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