683                 if (unlikely(anon_vma_prepare(vma)))
 684                         return VM_FAULT_OOM;
 685                 if (unlikely(khugepaged_enter(vma)))
 686                         return VM_FAULT_OOM;
 687                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 688                                           vma, haddr, numa_node_id(), 0);
 689                 if (unlikely(!page)) {
 690                         count_vm_event(THP_FAULT_FALLBACK);
 691                         goto out;
 692                 }
 693                 count_vm_event(THP_FAULT_ALLOC);
 694                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 695                         put_page(page);
 696                         goto out;
 697                 }
 698                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 699                                                           page))) {
 700                         mem_cgroup_uncharge_page(page);
 701                         put_page(page);
 702                         goto out;
 703                 }
 704
 705                 return 0;
 706         }
 707 out:
 708         /*
 709          * Use __pte_alloc instead of pte_alloc_map, because we can't
 710          * run pte_offset_map on the pmd, if an huge pmd could
 711          * materialize from under us from a different thread.
 712          */
 713         if (unlikely(pmd_none(*pmd)) &&
 714             unlikely(__pte_alloc(mm, vma, pmd, address)))
 715                 return VM_FAULT_OOM;
 716         /* if an huge pmd materialized from under us just retry later */
 717         if (unlikely(pmd_trans_huge(*pmd)))
 718                 return 0;
 719         /*
 720          * A regular pmd is established and it can't morph into a huge pmd
 721          * from under us anymore at this point because we hold the mmap_sem
 722          * read mode and khugepaged takes it in write mode. So now it's
 723          * safe to run pte_offset_map().
 724          */
 725         pte = pte_offset_map(pmd, address);
 726         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 727 }
 728
 729 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 730                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 731                   struct vm_area_struct *vma)
 732 {
 733         struct page *src_page;
 734         pmd_t pmd;
 735         pgtable_t pgtable;
 736         int ret;
 737
 738         ret = -ENOMEM;
 739         pgtable = pte_alloc_one(dst_mm, addr);
 740         if (unlikely(!pgtable))
 741                 goto out;
 742
 743         spin_lock(&dst_mm->page_table_lock);
 744         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 745
 746         ret = -EAGAIN;
 747         pmd = *src_pmd;
 748         if (unlikely(!pmd_trans_huge(pmd))) {
 749                 pte_free(dst_mm, pgtable);
 750                 goto out_unlock;
 751         }
 752         if (unlikely(pmd_trans_splitting(pmd))) {
 753                 /* split huge page running from under us */
 754                 spin_unlock(&src_mm->page_table_lock);
 755                 spin_unlock(&dst_mm->page_table_lock);
 756                 pte_free(dst_mm, pgtable);
 757
 758                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 759                 goto out;
 760         }
 761         src_page = pmd_page(pmd);
 762         VM_BUG_ON(!PageHead(src_page));
 763         get_page(src_page);
 764         page_dup_rmap(src_page);
 765         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 766
 767         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 768         pmd = pmd_mkold(pmd_wrprotect(pmd));
 769         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 770         pgtable_trans_huge_deposit(dst_mm, pgtable);
 771         dst_mm->nr_ptes++;
 772
 773         ret = 0;
 774 out_unlock:
 775         spin_unlock(&src_mm->page_table_lock);
 776         spin_unlock(&dst_mm->page_table_lock);
 777 out:
 778         return ret;
 779 }
 780
 781 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 782                                         struct vm_area_struct *vma,
 783                                         unsigned long address,
 784                                         pmd_t *pmd, pmd_t orig_pmd,
 785                                         struct page *page,
 786                                         unsigned long haddr)
 787 {
 788         pgtable_t pgtable;
 789         pmd_t _pmd;
 790         int ret = 0, i;
 791         struct page **pages;
 792         unsigned long mmun_start;       /* For mmu_notifiers */
 793         unsigned long mmun_end;         /* For mmu_notifiers */
 794
 795         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 796                         GFP_KERNEL);
 797         if (unlikely(!pages)) {
 798                 ret |= VM_FAULT_OOM;
 799                 goto out;
 800         }
 801
 802         for (i = 0; i < HPAGE_PMD_NR; i++) {
 803                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 804                                                __GFP_OTHER_NODE,
 805                                                vma, address, page_to_nid(page));
 806                 if (unlikely(!pages[i] ||
 807                              mem_cgroup_newpage_charge(pages[i], mm,
 808                                                        GFP_KERNEL))) {
 809                         if (pages[i])
 810                                 put_page(pages[i]);
 811                         mem_cgroup_uncharge_start();
 812                         while (--i >= 0) {
 813                                 mem_cgroup_uncharge_page(pages[i]);
 814                                 put_page(pages[i]);
 815                         }
 816                         mem_cgroup_uncharge_end();
 817                         kfree(pages);
 818                         ret |= VM_FAULT_OOM;
 819                         goto out;
 820                 }
 821         }
 822
 823         for (i = 0; i < HPAGE_PMD_NR; i++) {
 824                 copy_user_highpage(pages[i], page + i,
 825                                    haddr + PAGE_SIZE * i, vma);
 826                 __SetPageUptodate(pages[i]);
 827                 cond_resched();
 828         }
 829
 830         mmun_start = haddr;
 831         mmun_end   = haddr + HPAGE_PMD_SIZE;
 832         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 833
 834         spin_lock(&mm->page_table_lock);
 835         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 836                 goto out_free_pages;
 837         VM_BUG_ON(!PageHead(page));
 838
 839         pmdp_clear_flush(vma, haddr, pmd);
 840         /* leave pmd empty until pte is filled */
 841
 842         pgtable = pgtable_trans_huge_withdraw(mm);
 843         pmd_populate(mm, &_pmd, pgtable);
 844
 845         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 846                 pte_t *pte, entry;
 847                 entry = mk_pte(pages[i], vma->vm_page_prot);
 848                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 849                 page_add_new_anon_rmap(pages[i], vma, haddr);
 850                 pte = pte_offset_map(&_pmd, haddr);
 851                 VM_BUG_ON(!pte_none(*pte));
 852                 set_pte_at(mm, haddr, pte, entry);
 853                 pte_unmap(pte);
 854         }
 855         kfree(pages);
 856
 857         smp_wmb(); /* make pte visible before pmd */
 858         pmd_populate(mm, pmd, pgtable);
 859         page_remove_rmap(page);
 860         spin_unlock(&mm->page_table_lock);
 861
 862         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 863
 864         ret |= VM_FAULT_WRITE;
 865         put_page(page);
 866
 867 out:
 868         return ret;
 869
 870 out_free_pages:
 871         spin_unlock(&mm->page_table_lock);
 872         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 873         mem_cgroup_uncharge_start();
 874         for (i = 0; i < HPAGE_PMD_NR; i++) {
 875                 mem_cgroup_uncharge_page(pages[i]);
 876                 put_page(pages[i]);
 877         }
 878         mem_cgroup_uncharge_end();
 879         kfree(pages);
 880         goto out;
 881 }
 882
 883 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 884                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 885 {
 886         int ret = 0;
 887         struct page *page, *new_page;
 888         unsigned long haddr;
 889         unsigned long mmun_start;       /* For mmu_notifiers */
 890         unsigned long mmun_end;         /* For mmu_notifiers */
 891
