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         int ret = 0;
 640         pgtable_t pgtable;
 641
 642         VM_BUG_ON(!PageCompound(page));
 643         pgtable = pte_alloc_one(mm, haddr);
 644         if (unlikely(!pgtable)) {
 645                 mem_cgroup_uncharge_page(page);
 646                 put_page(page);
 647                 return VM_FAULT_OOM;
 648         }
 649
 650         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 651         __SetPageUptodate(page);
 652
 653         spin_lock(&mm->page_table_lock);
 654         if (unlikely(!pmd_none(*pmd))) {
 655                 spin_unlock(&mm->page_table_lock);
 656                 mem_cgroup_uncharge_page(page);
 657                 put_page(page);
 658                 pte_free(mm, pgtable);
 659         } else {
 660                 pmd_t entry;
 661                 entry = mk_pmd(page, vma->vm_page_prot);
 662                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 663                 entry = pmd_mkhuge(entry);
 664                 /*
 665                  * The spinlocking to take the lru_lock inside
 666                  * page_add_new_anon_rmap() acts as a full memory
 667                  * barrier to be sure clear_huge_page writes become
 668                  * visible after the set_pmd_at() write.
 669                  */
 670                 page_add_new_anon_rmap(page, vma, haddr);
 671                 set_pmd_at(mm, haddr, pmd, entry);
 672                 prepare_pmd_huge_pte(pgtable, mm);
 673                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 674                 mm->nr_ptes++;
 675                 spin_unlock(&mm->page_table_lock);
 676         }
 677
 678         return ret;
 679 }
 680
 681 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 682 {
 683         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 684 }
 685
 686 static inline struct page *alloc_hugepage_vma(int defrag,
 687                                               struct vm_area_struct *vma,
 688                                               unsigned long haddr, int nd,
 689                                               gfp_t extra_gfp)
 690 {
 691         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 692                                HPAGE_PMD_ORDER, vma, haddr, nd);
 693 }
 694
 695 #ifndef CONFIG_NUMA
 696 static inline struct page *alloc_hugepage(int defrag)
 697 {
 698         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 699                            HPAGE_PMD_ORDER);
 700 }
 701 #endif
 702
 703 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 704                                unsigned long address, pmd_t *pmd,
 705                                unsigned int flags)
 706 {
 707         struct page *page;
 708         unsigned long haddr = address & HPAGE_PMD_MASK;
 709         pte_t *pte;
 710
 711         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 712                 if (unlikely(anon_vma_prepare(vma)))
 713                         return VM_FAULT_OOM;
 714                 if (unlikely(khugepaged_enter(vma)))
 715                         return VM_FAULT_OOM;
 716                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 717                                           vma, haddr, numa_node_id(), 0);
 718                 if (unlikely(!page)) {
 719                         count_vm_event(THP_FAULT_FALLBACK);
 720                         goto out;
 721                 }
 722                 count_vm_event(THP_FAULT_ALLOC);
 723                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 724                         put_page(page);
 725                         goto out;
 726                 }
 727
 728                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
 729         }
 730 out:
 731         /*
 732          * Use __pte_alloc instead of pte_alloc_map, because we can't
 733          * run pte_offset_map on the pmd, if an huge pmd could
 734          * materialize from under us from a different thread.
 735          */
 736         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 737                 return VM_FAULT_OOM;
 738         /* if an huge pmd materialized from under us just retry later */
 739         if (unlikely(pmd_trans_huge(*pmd)))
 740                 return 0;
 741         /*
 742          * A regular pmd is established and it can't morph into a huge pmd
 743          * from under us anymore at this point because we hold the mmap_sem
 744          * read mode and khugepaged takes it in write mode. So now it's
 745          * safe to run pte_offset_map().
 746          */
 747         pte = pte_offset_map(pmd, address);
 748         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 749 }
 750
 751 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 752                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 753                   struct vm_area_struct *vma)
 754 {
 755         struct page *src_page;
 756         pmd_t pmd;
 757         pgtable_t pgtable;
 758         int ret;
 759
 760         ret = -ENOMEM;
 761         pgtable = pte_alloc_one(dst_mm, addr);
 762         if (unlikely(!pgtable))
 763                 goto out;
 764
 765         spin_lock(&dst_mm->page_table_lock);
 766         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 767
 768         ret = -EAGAIN;
 769         pmd = *src_pmd;
 770         if (unlikely(!pmd_trans_huge(pmd))) {
 771                 pte_free(dst_mm, pgtable);
 772                 goto out_unlock;
 773         }
 774         if (unlikely(pmd_trans_splitting(pmd))) {
 775                 /* split huge page running from under us */
 776                 spin_unlock(&src_mm->page_table_lock);
 777                 spin_unlock(&dst_mm->page_table_lock);
 778                 pte_free(dst_mm, pgtable);
 779
 780                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 781                 goto out;
 782         }
 783         src_page = pmd_page(pmd);
 784         VM_BUG_ON(!PageHead(src_page));
 785         get_page(src_page);
 786         page_dup_rmap(src_page);
 787         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 788
 789         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 790         pmd = pmd_mkold(pmd_wrprotect(pmd));
 791         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 792         prepare_pmd_huge_pte(pgtable, dst_mm);
 793         dst_mm->nr_ptes++;
 794
 795         ret = 0;
 796 out_unlock:
 797         spin_unlock(&src_mm->page_table_lock);
 798         spin_unlock(&dst_mm->page_table_lock);
 799 out:
 800         return ret;
 801 }
 802
 803 /* no "address" argument so destroys page coloring of some arch */
 804 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 805 {
 806         pgtable_t pgtable;
 807
 808         assert_spin_locked(&mm->page_table_lock);
 809
 810         /* FIFO */
 811         pgtable = mm->pmd_huge_pte;
 812         if (list_empty(&pgtable->lru))
 813                 mm->pmd_huge_pte = NULL;
 814         else {
 815                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 816                                               struct page, lru);
 817                 list_del(&pgtable->lru);
 818         }
 819         return pgtable;
 820 }
 821
 822 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 823                                         struct vm_area_struct *vma,
 824                                         unsigned long address,
 825                                         pmd_t *pmd, pmd_t orig_pmd,
 826                                         struct page *page,
 827                                         unsigned long haddr)
 828 {
 829         pgtable_t pgtable;
 830         pmd_t _pmd;
 831         int ret = 0, i;
 832         struct page **pages;
 833
 834         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 835                         GFP_KERNEL);
 836         if (unlikely(!pages)) {
 837                 ret |= VM_FAULT_OOM;
 838                 goto out;
 839         }
 840
 841         for (i = 0; i < HPAGE_PMD_NR; i++) {
 842                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 843                                                __GFP_OTHER_NODE,
 844                                                vma, address, page_to_nid(page));
 845                 if (unlikely(!pages[i] ||
 846                              mem_cgroup_newpage_charge(pages[i], mm,
 847                                                        GFP_KERNEL))) {
 848                         if (pages[i])
 849                                 put_page(pages[i]);
 850                         mem_cgroup_uncharge_start();
 851                         while (--i >= 0) {
 852                                 mem_cgroup_uncharge_page(pages[i]);
 853                                 put_page(pages[i]);
 854                         }
 855                         mem_cgroup_uncharge_end();
 856                         kfree(pages);
 857                         ret |= VM_FAULT_OOM;
