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 <linux/migrate.h>
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 /*
  27  * By default transparent hugepage support is enabled for all mappings
  28  * and khugepaged scans all mappings. Defrag is only invoked by
  29  * khugepaged hugepage allocations and by page faults inside
  30  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  31  * allocations.
  32  */
  33 unsigned long transparent_hugepage_flags __read_mostly =
  34 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  35         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  36 #endif
  37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  38         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  39 #endif
  40         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  41         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  42
  43 /* default scan 8*512 pte (or vmas) every 30 second */
  44 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  45 static unsigned int khugepaged_pages_collapsed;
  46 static unsigned int khugepaged_full_scans;
  47 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  48 /* during fragmentation poll the hugepage allocator once every minute */
  49 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  50 static struct task_struct *khugepaged_thread __read_mostly;
  51 static DEFINE_MUTEX(khugepaged_mutex);
  52 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  53 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  54 /*
  55  * default collapse hugepages if there is at least one pte mapped like
  56  * it would have happened if the vma was large enough during page
  57  * fault.
  58  */
  59 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  60
  61 static int khugepaged(void *none);
  62 static int mm_slots_hash_init(void);
  63 static int khugepaged_slab_init(void);
  64 static void khugepaged_slab_free(void);
  65
  66 #define MM_SLOTS_HASH_HEADS 1024
  67 static struct hlist_head *mm_slots_hash __read_mostly;
  68 static struct kmem_cache *mm_slot_cache __read_mostly;
  69
  70 /**
  71  * struct mm_slot - hash lookup from mm to mm_slot
  72  * @hash: hash collision list
  73  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  74  * @mm: the mm that this information is valid for
  75  */
  76 struct mm_slot {
  77         struct hlist_node hash;
  78         struct list_head mm_node;
  79         struct mm_struct *mm;
  80 };
  81
  82 /**
  83  * struct khugepaged_scan - cursor for scanning
  84  * @mm_head: the head of the mm list to scan
  85  * @mm_slot: the current mm_slot we are scanning
  86  * @address: the next address inside that to be scanned
  87  *
  88  * There is only the one khugepaged_scan instance of this cursor structure.
  89  */
  90 struct khugepaged_scan {
  91         struct list_head mm_head;
  92         struct mm_slot *mm_slot;
  93         unsigned long address;
  94 };
  95 static struct khugepaged_scan khugepaged_scan = {
  96         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  97 };
  98
  99
 100 static int set_recommended_min_free_kbytes(void)
 101 {
 102         struct zone *zone;
 103         int nr_zones = 0;
 104         unsigned long recommended_min;
 105         extern int min_free_kbytes;
 106
 107         if (!khugepaged_enabled())
 108                 return 0;
 109
 110         for_each_populated_zone(zone)
 111                 nr_zones++;
 112
 113         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 114         recommended_min = pageblock_nr_pages * nr_zones * 2;
 115
 116         /*
 117          * Make sure that on average at least two pageblocks are almost free
 118          * of another type, one for a migratetype to fall back to and a
 119          * second to avoid subsequent fallbacks of other types There are 3
 120          * MIGRATE_TYPES we care about.
 121          */
 122         recommended_min += pageblock_nr_pages * nr_zones *
 123                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 124
 125         /* don't ever allow to reserve more than 5% of the lowmem */
 126         recommended_min = min(recommended_min,
 127                               (unsigned long) nr_free_buffer_pages() / 20);
 128         recommended_min <<= (PAGE_SHIFT-10);
 129
 130         if (recommended_min > min_free_kbytes)
 131                 min_free_kbytes = recommended_min;
 132         setup_per_zone_wmarks();
 133         return 0;
 134 }
 135 late_initcall(set_recommended_min_free_kbytes);
 136
 137 static int start_khugepaged(void)
 138 {
 139         int err = 0;
 140         if (khugepaged_enabled()) {
 141                 if (!khugepaged_thread)
 142                         khugepaged_thread = kthread_run(khugepaged, NULL,
 143                                                         "khugepaged");
 144                 if (unlikely(IS_ERR(khugepaged_thread))) {
 145                         printk(KERN_ERR
 146                                "khugepaged: kthread_run(khugepaged) failed\n");
 147                         err = PTR_ERR(khugepaged_thread);
 148                         khugepaged_thread = NULL;
 149                 }
 150
 151                 if (!list_empty(&khugepaged_scan.mm_head))
 152                         wake_up_interruptible(&khugepaged_wait);
 153
 154                 set_recommended_min_free_kbytes();
 155         } else if (khugepaged_thread) {
 156                 kthread_stop(khugepaged_thread);
 157                 khugepaged_thread = NULL;
 158         }
 159
 160         return err;
 161 }
 162
 163 #ifdef CONFIG_SYSFS
 164
 165 static ssize_t double_flag_show(struct kobject *kobj,
 166                                 struct kobj_attribute *attr, char *buf,
 167                                 enum transparent_hugepage_flag enabled,
 168                                 enum transparent_hugepage_flag req_madv)
 169 {
 170         if (test_bit(enabled, &transparent_hugepage_flags)) {
 171                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 172                 return sprintf(buf, "[always] madvise never\n");
 173         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 174                 return sprintf(buf, "always [madvise] never\n");
 175         else
 176                 return sprintf(buf, "always madvise [never]\n");
 177 }
 178 static ssize_t double_flag_store(struct kobject *kobj,
 179                                  struct kobj_attribute *attr,
 180                                  const char *buf, size_t count,
 181                                  enum transparent_hugepage_flag enabled,
 182                                  enum transparent_hugepage_flag req_madv)
 183 {
 184         if (!memcmp("always", buf,
 185                     min(sizeof("always")-1, count))) {
 186                 set_bit(enabled, &transparent_hugepage_flags);
 187                 clear_bit(req_madv, &transparent_hugepage_flags);
 188         } else if (!memcmp("madvise", buf,
 189                            min(sizeof("madvise")-1, count))) {
 190                 clear_bit(enabled, &transparent_hugepage_flags);
 191                 set_bit(req_madv, &transparent_hugepage_flags);
 192         } else if (!memcmp("never", buf,
 193                            min(sizeof("never")-1, count))) {
 194                 clear_bit(enabled, &transparent_hugepage_flags);
 195                 clear_bit(req_madv, &transparent_hugepage_flags);
 196         } else
 197                 return -EINVAL;
 198
 199         return count;
 200 }
 201
 202 static ssize_t enabled_show(struct kobject *kobj,
 203                             struct kobj_attribute *attr, char *buf)
 204 {
 205         return double_flag_show(kobj, attr, buf,
 206                                 TRANSPARENT_HUGEPAGE_FLAG,
 207                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 208 }
 209 static ssize_t enabled_store(struct kobject *kobj,
 210                              struct kobj_attribute *attr,
 211                              const char *buf, size_t count)
 212 {
 213         ssize_t ret;
 214
 215         ret = double_flag_store(kobj, attr, buf, count,
 216                                 TRANSPARENT_HUGEPAGE_FLAG,
 217                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 218
 219         if (ret > 0) {
 220                 int err;
 221
 222                 mutex_lock(&khugepaged_mutex);
 223                 err = start_khugepaged();
 224                 mutex_unlock(&khugepaged_mutex);
 225
 226                 if (err)
 227                         ret = err;
 228         }
 229
 230         return ret;
 231 }
 232 static struct kobj_attribute enabled_attr =
 233         __ATTR(enabled, 0644, enabled_show, enabled_store);
 234
 235 static ssize_t single_flag_show(struct kobject *kobj,
 236                                 struct kobj_attribute *attr, char *buf,
 237                                 enum transparent_hugepage_flag flag)
 238 {
 239         return sprintf(buf, "%d\n",
 240                        !!test_bit(flag, &transparent_hugepage_flags));
 241 }
 242
 243 static ssize_t single_flag_store(struct kobject *kobj,
 244                                  struct kobj_attribute *attr,
 245                                  const char *buf, size_t count,
 246                                  enum transparent_hugepage_flag flag)
 247 {
 248         unsigned long value;
 249         int ret;
 250
 251         ret = kstrtoul(buf, 10, &value);
 252         if (ret < 0)
 253                 return ret;
 254         if (value > 1)
 255                 return -EINVAL;
 256
 257         if (value)
 258                 set_bit(flag, &transparent_hugepage_flags);
 259         else
 260                 clear_bit(flag, &transparent_hugepage_flags);
 261
 262         return count;
 263 }
 264
 265 /*
 266  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 267  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 268  * memory just to allocate one more hugepage.
