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                 spin_unlock(&mm->page_table_lock);
 675         }
 676
 677         return ret;
 678 }
 679
 680 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 681 {
 682         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 683 }
 684
 685 static inline struct page *alloc_hugepage_vma(int defrag,
 686                                               struct vm_area_struct *vma,
 687                                               unsigned long haddr, int nd,
 688                                               gfp_t extra_gfp)
 689 {
 690         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 691                                HPAGE_PMD_ORDER, vma, haddr, nd);
 692 }
 693
 694 #ifndef CONFIG_NUMA
 695 static inline struct page *alloc_hugepage(int defrag)
 696 {
 697         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 698                            HPAGE_PMD_ORDER);
 699 }
 700 #endif
 701
 702 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 703                                unsigned long address, pmd_t *pmd,
 704                                unsigned int flags)
 705 {
 706         struct page *page;
 707         unsigned long haddr = address & HPAGE_PMD_MASK;
 708         pte_t *pte;
 709
 710         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 711                 if (unlikely(anon_vma_prepare(vma)))
 712                         return VM_FAULT_OOM;
 713                 if (unlikely(khugepaged_enter(vma)))
 714                         return VM_FAULT_OOM;
 715                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 716                                           vma, haddr, numa_node_id(), 0);
 717                 if (unlikely(!page)) {
 718                         count_vm_event(THP_FAULT_FALLBACK);
 719                         goto out;
 720                 }
 721                 count_vm_event(THP_FAULT_ALLOC);
 722                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 723                         put_page(page);
 724                         goto out;
 725                 }
 726
 727                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
 728         }
 729 out:
 730         /*
 731          * Use __pte_alloc instead of pte_alloc_map, because we can't
 732          * run pte_offset_map on the pmd, if an huge pmd could
 733          * materialize from under us from a different thread.
 734          */
 735         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 736                 return VM_FAULT_OOM;
 737         /* if an huge pmd materialized from under us just retry later */
 738         if (unlikely(pmd_trans_huge(*pmd)))
 739                 return 0;
 740         /*
 741          * A regular pmd is established and it can't morph into a huge pmd
 742          * from under us anymore at this point because we hold the mmap_sem
 743          * read mode and khugepaged takes it in write mode. So now it's
 744          * safe to run pte_offset_map().
 745          */
 746         pte = pte_offset_map(pmd, address);
 747         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 748 }
 749
 750 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 751                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 752                   struct vm_area_struct *vma)
 753 {
 754         struct page *src_page;
 755         pmd_t pmd;
 756         pgtable_t pgtable;
 757         int ret;
 758
 759         ret = -ENOMEM;
 760         pgtable = pte_alloc_one(dst_mm, addr);
 761         if (unlikely(!pgtable))
 762                 goto out;
 763
 764         spin_lock(&dst_mm->page_table_lock);
 765         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 766
 767         ret = -EAGAIN;
 768         pmd = *src_pmd;
 769         if (unlikely(!pmd_trans_huge(pmd))) {
 770                 pte_free(dst_mm, pgtable);
 771                 goto out_unlock;
 772         }
 773         if (unlikely(pmd_trans_splitting(pmd))) {
 774                 /* split huge page running from under us */
 775                 spin_unlock(&src_mm->page_table_lock);
 776                 spin_unlock(&dst_mm->page_table_lock);
 777                 pte_free(dst_mm, pgtable);
 778
 779                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 780                 goto out;
 781         }
 782         src_page = pmd_page(pmd);
 783         VM_BUG_ON(!PageHead(src_page));
 784         get_page(src_page);
 785         page_dup_rmap(src_page);
 786         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 787