 858                         goto out;
 859                 }
 860         }
 861
 862         for (i = 0; i < HPAGE_PMD_NR; i++) {
 863                 copy_user_highpage(pages[i], page + i,
 864                                    haddr + PAGE_SIZE * i, vma);
 865                 __SetPageUptodate(pages[i]);
 866                 cond_resched();
 867         }
 868
 869         spin_lock(&mm->page_table_lock);
 870         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 871                 goto out_free_pages;
 872         VM_BUG_ON(!PageHead(page));
 873
 874         pmdp_clear_flush_notify(vma, haddr, pmd);
 875         /* leave pmd empty until pte is filled */
 876
 877         pgtable = get_pmd_huge_pte(mm);
 878         pmd_populate(mm, &_pmd, pgtable);
 879
 880         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 881                 pte_t *pte, entry;
 882                 entry = mk_pte(pages[i], vma->vm_page_prot);
 883                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 884                 page_add_new_anon_rmap(pages[i], vma, haddr);
 885                 pte = pte_offset_map(&_pmd, haddr);
 886                 VM_BUG_ON(!pte_none(*pte));
 887                 set_pte_at(mm, haddr, pte, entry);
 888                 pte_unmap(pte);
 889         }
 890         kfree(pages);
 891
 892         smp_wmb(); /* make pte visible before pmd */
 893         pmd_populate(mm, pmd, pgtable);
 894         page_remove_rmap(page);
 895         spin_unlock(&mm->page_table_lock);
 896
 897         ret |= VM_FAULT_WRITE;
 898         put_page(page);
 899
 900 out:
 901         return ret;
 902
 903 out_free_pages:
 904         spin_unlock(&mm->page_table_lock);
 905         mem_cgroup_uncharge_start();
 906         for (i = 0; i < HPAGE_PMD_NR; i++) {
 907                 mem_cgroup_uncharge_page(pages[i]);
 908                 put_page(pages[i]);
 909         }
 910         mem_cgroup_uncharge_end();
 911         kfree(pages);
 912         goto out;
 913 }
 914
 915 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 916                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 917 {
 918         int ret = 0;
 919         struct page *page, *new_page;
 920         unsigned long haddr;
 921
 922         VM_BUG_ON(!vma->anon_vma);
 923         spin_lock(&mm->page_table_lock);
 924         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 925                 goto out_unlock;
 926
 927         page = pmd_page(orig_pmd);
 928         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 929         haddr = address & HPAGE_PMD_MASK;
 930         if (page_mapcount(page) == 1) {
 931                 pmd_t entry;
 932                 entry = pmd_mkyoung(orig_pmd);
 933                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 934                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 935                         update_mmu_cache(vma, address, entry);
 936                 ret |= VM_FAULT_WRITE;
 937                 goto out_unlock;
 938         }
 939         get_page(page);
 940         spin_unlock(&mm->page_table_lock);
 941
 942         if (transparent_hugepage_enabled(vma) &&
 943             !transparent_hugepage_debug_cow())
 944                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 945                                               vma, haddr, numa_node_id(), 0);
 946         else
 947                 new_page = NULL;
 948
 949         if (unlikely(!new_page)) {
 950                 count_vm_event(THP_FAULT_FALLBACK);
 951                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 952                                                    pmd, orig_pmd, page, haddr);
 953                 put_page(page);
 954                 goto out;
 955         }
 956         count_vm_event(THP_FAULT_ALLOC);
 957
 958         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 959                 put_page(new_page);
 960                 put_page(page);
 961                 ret |= VM_FAULT_OOM;
 962                 goto out;
 963         }
 964
 965         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 966         __SetPageUptodate(new_page);
 967
 968         spin_lock(&mm->page_table_lock);
 969         put_page(page);
 970         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 971                 mem_cgroup_uncharge_page(new_page);
 972                 put_page(new_page);
 973         } else {
 974                 pmd_t entry;
 975                 VM_BUG_ON(!PageHead(page));
 976                 entry = mk_pmd(new_page, vma->vm_page_prot);
 977                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 978                 entry = pmd_mkhuge(entry);
 979                 pmdp_clear_flush_notify(vma, haddr, pmd);
 980                 page_add_new_anon_rmap(new_page, vma, haddr);
 981                 set_pmd_at(mm, haddr, pmd, entry);
 982                 update_mmu_cache(vma, address, entry);
 983                 page_remove_rmap(page);
 984                 put_page(page);
 985                 ret |= VM_FAULT_WRITE;
 986         }
 987 out_unlock:
 988         spin_unlock(&mm->page_table_lock);
 989 out:
 990         return ret;
 991 }
 992
 993 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 994                                    unsigned long addr,
 995                                    pmd_t *pmd,
 996                                    unsigned int flags)
 997 {
 998         struct page *page = NULL;
 999
1000         assert_spin_locked(&mm->page_table_lock);
1001
1002         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1003                 goto out;
1004
1005         page = pmd_page(*pmd);
1006         VM_BUG_ON(!PageHead(page));
1007         if (flags & FOLL_TOUCH) {
1008                 pmd_t _pmd;
1009                 /*
1010                  * We should set the dirty bit only for FOLL_WRITE but
1011                  * for now the dirty bit in the pmd is meaningless.
1012                  * And if the dirty bit will become meaningful and
1013                  * we'll only set it with FOLL_WRITE, an atomic
1014                  * set_bit will be required on the pmd to set the
1015                  * young bit, instead of the current set_pmd_at.
1016                  */
1017                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1018                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1019         }
1020         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1021         VM_BUG_ON(!PageCompound(page));
1022         if (flags & FOLL_GET)
1023                 get_page_foll(page);
1024
1025 out:
1026         return page;
1027 }
1028
1029 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1030                  pmd_t *pmd, unsigned long addr)
1031 {
1032         int ret = 0;
1033
1034         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1035                 struct page *page;
1036                 pgtable_t pgtable;
1037                 pgtable = get_pmd_huge_pte(tlb->mm);
1038                 page = pmd_page(*pmd);
1039                 pmd_clear(pmd);
1040                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1041                 page_remove_rmap(page);
1042                 VM_BUG_ON(page_mapcount(page) < 0);
1043                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1044                 VM_BUG_ON(!PageHead(page));
1045                 tlb->mm->nr_ptes--;
1046                 spin_unlock(&tlb->mm->page_table_lock);
1047                 tlb_remove_page(tlb, page);
1048                 pte_free(tlb->mm, pgtable);
1049                 ret = 1;
1050         }
1051         return ret;
1052 }
1053
1054 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1055                 unsigned long addr, unsigned long end,
1056                 unsigned char *vec)
1057 {
1058         int ret = 0;
1059
1060         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1061                 /*
1062                  * All logical pages in the range are present
1063                  * if backed by a huge page.
1064                  */
1065                 spin_unlock(&vma->vm_mm->page_table_lock);
1066                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1067                 ret = 1;
1068         }
1069
1070         return ret;
1071 }
1072
1073 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1074                   unsigned long old_addr,
1075                   unsigned long new_addr, unsigned long old_end,
1076                   pmd_t *old_pmd, pmd_t *new_pmd)
1077 {
1078         int ret = 0;
1079         pmd_t pmd;
1080
1081         struct mm_struct *mm = vma->vm_mm;
1082
1083         if ((old_addr & ~HPAGE_PMD_MASK) ||
1084             (new_addr & ~HPAGE_PMD_MASK) ||
1085             old_end - old_addr < HPAGE_PMD_SIZE ||
1086             (new_vma->vm_flags & VM_NOHUGEPAGE))
1087                 goto out;
1088
1089         /*
1090          * The destination pmd shouldn't be established, free_pgtables()
1091          * should have release it.