 269  */
 270 static ssize_t defrag_show(struct kobject *kobj,
 271                            struct kobj_attribute *attr, char *buf)
 272 {
 273         return double_flag_show(kobj, attr, buf,
 274                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 275                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 276 }
 277 static ssize_t defrag_store(struct kobject *kobj,
 278                             struct kobj_attribute *attr,
 279                             const char *buf, size_t count)
 280 {
 281         return double_flag_store(kobj, attr, buf, count,
 282                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 283                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 284 }
 285 static struct kobj_attribute defrag_attr =
 286         __ATTR(defrag, 0644, defrag_show, defrag_store);
 287
 288 #ifdef CONFIG_DEBUG_VM
 289 static ssize_t debug_cow_show(struct kobject *kobj,
 290                                 struct kobj_attribute *attr, char *buf)
 291 {
 292         return single_flag_show(kobj, attr, buf,
 293                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 294 }
 295 static ssize_t debug_cow_store(struct kobject *kobj,
 296                                struct kobj_attribute *attr,
 297                                const char *buf, size_t count)
 298 {
 299         return single_flag_store(kobj, attr, buf, count,
 300                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 301 }
 302 static struct kobj_attribute debug_cow_attr =
 303         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 304 #endif /* CONFIG_DEBUG_VM */
 305
 306 static struct attribute *hugepage_attr[] = {
 307         &enabled_attr.attr,
 308         &defrag_attr.attr,
 309 #ifdef CONFIG_DEBUG_VM
 310         &debug_cow_attr.attr,
 311 #endif
 312         NULL,
 313 };
 314
 315 static struct attribute_group hugepage_attr_group = {
 316         .attrs = hugepage_attr,
 317 };
 318
 319 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 320                                          struct kobj_attribute *attr,
 321                                          char *buf)
 322 {
 323         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 324 }
 325
 326 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 327                                           struct kobj_attribute *attr,
 328                                           const char *buf, size_t count)
 329 {
 330         unsigned long msecs;
 331         int err;
 332
 333         err = strict_strtoul(buf, 10, &msecs);
 334         if (err || msecs > UINT_MAX)
 335                 return -EINVAL;
 336
 337         khugepaged_scan_sleep_millisecs = msecs;
 338         wake_up_interruptible(&khugepaged_wait);
 339
 340         return count;
 341 }
 342 static struct kobj_attribute scan_sleep_millisecs_attr =
 343         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 344                scan_sleep_millisecs_store);
 345
 346 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 347                                           struct kobj_attribute *attr,
 348                                           char *buf)
 349 {
 350         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 351 }
 352
 353 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 354                                            struct kobj_attribute *attr,
 355                                            const char *buf, size_t count)
 356 {
 357         unsigned long msecs;
 358         int err;
 359
 360         err = strict_strtoul(buf, 10, &msecs);
 361         if (err || msecs > UINT_MAX)
 362                 return -EINVAL;
 363
 364         khugepaged_alloc_sleep_millisecs = msecs;
 365         wake_up_interruptible(&khugepaged_wait);
 366
 367         return count;
 368 }
 369 static struct kobj_attribute alloc_sleep_millisecs_attr =
 370         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 371                alloc_sleep_millisecs_store);
 372
 373 static ssize_t pages_to_scan_show(struct kobject *kobj,
 374                                   struct kobj_attribute *attr,
 375                                   char *buf)
 376 {
 377         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 378 }
 379 static ssize_t pages_to_scan_store(struct kobject *kobj,
 380                                    struct kobj_attribute *attr,
 381                                    const char *buf, size_t count)
 382 {
 383         int err;
 384         unsigned long pages;
 385
 386         err = strict_strtoul(buf, 10, &pages);
 387         if (err || !pages || pages > UINT_MAX)
 388                 return -EINVAL;
 389
 390         khugepaged_pages_to_scan = pages;
 391
 392         return count;
 393 }
 394 static struct kobj_attribute pages_to_scan_attr =
 395         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 396                pages_to_scan_store);
 397
 398 static ssize_t pages_collapsed_show(struct kobject *kobj,
 399                                     struct kobj_attribute *attr,
 400                                     char *buf)
 401 {
 402         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 403 }
 404 static struct kobj_attribute pages_collapsed_attr =
 405         __ATTR_RO(pages_collapsed);
 406
 407 static ssize_t full_scans_show(struct kobject *kobj,
 408                                struct kobj_attribute *attr,
 409                                char *buf)
 410 {
 411         return sprintf(buf, "%u\n", khugepaged_full_scans);
 412 }
 413 static struct kobj_attribute full_scans_attr =
 414         __ATTR_RO(full_scans);
 415
 416 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 417                                       struct kobj_attribute *attr, char *buf)
 418 {
 419         return single_flag_show(kobj, attr, buf,
 420                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 421 }
 422 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 423                                        struct kobj_attribute *attr,
 424                                        const char *buf, size_t count)
 425 {
 426         return single_flag_store(kobj, attr, buf, count,
 427                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 428 }
 429 static struct kobj_attribute khugepaged_defrag_attr =
 430         __ATTR(defrag, 0644, khugepaged_defrag_show,
 431                khugepaged_defrag_store);
 432
 433 /*
 434  * max_ptes_none controls if khugepaged should collapse hugepages over
 435  * any unmapped ptes in turn potentially increasing the memory
 436  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 437  * reduce the available free memory in the system as it
 438  * runs. Increasing max_ptes_none will instead potentially reduce the
 439  * free memory in the system during the khugepaged scan.
 440  */
 441 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 442                                              struct kobj_attribute *attr,
 443                                              char *buf)
 444 {
 445         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 446 }
 447 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 448                                               struct kobj_attribute *attr,
 449                                               const char *buf, size_t count)
 450 {
 451         int err;
 452         unsigned long max_ptes_none;
 453
 454         err = strict_strtoul(buf, 10, &max_ptes_none);
 455         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 456                 return -EINVAL;
 457
 458         khugepaged_max_ptes_none = max_ptes_none;
 459
 460         return count;
 461 }
 462 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 463         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 464                khugepaged_max_ptes_none_store);
 465
 466 static struct attribute *khugepaged_attr[] = {
 467         &khugepaged_defrag_attr.attr,
 468         &khugepaged_max_ptes_none_attr.attr,
 469         &pages_to_scan_attr.attr,
 470         &pages_collapsed_attr.attr,
 471         &full_scans_attr.attr,
 472         &scan_sleep_millisecs_attr.attr,
 473         &alloc_sleep_millisecs_attr.attr,
 474         NULL,
 475 };
 476
 477 static struct attribute_group khugepaged_attr_group = {
 478         .attrs = khugepaged_attr,
 479         .name = "khugepaged",
 480 };
 481
 482 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 483 {
 484         int err;
 485
 486         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 487         if (unlikely(!*hugepage_kobj)) {
 488                 printk(KERN_ERR "hugepage: failed kobject create\n");
 489                 return -ENOMEM;
 490         }
 491
 492         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 493         if (err) {
 494                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 495                 goto delete_obj;
 496         }
 497
 498         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 499         if (err) {
 500                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 501                 goto remove_hp_group;
 502         }
 503
 504         return 0;
 505
 506 remove_hp_group:
 507         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 508 delete_obj:
 509         kobject_put(*hugepage_kobj);
 510         return err;
 511 }
 512
 513 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 514 {
 515         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 516         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 517         kobject_put(hugepage_kobj);
 518 }
 519 #else
 520 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 521 {
 522         return 0;
 523 }
 524
 525 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 526 {
 527 }
 528 #endif /* CONFIG_SYSFS */
 529
 530 static int __init hugepage_init(void)
 531 {
 532         int err;
 533         struct kobject *hugepage_kobj;
 534
 535         if (!has_transparent_hugepage()) {
 536                 transparent_hugepage_flags = 0;
 537                 return -EINVAL;
 538         }
 539
 540         err = hugepage_init_sysfs(&hugepage_kobj);
 541         if (err)
 542                 return err;
 543
 544         err = khugepaged_slab_init();
 545         if (err)
 546                 goto out;
 547
 548         err = mm_slots_hash_init();
 549         if (err) {
 550                 khugepaged_slab_free();
 551                 goto out;
 552         }
 553
 554         /*
 555          * By default disable transparent hugepages on smaller systems,
 556          * where the extra memory used could hurt more than TLB overhead
 557          * is likely to save.  The admin can still enable it through /sys.
 558          */
 559         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 560                 transparent_hugepage_flags = 0;
 561
 562         start_khugepaged();
 563
 564         return 0;
 565 out:
 566         hugepage_exit_sysfs(hugepage_kobj);
 567         return err;
 568 }
 569 module_init(hugepage_init)
 570
 571 static int __init setup_transparent_hugepage(char *str)
 572 {
 573         int ret = 0;
 574         if (!str)
 575                 goto out;
 576         if (!strcmp(str, "always")) {
 577                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 578                         &transparent_hugepage_flags);
 579                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 580                           &transparent_hugepage_flags);
 581                 ret = 1;
 582         } else if (!strcmp(str, "madvise")) {
 583                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 584                           &transparent_hugepage_flags);
 585                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 586                         &transparent_hugepage_flags);
 587                 ret = 1;
 588         } else if (!strcmp(str, "never")) {
 589                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 590                           &transparent_hugepage_flags);
 591                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 592                           &transparent_hugepage_flags);
 593                 ret = 1;
 594         }
 595 out:
 596         if (!ret)
 597                 printk(KERN_WARNING
 598                        "transparent_hugepage= cannot parse, ignored\n");
 599         return ret;
 600 }
 601 __setup("transparent_hugepage=", setup_transparent_hugepage);
 602
 603 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 604 {
 605         if (likely(vma->vm_flags & VM_WRITE))
 606                 pmd = pmd_mkwrite(pmd);
 607         return pmd;
 608 }
 609
 610 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 611                                         struct vm_area_struct *vma,
 612                                         unsigned long haddr, pmd_t *pmd,
 613                                         struct page *page)
 614 {
 615         pgtable_t pgtable;
 616
 617         VM_BUG_ON(!PageCompound(page));
 618         pgtable = pte_alloc_one(mm, haddr);
 619         if (unlikely(!pgtable))
 620                 return VM_FAULT_OOM;
 621
 622         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 623         __SetPageUptodate(page);
 624
 625         spin_lock(&mm->page_table_lock);
 626         if (unlikely(!pmd_none(*pmd))) {
 627                 spin_unlock(&mm->page_table_lock);
 628                 mem_cgroup_uncharge_page(page);
 629                 put_page(page);
 630                 pte_free(mm, pgtable);
 631         } else {
 632                 pmd_t entry;
 633                 entry = mk_pmd(page, vma->vm_page_prot);
 634                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 635                 entry = pmd_mkhuge(entry);
 636                 /*
 637                  * The spinlocking to take the lru_lock inside
 638                  * page_add_new_anon_rmap() acts as a full memory
 639                  * barrier to be sure clear_huge_page writes become
 640                  * visible after the set_pmd_at() write.
 641                  */
 642                 page_add_new_anon_rmap(page, vma, haddr);
 643                 set_pmd_at(mm, haddr, pmd, entry);
 644                 pgtable_trans_huge_deposit(mm, pgtable);
 645                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 646                 mm->nr_ptes++;
 647                 spin_unlock(&mm->page_table_lock);
 648         }
 649
 650         return 0;
 651 }
 652
 653 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 654 {
 655         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 656 }
 657
 658 static inline struct page *alloc_hugepage_vma(int defrag,
 659                                               struct vm_area_struct *vma,
 660                                               unsigned long haddr, int nd,
 661                                               gfp_t extra_gfp)
 662 {
 663         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 664                                HPAGE_PMD_ORDER, vma, haddr, nd);
 665 }
 666
 667 #ifndef CONFIG_NUMA
 668 static inline struct page *alloc_hugepage(int defrag)
 669 {
 670         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 671                            HPAGE_PMD_ORDER);
 672 }
 673 #endif
 674
 675 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 676                                unsigned long address, pmd_t *pmd,
 677                                unsigned int flags)
 678 {
 679         struct page *page;
 680         unsigned long haddr = address & HPAGE_PMD_MASK;
 681         pte_t *pte;
 682
 683         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 684                 if (unlikely(anon_vma_prepare(vma)))
 685                         return VM_FAULT_OOM;
 686                 if (unlikely(khugepaged_enter(vma)))
 687                         return VM_FAULT_OOM;
 688                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 689                                           vma, haddr, numa_node_id(), 0);
 690                 if (unlikely(!page)) {
 691                         count_vm_event(THP_FAULT_FALLBACK);
 692                         goto out;
 693                 }
 694                 count_vm_event(THP_FAULT_ALLOC);
 695                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 696                         put_page(page);
 697                         goto out;
 698                 }
 699                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 700                                                           page))) {
 701                         mem_cgroup_uncharge_page(page);
 702                         put_page(page);
 703                         goto out;
 704                 }
 705
 706                 return 0;
 707         }
 708 out:
 709         /*
 710          * Use __pte_alloc instead of pte_alloc_map, because we can't
 711          * run pte_offset_map on the pmd, if an huge pmd could
 712          * materialize from under us from a different thread.