 892         VM_BUG_ON(!vma->anon_vma);
 893         spin_lock(&mm->page_table_lock);
 894         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 895                 goto out_unlock;
 896
 897         page = pmd_page(orig_pmd);
 898         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 899         haddr = address & HPAGE_PMD_MASK;
 900         if (page_mapcount(page) == 1) {
 901                 pmd_t entry;
 902                 entry = pmd_mkyoung(orig_pmd);
 903                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 904                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 905                         update_mmu_cache_pmd(vma, address, pmd);
 906                 ret |= VM_FAULT_WRITE;
 907                 goto out_unlock;
 908         }
 909         get_page(page);
 910         spin_unlock(&mm->page_table_lock);
 911
 912         if (transparent_hugepage_enabled(vma) &&
 913             !transparent_hugepage_debug_cow())
 914                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 915                                               vma, haddr, numa_node_id(), 0);
 916         else
 917                 new_page = NULL;
 918
 919         if (unlikely(!new_page)) {
 920                 count_vm_event(THP_FAULT_FALLBACK);
 921                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 922                                                    pmd, orig_pmd, page, haddr);
 923                 if (ret & VM_FAULT_OOM)
 924                         split_huge_page(page);
 925                 put_page(page);
 926                 goto out;
 927         }
 928         count_vm_event(THP_FAULT_ALLOC);
 929
 930         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 931                 put_page(new_page);
 932                 split_huge_page(page);
 933                 put_page(page);
 934                 ret |= VM_FAULT_OOM;
 935                 goto out;
 936         }
 937
 938         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 939         __SetPageUptodate(new_page);
 940
 941         mmun_start = haddr;
 942         mmun_end   = haddr + HPAGE_PMD_SIZE;
 943         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 944
 945         spin_lock(&mm->page_table_lock);
 946         put_page(page);
 947         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 948                 spin_unlock(&mm->page_table_lock);
 949                 mem_cgroup_uncharge_page(new_page);
 950                 put_page(new_page);
 951                 goto out_mn;
 952         } else {
 953                 pmd_t entry;
 954                 VM_BUG_ON(!PageHead(page));
 955                 entry = mk_pmd(new_page, vma->vm_page_prot);
 956                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 957                 entry = pmd_mkhuge(entry);
 958                 pmdp_clear_flush(vma, haddr, pmd);
 959                 page_add_new_anon_rmap(new_page, vma, haddr);
 960                 set_pmd_at(mm, haddr, pmd, entry);
 961                 update_mmu_cache_pmd(vma, address, pmd);
 962                 page_remove_rmap(page);
 963                 put_page(page);
 964                 ret |= VM_FAULT_WRITE;
 965         }
 966         spin_unlock(&mm->page_table_lock);
 967 out_mn:
 968         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 969 out:
 970         return ret;
 971 out_unlock:
 972         spin_unlock(&mm->page_table_lock);
 973         return ret;
 974 }
 975
 976 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
 977                                    unsigned long addr,
 978                                    pmd_t *pmd,
 979                                    unsigned int flags)
 980 {
 981         struct mm_struct *mm = vma->vm_mm;
 982         struct page *page = NULL;
 983
 984         assert_spin_locked(&mm->page_table_lock);
 985
 986         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 987                 goto out;
 988
 989         page = pmd_page(*pmd);
 990         VM_BUG_ON(!PageHead(page));
 991         if (flags & FOLL_TOUCH) {
 992                 pmd_t _pmd;
 993                 /*
 994                  * We should set the dirty bit only for FOLL_WRITE but
 995                  * for now the dirty bit in the pmd is meaningless.
 996                  * And if the dirty bit will become meaningful and
 997                  * we'll only set it with FOLL_WRITE, an atomic
 998                  * set_bit will be required on the pmd to set the
 999                  * young bit, instead of the current set_pmd_at.
1000                  */
1001                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1002                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1003         }
1004         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1005                 if (page->mapping && trylock_page(page)) {
1006                         lru_add_drain();
1007                         if (page->mapping)
1008                                 mlock_vma_page(page);
1009                         unlock_page(page);
1010                 }
1011         }
1012         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1013         VM_BUG_ON(!PageCompound(page));
1014         if (flags & FOLL_GET)
1015                 get_page_foll(page);
1016
1017 out:
1018         return page;
1019 }
1020
1021 /* NUMA hinting page fault entry point for trans huge pmds */
1022 int do_huge_pmd_numa_page(struct mm_struct *mm, unsigned long addr,
1023                                 pmd_t pmd, pmd_t *pmdp)
1024 {
1025         struct page *page;
1026         unsigned long haddr = addr & HPAGE_PMD_MASK;
1027
1028         spin_lock(&mm->page_table_lock);
1029         if (unlikely(!pmd_same(pmd, *pmdp)))
1030                 goto out_unlock;
1031
1032         page = pmd_page(pmd);
1033         pmd = pmd_mknonnuma(pmd);
1034         set_pmd_at(mm, haddr, pmdp, pmd);
1035         VM_BUG_ON(pmd_numa(*pmdp));
1036         update_mmu_cache_pmd(vma, addr, pmdp);
1037
1038 out_unlock:
1039         spin_unlock(&mm->page_table_lock);
1040         return 0;
1041 }
1042
1043 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1044                  pmd_t *pmd, unsigned long addr)
1045 {
1046         int ret = 0;
1047
1048         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1049                 struct page *page;
1050                 pgtable_t pgtable;
1051                 pmd_t orig_pmd;
1052                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1053                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1054                 page = pmd_page(orig_pmd);
1055                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1056                 page_remove_rmap(page);
1057                 VM_BUG_ON(page_mapcount(page) < 0);
1058                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1059                 VM_BUG_ON(!PageHead(page));
1060                 tlb->mm->nr_ptes--;
1061                 spin_unlock(&tlb->mm->page_table_lock);
1062                 tlb_remove_page(tlb, page);
1063                 pte_free(tlb->mm, pgtable);
1064                 ret = 1;
1065         }
1066         return ret;
1067 }
1068
1069 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1070                 unsigned long addr, unsigned long end,
1071                 unsigned char *vec)
1072 {
1073         int ret = 0;
1074
1075         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1076                 /*
1077                  * All logical pages in the range are present
1078                  * if backed by a huge page.
1079                  */
1080                 spin_unlock(&vma->vm_mm->page_table_lock);
1081                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1082                 ret = 1;
1083         }
1084
1085         return ret;
1086 }
1087
1088 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1089                   unsigned long old_addr,
1090                   unsigned long new_addr, unsigned long old_end,
1091                   pmd_t *old_pmd, pmd_t *new_pmd)
1092 {
1093         int ret = 0;
1094         pmd_t pmd;
1095
1096         struct mm_struct *mm = vma->vm_mm;
1097
1098         if ((old_addr & ~HPAGE_PMD_MASK) ||
1099             (new_addr & ~HPAGE_PMD_MASK) ||
1100             old_end - old_addr < HPAGE_PMD_SIZE ||
1101             (new_vma->vm_flags & VM_NOHUGEPAGE))
1102                 goto out;
1103
1104         /*
1105          * The destination pmd shouldn't be established, free_pgtables()
1106          * should have release it.