 788         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 789         pmd = pmd_mkold(pmd_wrprotect(pmd));
 790         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 791         prepare_pmd_huge_pte(pgtable, dst_mm);
 792
 793         ret = 0;
 794 out_unlock:
 795         spin_unlock(&src_mm->page_table_lock);
 796         spin_unlock(&dst_mm->page_table_lock);
 797 out:
 798         return ret;
 799 }
 800
 801 /* no "address" argument so destroys page coloring of some arch */
 802 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 803 {
 804         pgtable_t pgtable;
 805
 806         assert_spin_locked(&mm->page_table_lock);
 807
 808         /* FIFO */
 809         pgtable = mm->pmd_huge_pte;
 810         if (list_empty(&pgtable->lru))
 811                 mm->pmd_huge_pte = NULL;
 812         else {
 813                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 814                                               struct page, lru);
 815                 list_del(&pgtable->lru);
 816         }
 817         return pgtable;
 818 }
 819
 820 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 821                                         struct vm_area_struct *vma,
 822                                         unsigned long address,
 823                                         pmd_t *pmd, pmd_t orig_pmd,
 824                                         struct page *page,
 825                                         unsigned long haddr)
 826 {
 827         pgtable_t pgtable;
 828         pmd_t _pmd;
 829         int ret = 0, i;
 830         struct page **pages;
 831
 832         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 833                         GFP_KERNEL);
 834         if (unlikely(!pages)) {
 835                 ret |= VM_FAULT_OOM;
 836                 goto out;
 837         }
 838
 839         for (i = 0; i < HPAGE_PMD_NR; i++) {
 840                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 841                                                __GFP_OTHER_NODE,
 842                                                vma, address, page_to_nid(page));
 843                 if (unlikely(!pages[i] ||
 844                              mem_cgroup_newpage_charge(pages[i], mm,
 845                                                        GFP_KERNEL))) {
 846                         if (pages[i])
 847                                 put_page(pages[i]);
 848                         mem_cgroup_uncharge_start();
 849                         while (--i >= 0) {
 850                                 mem_cgroup_uncharge_page(pages[i]);
 851                                 put_page(pages[i]);
 852                         }
 853                         mem_cgroup_uncharge_end();
 854                         kfree(pages);
 855                         ret |= VM_FAULT_OOM;
 856                         goto out;
 857                 }
 858         }
 859
 860         for (i = 0; i < HPAGE_PMD_NR; i++) {
 861                 copy_user_highpage(pages[i], page + i,
 862                                    haddr + PAGE_SIZE * i, vma);
 863                 __SetPageUptodate(pages[i]);
 864                 cond_resched();
 865         }
 866
 867         spin_lock(&mm->page_table_lock);
 868         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 869                 goto out_free_pages;
 870         VM_BUG_ON(!PageHead(page));
 871
 872         pmdp_clear_flush_notify(vma, haddr, pmd);
 873         /* leave pmd empty until pte is filled */
 874
 875         pgtable = get_pmd_huge_pte(mm);
 876         pmd_populate(mm, &_pmd, pgtable);
 877
 878         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 879                 pte_t *pte, entry;
 880                 entry = mk_pte(pages[i], vma->vm_page_prot);
 881                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 882                 page_add_new_anon_rmap(pages[i], vma, haddr);
 883                 pte = pte_offset_map(&_pmd, haddr);
 884                 VM_BUG_ON(!pte_none(*pte));
 885                 set_pte_at(mm, haddr, pte, entry);
 886                 pte_unmap(pte);
 887         }
 888         kfree(pages);
 889
 890         mm->nr_ptes++;
 891         smp_wmb(); /* make pte visible before pmd */
 892         pmd_populate(mm, pmd, pgtable);
 893         page_remove_rmap(page);
 894         spin_unlock(&mm->page_table_lock);
 895
 896         ret |= VM_FAULT_WRITE;
 897         put_page(page);
 898
 899 out:
 900         return ret;
 901
 902 out_free_pages:
 903         spin_unlock(&mm->page_table_lock);
 904         mem_cgroup_uncharge_start();
 905         for (i = 0; i < HPAGE_PMD_NR; i++) {
 906                 mem_cgroup_uncharge_page(pages[i]);
 907                 put_page(pages[i]);
 908         }
 909         mem_cgroup_uncharge_end();
 910         kfree(pages);
 911         goto out;
 912 }
 913
 914 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 915                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 916 {
 917         int ret = 0;
 918         struct page *page, *new_page;
 919         unsigned long haddr;
 920
 921         VM_BUG_ON(!vma->anon_vma);
 922         spin_lock(&mm->page_table_lock);
 923         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 924                 goto out_unlock;
 925
 926         page = pmd_page(orig_pmd);
 927         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 928         haddr = address & HPAGE_PMD_MASK;
 929         if (page_mapcount(page) == 1) {
 930                 pmd_t entry;
 931                 entry = pmd_mkyoung(orig_pmd);
 932                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 933                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 934                         update_mmu_cache(vma, address, entry);
 935                 ret |= VM_FAULT_WRITE;
 936                 goto out_unlock;
 937         }
 938         get_page(page);
 939         spin_unlock(&mm->page_table_lock);
 940
 941         if (transparent_hugepage_enabled(vma) &&
 942             !transparent_hugepage_debug_cow())
 943                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 944                                               vma, haddr, numa_node_id(), 0);
 945         else
 946                 new_page = NULL;
 947
 948         if (unlikely(!new_page)) {
 949                 count_vm_event(THP_FAULT_FALLBACK);
 950                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 951                                                    pmd, orig_pmd, page, haddr);
 952                 put_page(page);
 953                 goto out;
 954         }
 955         count_vm_event(THP_FAULT_ALLOC);
 956
 957         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 958                 put_page(new_page);
 959                 put_page(page);
 960                 ret |= VM_FAULT_OOM;
 961                 goto out;
 962         }
 963
 964         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 965         __SetPageUptodate(new_page);
 966
 967         spin_lock(&mm->page_table_lock);
 968         put_page(page);
 969         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 970                 mem_cgroup_uncharge_page(new_page);
 971                 put_page(new_page);
 972         } else {
 973                 pmd_t entry;
 974                 VM_BUG_ON(!PageHead(page));
 975                 entry = mk_pmd(new_page, vma->vm_page_prot);
 976                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 977                 entry = pmd_mkhuge(entry);
 978                 pmdp_clear_flush_notify(vma, haddr, pmd);
 979                 page_add_new_anon_rmap(new_page, vma, haddr);
 980                 set_pmd_at(mm, haddr, pmd, entry);
 981                 update_mmu_cache(vma, address, entry);
 982                 page_remove_rmap(page);
 983                 put_page(page);
 984                 ret |= VM_FAULT_WRITE;
 985         }
 986 out_unlock:
 987         spin_unlock(&mm->page_table_lock);
 988 out:
 989         return ret;
 990 }
 991
 992 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 993                                    unsigned long addr,
 994                                    pmd_t *pmd,
 995                                    unsigned int flags)
 996 {
 997         struct page *page = NULL;
 998
 999         assert_spin_locked(&mm->page_table_lock);
1000
1001         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1002                 goto out;
1003
1004         page = pmd_page(*pmd);
1005         VM_BUG_ON(!PageHead(page));
1006         if (flags & FOLL_TOUCH) {
1007                 pmd_t _pmd;
1008                 /*
1009                  * We should set the dirty bit only for FOLL_WRITE but
1010                  * for now the dirty bit in the pmd is meaningless.
1011                  * And if the dirty bit will become meaningful and
1012                  * we'll only set it with FOLL_WRITE, an atomic
1013                  * set_bit will be required on the pmd to set the
1014                  * young bit, instead of the current set_pmd_at.