1092          */
1093         if (WARN_ON(!pmd_none(*new_pmd))) {
1094                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1095                 goto out;
1096         }
1097
1098         ret = __pmd_trans_huge_lock(old_pmd, vma);
1099         if (ret == 1) {
1100                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1101                 VM_BUG_ON(!pmd_none(*new_pmd));
1102                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1103                 spin_unlock(&mm->page_table_lock);
1104         }
1105 out:
1106         return ret;
1107 }
1108
1109 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1110                 unsigned long addr, pgprot_t newprot)
1111 {
1112         struct mm_struct *mm = vma->vm_mm;
1113         int ret = 0;
1114
1115         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1116                 pmd_t entry;
1117                 entry = pmdp_get_and_clear(mm, addr, pmd);
1118                 entry = pmd_modify(entry, newprot);
1119                 set_pmd_at(mm, addr, pmd, entry);
1120                 spin_unlock(&vma->vm_mm->page_table_lock);
1121                 ret = 1;
1122         }
1123
1124         return ret;
1125 }
1126
1127 /*
1128  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1129  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1130  *
1131  * Note that if it returns 1, this routine returns without unlocking page
1132  * table locks. So callers must unlock them.
1133  */
1134 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1135 {
1136         spin_lock(&vma->vm_mm->page_table_lock);
1137         if (likely(pmd_trans_huge(*pmd))) {
1138                 if (unlikely(pmd_trans_splitting(*pmd))) {
1139                         spin_unlock(&vma->vm_mm->page_table_lock);
1140                         wait_split_huge_page(vma->anon_vma, pmd);
1141                         return -1;
1142                 } else {
1143                         /* Thp mapped by 'pmd' is stable, so we can
1144                          * handle it as it is. */
1145                         return 1;
1146                 }
1147         }
1148         spin_unlock(&vma->vm_mm->page_table_lock);
1149         return 0;
1150 }
1151
1152 pmd_t *page_check_address_pmd(struct page *page,
1153                               struct mm_struct *mm,
1154                               unsigned long address,
1155                               enum page_check_address_pmd_flag flag)
1156 {
1157         pgd_t *pgd;
1158         pud_t *pud;
1159         pmd_t *pmd, *ret = NULL;
1160
1161         if (address & ~HPAGE_PMD_MASK)
1162                 goto out;
1163
1164         pgd = pgd_offset(mm, address);
1165         if (!pgd_present(*pgd))
1166                 goto out;
1167
1168         pud = pud_offset(pgd, address);
1169         if (!pud_present(*pud))
1170                 goto out;
1171
1172         pmd = pmd_offset(pud, address);
1173         if (pmd_none(*pmd))
1174                 goto out;
1175         if (pmd_page(*pmd) != page)
1176                 goto out;
1177         /*
1178          * split_vma() may create temporary aliased mappings. There is
1179          * no risk as long as all huge pmd are found and have their
1180          * splitting bit set before __split_huge_page_refcount
1181          * runs. Finding the same huge pmd more than once during the
1182          * same rmap walk is not a problem.
1183          */
1184         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1185             pmd_trans_splitting(*pmd))
1186                 goto out;
1187         if (pmd_trans_huge(*pmd)) {
1188                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1189                           !pmd_trans_splitting(*pmd));
1190                 ret = pmd;
1191         }
1192 out:
1193         return ret;
1194 }
1195
1196 static int __split_huge_page_splitting(struct page *page,
1197                                        struct vm_area_struct *vma,
1198                                        unsigned long address)
1199 {
1200         struct mm_struct *mm = vma->vm_mm;
1201         pmd_t *pmd;
1202         int ret = 0;
1203
1204         spin_lock(&mm->page_table_lock);
1205         pmd = page_check_address_pmd(page, mm, address,
1206                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1207         if (pmd) {
1208                 /*
1209                  * We can't temporarily set the pmd to null in order
1210                  * to split it, the pmd must remain marked huge at all
1211                  * times or the VM won't take the pmd_trans_huge paths
1212                  * and it won't wait on the anon_vma->root->mutex to
1213                  * serialize against split_huge_page*.
1214                  */
1215                 pmdp_splitting_flush_notify(vma, address, pmd);
1216                 ret = 1;
1217         }
1218         spin_unlock(&mm->page_table_lock);
1219
1220         return ret;
1221 }
1222
1223 static void __split_huge_page_refcount(struct page *page)
1224 {
1225         int i;
1226         struct zone *zone = page_zone(page);
1227         int tail_count = 0;
1228
1229         /* prevent PageLRU to go away from under us, and freeze lru stats */
1230         spin_lock_irq(&zone->lru_lock);
1231         compound_lock(page);
1232         /* complete memcg works before add pages to LRU */
1233         mem_cgroup_split_huge_fixup(page);
1234
1235         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1236                 struct page *page_tail = page + i;
1237
1238                 /* tail_page->_mapcount cannot change */
1239                 BUG_ON(page_mapcount(page_tail) < 0);
1240                 tail_count += page_mapcount(page_tail);
1241                 /* check for overflow */
1242                 BUG_ON(tail_count < 0);
1243                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1244                 /*
1245                  * tail_page->_count is zero and not changing from
1246                  * under us. But get_page_unless_zero() may be running
1247                  * from under us on the tail_page. If we used
1248                  * atomic_set() below instead of atomic_add(), we
1249                  * would then run atomic_set() concurrently with
1250                  * get_page_unless_zero(), and atomic_set() is
1251                  * implemented in C not using locked ops. spin_unlock
1252                  * on x86 sometime uses locked ops because of PPro
1253                  * errata 66, 92, so unless somebody can guarantee
1254                  * atomic_set() here would be safe on all archs (and
1255                  * not only on x86), it's safer to use atomic_add().
1256                  */
1257                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1258                            &page_tail->_count);
1259
1260                 /* after clearing PageTail the gup refcount can be released */
1261                 smp_mb();
1262
1263                 /*
1264                  * retain hwpoison flag of the poisoned tail page:
1265                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1266                  *   by the memory-failure.
1267                  */
1268                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1269                 page_tail->flags |= (page->flags &
1270                                      ((1L << PG_referenced) |
1271                                       (1L << PG_swapbacked) |
1272                                       (1L << PG_mlocked) |
1273                                       (1L << PG_uptodate)));
1274                 page_tail->flags |= (1L << PG_dirty);
1275
1276                 /* clear PageTail before overwriting first_page */
1277                 smp_wmb();
1278
1279                 /*
1280                  * __split_huge_page_splitting() already set the
1281                  * splitting bit in all pmd that could map this
1282                  * hugepage, that will ensure no CPU can alter the
1283                  * mapcount on the head page. The mapcount is only
1284                  * accounted in the head page and it has to be
1285                  * transferred to all tail pages in the below code. So
1286                  * for this code to be safe, the split the mapcount
1287                  * can't change. But that doesn't mean userland can't
1288                  * keep changing and reading the page contents while
1289                  * we transfer the mapcount, so the pmd splitting
1290                  * status is achieved setting a reserved bit in the
1291                  * pmd, not by clearing the present bit.