 713          */
 714         if (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 bool pmd_prot_none(struct vm_area_struct *vma, pmd_t pmd)
 730 {
 731         /*
 732          * See pte_prot_none().
 733          */
 734         if (pmd_same(pmd, pmd_modify(pmd, vma->vm_page_prot)))
 735                 return false;
 736
 737         return pmd_same(pmd, pmd_modify(pmd, vma_prot_none(vma)));
 738 }
 739
 740 void do_huge_pmd_prot_none(struct mm_struct *mm, struct vm_area_struct *vma,
 741                            unsigned long address, pmd_t *pmd,
 742                            unsigned int flags, pmd_t entry)
 743 {
 744         unsigned long haddr = address & HPAGE_PMD_MASK;
 745         struct page *new_page = NULL;
 746         struct page *page = NULL;
 747         int node, lru;
 748
 749         spin_lock(&mm->page_table_lock);
 750         if (unlikely(!pmd_same(*pmd, entry)))
 751                 goto unlock;
 752
 753         if (unlikely(pmd_trans_splitting(entry))) {
 754                 spin_unlock(&mm->page_table_lock);
 755                 wait_split_huge_page(vma->anon_vma, pmd);
 756                 return;
 757         }
 758
 759         page = pmd_page(entry);
 760         if (page) {
 761                 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 762
 763                 get_page(page);
 764                 node = mpol_misplaced(page, vma, haddr);
 765                 if (node != -1)
 766                         goto migrate;
 767         }
 768
 769 fixup:
 770         /* change back to regular protection */
 771         entry = pmd_modify(entry, vma->vm_page_prot);
 772         set_pmd_at(mm, haddr, pmd, entry);
 773         update_mmu_cache_pmd(vma, address, entry);
 774
 775 unlock:
 776         spin_unlock(&mm->page_table_lock);
 777         if (page) {
 778                 task_numa_fault(page_to_nid(page), HPAGE_PMD_NR);
 779                 put_page(page);
 780         }
 781         return;
 782
 783 migrate:
 784         WARN_ON(!(((unsigned long)page->mapping & PAGE_MAPPING_ANON)));
 785         WARN_ON((((unsigned long)page->mapping & PAGE_MAPPING_KSM)));
 786         BUG_ON(PageSwapCache(page));
 787
 788         spin_unlock(&mm->page_table_lock);
 789
 790         lock_page(page);
 791         spin_lock(&mm->page_table_lock);
 792         if (unlikely(!pmd_same(*pmd, entry))) {
 793                 spin_unlock(&mm->page_table_lock);
 794                 unlock_page(page);
 795                 put_page(page);
 796                 return;
 797         }
 798         spin_unlock(&mm->page_table_lock);
 799
 800         new_page = alloc_pages_node(node,
 801             (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT,
 802             HPAGE_PMD_ORDER);
 803
 804         WARN_ON(PageLRU(new_page));
 805
 806         if (!new_page)
 807                 goto alloc_fail;
 808
 809         lru = PageLRU(page);
 810
 811         if (lru && isolate_lru_page(page)) /* does an implicit get_page() */
 812                 goto alloc_fail;
 813
 814         if (!trylock_page(new_page))
 815                 BUG();
 816
 817         /* anon mapping, we can simply copy page->mapping to the new page: */
 818         new_page->mapping = page->mapping;
 819         new_page->index = page->index;
 820
 821         migrate_page_copy(new_page, page);
 822
 823         WARN_ON(PageLRU(new_page));
 824
 825         spin_lock(&mm->page_table_lock);
 826         if (unlikely(!pmd_same(*pmd, entry))) {
 827                 spin_unlock(&mm->page_table_lock);
 828                 if (lru)
 829                         putback_lru_page(page);
 830
 831                 unlock_page(new_page);
 832                 ClearPageActive(new_page);      /* Set by migrate_page_copy() */
 833                 new_page->mapping = NULL;
 834                 put_page(new_page);             /* Free it */
 835
 836                 unlock_page(page);
 837                 put_page(page);                 /* Drop the local reference */
 838
 839                 return;
 840         }
 841
 842         entry = mk_pmd(new_page, vma->vm_page_prot);
 843         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 844         entry = pmd_mkhuge(entry);
 845
 846         page_add_new_anon_rmap(new_page, vma, haddr);
 847
 848         set_pmd_at(mm, haddr, pmd, entry);
 849         update_mmu_cache_pmd(vma, address, entry);
 850         page_remove_rmap(page);
 851         spin_unlock(&mm->page_table_lock);
 852
 853         put_page(page);                 /* Drop the rmap reference */
 854
 855         task_numa_fault(node, HPAGE_PMD_NR);
 856
 857         if (lru)
 858                 put_page(page);         /* drop the LRU isolation reference */
 859
 860         unlock_page(new_page);
 861         unlock_page(page);
 862         put_page(page);                 /* Drop the local reference */
 863
 864         return;
 865
 866 alloc_fail:
 867         if (new_page)
 868                 put_page(new_page);
 869
 870         unlock_page(page);
 871
 872         spin_lock(&mm->page_table_lock);
 873         if (unlikely(!pmd_same(*pmd, entry))) {
 874                 put_page(page);
 875                 page = NULL;
 876                 goto unlock;
 877         }
 878         goto fixup;
 879 }
 880
 881 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 882                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 883                   struct vm_area_struct *vma)
 884 {
 885         struct page *src_page;
 886         pmd_t pmd;
 887         pgtable_t pgtable;
 888         int ret;
 889
 890         ret = -ENOMEM;
 891         pgtable = pte_alloc_one(dst_mm, addr);
 892         if (unlikely(!pgtable))
 893                 goto out;
 894
 895         spin_lock(&dst_mm->page_table_lock);
 896         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 897
 898         ret = -EAGAIN;
 899         pmd = *src_pmd;
 900         if (unlikely(!pmd_trans_huge(pmd))) {
 901                 pte_free(dst_mm, pgtable);
 902                 goto out_unlock;
 903         }
 904         if (unlikely(pmd_trans_splitting(pmd))) {
 905                 /* split huge page running from under us */
 906                 spin_unlock(&src_mm->page_table_lock);
 907                 spin_unlock(&dst_mm->page_table_lock);
 908                 pte_free(dst_mm, pgtable);
 909
 910                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 911                 goto out;
 912         }
 913         src_page = pmd_page(pmd);
 914         VM_BUG_ON(!PageHead(src_page));
 915         get_page(src_page);
 916         page_dup_rmap(src_page);
 917         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 918
 919         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 920         pmd = pmd_mkold(pmd_wrprotect(pmd));
 921         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 922         pgtable_trans_huge_deposit(dst_mm, pgtable);
 923         dst_mm->nr_ptes++;
 924
 925         ret = 0;
 926 out_unlock:
 927         spin_unlock(&src_mm->page_table_lock);
 928         spin_unlock(&dst_mm->page_table_lock);
 929 out:
 930         return ret;
 931 }
 932
 933 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 934                                         struct vm_area_struct *vma,
 935                                         unsigned long address,
 936                                         pmd_t *pmd, pmd_t orig_pmd,
 937                                         struct page *page,
 938                                         unsigned long haddr)
 939 {
 940         pgtable_t pgtable;
 941         pmd_t _pmd;
 942         int ret = 0, i;
 943         struct page **pages;
 944         unsigned long mmun_start;       /* For mmu_notifiers */
 945         unsigned long mmun_end;         /* For mmu_notifiers */
 946
 947         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 948                         GFP_KERNEL);
 949         if (unlikely(!pages)) {
 950                 ret |= VM_FAULT_OOM;
 951                 goto out;
 952         }
 953
 954         for (i = 0; i < HPAGE_PMD_NR; i++) {
 955                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 956                                                __GFP_OTHER_NODE,
 957                                                vma, address, page_to_nid(page));
 958                 if (unlikely(!pages[i] ||
 959                              mem_cgroup_newpage_charge(pages[i], mm,
 960                                                        GFP_KERNEL))) {
 961                         if (pages[i])
 962                                 put_page(pages[i]);
 963                         mem_cgroup_uncharge_start();
 964                         while (--i >= 0) {
 965                                 mem_cgroup_uncharge_page(pages[i]);
 966                                 put_page(pages[i]);
 967                         }
 968                         mem_cgroup_uncharge_end();
 969                         kfree(pages);
 970                         ret |= VM_FAULT_OOM;
 971                         goto out;
 972                 }
 973         }
 974
 975         for (i = 0; i < HPAGE_PMD_NR; i++) {
 976                 copy_user_highpage(pages[i], page + i,
 977                                    haddr + PAGE_SIZE * i, vma);
 978                 __SetPageUptodate(pages[i]);
 979                 cond_resched();
 980         }
 981
 982         mmun_start = haddr;
 983         mmun_end   = haddr + HPAGE_PMD_SIZE;
 984         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 985
 986         spin_lock(&mm->page_table_lock);
 987         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 988                 goto out_free_pages;
 989         VM_BUG_ON(!PageHead(page));
 990
 991         pmdp_clear_flush(vma, haddr, pmd);
 992         /* leave pmd empty until pte is filled */
 993
 994         pgtable = pgtable_trans_huge_withdraw(mm);
 995         pmd_populate(mm, &_pmd, pgtable);
 996
 997         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 998                 pte_t *pte, entry;
 999                 entry = mk_pte(pages[i], vma->vm_page_prot);
1000                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1001                 page_add_new_anon_rmap(pages[i], vma, haddr);
1002                 pte = pte_offset_map(&_pmd, haddr);
1003                 VM_BUG_ON(!pte_none(*pte));
1004                 set_pte_at(mm, haddr, pte, entry);
1005                 pte_unmap(pte);
1006         }
1007         kfree(pages);
1008
1009         smp_wmb(); /* make pte visible before pmd */
1010         pmd_populate(mm, pmd, pgtable);
1011         page_remove_rmap(page);
1012         spin_unlock(&mm->page_table_lock);
1013
1014         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1015
1016         ret |= VM_FAULT_WRITE;
1017         put_page(page);
1018
1019 out:
1020         return ret;
1021
1022 out_free_pages:
1023         spin_unlock(&mm->page_table_lock);
1024         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1025         mem_cgroup_uncharge_start();
1026         for (i = 0; i < HPAGE_PMD_NR; i++) {
1027                 mem_cgroup_uncharge_page(pages[i]);
1028                 put_page(pages[i]);
1029         }
1030         mem_cgroup_uncharge_end();
1031         kfree(pages);
1032         goto out;
1033 }
1034
1035 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1036                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1037 {