1107          */
1108         if (WARN_ON(!pmd_none(*new_pmd))) {
1109                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1110                 goto out;
1111         }
1112
1113         ret = __pmd_trans_huge_lock(old_pmd, vma);
1114         if (ret == 1) {
1115                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1116                 VM_BUG_ON(!pmd_none(*new_pmd));
1117                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1118                 spin_unlock(&mm->page_table_lock);
1119         }
1120 out:
1121         return ret;
1122 }
1123
1124 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1125                 unsigned long addr, pgprot_t newprot)
1126 {
1127         struct mm_struct *mm = vma->vm_mm;
1128         int ret = 0;
1129
1130         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1131                 pmd_t entry;
1132                 entry = pmdp_get_and_clear(mm, addr, pmd);
1133                 entry = pmd_modify(entry, newprot);
1134                 set_pmd_at(mm, addr, pmd, entry);
1135                 spin_unlock(&vma->vm_mm->page_table_lock);
1136                 ret = 1;
1137         }
1138
1139         return ret;
1140 }
1141
1142 /*
1143  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1144  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1145  *
1146  * Note that if it returns 1, this routine returns without unlocking page
1147  * table locks. So callers must unlock them.
1148  */
1149 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1150 {
1151         spin_lock(&vma->vm_mm->page_table_lock);
1152         if (likely(pmd_trans_huge(*pmd))) {
1153                 if (unlikely(pmd_trans_splitting(*pmd))) {
1154                         spin_unlock(&vma->vm_mm->page_table_lock);
1155                         wait_split_huge_page(vma->anon_vma, pmd);
1156                         return -1;
1157                 } else {
1158                         /* Thp mapped by 'pmd' is stable, so we can
1159                          * handle it as it is. */
1160                         return 1;
1161                 }
1162         }
1163         spin_unlock(&vma->vm_mm->page_table_lock);
1164         return 0;
1165 }
1166
1167 pmd_t *page_check_address_pmd(struct page *page,
1168                               struct mm_struct *mm,
1169                               unsigned long address,
1170                               enum page_check_address_pmd_flag flag)
1171 {
1172         pgd_t *pgd;
1173         pud_t *pud;
1174         pmd_t *pmd, *ret = NULL;
1175
1176         if (address & ~HPAGE_PMD_MASK)
1177                 goto out;
1178
1179         pgd = pgd_offset(mm, address);
1180         if (!pgd_present(*pgd))
1181                 goto out;
1182
1183         pud = pud_offset(pgd, address);
1184         if (!pud_present(*pud))
1185                 goto out;
1186
1187         pmd = pmd_offset(pud, address);
1188         if (pmd_none(*pmd))
1189                 goto out;
1190         if (pmd_page(*pmd) != page)
1191                 goto out;
1192         /*
1193          * split_vma() may create temporary aliased mappings. There is
1194          * no risk as long as all huge pmd are found and have their
1195          * splitting bit set before __split_huge_page_refcount
1196          * runs. Finding the same huge pmd more than once during the
1197          * same rmap walk is not a problem.
1198          */
1199         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1200             pmd_trans_splitting(*pmd))
1201                 goto out;
1202         if (pmd_trans_huge(*pmd)) {
1203                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1204                           !pmd_trans_splitting(*pmd));
1205                 ret = pmd;
1206         }
1207 out:
1208         return ret;
1209 }
1210
1211 static int __split_huge_page_splitting(struct page *page,
1212                                        struct vm_area_struct *vma,
1213                                        unsigned long address)
1214 {
1215         struct mm_struct *mm = vma->vm_mm;
1216         pmd_t *pmd;
1217         int ret = 0;
1218         /* For mmu_notifiers */
1219         const unsigned long mmun_start = address;
1220         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1221
1222         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1223         spin_lock(&mm->page_table_lock);
1224         pmd = page_check_address_pmd(page, mm, address,
1225                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1226         if (pmd) {
1227                 /*
1228                  * We can't temporarily set the pmd to null in order
1229                  * to split it, the pmd must remain marked huge at all
1230                  * times or the VM won't take the pmd_trans_huge paths
1231                  * and it won't wait on the anon_vma->root->mutex to
1232                  * serialize against split_huge_page*.
1233                  */
1234                 pmdp_splitting_flush(vma, address, pmd);
1235                 ret = 1;
1236         }
1237         spin_unlock(&mm->page_table_lock);
1238         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1239
1240         return ret;
1241 }
1242
1243 static void __split_huge_page_refcount(struct page *page)
1244 {
1245         int i;
1246         struct zone *zone = page_zone(page);
1247         struct lruvec *lruvec;
1248         int tail_count = 0;
1249
1250         /* prevent PageLRU to go away from under us, and freeze lru stats */
1251         spin_lock_irq(&zone->lru_lock);
1252         lruvec = mem_cgroup_page_lruvec(page, zone);
1253
1254         compound_lock(page);
1255         /* complete memcg works before add pages to LRU */
1256         mem_cgroup_split_huge_fixup(page);
1257
1258         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1259                 struct page *page_tail = page + i;
1260
1261                 /* tail_page->_mapcount cannot change */
1262                 BUG_ON(page_mapcount(page_tail) < 0);
1263                 tail_count += page_mapcount(page_tail);
1264                 /* check for overflow */
1265                 BUG_ON(tail_count < 0);
1266                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1267                 /*
1268                  * tail_page->_count is zero and not changing from
1269                  * under us. But get_page_unless_zero() may be running
1270                  * from under us on the tail_page. If we used
1271                  * atomic_set() below instead of atomic_add(), we
1272                  * would then run atomic_set() concurrently with
1273                  * get_page_unless_zero(), and atomic_set() is
1274                  * implemented in C not using locked ops. spin_unlock
1275                  * on x86 sometime uses locked ops because of PPro
1276                  * errata 66, 92, so unless somebody can guarantee
1277                  * atomic_set() here would be safe on all archs (and
1278                  * not only on x86), it's safer to use atomic_add().
1279                  */
1280                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1281                            &page_tail->_count);
1282
1283                 /* after clearing PageTail the gup refcount can be released */
1284                 smp_mb();
1285
1286                 /*
1287                  * retain hwpoison flag of the poisoned tail page:
1288                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1289                  *   by the memory-failure.
1290                  */
1291                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1292                 page_tail->flags |= (page->flags &
1293                                      ((1L << PG_referenced) |
1294                                       (1L << PG_swapbacked) |
1295                                       (1L << PG_mlocked) |
1296                                       (1L << PG_uptodate)));
1297                 page_tail->flags |= (1L << PG_dirty);
1298
1299                 /* clear PageTail before overwriting first_page */
1300                 smp_wmb();
1301
1302                 /*
1303                  * __split_huge_page_splitting() already set the
1304                  * splitting bit in all pmd that could map this
1305                  * hugepage, that will ensure no CPU can alter the
1306                  * mapcount on the head page. The mapcount is only
1307                  * accounted in the head page and it has to be
1308                  * transferred to all tail pages in the below code. So
1309                  * for this code to be safe, the split the mapcount
1310                  * can't change. But that doesn't mean userland can't
1311                  * keep changing and reading the page contents while
1312                  * we transfer the mapcount, so the pmd splitting
1313                  * status is achieved setting a reserved bit in the
1314                  * pmd, not by clearing the present bit.