1292                 */
1293                 page_tail->_mapcount = page->_mapcount;
1294
1295                 BUG_ON(page_tail->mapping);
1296                 page_tail->mapping = page->mapping;
1297
1298                 page_tail->index = page->index + i;
1299
1300                 BUG_ON(!PageAnon(page_tail));
1301                 BUG_ON(!PageUptodate(page_tail));
1302                 BUG_ON(!PageDirty(page_tail));
1303                 BUG_ON(!PageSwapBacked(page_tail));
1304
1305
1306                 lru_add_page_tail(zone, page, page_tail);
1307         }
1308         atomic_sub(tail_count, &page->_count);
1309         BUG_ON(atomic_read(&page->_count) <= 0);
1310
1311         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1312         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1313
1314         ClearPageCompound(page);
1315         compound_unlock(page);
1316         spin_unlock_irq(&zone->lru_lock);
1317
1318         for (i = 1; i < HPAGE_PMD_NR; i++) {
1319                 struct page *page_tail = page + i;
1320                 BUG_ON(page_count(page_tail) <= 0);
1321                 /*
1322                  * Tail pages may be freed if there wasn't any mapping
1323                  * like if add_to_swap() is running on a lru page that
1324                  * had its mapping zapped. And freeing these pages
1325                  * requires taking the lru_lock so we do the put_page
1326                  * of the tail pages after the split is complete.
1327                  */
1328                 put_page(page_tail);
1329         }
1330
1331         /*
1332          * Only the head page (now become a regular page) is required
1333          * to be pinned by the caller.
1334          */
1335         BUG_ON(page_count(page) <= 0);
1336 }
1337
1338 static int __split_huge_page_map(struct page *page,
1339                                  struct vm_area_struct *vma,
1340                                  unsigned long address)
1341 {
1342         struct mm_struct *mm = vma->vm_mm;
1343         pmd_t *pmd, _pmd;
1344         int ret = 0, i;
1345         pgtable_t pgtable;
1346         unsigned long haddr;
1347
1348         spin_lock(&mm->page_table_lock);
1349         pmd = page_check_address_pmd(page, mm, address,
1350                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1351         if (pmd) {
1352                 pgtable = get_pmd_huge_pte(mm);
1353                 pmd_populate(mm, &_pmd, pgtable);
1354
1355                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1356                      i++, haddr += PAGE_SIZE) {
1357                         pte_t *pte, entry;
1358                         BUG_ON(PageCompound(page+i));
1359                         entry = mk_pte(page + i, vma->vm_page_prot);
1360                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1361                         if (!pmd_write(*pmd))
1362                                 entry = pte_wrprotect(entry);
1363                         else
1364                                 BUG_ON(page_mapcount(page) != 1);
1365                         if (!pmd_young(*pmd))
1366                                 entry = pte_mkold(entry);
1367                         pte = pte_offset_map(&_pmd, haddr);
1368                         BUG_ON(!pte_none(*pte));
1369                         set_pte_at(mm, haddr, pte, entry);
1370                         pte_unmap(pte);
1371                 }
1372
1373                 smp_wmb(); /* make pte visible before pmd */
1374                 /*
1375                  * Up to this point the pmd is present and huge and
1376                  * userland has the whole access to the hugepage
1377                  * during the split (which happens in place). If we
1378                  * overwrite the pmd with the not-huge version
1379                  * pointing to the pte here (which of course we could
1380                  * if all CPUs were bug free), userland could trigger
1381                  * a small page size TLB miss on the small sized TLB
1382                  * while the hugepage TLB entry is still established
1383                  * in the huge TLB. Some CPU doesn't like that. See
1384                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1385                  * Erratum 383 on page 93. Intel should be safe but is
1386                  * also warns that it's only safe if the permission
1387                  * and cache attributes of the two entries loaded in
1388                  * the two TLB is identical (which should be the case
1389                  * here). But it is generally safer to never allow
1390                  * small and huge TLB entries for the same virtual
1391                  * address to be loaded simultaneously. So instead of
1392                  * doing "pmd_populate(); flush_tlb_range();" we first
1393                  * mark the current pmd notpresent (atomically because
1394                  * here the pmd_trans_huge and pmd_trans_splitting
1395                  * must remain set at all times on the pmd until the
1396                  * split is complete for this pmd), then we flush the
1397                  * SMP TLB and finally we write the non-huge version
1398                  * of the pmd entry with pmd_populate.
1399                  */
1400                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1401                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1402                 pmd_populate(mm, pmd, pgtable);
1403                 ret = 1;
1404         }
1405         spin_unlock(&mm->page_table_lock);
1406
1407         return ret;
1408 }
1409
1410 /* must be called with anon_vma->root->mutex hold */
1411 static void __split_huge_page(struct page *page,
1412                               struct anon_vma *anon_vma)
1413 {
1414         int mapcount, mapcount2;
1415         struct anon_vma_chain *avc;
1416
1417         BUG_ON(!PageHead(page));
1418         BUG_ON(PageTail(page));
1419
1420         mapcount = 0;
1421         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1422                 struct vm_area_struct *vma = avc->vma;
1423                 unsigned long addr = vma_address(page, vma);
1424                 BUG_ON(is_vma_temporary_stack(vma));
1425                 if (addr == -EFAULT)
1426                         continue;
1427                 mapcount += __split_huge_page_splitting(page, vma, addr);
1428         }
1429         /*
1430          * It is critical that new vmas are added to the tail of the
1431          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1432          * and establishes a child pmd before
1433          * __split_huge_page_splitting() freezes the parent pmd (so if
1434          * we fail to prevent copy_huge_pmd() from running until the
1435          * whole __split_huge_page() is complete), we will still see
1436          * the newly established pmd of the child later during the
1437          * walk, to be able to set it as pmd_trans_splitting too.
1438          */
1439         if (mapcount != page_mapcount(page))
1440                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1441                        mapcount, page_mapcount(page));
1442         BUG_ON(mapcount != page_mapcount(page));
1443
1444         __split_huge_page_refcount(page);
1445
1446         mapcount2 = 0;
1447         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1448                 struct vm_area_struct *vma = avc->vma;
1449                 unsigned long addr = vma_address(page, vma);
1450                 BUG_ON(is_vma_temporary_stack(vma));
1451                 if (addr == -EFAULT)
1452                         continue;
1453                 mapcount2 += __split_huge_page_map(page, vma, addr);
1454         }
1455         if (mapcount != mapcount2)
1456                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1457                        mapcount, mapcount2, page_mapcount(page));
1458         BUG_ON(mapcount != mapcount2);
1459 }
1460
1461 int split_huge_page(struct page *page)
1462 {
1463         struct anon_vma *anon_vma;
1464         int ret = 1;
1465
1466         BUG_ON(!PageAnon(page));
1467         anon_vma = page_lock_anon_vma(page);
1468         if (!anon_vma)
1469                 goto out;
1470         ret = 0;
1471         if (!PageCompound(page))
1472                 goto out_unlock;
1473
1474         BUG_ON(!PageSwapBacked(page));
1475         __split_huge_page(page, anon_vma);
1476         count_vm_event(THP_SPLIT);
1477
1478         BUG_ON(PageCompound(page));
1479 out_unlock:
1480         page_unlock_anon_vma(anon_vma);
1481 out:
1482         return ret;
1483 }
1484
1485 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1486                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1487
1488 int hugepage_madvise(struct vm_area_struct *vma,
1489                      unsigned long *vm_flags, int advice)
1490 {
1491         switch (advice) {
1492         case MADV_HUGEPAGE:
1493                 /*
1494                  * Be somewhat over-protective like KSM for now!
1495                  */
1496                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1497                         return -EINVAL;
1498                 *vm_flags &= ~VM_NOHUGEPAGE;
1499                 *vm_flags |= VM_HUGEPAGE;
1500                 /*
1501                  * If the vma become good for khugepaged to scan,
1502                  * register it here without waiting a page fault that
1503                  * may not happen any time soon.
1504                  */
1505                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1506                         return -ENOMEM;
1507                 break;
1508         case MADV_NOHUGEPAGE:
1509                 /*
1510                  * Be somewhat over-protective like KSM for now!