1038         int ret = 0;
1039         struct page *page, *new_page;
1040         unsigned long haddr;
1041         unsigned long mmun_start;       /* For mmu_notifiers */
1042         unsigned long mmun_end;         /* For mmu_notifiers */
1043
1044         VM_BUG_ON(!vma->anon_vma);
1045         spin_lock(&mm->page_table_lock);
1046         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1047                 goto out_unlock;
1048
1049         page = pmd_page(orig_pmd);
1050         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1051         haddr = address & HPAGE_PMD_MASK;
1052         if (page_mapcount(page) == 1) {
1053                 pmd_t entry;
1054                 entry = pmd_mkyoung(orig_pmd);
1055                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1056                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1057                         update_mmu_cache_pmd(vma, address, pmd);
1058                 ret |= VM_FAULT_WRITE;
1059                 goto out_unlock;
1060         }
1061         get_page(page);
1062         spin_unlock(&mm->page_table_lock);
1063
1064         if (transparent_hugepage_enabled(vma) &&
1065             !transparent_hugepage_debug_cow())
1066                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1067                                               vma, haddr, numa_node_id(), 0);
1068         else
1069                 new_page = NULL;
1070
1071         if (unlikely(!new_page)) {
1072                 count_vm_event(THP_FAULT_FALLBACK);
1073                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1074                                                    pmd, orig_pmd, page, haddr);
1075                 if (ret & VM_FAULT_OOM)
1076                         split_huge_page(page);
1077                 put_page(page);
1078                 goto out;
1079         }
1080         count_vm_event(THP_FAULT_ALLOC);
1081
1082         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1083                 put_page(new_page);
1084                 split_huge_page(page);
1085                 put_page(page);
1086                 ret |= VM_FAULT_OOM;
1087                 goto out;
1088         }
1089
1090         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1091         __SetPageUptodate(new_page);
1092
1093         mmun_start = haddr;
1094         mmun_end   = haddr + HPAGE_PMD_SIZE;
1095         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1096
1097         spin_lock(&mm->page_table_lock);
1098         put_page(page);
1099         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1100                 spin_unlock(&mm->page_table_lock);
1101                 mem_cgroup_uncharge_page(new_page);
1102                 put_page(new_page);
1103                 goto out_mn;
1104         } else {
1105                 pmd_t entry;
1106                 VM_BUG_ON(!PageHead(page));
1107                 entry = mk_pmd(new_page, vma->vm_page_prot);
1108                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1109                 entry = pmd_mkhuge(entry);
1110                 pmdp_clear_flush(vma, haddr, pmd);
1111                 page_add_new_anon_rmap(new_page, vma, haddr);
1112                 set_pmd_at(mm, haddr, pmd, entry);
1113                 update_mmu_cache_pmd(vma, address, pmd);
1114                 page_remove_rmap(page);
1115                 put_page(page);
1116                 ret |= VM_FAULT_WRITE;
1117         }
1118         spin_unlock(&mm->page_table_lock);
1119 out_mn:
1120         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1121 out:
1122         return ret;
1123 out_unlock:
1124         spin_unlock(&mm->page_table_lock);
1125         return ret;
1126 }
1127
1128 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1129                                    unsigned long addr,
1130                                    pmd_t *pmd,
1131                                    unsigned int flags)
1132 {
1133         struct mm_struct *mm = vma->vm_mm;
1134         struct page *page = NULL;
1135
1136         assert_spin_locked(&mm->page_table_lock);
1137
1138         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1139                 goto out;
1140
1141         page = pmd_page(*pmd);
1142         VM_BUG_ON(!PageHead(page));
1143         if (flags & FOLL_TOUCH) {
1144                 pmd_t _pmd;
1145                 /*
1146                  * We should set the dirty bit only for FOLL_WRITE but
1147                  * for now the dirty bit in the pmd is meaningless.
1148                  * And if the dirty bit will become meaningful and
1149                  * we'll only set it with FOLL_WRITE, an atomic
1150                  * set_bit will be required on the pmd to set the
1151                  * young bit, instead of the current set_pmd_at.
1152                  */
1153                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1154                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1155         }
1156         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1157                 if (page->mapping && trylock_page(page)) {
1158                         lru_add_drain();
1159                         if (page->mapping)
1160                                 mlock_vma_page(page);
1161                         unlock_page(page);
1162                 }
1163         }
1164         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1165         VM_BUG_ON(!PageCompound(page));
1166         if (flags & FOLL_GET)
1167                 get_page_foll(page);
1168
1169 out:
1170         return page;
1171 }
1172
1173 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1174                  pmd_t *pmd, unsigned long addr)
1175 {
1176         int ret = 0;
1177
1178         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1179                 struct page *page;
1180                 pgtable_t pgtable;
1181                 pmd_t orig_pmd;
1182                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1183                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1184                 page = pmd_page(orig_pmd);
1185                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1186                 page_remove_rmap(page);
1187                 VM_BUG_ON(page_mapcount(page) < 0);
1188                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1189                 VM_BUG_ON(!PageHead(page));
1190                 tlb->mm->nr_ptes--;
1191                 spin_unlock(&tlb->mm->page_table_lock);
1192                 tlb_remove_page(tlb, page);
1193                 pte_free(tlb->mm, pgtable);
1194                 ret = 1;
1195         }
1196         return ret;
1197 }
1198
1199 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1200                 unsigned long addr, unsigned long end,
1201                 unsigned char *vec)
1202 {
1203         int ret = 0;
1204
1205         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1206                 /*
1207                  * All logical pages in the range are present
1208                  * if backed by a huge page.
1209                  */
1210                 spin_unlock(&vma->vm_mm->page_table_lock);
1211                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1212                 ret = 1;
1213         }
1214
1215         return ret;
1216 }
1217
1218 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1219                   unsigned long old_addr,
1220                   unsigned long new_addr, unsigned long old_end,
1221                   pmd_t *old_pmd, pmd_t *new_pmd)
1222 {
1223         int ret = 0;
1224         pmd_t pmd;
1225
1226         struct mm_struct *mm = vma->vm_mm;
1227
1228         if ((old_addr & ~HPAGE_PMD_MASK) ||
1229             (new_addr & ~HPAGE_PMD_MASK) ||
1230             old_end - old_addr < HPAGE_PMD_SIZE ||
1231             (new_vma->vm_flags & VM_NOHUGEPAGE))
1232                 goto out;
1233
1234         /*
1235          * The destination pmd shouldn't be established, free_pgtables()
1236          * should have release it.
1237          */
1238         if (WARN_ON(!pmd_none(*new_pmd))) {
1239                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1240                 goto out;
1241         }
1242
1243         ret = __pmd_trans_huge_lock(old_pmd, vma);
1244         if (ret == 1) {
1245                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1246                 VM_BUG_ON(!pmd_none(*new_pmd));
1247                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1248                 spin_unlock(&mm->page_table_lock);
1249         }
1250 out:
1251         return ret;
1252 }
1253
1254 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1255                 unsigned long addr, pgprot_t newprot)
1256 {
1257         struct mm_struct *mm = vma->vm_mm;
1258         int ret = 0;
1259
1260         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1261                 pmd_t entry;
1262                 entry = pmdp_get_and_clear(mm, addr, pmd);
1263                 entry = pmd_modify(entry, newprot);
1264                 set_pmd_at(mm, addr, pmd, entry);
1265                 spin_unlock(&vma->vm_mm->page_table_lock);
1266                 ret = 1;
1267         }
1268
1269         return ret;
1270 }
1271
1272 /*
1273  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1274  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1275  *
1276  * Note that if it returns 1, this routine returns without unlocking page
1277  * table locks. So callers must unlock them.
1278  */
1279 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1280 {
1281         spin_lock(&vma->vm_mm->page_table_lock);
1282         if (likely(pmd_trans_huge(*pmd))) {
1283                 if (unlikely(pmd_trans_splitting(*pmd))) {
1284                         spin_unlock(&vma->vm_mm->page_table_lock);
1285                         wait_split_huge_page(vma->anon_vma, pmd);
1286                         return -1;
1287                 } else {
1288                         /* Thp mapped by 'pmd' is stable, so we can
1289                          * handle it as it is. */
1290                         return 1;
1291                 }
1292         }
1293         spin_unlock(&vma->vm_mm->page_table_lock);
1294         return 0;
1295 }
1296
1297 pmd_t *page_check_address_pmd(struct page *page,
1298                               struct mm_struct *mm,
1299                               unsigned long address,
1300                               enum page_check_address_pmd_flag flag)
1301 {
1302         pgd_t *pgd;
1303         pud_t *pud;
1304         pmd_t *pmd, *ret = NULL;
1305
1306         if (address & ~HPAGE_PMD_MASK)
1307                 goto out;
1308
1309         pgd = pgd_offset(mm, address);
1310         if (!pgd_present(*pgd))
1311                 goto out;
1312
1313         pud = pud_offset(pgd, address);
1314         if (!pud_present(*pud))
1315                 goto out;
1316
1317         pmd = pmd_offset(pud, address);
1318         if (pmd_none(*pmd))
1319                 goto out;
1320         if (pmd_page(*pmd) != page)
1321                 goto out;
1322         /*
1323          * split_vma() may create temporary aliased mappings. There is
1324          * no risk as long as all huge pmd are found and have their
1325          * splitting bit set before __split_huge_page_refcount
1326          * runs. Finding the same huge pmd more than once during the
1327          * same rmap walk is not a problem.