1015                  */
1016                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1017                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1018         }
1019         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1020         VM_BUG_ON(!PageCompound(page));
1021         if (flags & FOLL_GET)
1022                 get_page_foll(page);
1023
1024 out:
1025         return page;
1026 }
1027
1028 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1029                  pmd_t *pmd, unsigned long addr)
1030 {
1031         int ret = 0;
1032
1033         spin_lock(&tlb->mm->page_table_lock);
1034         if (likely(pmd_trans_huge(*pmd))) {
1035                 if (unlikely(pmd_trans_splitting(*pmd))) {
1036                         spin_unlock(&tlb->mm->page_table_lock);
1037                         wait_split_huge_page(vma->anon_vma,
1038                                              pmd);
1039                 } else {
1040                         struct page *page;
1041                         pgtable_t pgtable;
1042                         pgtable = get_pmd_huge_pte(tlb->mm);
1043                         page = pmd_page(*pmd);
1044                         pmd_clear(pmd);
1045                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1046                         page_remove_rmap(page);
1047                         VM_BUG_ON(page_mapcount(page) < 0);
1048                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1049                         VM_BUG_ON(!PageHead(page));
1050                         spin_unlock(&tlb->mm->page_table_lock);
1051                         tlb_remove_page(tlb, page);
1052                         pte_free(tlb->mm, pgtable);
1053                         ret = 1;
1054                 }
1055         } else
1056                 spin_unlock(&tlb->mm->page_table_lock);
1057
1058         return ret;
1059 }
1060
1061 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1062                 unsigned long addr, unsigned long end,
1063                 unsigned char *vec)
1064 {
1065         int ret = 0;
1066
1067         spin_lock(&vma->vm_mm->page_table_lock);
1068         if (likely(pmd_trans_huge(*pmd))) {
1069                 ret = !pmd_trans_splitting(*pmd);
1070                 spin_unlock(&vma->vm_mm->page_table_lock);
1071                 if (unlikely(!ret))
1072                         wait_split_huge_page(vma->anon_vma, pmd);
1073                 else {
1074                         /*
1075                          * All logical pages in the range are present
1076                          * if backed by a huge page.
1077                          */
1078                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1079                 }
1080         } else
1081                 spin_unlock(&vma->vm_mm->page_table_lock);
1082
1083         return ret;
1084 }
1085
1086 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1087                   unsigned long old_addr,
1088                   unsigned long new_addr, unsigned long old_end,
1089                   pmd_t *old_pmd, pmd_t *new_pmd)
1090 {
1091         int ret = 0;
1092         pmd_t pmd;
1093
1094         struct mm_struct *mm = vma->vm_mm;
1095
1096         if ((old_addr & ~HPAGE_PMD_MASK) ||
1097             (new_addr & ~HPAGE_PMD_MASK) ||
1098             old_end - old_addr < HPAGE_PMD_SIZE ||
1099             (new_vma->vm_flags & VM_NOHUGEPAGE))
1100                 goto out;
1101
1102         /*
1103          * The destination pmd shouldn't be established, free_pgtables()
1104          * should have release it.
1105          */
1106         if (WARN_ON(!pmd_none(*new_pmd))) {
1107                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1108                 goto out;
1109         }
1110
1111         spin_lock(&mm->page_table_lock);
1112         if (likely(pmd_trans_huge(*old_pmd))) {
1113                 if (pmd_trans_splitting(*old_pmd)) {
1114                         spin_unlock(&mm->page_table_lock);
1115                         wait_split_huge_page(vma->anon_vma, old_pmd);
1116                         ret = -1;
1117                 } else {
1118                         pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1119                         VM_BUG_ON(!pmd_none(*new_pmd));
1120                         set_pmd_at(mm, new_addr, new_pmd, pmd);
1121                         spin_unlock(&mm->page_table_lock);
1122                         ret = 1;
1123                 }
1124         } else {
1125                 spin_unlock(&mm->page_table_lock);
1126         }
1127 out:
1128         return ret;
1129 }
1130
1131 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1132                 unsigned long addr, pgprot_t newprot)
1133 {
1134         struct mm_struct *mm = vma->vm_mm;
1135         int ret = 0;
1136
1137         spin_lock(&mm->page_table_lock);
1138         if (likely(pmd_trans_huge(*pmd))) {
1139                 if (unlikely(pmd_trans_splitting(*pmd))) {
1140                         spin_unlock(&mm->page_table_lock);
1141                         wait_split_huge_page(vma->anon_vma, pmd);
1142                 } else {
1143                         pmd_t entry;
1144
1145                         entry = pmdp_get_and_clear(mm, addr, pmd);
1146                         entry = pmd_modify(entry, newprot);
1147                         set_pmd_at(mm, addr, pmd, entry);
1148                         spin_unlock(&vma->vm_mm->page_table_lock);
1149                         ret = 1;
1150                 }
1151         } else
1152                 spin_unlock(&vma->vm_mm->page_table_lock);
1153
1154         return ret;
1155 }
1156
1157 pmd_t *page_check_address_pmd(struct page *page,
1158                               struct mm_struct *mm,
1159                               unsigned long address,
1160                               enum page_check_address_pmd_flag flag)
1161 {
1162         pgd_t *pgd;
1163         pud_t *pud;
1164         pmd_t *pmd, *ret = NULL;
1165
1166         if (address & ~HPAGE_PMD_MASK)
1167                 goto out;
1168
1169         pgd = pgd_offset(mm, address);
1170         if (!pgd_present(*pgd))
1171                 goto out;
1172
1173         pud = pud_offset(pgd, address);
1174         if (!pud_present(*pud))
1175                 goto out;
1176
1177         pmd = pmd_offset(pud, address);
1178         if (pmd_none(*pmd))
1179                 goto out;
1180         if (pmd_page(*pmd) != page)
1181                 goto out;
1182         /*
1183          * split_vma() may create temporary aliased mappings. There is
1184          * no risk as long as all huge pmd are found and have their
1185          * splitting bit set before __split_huge_page_refcount
1186          * runs. Finding the same huge pmd more than once during the
1187          * same rmap walk is not a problem.