1511                  */
1512                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1513                         return -EINVAL;
1514                 *vm_flags &= ~VM_HUGEPAGE;
1515                 *vm_flags |= VM_NOHUGEPAGE;
1516                 /*
1517                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1518                  * this vma even if we leave the mm registered in khugepaged if
1519                  * it got registered before VM_NOHUGEPAGE was set.
1520                  */
1521                 break;
1522         }
1523
1524         return 0;
1525 }
1526
1527 static int __init khugepaged_slab_init(void)
1528 {
1529         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1530                                           sizeof(struct mm_slot),
1531                                           __alignof__(struct mm_slot), 0, NULL);
1532         if (!mm_slot_cache)
1533                 return -ENOMEM;
1534
1535         return 0;
1536 }
1537
1538 static void __init khugepaged_slab_free(void)
1539 {
1540         kmem_cache_destroy(mm_slot_cache);
1541         mm_slot_cache = NULL;
1542 }
1543
1544 static inline struct mm_slot *alloc_mm_slot(void)
1545 {
1546         if (!mm_slot_cache)     /* initialization failed */
1547                 return NULL;
1548         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1549 }
1550
1551 static inline void free_mm_slot(struct mm_slot *mm_slot)
1552 {
1553         kmem_cache_free(mm_slot_cache, mm_slot);
1554 }
1555
1556 static int __init mm_slots_hash_init(void)
1557 {
1558         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1559                                 GFP_KERNEL);
1560         if (!mm_slots_hash)
1561                 return -ENOMEM;
1562         return 0;
1563 }
1564
1565 #if 0
1566 static void __init mm_slots_hash_free(void)
1567 {
1568         kfree(mm_slots_hash);
1569         mm_slots_hash = NULL;
1570 }
1571 #endif
1572
1573 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1574 {
1575         struct mm_slot *mm_slot;
1576         struct hlist_head *bucket;
1577         struct hlist_node *node;
1578
1579         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1580                                 % MM_SLOTS_HASH_HEADS];
1581         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1582                 if (mm == mm_slot->mm)
1583                         return mm_slot;
1584         }
1585         return NULL;
1586 }
1587
1588 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1589                                     struct mm_slot *mm_slot)
1590 {
1591         struct hlist_head *bucket;
1592
1593         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1594                                 % MM_SLOTS_HASH_HEADS];
1595         mm_slot->mm = mm;
1596         hlist_add_head(&mm_slot->hash, bucket);
1597 }
1598
1599 static inline int khugepaged_test_exit(struct mm_struct *mm)
1600 {
1601         return atomic_read(&mm->mm_users) == 0;
1602 }
1603
1604 int __khugepaged_enter(struct mm_struct *mm)
1605 {
1606         struct mm_slot *mm_slot;
1607         int wakeup;
1608
1609         mm_slot = alloc_mm_slot();
1610         if (!mm_slot)
1611                 return -ENOMEM;
1612
1613         /* __khugepaged_exit() must not run from under us */
1614         VM_BUG_ON(khugepaged_test_exit(mm));
1615         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1616                 free_mm_slot(mm_slot);
1617                 return 0;
1618         }
1619
1620         spin_lock(&khugepaged_mm_lock);
1621         insert_to_mm_slots_hash(mm, mm_slot);
1622         /*
1623          * Insert just behind the scanning cursor, to let the area settle
1624          * down a little.
1625          */
1626         wakeup = list_empty(&khugepaged_scan.mm_head);
1627         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1628         spin_unlock(&khugepaged_mm_lock);
1629
1630         atomic_inc(&mm->mm_count);
1631         if (wakeup)
1632                 wake_up_interruptible(&khugepaged_wait);
1633
1634         return 0;
1635 }
1636
1637 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1638 {
1639         unsigned long hstart, hend;
1640         if (!vma->anon_vma)
1641                 /*
1642                  * Not yet faulted in so we will register later in the
1643                  * page fault if needed.
1644                  */
1645                 return 0;
1646         if (vma->vm_ops)
1647                 /* khugepaged not yet working on file or special mappings */
1648                 return 0;
1649         /*
1650          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1651          * true too, verify it here.
1652          */
1653         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1654         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1655         hend = vma->vm_end & HPAGE_PMD_MASK;
1656         if (hstart < hend)
1657                 return khugepaged_enter(vma);
1658         return 0;
1659 }
1660
1661 void __khugepaged_exit(struct mm_struct *mm)
1662 {
1663         struct mm_slot *mm_slot;
1664         int free = 0;
1665
1666         spin_lock(&khugepaged_mm_lock);
1667         mm_slot = get_mm_slot(mm);
1668         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1669                 hlist_del(&mm_slot->hash);
1670                 list_del(&mm_slot->mm_node);
1671                 free = 1;
1672         }
1673         spin_unlock(&khugepaged_mm_lock);
1674
1675         if (free) {
1676                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1677                 free_mm_slot(mm_slot);
1678                 mmdrop(mm);
1679         } else if (mm_slot) {
1680                 /*
1681                  * This is required to serialize against
1682                  * khugepaged_test_exit() (which is guaranteed to run
1683                  * under mmap sem read mode). Stop here (after we
1684                  * return all pagetables will be destroyed) until
1685                  * khugepaged has finished working on the pagetables
1686                  * under the mmap_sem.
1687                  */
1688                 down_write(&mm->mmap_sem);
1689                 up_write(&mm->mmap_sem);
1690         }
1691 }
1692
1693 static void release_pte_page(struct page *page)
1694 {
1695         /* 0 stands for page_is_file_cache(page) == false */
1696         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1697         unlock_page(page);
1698         putback_lru_page(page);
1699 }
1700
1701 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1702 {
1703         while (--_pte >= pte) {
1704                 pte_t pteval = *_pte;
1705                 if (!pte_none(pteval))
1706                         release_pte_page(pte_page(pteval));
1707         }
1708 }
1709
1710 static void release_all_pte_pages(pte_t *pte)
1711 {
1712         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1713 }
1714
1715 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1716                                         unsigned long address,
1717                                         pte_t *pte)
1718 {
1719         struct page *page;
1720         pte_t *_pte;
1721         int referenced = 0, isolated = 0, none = 0;
1722         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1723              _pte++, address += PAGE_SIZE) {
1724                 pte_t pteval = *_pte;
1725                 if (pte_none(pteval)) {
1726                         if (++none <= khugepaged_max_ptes_none)
1727                                 continue;
1728                         else {
1729                                 release_pte_pages(pte, _pte);
1730                                 goto out;
1731                         }
1732                 }
1733                 if (!pte_present(pteval) || !pte_write(pteval)) {
1734                         release_pte_pages(pte, _pte);
1735                         goto out;
1736                 }
1737                 page = vm_normal_page(vma, address, pteval);
1738                 if (unlikely(!page)) {
1739                         release_pte_pages(pte, _pte);
1740                         goto out;
1741                 }
1742                 VM_BUG_ON(PageCompound(page));
1743                 BUG_ON(!PageAnon(page));
1744                 VM_BUG_ON(!PageSwapBacked(page));
1745
1746                 /* cannot use mapcount: can't collapse if there's a gup pin */
1747                 if (page_count(page) != 1) {
1748                         release_pte_pages(pte, _pte);
1749                         goto out;
1750                 }
1751                 /*
1752                  * We can do it before isolate_lru_page because the
1753                  * page can't be freed from under us. NOTE: PG_lock
1754                  * is needed to serialize against split_huge_page
1755                  * when invoked from the VM.
1756                  */
1757                 if (!trylock_page(page)) {
1758                         release_pte_pages(pte, _pte);
1759                         goto out;
1760                 }
1761                 /*
1762                  * Isolate the page to avoid collapsing an hugepage
1763                  * currently in use by the VM.