1328          */
1329         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1330             pmd_trans_splitting(*pmd))
1331                 goto out;
1332         if (pmd_trans_huge(*pmd)) {
1333                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1334                           !pmd_trans_splitting(*pmd));
1335                 ret = pmd;
1336         }
1337 out:
1338         return ret;
1339 }
1340
1341 static int __split_huge_page_splitting(struct page *page,
1342                                        struct vm_area_struct *vma,
1343                                        unsigned long address)
1344 {
1345         struct mm_struct *mm = vma->vm_mm;
1346         pmd_t *pmd;
1347         int ret = 0;
1348         /* For mmu_notifiers */
1349         const unsigned long mmun_start = address;
1350         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1351
1352         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1353         spin_lock(&mm->page_table_lock);
1354         pmd = page_check_address_pmd(page, mm, address,
1355                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1356         if (pmd) {
1357                 /*
1358                  * We can't temporarily set the pmd to null in order
1359                  * to split it, the pmd must remain marked huge at all
1360                  * times or the VM won't take the pmd_trans_huge paths
1361                  * and it won't wait on the anon_vma->root->mutex to
1362                  * serialize against split_huge_page*.
1363                  */
1364                 pmdp_splitting_flush(vma, address, pmd);
1365                 ret = 1;
1366         }
1367         spin_unlock(&mm->page_table_lock);
1368         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1369
1370         return ret;
1371 }
1372
1373 static void __split_huge_page_refcount(struct page *page)
1374 {
1375         int i;
1376         struct zone *zone = page_zone(page);
1377         struct lruvec *lruvec;
1378         int tail_count = 0;
1379
1380         /* prevent PageLRU to go away from under us, and freeze lru stats */
1381         spin_lock_irq(&zone->lru_lock);
1382         lruvec = mem_cgroup_page_lruvec(page, zone);
1383
1384         compound_lock(page);
1385         /* complete memcg works before add pages to LRU */
1386         mem_cgroup_split_huge_fixup(page);
1387
1388         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1389                 struct page *page_tail = page + i;
1390
1391                 /* tail_page->_mapcount cannot change */
1392                 BUG_ON(page_mapcount(page_tail) < 0);
1393                 tail_count += page_mapcount(page_tail);
1394                 /* check for overflow */
1395                 BUG_ON(tail_count < 0);
1396                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1397                 /*
1398                  * tail_page->_count is zero and not changing from
1399                  * under us. But get_page_unless_zero() may be running
1400                  * from under us on the tail_page. If we used
1401                  * atomic_set() below instead of atomic_add(), we
1402                  * would then run atomic_set() concurrently with
1403                  * get_page_unless_zero(), and atomic_set() is
1404                  * implemented in C not using locked ops. spin_unlock
1405                  * on x86 sometime uses locked ops because of PPro
1406                  * errata 66, 92, so unless somebody can guarantee
1407                  * atomic_set() here would be safe on all archs (and
1408                  * not only on x86), it's safer to use atomic_add().
1409                  */
1410                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1411                            &page_tail->_count);
1412
1413                 /* after clearing PageTail the gup refcount can be released */
1414                 smp_mb();
1415
1416                 /*
1417                  * retain hwpoison flag of the poisoned tail page:
1418                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1419                  *   by the memory-failure.
1420                  */
1421                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1422                 page_tail->flags |= (page->flags &
1423                                      ((1L << PG_referenced) |
1424                                       (1L << PG_swapbacked) |
1425                                       (1L << PG_mlocked) |
1426                                       (1L << PG_uptodate)));
1427                 page_tail->flags |= (1L << PG_dirty);
1428
1429                 /* clear PageTail before overwriting first_page */
1430                 smp_wmb();
1431
1432                 /*
1433                  * __split_huge_page_splitting() already set the
1434                  * splitting bit in all pmd that could map this
1435                  * hugepage, that will ensure no CPU can alter the
1436                  * mapcount on the head page. The mapcount is only
1437                  * accounted in the head page and it has to be
1438                  * transferred to all tail pages in the below code. So
1439                  * for this code to be safe, the split the mapcount
1440                  * can't change. But that doesn't mean userland can't
1441                  * keep changing and reading the page contents while
1442                  * we transfer the mapcount, so the pmd splitting
1443                  * status is achieved setting a reserved bit in the
1444                  * pmd, not by clearing the present bit.
1445                 */
1446                 page_tail->_mapcount = page->_mapcount;
1447
1448                 BUG_ON(page_tail->mapping);
1449                 page_tail->mapping = page->mapping;
1450
1451                 page_tail->index = page->index + i;
1452                 page_xchg_last_nid(page, page_last_nid(page_tail));
1453
1454                 BUG_ON(!PageAnon(page_tail));
1455                 BUG_ON(!PageUptodate(page_tail));
1456                 BUG_ON(!PageDirty(page_tail));
1457                 BUG_ON(!PageSwapBacked(page_tail));
1458
1459                 lru_add_page_tail(page, page_tail, lruvec);
1460         }
1461         atomic_sub(tail_count, &page->_count);
1462         BUG_ON(atomic_read(&page->_count) <= 0);
1463
1464         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1465         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1466
1467         ClearPageCompound(page);
1468         compound_unlock(page);
1469         spin_unlock_irq(&zone->lru_lock);
1470
1471         for (i = 1; i < HPAGE_PMD_NR; i++) {
1472                 struct page *page_tail = page + i;
1473                 BUG_ON(page_count(page_tail) <= 0);
1474                 /*
1475                  * Tail pages may be freed if there wasn't any mapping
1476                  * like if add_to_swap() is running on a lru page that
1477                  * had its mapping zapped. And freeing these pages
1478                  * requires taking the lru_lock so we do the put_page
1479                  * of the tail pages after the split is complete.
1480                  */
1481                 put_page(page_tail);
1482         }
1483
1484         /*
1485          * Only the head page (now become a regular page) is required
1486          * to be pinned by the caller.
1487          */
1488         BUG_ON(page_count(page) <= 0);
1489 }
1490
1491 static int __split_huge_page_map(struct page *page,
1492                                  struct vm_area_struct *vma,
1493                                  unsigned long address)
1494 {
1495         struct mm_struct *mm = vma->vm_mm;
1496         pmd_t *pmd, _pmd;
1497         int ret = 0, i;
1498         pgtable_t pgtable;
1499         unsigned long haddr;
1500         pgprot_t prot;
1501
1502         spin_lock(&mm->page_table_lock);
1503         pmd = page_check_address_pmd(page, mm, address,
1504                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1505         if (!pmd)
1506                 goto unlock;
1507
1508         prot = pmd_pgprot(*pmd);
1509         pgtable = pgtable_trans_huge_withdraw(mm);
1510         pmd_populate(mm, &_pmd, pgtable);
1511
1512         for (i = 0, haddr = address; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1513                 pte_t *pte, entry;
1514
1515                 BUG_ON(PageCompound(page+i));
1516                 entry = mk_pte(page + i, prot);
1517                 entry = pte_mkdirty(entry);
1518                 if (!pmd_young(*pmd))
1519                         entry = pte_mkold(entry);
1520                 pte = pte_offset_map(&_pmd, haddr);
1521                 BUG_ON(!pte_none(*pte));
1522                 set_pte_at(mm, haddr, pte, entry);
1523                 pte_unmap(pte);
1524         }
1525
1526         smp_wmb(); /* make ptes visible before pmd, see __pte_alloc */
1527         /*
1528          * Up to this point the pmd is present and huge.
1529          *
1530          * If we overwrite the pmd with the not-huge version, we could trigger
1531          * a small page size TLB miss on the small sized TLB while the hugepage
1532          * TLB entry is still established in the huge TLB.
1533          *
1534          * Some CPUs don't like that. See
1535          * http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 383
1536          * on page 93.
1537          *
1538          * Thus it is generally safer to never allow small and huge TLB entries
1539          * for overlapping virtual addresses to be loaded. So we first mark the
1540          * current pmd not present, then we flush the TLB and finally we write
1541          * the non-huge version of the pmd entry with pmd_populate.
1542          *
1543          * The above needs to be done under the ptl because pmd_trans_huge and
1544          * pmd_trans_splitting must remain set on the pmd until the split is
1545          * complete. The ptl also protects against concurrent faults due to
1546          * making the pmd not-present.
1547          */
1548         set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1549         flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1550         pmd_populate(mm, pmd, pgtable);
1551         ret = 1;
1552
1553 unlock:
1554         spin_unlock(&mm->page_table_lock);
1555
1556         return ret;
1557 }
1558
1559 /* must be called with anon_vma->root->mutex hold */
1560 static void __split_huge_page(struct page *page,
1561                               struct anon_vma *anon_vma)
1562 {
1563         int mapcount, mapcount2;
1564         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1565         struct anon_vma_chain *avc;
1566
1567         BUG_ON(!PageHead(page));
1568         BUG_ON(PageTail(page));
1569
1570         mapcount = 0;
1571         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1572                 struct vm_area_struct *vma = avc->vma;
1573                 unsigned long addr = vma_address(page, vma);
1574                 BUG_ON(is_vma_temporary_stack(vma));
1575                 mapcount += __split_huge_page_splitting(page, vma, addr);
1576         }
1577         /*
1578          * It is critical that new vmas are added to the tail of the
1579          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1580          * and establishes a child pmd before
1581          * __split_huge_page_splitting() freezes the parent pmd (so if
1582          * we fail to prevent copy_huge_pmd() from running until the
1583          * whole __split_huge_page() is complete), we will still see
1584          * the newly established pmd of the child later during the
1585          * walk, to be able to set it as pmd_trans_splitting too.