1315                 */
1316                 page_tail->_mapcount = page->_mapcount;
1317
1318                 BUG_ON(page_tail->mapping);
1319                 page_tail->mapping = page->mapping;
1320
1321                 page_tail->index = page->index + i;
1322
1323                 BUG_ON(!PageAnon(page_tail));
1324                 BUG_ON(!PageUptodate(page_tail));
1325                 BUG_ON(!PageDirty(page_tail));
1326                 BUG_ON(!PageSwapBacked(page_tail));
1327
1328                 lru_add_page_tail(page, page_tail, lruvec);
1329         }
1330         atomic_sub(tail_count, &page->_count);
1331         BUG_ON(atomic_read(&page->_count) <= 0);
1332
1333         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1334         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1335
1336         ClearPageCompound(page);
1337         compound_unlock(page);
1338         spin_unlock_irq(&zone->lru_lock);
1339
1340         for (i = 1; i < HPAGE_PMD_NR; i++) {
1341                 struct page *page_tail = page + i;
1342                 BUG_ON(page_count(page_tail) <= 0);
1343                 /*
1344                  * Tail pages may be freed if there wasn't any mapping
1345                  * like if add_to_swap() is running on a lru page that
1346                  * had its mapping zapped. And freeing these pages
1347                  * requires taking the lru_lock so we do the put_page
1348                  * of the tail pages after the split is complete.
1349                  */
1350                 put_page(page_tail);
1351         }
1352
1353         /*
1354          * Only the head page (now become a regular page) is required
1355          * to be pinned by the caller.
1356          */
1357         BUG_ON(page_count(page) <= 0);
1358 }
1359
1360 static int __split_huge_page_map(struct page *page,
1361                                  struct vm_area_struct *vma,
1362                                  unsigned long address)
1363 {
1364         struct mm_struct *mm = vma->vm_mm;
1365         pmd_t *pmd, _pmd;
1366         int ret = 0, i;
1367         pgtable_t pgtable;
1368         unsigned long haddr;
1369
1370         spin_lock(&mm->page_table_lock);
1371         pmd = page_check_address_pmd(page, mm, address,
1372                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1373         if (pmd) {
1374                 pgtable = pgtable_trans_huge_withdraw(mm);
1375                 pmd_populate(mm, &_pmd, pgtable);
1376
1377                 haddr = address;
1378                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1379                         pte_t *pte, entry;
1380                         BUG_ON(PageCompound(page+i));
1381                         entry = mk_pte(page + i, vma->vm_page_prot);
1382                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1383                         if (!pmd_write(*pmd))
1384                                 entry = pte_wrprotect(entry);
1385                         else
1386                                 BUG_ON(page_mapcount(page) != 1);
1387                         if (!pmd_young(*pmd))
1388                                 entry = pte_mkold(entry);
1389                         if (pmd_numa(*pmd))
1390                                 entry = pte_mknuma(entry);
1391                         pte = pte_offset_map(&_pmd, haddr);
1392                         BUG_ON(!pte_none(*pte));
1393                         set_pte_at(mm, haddr, pte, entry);
1394                         pte_unmap(pte);
1395                 }
1396
1397                 smp_wmb(); /* make pte visible before pmd */
1398                 /*
1399                  * Up to this point the pmd is present and huge and
1400                  * userland has the whole access to the hugepage
1401                  * during the split (which happens in place). If we
1402                  * overwrite the pmd with the not-huge version
1403                  * pointing to the pte here (which of course we could
1404                  * if all CPUs were bug free), userland could trigger
1405                  * a small page size TLB miss on the small sized TLB
1406                  * while the hugepage TLB entry is still established
1407                  * in the huge TLB. Some CPU doesn't like that. See
1408                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1409                  * Erratum 383 on page 93. Intel should be safe but is
1410                  * also warns that it's only safe if the permission
1411                  * and cache attributes of the two entries loaded in
1412                  * the two TLB is identical (which should be the case
1413                  * here). But it is generally safer to never allow
1414                  * small and huge TLB entries for the same virtual
1415                  * address to be loaded simultaneously. So instead of
1416                  * doing "pmd_populate(); flush_tlb_range();" we first
1417                  * mark the current pmd notpresent (atomically because
1418                  * here the pmd_trans_huge and pmd_trans_splitting
1419                  * must remain set at all times on the pmd until the
1420                  * split is complete for this pmd), then we flush the
1421                  * SMP TLB and finally we write the non-huge version
1422                  * of the pmd entry with pmd_populate.
1423                  */
1424                 pmdp_invalidate(vma, address, pmd);
1425                 pmd_populate(mm, pmd, pgtable);
1426                 ret = 1;
1427         }
1428         spin_unlock(&mm->page_table_lock);
1429
1430         return ret;
1431 }
1432
1433 /* must be called with anon_vma->root->mutex hold */
1434 static void __split_huge_page(struct page *page,
1435                               struct anon_vma *anon_vma)
1436 {
1437         int mapcount, mapcount2;
1438         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1439         struct anon_vma_chain *avc;
1440
1441         BUG_ON(!PageHead(page));
1442         BUG_ON(PageTail(page));
1443
1444         mapcount = 0;
1445         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1446                 struct vm_area_struct *vma = avc->vma;
1447                 unsigned long addr = vma_address(page, vma);
1448                 BUG_ON(is_vma_temporary_stack(vma));
1449                 mapcount += __split_huge_page_splitting(page, vma, addr);
1450         }
1451         /*
1452          * It is critical that new vmas are added to the tail of the
1453          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1454          * and establishes a child pmd before
1455          * __split_huge_page_splitting() freezes the parent pmd (so if
1456          * we fail to prevent copy_huge_pmd() from running until the
1457          * whole __split_huge_page() is complete), we will still see
1458          * the newly established pmd of the child later during the
1459          * walk, to be able to set it as pmd_trans_splitting too.
1460          */
1461         if (mapcount != page_mapcount(page))
1462                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1463                        mapcount, page_mapcount(page));
1464         BUG_ON(mapcount != page_mapcount(page));
1465
1466         __split_huge_page_refcount(page);
1467
1468         mapcount2 = 0;
1469         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1470                 struct vm_area_struct *vma = avc->vma;
1471                 unsigned long addr = vma_address(page, vma);
1472                 BUG_ON(is_vma_temporary_stack(vma));
1473                 mapcount2 += __split_huge_page_map(page, vma, addr);
1474         }
1475         if (mapcount != mapcount2)
1476                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1477                        mapcount, mapcount2, page_mapcount(page));
1478         BUG_ON(mapcount != mapcount2);
1479 }
1480
1481 int split_huge_page(struct page *page)
1482 {
1483         struct anon_vma *anon_vma;
1484         int ret = 1;
1485
1486         BUG_ON(!PageAnon(page));
1487         anon_vma = page_lock_anon_vma(page);
1488         if (!anon_vma)
1489                 goto out;
1490         ret = 0;
1491         if (!PageCompound(page))
1492                 goto out_unlock;
1493
1494         BUG_ON(!PageSwapBacked(page));
1495         __split_huge_page(page, anon_vma);
1496         count_vm_event(THP_SPLIT);
1497
1498         BUG_ON(PageCompound(page));
1499 out_unlock:
1500         page_unlock_anon_vma(anon_vma);
1501 out:
1502         return ret;
1503 }
1504
1505 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1506
1507 int hugepage_madvise(struct vm_area_struct *vma,
1508                      unsigned long *vm_flags, int advice)
1509 {
1510         struct mm_struct *mm = vma->vm_mm;
1511
1512         switch (advice) {
1513         case MADV_HUGEPAGE:
1514                 /*
1515                  * Be somewhat over-protective like KSM for now!
1516                  */
1517                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1518                         return -EINVAL;
1519                 if (mm->def_flags & VM_NOHUGEPAGE)
1520                         return -EINVAL;
1521                 *vm_flags &= ~VM_NOHUGEPAGE;
1522                 *vm_flags |= VM_HUGEPAGE;
1523                 /*
1524                  * If the vma become good for khugepaged to scan,
1525                  * register it here without waiting a page fault that
1526                  * may not happen any time soon.