1764                  */
1765                 if (isolate_lru_page(page)) {
1766                         unlock_page(page);
1767                         release_pte_pages(pte, _pte);
1768                         goto out;
1769                 }
1770                 /* 0 stands for page_is_file_cache(page) == false */
1771                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1772                 VM_BUG_ON(!PageLocked(page));
1773                 VM_BUG_ON(PageLRU(page));
1774
1775                 /* If there is no mapped pte young don't collapse the page */
1776                 if (pte_young(pteval) || PageReferenced(page) ||
1777                     mmu_notifier_test_young(vma->vm_mm, address))
1778                         referenced = 1;
1779         }
1780         if (unlikely(!referenced))
1781                 release_all_pte_pages(pte);
1782         else
1783                 isolated = 1;
1784 out:
1785         return isolated;
1786 }
1787
1788 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1789                                       struct vm_area_struct *vma,
1790                                       unsigned long address,
1791                                       spinlock_t *ptl)
1792 {
1793         pte_t *_pte;
1794         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1795                 pte_t pteval = *_pte;
1796                 struct page *src_page;
1797
1798                 if (pte_none(pteval)) {
1799                         clear_user_highpage(page, address);
1800                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1801                 } else {
1802                         src_page = pte_page(pteval);
1803                         copy_user_highpage(page, src_page, address, vma);
1804                         VM_BUG_ON(page_mapcount(src_page) != 1);
1805                         VM_BUG_ON(page_count(src_page) != 2);
1806                         release_pte_page(src_page);
1807                         /*
1808                          * ptl mostly unnecessary, but preempt has to
1809                          * be disabled to update the per-cpu stats
1810                          * inside page_remove_rmap().
1811                          */
1812                         spin_lock(ptl);
1813                         /*
1814                          * paravirt calls inside pte_clear here are
1815                          * superfluous.
1816                          */
1817                         pte_clear(vma->vm_mm, address, _pte);
1818                         page_remove_rmap(src_page);
1819                         spin_unlock(ptl);
1820                         free_page_and_swap_cache(src_page);
1821                 }
1822
1823                 address += PAGE_SIZE;
1824                 page++;
1825         }
1826 }
1827
1828 static void collapse_huge_page(struct mm_struct *mm,
1829                                unsigned long address,
1830                                struct page **hpage,
1831                                struct vm_area_struct *vma,
1832                                int node)
1833 {
1834         pgd_t *pgd;
1835         pud_t *pud;
1836         pmd_t *pmd, _pmd;
1837         pte_t *pte;
1838         pgtable_t pgtable;
1839         struct page *new_page;
1840         spinlock_t *ptl;
1841         int isolated;
1842         unsigned long hstart, hend;
1843
1844         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1845 #ifndef CONFIG_NUMA
1846         up_read(&mm->mmap_sem);
1847         VM_BUG_ON(!*hpage);
1848         new_page = *hpage;
1849 #else
1850         VM_BUG_ON(*hpage);
1851         /*
1852          * Allocate the page while the vma is still valid and under
1853          * the mmap_sem read mode so there is no memory allocation
1854          * later when we take the mmap_sem in write mode. This is more
1855          * friendly behavior (OTOH it may actually hide bugs) to
1856          * filesystems in userland with daemons allocating memory in
1857          * the userland I/O paths.  Allocating memory with the
1858          * mmap_sem in read mode is good idea also to allow greater
1859          * scalability.
1860          */
1861         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1862                                       node, __GFP_OTHER_NODE);
1863
1864         /*
1865          * After allocating the hugepage, release the mmap_sem read lock in
1866          * preparation for taking it in write mode.
1867          */
1868         up_read(&mm->mmap_sem);
1869         if (unlikely(!new_page)) {
1870                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1871                 *hpage = ERR_PTR(-ENOMEM);
1872                 return;
1873         }
1874 #endif
1875
1876         count_vm_event(THP_COLLAPSE_ALLOC);
1877         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1878 #ifdef CONFIG_NUMA
1879                 put_page(new_page);
1880 #endif
1881                 return;
1882         }
1883
1884         /*
1885          * Prevent all access to pagetables with the exception of
1886          * gup_fast later hanlded by the ptep_clear_flush and the VM
1887          * handled by the anon_vma lock + PG_lock.
1888          */
1889         down_write(&mm->mmap_sem);
1890         if (unlikely(khugepaged_test_exit(mm)))
1891                 goto out;
1892
1893         vma = find_vma(mm, address);
1894         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1895         hend = vma->vm_end & HPAGE_PMD_MASK;
1896         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1897                 goto out;
1898
1899         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1900             (vma->vm_flags & VM_NOHUGEPAGE))
1901                 goto out;
1902
1903         if (!vma->anon_vma || vma->vm_ops)
1904                 goto out;
1905         if (is_vma_temporary_stack(vma))
1906                 goto out;
1907         /*
1908          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1909          * true too, verify it here.
1910          */
1911         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1912
1913         pgd = pgd_offset(mm, address);
1914         if (!pgd_present(*pgd))
1915                 goto out;
1916
1917         pud = pud_offset(pgd, address);
1918         if (!pud_present(*pud))
1919                 goto out;
1920
1921         pmd = pmd_offset(pud, address);
1922         /* pmd can't go away or become huge under us */
1923         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1924                 goto out;
1925
1926         anon_vma_lock(vma->anon_vma);
1927
1928         pte = pte_offset_map(pmd, address);
1929         ptl = pte_lockptr(mm, pmd);
1930
1931         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1932         /*
1933          * After this gup_fast can't run anymore. This also removes
1934          * any huge TLB entry from the CPU so we won't allow
1935          * huge and small TLB entries for the same virtual address
1936          * to avoid the risk of CPU bugs in that area.
1937          */
1938         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1939         spin_unlock(&mm->page_table_lock);
1940
1941         spin_lock(ptl);
1942         isolated = __collapse_huge_page_isolate(vma, address, pte);
1943         spin_unlock(ptl);
1944
1945         if (unlikely(!isolated)) {
1946                 pte_unmap(pte);
1947                 spin_lock(&mm->page_table_lock);
1948                 BUG_ON(!pmd_none(*pmd));
1949                 set_pmd_at(mm, address, pmd, _pmd);
1950                 spin_unlock(&mm->page_table_lock);
1951                 anon_vma_unlock(vma->anon_vma);
1952                 goto out;
1953         }
1954
1955         /*
1956          * All pages are isolated and locked so anon_vma rmap
1957          * can't run anymore.
1958          */
1959         anon_vma_unlock(vma->anon_vma);
1960
1961         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1962         pte_unmap(pte);
1963         __SetPageUptodate(new_page);
1964         pgtable = pmd_pgtable(_pmd);
1965         VM_BUG_ON(page_count(pgtable) != 1);
1966         VM_BUG_ON(page_mapcount(pgtable) != 0);
1967
1968         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1969         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1970         _pmd = pmd_mkhuge(_pmd);
1971
1972         /*
1973          * spin_lock() below is not the equivalent of smp_wmb(), so
1974          * this is needed to avoid the copy_huge_page writes to become
1975          * visible after the set_pmd_at() write.