1586          */
1587         if (mapcount != page_mapcount(page))
1588                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1589                        mapcount, page_mapcount(page));
1590         BUG_ON(mapcount != page_mapcount(page));
1591
1592         __split_huge_page_refcount(page);
1593
1594         mapcount2 = 0;
1595         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1596                 struct vm_area_struct *vma = avc->vma;
1597                 unsigned long addr = vma_address(page, vma);
1598                 BUG_ON(is_vma_temporary_stack(vma));
1599                 mapcount2 += __split_huge_page_map(page, vma, addr);
1600         }
1601         if (mapcount != mapcount2)
1602                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1603                        mapcount, mapcount2, page_mapcount(page));
1604         BUG_ON(mapcount != mapcount2);
1605 }
1606
1607 int split_huge_page(struct page *page)
1608 {
1609         struct anon_vma *anon_vma;
1610         int ret = 1;
1611
1612         BUG_ON(!PageAnon(page));
1613         anon_vma = page_lock_anon_vma(page);
1614         if (!anon_vma)
1615                 goto out;
1616         ret = 0;
1617         if (!PageCompound(page))
1618                 goto out_unlock;
1619
1620         BUG_ON(!PageSwapBacked(page));
1621         __split_huge_page(page, anon_vma);
1622         count_vm_event(THP_SPLIT);
1623
1624         BUG_ON(PageCompound(page));
1625 out_unlock:
1626         page_unlock_anon_vma(anon_vma);
1627 out:
1628         return ret;
1629 }
1630
1631 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1632
1633 int hugepage_madvise(struct vm_area_struct *vma,
1634                      unsigned long *vm_flags, int advice)
1635 {
1636         struct mm_struct *mm = vma->vm_mm;
1637
1638         switch (advice) {
1639         case MADV_HUGEPAGE:
1640                 /*
1641                  * Be somewhat over-protective like KSM for now!
1642                  */
1643                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1644                         return -EINVAL;
1645                 if (mm->def_flags & VM_NOHUGEPAGE)
1646                         return -EINVAL;
1647                 *vm_flags &= ~VM_NOHUGEPAGE;
1648                 *vm_flags |= VM_HUGEPAGE;
1649                 /*
1650                  * If the vma become good for khugepaged to scan,
1651                  * register it here without waiting a page fault that
1652                  * may not happen any time soon.
1653                  */
1654                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1655                         return -ENOMEM;
1656                 break;
1657         case MADV_NOHUGEPAGE:
1658                 /*
1659                  * Be somewhat over-protective like KSM for now!
1660                  */
1661                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1662                         return -EINVAL;
1663                 *vm_flags &= ~VM_HUGEPAGE;
1664                 *vm_flags |= VM_NOHUGEPAGE;
1665                 /*
1666                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1667                  * this vma even if we leave the mm registered in khugepaged if
1668                  * it got registered before VM_NOHUGEPAGE was set.
1669                  */
1670                 break;
1671         }
1672
1673         return 0;
1674 }
1675
1676 static int __init khugepaged_slab_init(void)
1677 {
1678         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1679                                           sizeof(struct mm_slot),
1680                                           __alignof__(struct mm_slot), 0, NULL);
1681         if (!mm_slot_cache)
1682                 return -ENOMEM;
1683
1684         return 0;
1685 }
1686
1687 static void __init khugepaged_slab_free(void)
1688 {
1689         kmem_cache_destroy(mm_slot_cache);
1690         mm_slot_cache = NULL;
1691 }
1692
1693 static inline struct mm_slot *alloc_mm_slot(void)
1694 {
1695         if (!mm_slot_cache)     /* initialization failed */
1696                 return NULL;
1697         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1698 }
1699
1700 static inline void free_mm_slot(struct mm_slot *mm_slot)
1701 {
1702         kmem_cache_free(mm_slot_cache, mm_slot);
1703 }
1704
1705 static int __init mm_slots_hash_init(void)
1706 {
1707         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1708                                 GFP_KERNEL);
1709         if (!mm_slots_hash)
1710                 return -ENOMEM;
1711         return 0;
1712 }
1713
1714 #if 0
1715 static void __init mm_slots_hash_free(void)
1716 {
1717         kfree(mm_slots_hash);
1718         mm_slots_hash = NULL;
1719 }
1720 #endif
1721
1722 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1723 {
1724         struct mm_slot *mm_slot;
1725         struct hlist_head *bucket;
1726         struct hlist_node *node;
1727
1728         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1729                                 % MM_SLOTS_HASH_HEADS];
1730         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1731                 if (mm == mm_slot->mm)
1732                         return mm_slot;
1733         }
1734         return NULL;
1735 }
1736
1737 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1738                                     struct mm_slot *mm_slot)
1739 {
1740         struct hlist_head *bucket;
1741
1742         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1743                                 % MM_SLOTS_HASH_HEADS];
1744         mm_slot->mm = mm;
1745         hlist_add_head(&mm_slot->hash, bucket);
1746 }
1747
1748 static inline int khugepaged_test_exit(struct mm_struct *mm)
1749 {
1750         return atomic_read(&mm->mm_users) == 0;
1751 }
1752
1753 int __khugepaged_enter(struct mm_struct *mm)
1754 {
1755         struct mm_slot *mm_slot;
1756         int wakeup;
1757
1758         mm_slot = alloc_mm_slot();
1759         if (!mm_slot)
1760                 return -ENOMEM;
1761
1762         /* __khugepaged_exit() must not run from under us */
1763         VM_BUG_ON(khugepaged_test_exit(mm));
1764         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1765                 free_mm_slot(mm_slot);
1766                 return 0;
1767         }
1768
1769         spin_lock(&khugepaged_mm_lock);
1770         insert_to_mm_slots_hash(mm, mm_slot);
1771         /*
1772          * Insert just behind the scanning cursor, to let the area settle
1773          * down a little.
1774          */
1775         wakeup = list_empty(&khugepaged_scan.mm_head);
1776         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1777         spin_unlock(&khugepaged_mm_lock);
1778
1779         atomic_inc(&mm->mm_count);
1780         if (wakeup)
1781                 wake_up_interruptible(&khugepaged_wait);
1782
1783         return 0;
1784 }
1785
1786 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1787 {
1788         unsigned long hstart, hend;
1789         if (!vma->anon_vma)
1790                 /*
1791                  * Not yet faulted in so we will register later in the
1792                  * page fault if needed.
1793                  */
1794                 return 0;
1795         if (vma->vm_ops)
1796                 /* khugepaged not yet working on file or special mappings */
1797                 return 0;
1798         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1799         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1800         hend = vma->vm_end & HPAGE_PMD_MASK;
1801         if (hstart < hend)
1802                 return khugepaged_enter(vma);
1803         return 0;
1804 }
1805
1806 void __khugepaged_exit(struct mm_struct *mm)
1807 {
1808         struct mm_slot *mm_slot;
1809         int free = 0;
1810
1811         spin_lock(&khugepaged_mm_lock);
1812         mm_slot = get_mm_slot(mm);
1813         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1814                 hlist_del(&mm_slot->hash);
1815                 list_del(&mm_slot->mm_node);
1816                 free = 1;
1817         }
1818         spin_unlock(&khugepaged_mm_lock);
1819
1820         if (free) {
1821                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1822                 free_mm_slot(mm_slot);
1823                 mmdrop(mm);
1824         } else if (mm_slot) {
1825                 /*
1826                  * This is required to serialize against
1827                  * khugepaged_test_exit() (which is guaranteed to run
1828                  * under mmap sem read mode). Stop here (after we
1829                  * return all pagetables will be destroyed) until
1830                  * khugepaged has finished working on the pagetables
1831                  * under the mmap_sem.
1832                  */
1833                 down_write(&mm->mmap_sem);
1834                 up_write(&mm->mmap_sem);
1835         }
1836 }
1837
1838 static void release_pte_page(struct page *page)
1839 {
1840         /* 0 stands for page_is_file_cache(page) == false */
1841         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1842         unlock_page(page);
1843         putback_lru_page(page);
1844 }
1845
1846 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1847 {
1848         while (--_pte >= pte) {
1849                 pte_t pteval = *_pte;
1850                 if (!pte_none(pteval))
1851                         release_pte_page(pte_page(pteval));
1852         }
1853 }
1854
1855 static void release_all_pte_pages(pte_t *pte)
1856 {
1857         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1858 }
1859
1860 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1861                                         unsigned long address,
1862                                         pte_t *pte)
1863 {
1864         struct page *page;
1865         pte_t *_pte;
1866         int referenced = 0, isolated = 0, none = 0;
1867         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1868              _pte++, address += PAGE_SIZE) {
1869                 pte_t pteval = *_pte;
1870                 if (pte_none(pteval)) {
1871                         if (++none <= khugepaged_max_ptes_none)
1872                                 continue;
1873                         else {
1874                                 release_pte_pages(pte, _pte);
1875                                 goto out;
1876                         }
1877                 }
1878                 if (!pte_present(pteval) || !pte_write(pteval)) {
1879                         release_pte_pages(pte, _pte);
1880                         goto out;
1881                 }
1882                 page = vm_normal_page(vma, address, pteval);
1883                 if (unlikely(!page)) {
1884                         release_pte_pages(pte, _pte);
1885                         goto out;
1886                 }
1887                 VM_BUG_ON(PageCompound(page));
1888                 BUG_ON(!PageAnon(page));
1889                 VM_BUG_ON(!PageSwapBacked(page));
1890
1891                 /* cannot use mapcount: can't collapse if there's a gup pin */
1892                 if (page_count(page) != 1) {
1893                         release_pte_pages(pte, _pte);
1894                         goto out;
1895                 }
1896                 /*
1897                  * We can do it before isolate_lru_page because the
1898                  * page can't be freed from under us. NOTE: PG_lock
1899                  * is needed to serialize against split_huge_page
1900                  * when invoked from the VM.
1901                  */
1902                 if (!trylock_page(page)) {
1903                         release_pte_pages(pte, _pte);
1904                         goto out;
1905                 }
1906                 /*
1907                  * Isolate the page to avoid collapsing an hugepage
1908                  * currently in use by the VM.
1909                  */
1910                 if (isolate_lru_page(page)) {
1911                         unlock_page(page);
1912                         release_pte_pages(pte, _pte);
1913                         goto out;
1914                 }
1915                 /* 0 stands for page_is_file_cache(page) == false */
1916                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1917                 VM_BUG_ON(!PageLocked(page));
1918                 VM_BUG_ON(PageLRU(page));
1919
1920                 /* If there is no mapped pte young don't collapse the page */
1921                 if (pte_young(pteval) || PageReferenced(page) ||
1922                     mmu_notifier_test_young(vma->vm_mm, address))
1923                         referenced = 1;
1924         }
1925         if (unlikely(!referenced))
1926                 release_all_pte_pages(pte);
1927         else
1928                 isolated = 1;
1929 out:
1930         return isolated;
1931 }
1932
1933 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1934                                       struct vm_area_struct *vma,
1935                                       unsigned long address,
1936                                       spinlock_t *ptl)
1937 {
1938         pte_t *_pte;
1939         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1940                 pte_t pteval = *_pte;
1941                 struct page *src_page;
1942
1943                 if (pte_none(pteval)) {
1944                         clear_user_highpage(page, address);
1945                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1946                 } else {
1947                         src_page = pte_page(pteval);
1948                         copy_user_highpage(page, src_page, address, vma);
1949                         VM_BUG_ON(page_mapcount(src_page) != 1);
1950                         release_pte_page(src_page);
1951                         /*
1952                          * ptl mostly unnecessary, but preempt has to
1953                          * be disabled to update the per-cpu stats
1954                          * inside page_remove_rmap().