1527                  */
1528                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1529                         return -ENOMEM;
1530                 break;
1531         case MADV_NOHUGEPAGE:
1532                 /*
1533                  * Be somewhat over-protective like KSM for now!
1534                  */
1535                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1536                         return -EINVAL;
1537                 *vm_flags &= ~VM_HUGEPAGE;
1538                 *vm_flags |= VM_NOHUGEPAGE;
1539                 /*
1540                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1541                  * this vma even if we leave the mm registered in khugepaged if
1542                  * it got registered before VM_NOHUGEPAGE was set.
1543                  */
1544                 break;
1545         }
1546
1547         return 0;
1548 }
1549
1550 static int __init khugepaged_slab_init(void)
1551 {
1552         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1553                                           sizeof(struct mm_slot),
1554                                           __alignof__(struct mm_slot), 0, NULL);
1555         if (!mm_slot_cache)
1556                 return -ENOMEM;
1557
1558         return 0;
1559 }
1560
1561 static void __init khugepaged_slab_free(void)
1562 {
1563         kmem_cache_destroy(mm_slot_cache);
1564         mm_slot_cache = NULL;
1565 }
1566
1567 static inline struct mm_slot *alloc_mm_slot(void)
1568 {
1569         if (!mm_slot_cache)     /* initialization failed */
1570                 return NULL;
1571         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1572 }
1573
1574 static inline void free_mm_slot(struct mm_slot *mm_slot)
1575 {
1576         kmem_cache_free(mm_slot_cache, mm_slot);
1577 }
1578
1579 static int __init mm_slots_hash_init(void)
1580 {
1581         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1582                                 GFP_KERNEL);
1583         if (!mm_slots_hash)
1584                 return -ENOMEM;
1585         return 0;
1586 }
1587
1588 #if 0
1589 static void __init mm_slots_hash_free(void)
1590 {
1591         kfree(mm_slots_hash);
1592         mm_slots_hash = NULL;
1593 }
1594 #endif
1595
1596 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1597 {
1598         struct mm_slot *mm_slot;
1599         struct hlist_head *bucket;
1600         struct hlist_node *node;
1601
1602         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1603                                 % MM_SLOTS_HASH_HEADS];
1604         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1605                 if (mm == mm_slot->mm)
1606                         return mm_slot;
1607         }
1608         return NULL;
1609 }
1610
1611 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1612                                     struct mm_slot *mm_slot)
1613 {
1614         struct hlist_head *bucket;
1615
1616         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1617                                 % MM_SLOTS_HASH_HEADS];
1618         mm_slot->mm = mm;
1619         hlist_add_head(&mm_slot->hash, bucket);
1620 }
1621
1622 static inline int khugepaged_test_exit(struct mm_struct *mm)
1623 {
1624         return atomic_read(&mm->mm_users) == 0;
1625 }
1626
1627 int __khugepaged_enter(struct mm_struct *mm)
1628 {
1629         struct mm_slot *mm_slot;
1630         int wakeup;
1631
1632         mm_slot = alloc_mm_slot();
1633         if (!mm_slot)
1634                 return -ENOMEM;
1635
1636         /* __khugepaged_exit() must not run from under us */
1637         VM_BUG_ON(khugepaged_test_exit(mm));
1638         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1639                 free_mm_slot(mm_slot);
1640                 return 0;
1641         }
1642
1643         spin_lock(&khugepaged_mm_lock);
1644         insert_to_mm_slots_hash(mm, mm_slot);
1645         /*
1646          * Insert just behind the scanning cursor, to let the area settle
1647          * down a little.
1648          */
1649         wakeup = list_empty(&khugepaged_scan.mm_head);
1650         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1651         spin_unlock(&khugepaged_mm_lock);
1652
1653         atomic_inc(&mm->mm_count);
1654         if (wakeup)
1655                 wake_up_interruptible(&khugepaged_wait);
1656
1657         return 0;
1658 }
1659
1660 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1661 {
1662         unsigned long hstart, hend;
1663         if (!vma->anon_vma)
1664                 /*
1665                  * Not yet faulted in so we will register later in the
1666                  * page fault if needed.
1667                  */
1668                 return 0;
1669         if (vma->vm_ops)
1670                 /* khugepaged not yet working on file or special mappings */
1671                 return 0;
1672         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1673         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1674         hend = vma->vm_end & HPAGE_PMD_MASK;
1675         if (hstart < hend)
1676                 return khugepaged_enter(vma);
1677         return 0;
1678 }
1679
1680 void __khugepaged_exit(struct mm_struct *mm)
1681 {
1682         struct mm_slot *mm_slot;
1683         int free = 0;
1684
1685         spin_lock(&khugepaged_mm_lock);
1686         mm_slot = get_mm_slot(mm);
1687         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1688                 hlist_del(&mm_slot->hash);
1689                 list_del(&mm_slot->mm_node);
1690                 free = 1;
1691         }
1692         spin_unlock(&khugepaged_mm_lock);
1693
1694         if (free) {
1695                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1696                 free_mm_slot(mm_slot);
1697                 mmdrop(mm);
1698         } else if (mm_slot) {
1699                 /*
1700                  * This is required to serialize against
1701                  * khugepaged_test_exit() (which is guaranteed to run
1702                  * under mmap sem read mode). Stop here (after we
1703                  * return all pagetables will be destroyed) until
1704                  * khugepaged has finished working on the pagetables
1705                  * under the mmap_sem.
1706                  */
1707                 down_write(&mm->mmap_sem);
1708                 up_write(&mm->mmap_sem);
1709         }
1710 }
1711
1712 static void release_pte_page(struct page *page)
1713 {
1714         /* 0 stands for page_is_file_cache(page) == false */
1715         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1716         unlock_page(page);
1717         putback_lru_page(page);
1718 }
1719
1720 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1721 {
1722         while (--_pte >= pte) {
1723                 pte_t pteval = *_pte;
1724                 if (!pte_none(pteval))
1725                         release_pte_page(pte_page(pteval));
1726         }
1727 }
1728
1729 static void release_all_pte_pages(pte_t *pte)
1730 {
1731         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1732 }
1733
1734 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1735                                         unsigned long address,
1736                                         pte_t *pte)
1737 {
1738         struct page *page;
1739         pte_t *_pte;
1740         int referenced = 0, isolated = 0, none = 0;
1741         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1742              _pte++, address += PAGE_SIZE) {
1743                 pte_t pteval = *_pte;
1744                 if (pte_none(pteval)) {
1745                         if (++none <= khugepaged_max_ptes_none)
1746                                 continue;
1747                         else {
1748                                 release_pte_pages(pte, _pte);
1749                                 goto out;
1750                         }
1751                 }
1752                 if (!pte_present(pteval) || !pte_write(pteval)) {
1753                         release_pte_pages(pte, _pte);
1754                         goto out;
1755                 }
1756                 page = vm_normal_page(vma, address, pteval);
1757                 if (unlikely(!page)) {
1758                         release_pte_pages(pte, _pte);
1759                         goto out;
1760                 }
1761                 VM_BUG_ON(PageCompound(page));
1762                 BUG_ON(!PageAnon(page));
1763                 VM_BUG_ON(!PageSwapBacked(page));
1764
1765                 /* cannot use mapcount: can't collapse if there's a gup pin */
1766                 if (page_count(page) != 1) {
1767                         release_pte_pages(pte, _pte);
1768                         goto out;
1769                 }
1770                 /*
1771                  * We can do it before isolate_lru_page because the
1772                  * page can't be freed from under us. NOTE: PG_lock
1773                  * is needed to serialize against split_huge_page
1774                  * when invoked from the VM.