1976          */
1977         smp_wmb();
1978
1979         spin_lock(&mm->page_table_lock);
1980         BUG_ON(!pmd_none(*pmd));
1981         page_add_new_anon_rmap(new_page, vma, address);
1982         set_pmd_at(mm, address, pmd, _pmd);
1983         update_mmu_cache(vma, address, _pmd);
1984         prepare_pmd_huge_pte(pgtable, mm);
1985         spin_unlock(&mm->page_table_lock);
1986
1987 #ifndef CONFIG_NUMA
1988         *hpage = NULL;
1989 #endif
1990         khugepaged_pages_collapsed++;
1991 out_up_write:
1992         up_write(&mm->mmap_sem);
1993         return;
1994
1995 out:
1996         mem_cgroup_uncharge_page(new_page);
1997 #ifdef CONFIG_NUMA
1998         put_page(new_page);
1999 #endif
2000         goto out_up_write;
2001 }
2002
2003 static int khugepaged_scan_pmd(struct mm_struct *mm,
2004                                struct vm_area_struct *vma,
2005                                unsigned long address,
2006                                struct page **hpage)
2007 {
2008         pgd_t *pgd;
2009         pud_t *pud;
2010         pmd_t *pmd;
2011         pte_t *pte, *_pte;
2012         int ret = 0, referenced = 0, none = 0;
2013         struct page *page;
2014         unsigned long _address;
2015         spinlock_t *ptl;
2016         int node = -1;
2017
2018         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2019
2020         pgd = pgd_offset(mm, address);
2021         if (!pgd_present(*pgd))
2022                 goto out;
2023
2024         pud = pud_offset(pgd, address);
2025         if (!pud_present(*pud))
2026                 goto out;
2027
2028         pmd = pmd_offset(pud, address);
2029         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2030                 goto out;
2031
2032         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2033         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2034              _pte++, _address += PAGE_SIZE) {
2035                 pte_t pteval = *_pte;
2036                 if (pte_none(pteval)) {
2037                         if (++none <= khugepaged_max_ptes_none)
2038                                 continue;
2039                         else
2040                                 goto out_unmap;
2041                 }
2042                 if (!pte_present(pteval) || !pte_write(pteval))
2043                         goto out_unmap;
2044                 page = vm_normal_page(vma, _address, pteval);
2045                 if (unlikely(!page))
2046                         goto out_unmap;
2047                 /*
2048                  * Chose the node of the first page. This could
2049                  * be more sophisticated and look at more pages,
2050                  * but isn't for now.
2051                  */
2052                 if (node == -1)
2053                         node = page_to_nid(page);
2054                 VM_BUG_ON(PageCompound(page));
2055                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2056                         goto out_unmap;
2057                 /* cannot use mapcount: can't collapse if there's a gup pin */
2058                 if (page_count(page) != 1)
2059                         goto out_unmap;
2060                 if (pte_young(pteval) || PageReferenced(page) ||
2061                     mmu_notifier_test_young(vma->vm_mm, address))
2062                         referenced = 1;
2063         }
2064         if (referenced)
2065                 ret = 1;
2066 out_unmap:
2067         pte_unmap_unlock(pte, ptl);
2068         if (ret)
2069                 /* collapse_huge_page will return with the mmap_sem released */
2070                 collapse_huge_page(mm, address, hpage, vma, node);
2071 out:
2072         return ret;
2073 }
2074
2075 static void collect_mm_slot(struct mm_slot *mm_slot)
2076 {
2077         struct mm_struct *mm = mm_slot->mm;
2078
2079         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2080
2081         if (khugepaged_test_exit(mm)) {
2082                 /* free mm_slot */
2083                 hlist_del(&mm_slot->hash);
2084                 list_del(&mm_slot->mm_node);
2085
2086                 /*
2087                  * Not strictly needed because the mm exited already.
2088                  *
2089                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2090                  */
2091
2092                 /* khugepaged_mm_lock actually not necessary for the below */
2093                 free_mm_slot(mm_slot);
2094                 mmdrop(mm);
2095         }
2096 }
2097
2098 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2099                                             struct page **hpage)
2100         __releases(&khugepaged_mm_lock)
2101         __acquires(&khugepaged_mm_lock)
2102 {
2103         struct mm_slot *mm_slot;
2104         struct mm_struct *mm;
2105         struct vm_area_struct *vma;
2106         int progress = 0;
2107
2108         VM_BUG_ON(!pages);
2109         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2110
2111         if (khugepaged_scan.mm_slot)
2112                 mm_slot = khugepaged_scan.mm_slot;
2113         else {
2114                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2115                                      struct mm_slot, mm_node);
2116                 khugepaged_scan.address = 0;
2117                 khugepaged_scan.mm_slot = mm_slot;
2118         }
2119         spin_unlock(&khugepaged_mm_lock);
2120
2121         mm = mm_slot->mm;
2122         down_read(&mm->mmap_sem);
2123         if (unlikely(khugepaged_test_exit(mm)))
2124                 vma = NULL;
2125         else
2126                 vma = find_vma(mm, khugepaged_scan.address);
2127
2128         progress++;
2129         for (; vma; vma = vma->vm_next) {
2130                 unsigned long hstart, hend;
2131
2132                 cond_resched();
2133                 if (unlikely(khugepaged_test_exit(mm))) {
2134                         progress++;
2135                         break;
2136                 }
2137
2138                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2139                      !khugepaged_always()) ||
2140                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2141                 skip:
2142                         progress++;
2143                         continue;
2144                 }
2145                 if (!vma->anon_vma || vma->vm_ops)
2146                         goto skip;
2147                 if (is_vma_temporary_stack(vma))
2148                         goto skip;
2149                 /*
2150                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2151                  * must be true too, verify it here.
2152                  */
2153                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2154                           vma->vm_flags & VM_NO_THP);
2155
2156                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2157                 hend = vma->vm_end & HPAGE_PMD_MASK;
2158                 if (hstart >= hend)
2159                         goto skip;
2160                 if (khugepaged_scan.address > hend)
2161                         goto skip;
2162                 if (khugepaged_scan.address < hstart)
2163                         khugepaged_scan.address = hstart;
2164                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2165
2166                 while (khugepaged_scan.address < hend) {
2167                         int ret;
2168                         cond_resched();
2169                         if (unlikely(khugepaged_test_exit(mm)))
2170                                 goto breakouterloop;
2171
2172                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2173                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2174                                   hend);
2175                         ret = khugepaged_scan_pmd(mm, vma,
2176                                                   khugepaged_scan.address,
2177                                                   hpage);
2178                         /* move to next address */
2179                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2180                         progress += HPAGE_PMD_NR;
2181                         if (ret)
2182                                 /* we released mmap_sem so break loop */
2183                                 goto breakouterloop_mmap_sem;
2184                         if (progress >= pages)
2185                                 goto breakouterloop;
2186                 }
2187         }
2188 breakouterloop:
2189         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2190 breakouterloop_mmap_sem:
2191
2192         spin_lock(&khugepaged_mm_lock);
2193         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2194         /*
2195          * Release the current mm_slot if this mm is about to die, or
2196          * if we scanned all vmas of this mm.
2197          */
2198         if (khugepaged_test_exit(mm) || !vma) {
2199                 /*
2200                  * Make sure that if mm_users is reaching zero while
2201                  * khugepaged runs here, khugepaged_exit will find
2202                  * mm_slot not pointing to the exiting mm.