1955                          */
1956                         spin_lock(ptl);
1957                         /*
1958                          * paravirt calls inside pte_clear here are
1959                          * superfluous.
1960                          */
1961                         pte_clear(vma->vm_mm, address, _pte);
1962                         page_remove_rmap(src_page);
1963                         spin_unlock(ptl);
1964                         free_page_and_swap_cache(src_page);
1965                 }
1966
1967                 address += PAGE_SIZE;
1968                 page++;
1969         }
1970 }
1971
1972 static void khugepaged_alloc_sleep(void)
1973 {
1974         wait_event_freezable_timeout(khugepaged_wait, false,
1975                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1976 }
1977
1978 #ifdef CONFIG_NUMA
1979 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1980 {
1981         if (IS_ERR(*hpage)) {
1982                 if (!*wait)
1983                         return false;
1984
1985                 *wait = false;
1986                 *hpage = NULL;
1987                 khugepaged_alloc_sleep();
1988         } else if (*hpage) {
1989                 put_page(*hpage);
1990                 *hpage = NULL;
1991         }
1992
1993         return true;
1994 }
1995
1996 static struct page
1997 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1998                        struct vm_area_struct *vma, unsigned long address,
1999                        int node)
2000 {
2001         VM_BUG_ON(*hpage);
2002         /*
2003          * Allocate the page while the vma is still valid and under
2004          * the mmap_sem read mode so there is no memory allocation
2005          * later when we take the mmap_sem in write mode. This is more
2006          * friendly behavior (OTOH it may actually hide bugs) to
2007          * filesystems in userland with daemons allocating memory in
2008          * the userland I/O paths.  Allocating memory with the
2009          * mmap_sem in read mode is good idea also to allow greater
2010          * scalability.
2011          */
2012         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2013                                       node, __GFP_OTHER_NODE);
2014
2015         /*
2016          * After allocating the hugepage, release the mmap_sem read lock in
2017          * preparation for taking it in write mode.
2018          */
2019         up_read(&mm->mmap_sem);
2020         if (unlikely(!*hpage)) {
2021                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2022                 *hpage = ERR_PTR(-ENOMEM);
2023                 return NULL;
2024         }
2025
2026         count_vm_event(THP_COLLAPSE_ALLOC);
2027         return *hpage;
2028 }
2029 #else
2030 static struct page *khugepaged_alloc_hugepage(bool *wait)
2031 {
2032         struct page *hpage;
2033
2034         do {
2035                 hpage = alloc_hugepage(khugepaged_defrag());
2036                 if (!hpage) {
2037                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2038                         if (!*wait)
2039                                 return NULL;
2040
2041                         *wait = false;
2042                         khugepaged_alloc_sleep();
2043                 } else
2044                         count_vm_event(THP_COLLAPSE_ALLOC);
2045         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2046
2047         return hpage;
2048 }
2049
2050 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2051 {
2052         if (!*hpage)
2053                 *hpage = khugepaged_alloc_hugepage(wait);
2054
2055         if (unlikely(!*hpage))
2056                 return false;
2057
2058         return true;
2059 }
2060
2061 static struct page
2062 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2063                        struct vm_area_struct *vma, unsigned long address,
2064                        int node)
2065 {
2066         up_read(&mm->mmap_sem);
2067         VM_BUG_ON(!*hpage);
2068         return  *hpage;
2069 }
2070 #endif
2071
2072 static void collapse_huge_page(struct mm_struct *mm,
2073                                    unsigned long address,
2074                                    struct page **hpage,
2075                                    struct vm_area_struct *vma,
2076                                    int node)
2077 {
2078         pgd_t *pgd;
2079         pud_t *pud;
2080         pmd_t *pmd, _pmd;
2081         pte_t *pte;
2082         pgtable_t pgtable;
2083         struct page *new_page;
2084         spinlock_t *ptl;
2085         int isolated;
2086         unsigned long hstart, hend;
2087         unsigned long mmun_start;       /* For mmu_notifiers */
2088         unsigned long mmun_end;         /* For mmu_notifiers */
2089
2090         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2091
2092         /* release the mmap_sem read lock. */
2093         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2094         if (!new_page)
2095                 return;
2096
2097         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2098                 return;
2099
2100         /*
2101          * Prevent all access to pagetables with the exception of
2102          * gup_fast later hanlded by the ptep_clear_flush and the VM
2103          * handled by the anon_vma lock + PG_lock.
2104          */
2105         down_write(&mm->mmap_sem);
2106         if (unlikely(khugepaged_test_exit(mm)))
2107                 goto out;
2108
2109         vma = find_vma(mm, address);
2110         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2111         hend = vma->vm_end & HPAGE_PMD_MASK;
2112         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2113                 goto out;
2114
2115         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2116             (vma->vm_flags & VM_NOHUGEPAGE))
2117                 goto out;
2118
2119         if (!vma->anon_vma || vma->vm_ops)
2120                 goto out;
2121         if (is_vma_temporary_stack(vma))
2122                 goto out;
2123         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2124
2125         pgd = pgd_offset(mm, address);
2126         if (!pgd_present(*pgd))
2127                 goto out;
2128
2129         pud = pud_offset(pgd, address);
2130         if (!pud_present(*pud))
2131                 goto out;
2132
2133         pmd = pmd_offset(pud, address);
2134         /* pmd can't go away or become huge under us */
2135         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2136                 goto out;
2137
2138         anon_vma_lock(vma->anon_vma);
2139
2140         pte = pte_offset_map(pmd, address);
2141         ptl = pte_lockptr(mm, pmd);
2142
2143         mmun_start = address;
2144         mmun_end   = address + HPAGE_PMD_SIZE;
2145         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2146         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2147         /*
2148          * After this gup_fast can't run anymore. This also removes
2149          * any huge TLB entry from the CPU so we won't allow
2150          * huge and small TLB entries for the same virtual address
2151          * to avoid the risk of CPU bugs in that area.
2152          */
2153         _pmd = pmdp_clear_flush(vma, address, pmd);
2154         spin_unlock(&mm->page_table_lock);
2155         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2156
2157         spin_lock(ptl);
2158         isolated = __collapse_huge_page_isolate(vma, address, pte);
2159         spin_unlock(ptl);
2160
2161         if (unlikely(!isolated)) {
2162                 pte_unmap(pte);
2163                 spin_lock(&mm->page_table_lock);
2164                 BUG_ON(!pmd_none(*pmd));
2165                 set_pmd_at(mm, address, pmd, _pmd);
2166                 spin_unlock(&mm->page_table_lock);
2167                 anon_vma_unlock(vma->anon_vma);
2168                 goto out;
2169         }
2170
2171         /*
2172          * All pages are isolated and locked so anon_vma rmap
2173          * can't run anymore.
2174          */
2175         anon_vma_unlock(vma->anon_vma);
2176
2177         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2178         pte_unmap(pte);
2179         __SetPageUptodate(new_page);
2180         pgtable = pmd_pgtable(_pmd);
2181
2182         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2183         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2184         _pmd = pmd_mkhuge(_pmd);
2185
2186         /*
2187          * spin_lock() below is not the equivalent of smp_wmb(), so
2188          * this is needed to avoid the copy_huge_page writes to become
2189          * visible after the set_pmd_at() write.
2190          */
2191         smp_wmb();
2192
2193         spin_lock(&mm->page_table_lock);
2194         BUG_ON(!pmd_none(*pmd));
2195         page_add_new_anon_rmap(new_page, vma, address);
2196         set_pmd_at(mm, address, pmd, _pmd);
2197         update_mmu_cache_pmd(vma, address, pmd);
2198         pgtable_trans_huge_deposit(mm, pgtable);
2199         spin_unlock(&mm->page_table_lock);
2200
2201         *hpage = NULL;
2202
2203         khugepaged_pages_collapsed++;
2204 out_up_write:
2205         up_write(&mm->mmap_sem);
2206         return;
2207
2208 out:
2209         mem_cgroup_uncharge_page(new_page);
2210         goto out_up_write;
2211 }
2212
2213 static int khugepaged_scan_pmd(struct mm_struct *mm,
2214                                struct vm_area_struct *vma,
2215                                unsigned long address,
2216                                struct page **hpage)
2217 {
2218         pgd_t *pgd;
2219         pud_t *pud;
2220         pmd_t *pmd;
2221         pte_t *pte, *_pte;
2222         int ret = 0, referenced = 0, none = 0;
2223         struct page *page;
2224         unsigned long _address;
2225         spinlock_t *ptl;
2226         int node = -1;
2227
2228         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2229
2230         pgd = pgd_offset(mm, address);
2231         if (!pgd_present(*pgd))
2232                 goto out;
2233
2234         pud = pud_offset(pgd, address);
2235         if (!pud_present(*pud))
2236                 goto out;
2237
2238         pmd = pmd_offset(pud, address);
2239         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2240                 goto out;
2241
2242         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2243         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2244              _pte++, _address += PAGE_SIZE) {
2245                 pte_t pteval = *_pte;
2246                 if (pte_none(pteval)) {
2247                         if (++none <= khugepaged_max_ptes_none)
2248                                 continue;
2249                         else
2250                                 goto out_unmap;
2251                 }
2252                 if (!pte_present(pteval) || !pte_write(pteval))
2253                         goto out_unmap;
2254                 page = vm_normal_page(vma, _address, pteval);
2255                 if (unlikely(!page))
2256                         goto out_unmap;
2257                 /*
2258                  * Chose the node of the first page. This could
2259                  * be more sophisticated and look at more pages,
2260                  * but isn't for now.