1775                  */
1776                 if (!trylock_page(page)) {
1777                         release_pte_pages(pte, _pte);
1778                         goto out;
1779                 }
1780                 /*
1781                  * Isolate the page to avoid collapsing an hugepage
1782                  * currently in use by the VM.
1783                  */
1784                 if (isolate_lru_page(page)) {
1785                         unlock_page(page);
1786                         release_pte_pages(pte, _pte);
1787                         goto out;
1788                 }
1789                 /* 0 stands for page_is_file_cache(page) == false */
1790                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1791                 VM_BUG_ON(!PageLocked(page));
1792                 VM_BUG_ON(PageLRU(page));
1793
1794                 /* If there is no mapped pte young don't collapse the page */
1795                 if (pte_young(pteval) || PageReferenced(page) ||
1796                     mmu_notifier_test_young(vma->vm_mm, address))
1797                         referenced = 1;
1798         }
1799         if (unlikely(!referenced))
1800                 release_all_pte_pages(pte);
1801         else
1802                 isolated = 1;
1803 out:
1804         return isolated;
1805 }
1806
1807 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1808                                       struct vm_area_struct *vma,
1809                                       unsigned long address,
1810                                       spinlock_t *ptl)
1811 {
1812         pte_t *_pte;
1813         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1814                 pte_t pteval = *_pte;
1815                 struct page *src_page;
1816
1817                 if (pte_none(pteval)) {
1818                         clear_user_highpage(page, address);
1819                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1820                 } else {
1821                         src_page = pte_page(pteval);
1822                         copy_user_highpage(page, src_page, address, vma);
1823                         VM_BUG_ON(page_mapcount(src_page) != 1);
1824                         release_pte_page(src_page);
1825                         /*
1826                          * ptl mostly unnecessary, but preempt has to
1827                          * be disabled to update the per-cpu stats
1828                          * inside page_remove_rmap().
1829                          */
1830                         spin_lock(ptl);
1831                         /*
1832                          * paravirt calls inside pte_clear here are
1833                          * superfluous.
1834                          */
1835                         pte_clear(vma->vm_mm, address, _pte);
1836                         page_remove_rmap(src_page);
1837                         spin_unlock(ptl);
1838                         free_page_and_swap_cache(src_page);
1839                 }
1840
1841                 address += PAGE_SIZE;
1842                 page++;
1843         }
1844 }
1845
1846 static void khugepaged_alloc_sleep(void)
1847 {
1848         wait_event_freezable_timeout(khugepaged_wait, false,
1849                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1850 }
1851
1852 #ifdef CONFIG_NUMA
1853 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1854 {
1855         if (IS_ERR(*hpage)) {
1856                 if (!*wait)
1857                         return false;
1858
1859                 *wait = false;
1860                 *hpage = NULL;
1861                 khugepaged_alloc_sleep();
1862         } else if (*hpage) {
1863                 put_page(*hpage);
1864                 *hpage = NULL;
1865         }
1866
1867         return true;
1868 }
1869
1870 static struct page
1871 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1872                        struct vm_area_struct *vma, unsigned long address,
1873                        int node)
1874 {
1875         VM_BUG_ON(*hpage);
1876         /*
1877          * Allocate the page while the vma is still valid and under
1878          * the mmap_sem read mode so there is no memory allocation
1879          * later when we take the mmap_sem in write mode. This is more
1880          * friendly behavior (OTOH it may actually hide bugs) to
1881          * filesystems in userland with daemons allocating memory in
1882          * the userland I/O paths.  Allocating memory with the
1883          * mmap_sem in read mode is good idea also to allow greater
1884          * scalability.
1885          */
1886         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1887                                       node, __GFP_OTHER_NODE);
1888
1889         /*
1890          * After allocating the hugepage, release the mmap_sem read lock in
1891          * preparation for taking it in write mode.
1892          */
1893         up_read(&mm->mmap_sem);
1894         if (unlikely(!*hpage)) {
1895                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1896                 *hpage = ERR_PTR(-ENOMEM);
1897                 return NULL;
1898         }
1899
1900         count_vm_event(THP_COLLAPSE_ALLOC);
1901         return *hpage;
1902 }
1903 #else
1904 static struct page *khugepaged_alloc_hugepage(bool *wait)
1905 {
1906         struct page *hpage;
1907
1908         do {
1909                 hpage = alloc_hugepage(khugepaged_defrag());
1910                 if (!hpage) {
1911                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1912                         if (!*wait)
1913                                 return NULL;
1914
1915                         *wait = false;
1916                         khugepaged_alloc_sleep();
1917                 } else
1918                         count_vm_event(THP_COLLAPSE_ALLOC);
1919         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1920
1921         return hpage;
1922 }
1923
1924 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1925 {
1926         if (!*hpage)
1927                 *hpage = khugepaged_alloc_hugepage(wait);
1928
1929         if (unlikely(!*hpage))
1930                 return false;
1931
1932         return true;
1933 }
1934
1935 static struct page
1936 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1937                        struct vm_area_struct *vma, unsigned long address,
1938                        int node)
1939 {
1940         up_read(&mm->mmap_sem);
1941         VM_BUG_ON(!*hpage);
1942         return  *hpage;
1943 }
1944 #endif
1945
1946 static void collapse_huge_page(struct mm_struct *mm,
1947                                    unsigned long address,
1948                                    struct page **hpage,
1949                                    struct vm_area_struct *vma,
1950                                    int node)
1951 {
1952         pgd_t *pgd;
1953         pud_t *pud;
1954         pmd_t *pmd, _pmd;
1955         pte_t *pte;
1956         pgtable_t pgtable;
1957         struct page *new_page;
1958         spinlock_t *ptl;
1959         int isolated;
1960         unsigned long hstart, hend;
1961         unsigned long mmun_start;       /* For mmu_notifiers */
1962         unsigned long mmun_end;         /* For mmu_notifiers */
1963
1964         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1965
1966         /* release the mmap_sem read lock. */
1967         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1968         if (!new_page)
1969                 return;
1970
1971         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1972                 return;
1973
1974         /*
1975          * Prevent all access to pagetables with the exception of
1976          * gup_fast later hanlded by the ptep_clear_flush and the VM
1977          * handled by the anon_vma lock + PG_lock.
1978          */
1979         down_write(&mm->mmap_sem);
1980         if (unlikely(khugepaged_test_exit(mm)))
1981                 goto out;
1982
1983         vma = find_vma(mm, address);
1984         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1985         hend = vma->vm_end & HPAGE_PMD_MASK;
1986         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1987                 goto out;
1988
1989         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1990             (vma->vm_flags & VM_NOHUGEPAGE))
1991                 goto out;
1992
1993         if (!vma->anon_vma || vma->vm_ops)
1994                 goto out;
1995         if (is_vma_temporary_stack(vma))
1996                 goto out;
1997         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1998
1999         pgd = pgd_offset(mm, address);
2000         if (!pgd_present(*pgd))
2001                 goto out;
2002
2003         pud = pud_offset(pgd, address);
2004         if (!pud_present(*pud))
2005                 goto out;
2006
2007         pmd = pmd_offset(pud, address);
2008         /* pmd can't go away or become huge under us */
2009         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2010                 goto out;
2011
2012         anon_vma_lock(vma->anon_vma);
2013
2014         pte = pte_offset_map(pmd, address);
2015         ptl = pte_lockptr(mm, pmd);
2016
2017         mmun_start = address;
2018         mmun_end   = address + HPAGE_PMD_SIZE;
2019         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2020         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2021         /*
2022          * After this gup_fast can't run anymore. This also removes
2023          * any huge TLB entry from the CPU so we won't allow
2024          * huge and small TLB entries for the same virtual address
2025          * to avoid the risk of CPU bugs in that area.