2203                  */
2204                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2205                         khugepaged_scan.mm_slot = list_entry(
2206                                 mm_slot->mm_node.next,
2207                                 struct mm_slot, mm_node);
2208                         khugepaged_scan.address = 0;
2209                 } else {
2210                         khugepaged_scan.mm_slot = NULL;
2211                         khugepaged_full_scans++;
2212                 }
2213
2214                 collect_mm_slot(mm_slot);
2215         }
2216
2217         return progress;
2218 }
2219
2220 static int khugepaged_has_work(void)
2221 {
2222         return !list_empty(&khugepaged_scan.mm_head) &&
2223                 khugepaged_enabled();
2224 }
2225
2226 static int khugepaged_wait_event(void)
2227 {
2228         return !list_empty(&khugepaged_scan.mm_head) ||
2229                 !khugepaged_enabled();
2230 }
2231
2232 static void khugepaged_do_scan(struct page **hpage)
2233 {
2234         unsigned int progress = 0, pass_through_head = 0;
2235         unsigned int pages = khugepaged_pages_to_scan;
2236
2237         barrier(); /* write khugepaged_pages_to_scan to local stack */
2238
2239         while (progress < pages) {
2240                 cond_resched();
2241
2242 #ifndef CONFIG_NUMA
2243                 if (!*hpage) {
2244                         *hpage = alloc_hugepage(khugepaged_defrag());
2245                         if (unlikely(!*hpage)) {
2246                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2247                                 break;
2248                         }
2249                         count_vm_event(THP_COLLAPSE_ALLOC);
2250                 }
2251 #else
2252                 if (IS_ERR(*hpage))
2253                         break;
2254 #endif
2255
2256                 if (unlikely(kthread_should_stop() || freezing(current)))
2257                         break;
2258
2259                 spin_lock(&khugepaged_mm_lock);
2260                 if (!khugepaged_scan.mm_slot)
2261                         pass_through_head++;
2262                 if (khugepaged_has_work() &&
2263                     pass_through_head < 2)
2264                         progress += khugepaged_scan_mm_slot(pages - progress,
2265                                                             hpage);
2266                 else
2267                         progress = pages;
2268                 spin_unlock(&khugepaged_mm_lock);
2269         }
2270 }
2271
2272 static void khugepaged_alloc_sleep(void)
2273 {
2274         wait_event_freezable_timeout(khugepaged_wait, false,
2275                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2276 }
2277
2278 #ifndef CONFIG_NUMA
2279 static struct page *khugepaged_alloc_hugepage(void)
2280 {
2281         struct page *hpage;
2282
2283         do {
2284                 hpage = alloc_hugepage(khugepaged_defrag());
2285                 if (!hpage) {
2286                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2287                         khugepaged_alloc_sleep();
2288                 } else
2289                         count_vm_event(THP_COLLAPSE_ALLOC);
2290         } while (unlikely(!hpage) &&
2291                  likely(khugepaged_enabled()));
2292         return hpage;
2293 }
2294 #endif
2295
2296 static void khugepaged_loop(void)
2297 {
2298         struct page *hpage;
2299
2300 #ifdef CONFIG_NUMA
2301         hpage = NULL;
2302 #endif
2303         while (likely(khugepaged_enabled())) {
2304 #ifndef CONFIG_NUMA
2305                 hpage = khugepaged_alloc_hugepage();
2306                 if (unlikely(!hpage))
2307                         break;
2308 #else
2309                 if (IS_ERR(hpage)) {
2310                         khugepaged_alloc_sleep();
2311                         hpage = NULL;
2312                 }
2313 #endif
2314
2315                 khugepaged_do_scan(&hpage);
2316 #ifndef CONFIG_NUMA
2317                 if (hpage)
2318                         put_page(hpage);
2319 #endif
2320                 try_to_freeze();
2321                 if (unlikely(kthread_should_stop()))
2322                         break;
2323                 if (khugepaged_has_work()) {
2324                         if (!khugepaged_scan_sleep_millisecs)
2325                                 continue;
2326                         wait_event_freezable_timeout(khugepaged_wait, false,
2327                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2328                 } else if (khugepaged_enabled())
2329                         wait_event_freezable(khugepaged_wait,
2330                                              khugepaged_wait_event());
2331         }
2332 }
2333
2334 static int khugepaged(void *none)
2335 {
2336         struct mm_slot *mm_slot;
2337
2338         set_freezable();
2339         set_user_nice(current, 19);
2340
2341         /* serialize with start_khugepaged() */
2342         mutex_lock(&khugepaged_mutex);
2343
2344         for (;;) {
2345                 mutex_unlock(&khugepaged_mutex);
2346                 VM_BUG_ON(khugepaged_thread != current);
2347                 khugepaged_loop();
2348                 VM_BUG_ON(khugepaged_thread != current);
2349
2350                 mutex_lock(&khugepaged_mutex);
2351                 if (!khugepaged_enabled())
2352                         break;
2353                 if (unlikely(kthread_should_stop()))
2354                         break;
2355         }
2356
2357         spin_lock(&khugepaged_mm_lock);
2358         mm_slot = khugepaged_scan.mm_slot;
2359         khugepaged_scan.mm_slot = NULL;
2360         if (mm_slot)
2361                 collect_mm_slot(mm_slot);
2362         spin_unlock(&khugepaged_mm_lock);
2363
2364         khugepaged_thread = NULL;
2365         mutex_unlock(&khugepaged_mutex);
2366
2367         return 0;
2368 }
2369
2370 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2371 {
2372         struct page *page;
2373
2374         spin_lock(&mm->page_table_lock);
2375         if (unlikely(!pmd_trans_huge(*pmd))) {
2376                 spin_unlock(&mm->page_table_lock);
2377                 return;
2378         }
2379         page = pmd_page(*pmd);
2380         VM_BUG_ON(!page_count(page));
2381         get_page(page);
2382         spin_unlock(&mm->page_table_lock);
2383
2384         split_huge_page(page);
2385
2386         put_page(page);
2387         BUG_ON(pmd_trans_huge(*pmd));
2388 }
2389
2390 static void split_huge_page_address(struct mm_struct *mm,
2391                                     unsigned long address)
2392 {
2393         pgd_t *pgd;
2394         pud_t *pud;
2395         pmd_t *pmd;
2396
2397         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2398
2399         pgd = pgd_offset(mm, address);
2400         if (!pgd_present(*pgd))
2401                 return;
2402
2403         pud = pud_offset(pgd, address);
2404         if (!pud_present(*pud))
2405                 return;
2406
2407         pmd = pmd_offset(pud, address);
2408         if (!pmd_present(*pmd))
2409                 return;
2410         /*
2411          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2412          * materialize from under us.
2413          */
2414         split_huge_page_pmd(mm, pmd);
2415 }
2416
2417 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2418                              unsigned long start,
2419                              unsigned long end,
2420                              long adjust_next)
2421 {
2422         /*
2423          * If the new start address isn't hpage aligned and it could
2424          * previously contain an hugepage: check if we need to split
2425          * an huge pmd.
2426          */
2427         if (start & ~HPAGE_PMD_MASK &&
2428             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2429             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2430                 split_huge_page_address(vma->vm_mm, start);
2431
2432         /*
2433          * If the new end address isn't hpage aligned and it could
2434          * previously contain an hugepage: check if we need to split
2435          * an huge pmd.
2436          */
2437         if (end & ~HPAGE_PMD_MASK &&
2438             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2439             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2440                 split_huge_page_address(vma->vm_mm, end);
2441
2442         /*
2443          * If we're also updating the vma->vm_next->vm_start, if the new
2444          * vm_next->vm_start isn't page aligned and it could previously
2445          * contain an hugepage: check if we need to split an huge pmd.
2446          */
2447         if (adjust_next > 0) {
2448                 struct vm_area_struct *next = vma->vm_next;
2449                 unsigned long nstart = next->vm_start;
2450                 nstart += adjust_next << PAGE_SHIFT;
2451                 if (nstart & ~HPAGE_PMD_MASK &&
2452                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2453                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2454                         split_huge_page_address(next->vm_mm, nstart);
2455         }
2456 }