2261                  */
2262                 if (node == -1)
2263                         node = page_to_nid(page);
2264                 VM_BUG_ON(PageCompound(page));
2265                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2266                         goto out_unmap;
2267                 /* cannot use mapcount: can't collapse if there's a gup pin */
2268                 if (page_count(page) != 1)
2269                         goto out_unmap;
2270                 if (pte_young(pteval) || PageReferenced(page) ||
2271                     mmu_notifier_test_young(vma->vm_mm, address))
2272                         referenced = 1;
2273         }
2274         if (referenced)
2275                 ret = 1;
2276 out_unmap:
2277         pte_unmap_unlock(pte, ptl);
2278         if (ret)
2279                 /* collapse_huge_page will return with the mmap_sem released */
2280                 collapse_huge_page(mm, address, hpage, vma, node);
2281 out:
2282         return ret;
2283 }
2284
2285 static void collect_mm_slot(struct mm_slot *mm_slot)
2286 {
2287         struct mm_struct *mm = mm_slot->mm;
2288
2289         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2290
2291         if (khugepaged_test_exit(mm)) {
2292                 /* free mm_slot */
2293                 hlist_del(&mm_slot->hash);
2294                 list_del(&mm_slot->mm_node);
2295
2296                 /*
2297                  * Not strictly needed because the mm exited already.
2298                  *
2299                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2300                  */
2301
2302                 /* khugepaged_mm_lock actually not necessary for the below */
2303                 free_mm_slot(mm_slot);
2304                 mmdrop(mm);
2305         }
2306 }
2307
2308 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2309                                             struct page **hpage)
2310         __releases(&khugepaged_mm_lock)
2311         __acquires(&khugepaged_mm_lock)
2312 {
2313         struct mm_slot *mm_slot;
2314         struct mm_struct *mm;
2315         struct vm_area_struct *vma;
2316         int progress = 0;
2317
2318         VM_BUG_ON(!pages);
2319         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2320
2321         if (khugepaged_scan.mm_slot)
2322                 mm_slot = khugepaged_scan.mm_slot;
2323         else {
2324                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2325                                      struct mm_slot, mm_node);
2326                 khugepaged_scan.address = 0;
2327                 khugepaged_scan.mm_slot = mm_slot;
2328         }
2329         spin_unlock(&khugepaged_mm_lock);
2330
2331         mm = mm_slot->mm;
2332         down_read(&mm->mmap_sem);
2333         if (unlikely(khugepaged_test_exit(mm)))
2334                 vma = NULL;
2335         else
2336                 vma = find_vma(mm, khugepaged_scan.address);
2337
2338         progress++;
2339         for (; vma; vma = vma->vm_next) {
2340                 unsigned long hstart, hend;
2341
2342                 cond_resched();
2343                 if (unlikely(khugepaged_test_exit(mm))) {
2344                         progress++;
2345                         break;
2346                 }
2347
2348                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2349                      !khugepaged_always()) ||
2350                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2351                 skip:
2352                         progress++;
2353                         continue;
2354                 }
2355                 if (!vma->anon_vma || vma->vm_ops)
2356                         goto skip;
2357                 if (is_vma_temporary_stack(vma))
2358                         goto skip;
2359                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2360
2361                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2362                 hend = vma->vm_end & HPAGE_PMD_MASK;
2363                 if (hstart >= hend)
2364                         goto skip;
2365                 if (khugepaged_scan.address > hend)
2366                         goto skip;
2367                 if (khugepaged_scan.address < hstart)
2368                         khugepaged_scan.address = hstart;
2369                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2370
2371                 while (khugepaged_scan.address < hend) {
2372                         int ret;
2373                         cond_resched();
2374                         if (unlikely(khugepaged_test_exit(mm)))
2375                                 goto breakouterloop;
2376
2377                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2378                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2379                                   hend);
2380                         ret = khugepaged_scan_pmd(mm, vma,
2381                                                   khugepaged_scan.address,
2382                                                   hpage);
2383                         /* move to next address */
2384                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2385                         progress += HPAGE_PMD_NR;
2386                         if (ret)
2387                                 /* we released mmap_sem so break loop */
2388                                 goto breakouterloop_mmap_sem;
2389                         if (progress >= pages)
2390                                 goto breakouterloop;
2391                 }
2392         }
2393 breakouterloop:
2394         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2395 breakouterloop_mmap_sem:
2396
2397         spin_lock(&khugepaged_mm_lock);
2398         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2399         /*
2400          * Release the current mm_slot if this mm is about to die, or
2401          * if we scanned all vmas of this mm.
2402          */
2403         if (khugepaged_test_exit(mm) || !vma) {
2404                 /*
2405                  * Make sure that if mm_users is reaching zero while
2406                  * khugepaged runs here, khugepaged_exit will find
2407                  * mm_slot not pointing to the exiting mm.
2408                  */
2409                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2410                         khugepaged_scan.mm_slot = list_entry(
2411                                 mm_slot->mm_node.next,
2412                                 struct mm_slot, mm_node);
2413                         khugepaged_scan.address = 0;
2414                 } else {
2415                         khugepaged_scan.mm_slot = NULL;
2416                         khugepaged_full_scans++;
2417                 }
2418
2419                 collect_mm_slot(mm_slot);
2420         }
2421
2422         return progress;
2423 }
2424
2425 static int khugepaged_has_work(void)
2426 {
2427         return !list_empty(&khugepaged_scan.mm_head) &&
2428                 khugepaged_enabled();
2429 }
2430
2431 static int khugepaged_wait_event(void)
2432 {
2433         return !list_empty(&khugepaged_scan.mm_head) ||
2434                 kthread_should_stop();
2435 }
2436
2437 static void khugepaged_do_scan(void)
2438 {
2439         struct page *hpage = NULL;
2440         unsigned int progress = 0, pass_through_head = 0;
2441         bool wait = true;
2442         unsigned int pages = ACCESS_ONCE(khugepaged_pages_to_scan);
2443
2444         while (progress < pages) {
2445                 if (!khugepaged_prealloc_page(&hpage, &wait))
2446                         break;
2447
2448                 cond_resched();
2449
2450                 if (unlikely(kthread_should_stop() || freezing(current)))
2451                         break;
2452
2453                 spin_lock(&khugepaged_mm_lock);
2454                 if (!khugepaged_scan.mm_slot)
2455                         pass_through_head++;
2456                 if (khugepaged_has_work() &&
2457                     pass_through_head < 2)
2458                         progress += khugepaged_scan_mm_slot(pages - progress,
2459                                                             &hpage);
2460                 else
2461                         progress = pages;
2462                 spin_unlock(&khugepaged_mm_lock);
2463         }
2464
2465         if (!IS_ERR_OR_NULL(hpage))
2466                 put_page(hpage);
2467 }
2468
2469 static void khugepaged_wait_work(void)
2470 {
2471         try_to_freeze();
2472
2473         if (khugepaged_has_work()) {
2474                 if (!khugepaged_scan_sleep_millisecs)
2475                         return;
2476
2477                 wait_event_freezable_timeout(khugepaged_wait,
2478                                              kthread_should_stop(),
2479                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2480                 return;
2481         }
2482
2483         if (khugepaged_enabled())
2484                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2485 }
2486
2487 static int khugepaged(void *none)
2488 {
2489         struct mm_slot *mm_slot;
2490
2491         set_freezable();
2492         set_user_nice(current, 19);
2493
2494         while (!kthread_should_stop()) {
2495                 khugepaged_do_scan();
2496                 khugepaged_wait_work();
2497         }
2498
2499         spin_lock(&khugepaged_mm_lock);
2500         mm_slot = khugepaged_scan.mm_slot;
2501         khugepaged_scan.mm_slot = NULL;
2502         if (mm_slot)
2503                 collect_mm_slot(mm_slot);
2504         spin_unlock(&khugepaged_mm_lock);
2505         return 0;
2506 }
2507
2508 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2509 {
2510         struct page *page;
2511
2512         spin_lock(&mm->page_table_lock);
2513         if (unlikely(!pmd_trans_huge(*pmd))) {
2514                 spin_unlock(&mm->page_table_lock);
2515                 return;
2516         }
2517         page = pmd_page(*pmd);
2518         VM_BUG_ON(!page_count(page));
2519         get_page(page);
2520         spin_unlock(&mm->page_table_lock);
2521
2522         split_huge_page(page);
2523
2524         put_page(page);
2525         BUG_ON(pmd_trans_huge(*pmd));
2526 }
2527
2528 static void split_huge_page_address(struct mm_struct *mm,
2529                                     unsigned long address)
2530 {
2531         pgd_t *pgd;
2532         pud_t *pud;
2533         pmd_t *pmd;
2534
2535         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2536
2537         pgd = pgd_offset(mm, address);
2538         if (!pgd_present(*pgd))
2539                 return;
2540
2541         pud = pud_offset(pgd, address);
2542         if (!pud_present(*pud))
2543                 return;
2544
2545         pmd = pmd_offset(pud, address);
2546         if (!pmd_present(*pmd))
2547                 return;
2548         /*
2549          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2550          * materialize from under us.
2551          */
2552         split_huge_page_pmd(mm, pmd);
2553 }
2554
2555 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2556                              unsigned long start,
2557                              unsigned long end,
2558                              long adjust_next)
2559 {
2560         /*
2561          * If the new start address isn't hpage aligned and it could
2562          * previously contain an hugepage: check if we need to split
2563          * an huge pmd.
2564          */
2565         if (start & ~HPAGE_PMD_MASK &&
2566             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2567             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2568                 split_huge_page_address(vma->vm_mm, start);
2569
2570         /*
2571          * If the new end address isn't hpage aligned and it could
2572          * previously contain an hugepage: check if we need to split
2573          * an huge pmd.
2574          */
2575         if (end & ~HPAGE_PMD_MASK &&
2576             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2577             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2578                 split_huge_page_address(vma->vm_mm, end);
2579
2580         /*
2581          * If we're also updating the vma->vm_next->vm_start, if the new
2582          * vm_next->vm_start isn't page aligned and it could previously
2583          * contain an hugepage: check if we need to split an huge pmd.
2584          */
2585         if (adjust_next > 0) {
2586                 struct vm_area_struct *next = vma->vm_next;
2587                 unsigned long nstart = next->vm_start;
2588                 nstart += adjust_next << PAGE_SHIFT;
2589                 if (nstart & ~HPAGE_PMD_MASK &&
2590                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2591                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2592                         split_huge_page_address(next->vm_mm, nstart);
2593         }
2594 }