2026          */
2027         _pmd = pmdp_clear_flush(vma, address, pmd);
2028         spin_unlock(&mm->page_table_lock);
2029         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2030
2031         spin_lock(ptl);
2032         isolated = __collapse_huge_page_isolate(vma, address, pte);
2033         spin_unlock(ptl);
2034
2035         if (unlikely(!isolated)) {
2036                 pte_unmap(pte);
2037                 spin_lock(&mm->page_table_lock);
2038                 BUG_ON(!pmd_none(*pmd));
2039                 set_pmd_at(mm, address, pmd, _pmd);
2040                 spin_unlock(&mm->page_table_lock);
2041                 anon_vma_unlock(vma->anon_vma);
2042                 goto out;
2043         }
2044
2045         /*
2046          * All pages are isolated and locked so anon_vma rmap
2047          * can't run anymore.
2048          */
2049         anon_vma_unlock(vma->anon_vma);
2050
2051         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2052         pte_unmap(pte);
2053         __SetPageUptodate(new_page);
2054         pgtable = pmd_pgtable(_pmd);
2055
2056         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2057         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2058         _pmd = pmd_mkhuge(_pmd);
2059
2060         /*
2061          * spin_lock() below is not the equivalent of smp_wmb(), so
2062          * this is needed to avoid the copy_huge_page writes to become
2063          * visible after the set_pmd_at() write.
2064          */
2065         smp_wmb();
2066
2067         spin_lock(&mm->page_table_lock);
2068         BUG_ON(!pmd_none(*pmd));
2069         page_add_new_anon_rmap(new_page, vma, address);
2070         set_pmd_at(mm, address, pmd, _pmd);
2071         update_mmu_cache_pmd(vma, address, pmd);
2072         pgtable_trans_huge_deposit(mm, pgtable);
2073         spin_unlock(&mm->page_table_lock);
2074
2075         *hpage = NULL;
2076
2077         khugepaged_pages_collapsed++;
2078 out_up_write:
2079         up_write(&mm->mmap_sem);
2080         return;
2081
2082 out:
2083         mem_cgroup_uncharge_page(new_page);
2084         goto out_up_write;
2085 }
2086
2087 static int khugepaged_scan_pmd(struct mm_struct *mm,
2088                                struct vm_area_struct *vma,
2089                                unsigned long address,
2090                                struct page **hpage)
2091 {
2092         pgd_t *pgd;
2093         pud_t *pud;
2094         pmd_t *pmd;
2095         pte_t *pte, *_pte;
2096         int ret = 0, referenced = 0, none = 0;
2097         struct page *page;
2098         unsigned long _address;
2099         spinlock_t *ptl;
2100         int node = -1;
2101
2102         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2103
2104         pgd = pgd_offset(mm, address);
2105         if (!pgd_present(*pgd))
2106                 goto out;
2107
2108         pud = pud_offset(pgd, address);
2109         if (!pud_present(*pud))
2110                 goto out;
2111
2112         pmd = pmd_offset(pud, address);
2113         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2114                 goto out;
2115
2116         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2117         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2118              _pte++, _address += PAGE_SIZE) {
2119                 pte_t pteval = *_pte;
2120                 if (pte_none(pteval)) {
2121                         if (++none <= khugepaged_max_ptes_none)
2122                                 continue;
2123                         else
2124                                 goto out_unmap;
2125                 }
2126                 if (!pte_present(pteval) || !pte_write(pteval))
2127                         goto out_unmap;
2128                 page = vm_normal_page(vma, _address, pteval);
2129                 if (unlikely(!page))
2130                         goto out_unmap;
2131                 /*
2132                  * Chose the node of the first page. This could
2133                  * be more sophisticated and look at more pages,
2134                  * but isn't for now.
2135                  */
2136                 if (node == -1)
2137                         node = page_to_nid(page);
2138                 VM_BUG_ON(PageCompound(page));
2139                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2140                         goto out_unmap;
2141                 /* cannot use mapcount: can't collapse if there's a gup pin */
2142                 if (page_count(page) != 1)
2143                         goto out_unmap;
2144                 if (pte_young(pteval) || PageReferenced(page) ||
2145                     mmu_notifier_test_young(vma->vm_mm, address))
2146                         referenced = 1;
2147         }
2148         if (referenced)
2149                 ret = 1;
2150 out_unmap:
2151         pte_unmap_unlock(pte, ptl);
2152         if (ret)
2153                 /* collapse_huge_page will return with the mmap_sem released */
2154                 collapse_huge_page(mm, address, hpage, vma, node);
2155 out:
2156         return ret;
2157 }
2158
2159 static void collect_mm_slot(struct mm_slot *mm_slot)
2160 {
2161         struct mm_struct *mm = mm_slot->mm;
2162
2163         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2164
2165         if (khugepaged_test_exit(mm)) {
2166                 /* free mm_slot */
2167                 hlist_del(&mm_slot->hash);
2168                 list_del(&mm_slot->mm_node);
2169
2170                 /*
2171                  * Not strictly needed because the mm exited already.
2172                  *
2173                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2174                  */
2175
2176                 /* khugepaged_mm_lock actually not necessary for the below */
2177                 free_mm_slot(mm_slot);
2178                 mmdrop(mm);
2179         }
2180 }
2181
2182 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2183                                             struct page **hpage)
2184         __releases(&khugepaged_mm_lock)
2185         __acquires(&khugepaged_mm_lock)
2186 {
2187         struct mm_slot *mm_slot;
2188         struct mm_struct *mm;
2189         struct vm_area_struct *vma;
2190         int progress = 0;
2191
2192         VM_BUG_ON(!pages);
2193         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2194
2195         if (khugepaged_scan.mm_slot)
2196                 mm_slot = khugepaged_scan.mm_slot;
2197         else {
2198                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2199                                      struct mm_slot, mm_node);
2200                 khugepaged_scan.address = 0;
2201                 khugepaged_scan.mm_slot = mm_slot;
2202         }
2203         spin_unlock(&khugepaged_mm_lock);
2204
2205         mm = mm_slot->mm;
2206         down_read(&mm->mmap_sem);
2207         if (unlikely(khugepaged_test_exit(mm)))
2208                 vma = NULL;
2209         else
2210                 vma = find_vma(mm, khugepaged_scan.address);
2211
2212         progress++;
2213         for (; vma; vma = vma->vm_next) {
2214                 unsigned long hstart, hend;
2215
2216                 cond_resched();
2217                 if (unlikely(khugepaged_test_exit(mm))) {
2218                         progress++;
2219                         break;
2220                 }
2221
2222                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2223                      !khugepaged_always()) ||
2224                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2225                 skip:
2226                         progress++;
2227                         continue;
2228                 }
2229                 if (!vma->anon_vma || vma->vm_ops)
2230                         goto skip;
2231                 if (is_vma_temporary_stack(vma))
2232                         goto skip;
2233                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2234
2235                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2236                 hend = vma->vm_end & HPAGE_PMD_MASK;
2237                 if (hstart >= hend)
2238                         goto skip;
2239                 if (khugepaged_scan.address > hend)
2240                         goto skip;
2241                 if (khugepaged_scan.address < hstart)
2242                         khugepaged_scan.address = hstart;
2243                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2244
2245                 while (khugepaged_scan.address < hend) {
2246                         int ret;
2247                         cond_resched();
2248                         if (unlikely(khugepaged_test_exit(mm)))
2249                                 goto breakouterloop;
2250
2251                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2252                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2253                                   hend);
2254                         ret = khugepaged_scan_pmd(mm, vma,
2255                                                   khugepaged_scan.address,
2256                                                   hpage);
2257                         /* move to next address */
2258                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2259                         progress += HPAGE_PMD_NR;
2260                         if (ret)
2261                                 /* we released mmap_sem so break loop */
2262                                 goto breakouterloop_mmap_sem;
2263                         if (progress >= pages)
2264                                 goto breakouterloop;
2265                 }
2266         }
2267 breakouterloop:
2268         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2269 breakouterloop_mmap_sem:
2270
2271         spin_lock(&khugepaged_mm_lock);
2272         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2273         /*
2274          * Release the current mm_slot if this mm is about to die, or
2275          * if we scanned all vmas of this mm.
2276          */
2277         if (khugepaged_test_exit(mm) || !vma) {
2278                 /*
2279                  * Make sure that if mm_users is reaching zero while
2280                  * khugepaged runs here, khugepaged_exit will find
2281                  * mm_slot not pointing to the exiting mm.
2282                  */
2283                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2284                         khugepaged_scan.mm_slot = list_entry(
2285                                 mm_slot->mm_node.next,
2286                                 struct mm_slot, mm_node);
2287                         khugepaged_scan.address = 0;
2288                 } else {
2289                         khugepaged_scan.mm_slot = NULL;
2290                         khugepaged_full_scans++;
2291                 }
2292
2293                 collect_mm_slot(mm_slot);
2294         }
2295
2296         return progress;
2297 }
2298
2299 static int khugepaged_has_work(void)
2300 {
2301         return !list_empty(&khugepaged_scan.mm_head) &&
2302                 khugepaged_enabled();
2303 }
2304
2305 static int khugepaged_wait_event(void)
2306 {
2307         return !list_empty(&khugepaged_scan.mm_head) ||
2308                 kthread_should_stop();
2309 }
2310
2311 static void khugepaged_do_scan(void)
2312 {
2313         struct page *hpage = NULL;
2314         unsigned int progress = 0, pass_through_head = 0;
2315         unsigned int pages = khugepaged_pages_to_scan;
2316         bool wait = true;
2317
2318         barrier(); /* write khugepaged_pages_to_scan to local stack */
2319
2320         while (progress < pages) {
2321                 if (!khugepaged_prealloc_page(&hpage, &wait))
2322                         break;
2323
2324                 cond_resched();
2325
2326                 if (unlikely(kthread_should_stop() || freezing(current)))
2327                         break;
2328
2329                 spin_lock(&khugepaged_mm_lock);
2330                 if (!khugepaged_scan.mm_slot)
2331                         pass_through_head++;
2332                 if (khugepaged_has_work() &&
2333                     pass_through_head < 2)
2334                         progress += khugepaged_scan_mm_slot(pages - progress,
2335                                                             &hpage);
2336                 else
2337                         progress = pages;
2338                 spin_unlock(&khugepaged_mm_lock);
2339         }
2340
2341         if (!IS_ERR_OR_NULL(hpage))
2342                 put_page(hpage);
2343 }
2344
2345 static void khugepaged_wait_work(void)
2346 {
2347         try_to_freeze();
2348
2349         if (khugepaged_has_work()) {
2350                 if (!khugepaged_scan_sleep_millisecs)
2351                         return;
2352
2353                 wait_event_freezable_timeout(khugepaged_wait,
2354                                              kthread_should_stop(),
2355                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2356                 return;
2357         }
2358
2359         if (khugepaged_enabled())
2360                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2361 }
2362
2363 static int khugepaged(void *none)
2364 {
2365         struct mm_slot *mm_slot;
2366
2367         set_freezable();
2368         set_user_nice(current, 19);
2369
2370         while (!kthread_should_stop()) {
2371                 khugepaged_do_scan();
2372                 khugepaged_wait_work();
2373         }
2374
2375         spin_lock(&khugepaged_mm_lock);
2376         mm_slot = khugepaged_scan.mm_slot;
2377         khugepaged_scan.mm_slot = NULL;
2378         if (mm_slot)
2379                 collect_mm_slot(mm_slot);
2380         spin_unlock(&khugepaged_mm_lock);
2381         return 0;
2382 }
2383
2384 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2385 {
2386         struct page *page;
2387
2388         spin_lock(&mm->page_table_lock);
2389         if (unlikely(!pmd_trans_huge(*pmd))) {
2390                 spin_unlock(&mm->page_table_lock);
2391                 return;
2392         }
2393         page = pmd_page(*pmd);
2394         VM_BUG_ON(!page_count(page));
2395         get_page(page);
2396         spin_unlock(&mm->page_table_lock);
2397
2398         split_huge_page(page);
2399
2400         put_page(page);
2401         BUG_ON(pmd_trans_huge(*pmd));
2402 }
2403
2404 static void split_huge_page_address(struct mm_struct *mm,
2405                                     unsigned long address)
2406 {
2407         pgd_t *pgd;
2408         pud_t *pud;
2409         pmd_t *pmd;
2410
2411         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2412
2413         pgd = pgd_offset(mm, address);
2414         if (!pgd_present(*pgd))
2415                 return;
2416
2417         pud = pud_offset(pgd, address);
2418         if (!pud_present(*pud))
2419                 return;
2420
2421         pmd = pmd_offset(pud, address);
2422         if (!pmd_present(*pmd))
2423                 return;
2424         /*
2425          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2426          * materialize from under us.
2427          */
2428         split_huge_page_pmd(mm, pmd);
2429 }
2430
2431 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2432                              unsigned long start,
2433                              unsigned long end,
2434                              long adjust_next)
2435 {
2436         /*
2437          * If the new start address isn't hpage aligned and it could
2438          * previously contain an hugepage: check if we need to split
2439          * an huge pmd.
2440          */
2441         if (start & ~HPAGE_PMD_MASK &&
2442             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2443             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2444                 split_huge_page_address(vma->vm_mm, start);
2445
2446         /*
2447          * If the new end address isn't hpage aligned and it could
2448          * previously contain an hugepage: check if we need to split
2449          * an huge pmd.
2450          */
2451         if (end & ~HPAGE_PMD_MASK &&
2452             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2453             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2454                 split_huge_page_address(vma->vm_mm, end);
2455
2456         /*
2457          * If we're also updating the vma->vm_next->vm_start, if the new
2458          * vm_next->vm_start isn't page aligned and it could previously
2459          * contain an hugepage: check if we need to split an huge pmd.
2460          */
2461         if (adjust_next > 0) {
2462                 struct vm_area_struct *next = vma->vm_next;
2463                 unsigned long nstart = next->vm_start;
2464                 nstart += adjust_next << PAGE_SHIFT;
2465                 if (nstart & ~HPAGE_PMD_MASK &&
2466                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2467                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2468                         split_huge_page_address(next->vm_mm, nstart);
2469         }
2470 }