1188          */
1189         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1190             pmd_trans_splitting(*pmd))
1191                 goto out;
1192         if (pmd_trans_huge(*pmd)) {
1193                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1194                           !pmd_trans_splitting(*pmd));
1195                 ret = pmd;
1196         }
1197 out:
1198         return ret;
1199 }
1200
1201 static int __split_huge_page_splitting(struct page *page,
1202                                        struct vm_area_struct *vma,
1203                                        unsigned long address)
1204 {
1205         struct mm_struct *mm = vma->vm_mm;
1206         pmd_t *pmd;
1207         int ret = 0;
1208
1209         spin_lock(&mm->page_table_lock);
1210         pmd = page_check_address_pmd(page, mm, address,
1211                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1212         if (pmd) {
1213                 /*
1214                  * We can't temporarily set the pmd to null in order
1215                  * to split it, the pmd must remain marked huge at all
1216                  * times or the VM won't take the pmd_trans_huge paths
1217                  * and it won't wait on the anon_vma->root->mutex to
1218                  * serialize against split_huge_page*.
1219                  */
1220                 pmdp_splitting_flush_notify(vma, address, pmd);
1221                 ret = 1;
1222         }
1223         spin_unlock(&mm->page_table_lock);
1224
1225         return ret;
1226 }
1227
1228 static void __split_huge_page_refcount(struct page *page)
1229 {
1230         int i;
1231         struct zone *zone = page_zone(page);
1232         int zonestat;
1233         int tail_count = 0;
1234
1235         /* prevent PageLRU to go away from under us, and freeze lru stats */
1236         spin_lock_irq(&zone->lru_lock);
1237         compound_lock(page);
1238         /* complete memcg works before add pages to LRU */
1239         mem_cgroup_split_huge_fixup(page);
1240
1241         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1242                 struct page *page_tail = page + i;
1243
1244                 /* tail_page->_mapcount cannot change */
1245                 BUG_ON(page_mapcount(page_tail) < 0);
1246                 tail_count += page_mapcount(page_tail);
1247                 /* check for overflow */
1248                 BUG_ON(tail_count < 0);
1249                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1250                 /*
1251                  * tail_page->_count is zero and not changing from
1252                  * under us. But get_page_unless_zero() may be running
1253                  * from under us on the tail_page. If we used
1254                  * atomic_set() below instead of atomic_add(), we
1255                  * would then run atomic_set() concurrently with
1256                  * get_page_unless_zero(), and atomic_set() is
1257                  * implemented in C not using locked ops. spin_unlock
1258                  * on x86 sometime uses locked ops because of PPro
1259                  * errata 66, 92, so unless somebody can guarantee
1260                  * atomic_set() here would be safe on all archs (and
1261                  * not only on x86), it's safer to use atomic_add().
1262                  */
1263                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1264                            &page_tail->_count);
1265
1266                 /* after clearing PageTail the gup refcount can be released */
1267                 smp_mb();
1268
1269                 /*
1270                  * retain hwpoison flag of the poisoned tail page:
1271                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1272                  *   by the memory-failure.
1273                  */
1274                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1275                 page_tail->flags |= (page->flags &
1276                                      ((1L << PG_referenced) |
1277                                       (1L << PG_swapbacked) |
1278                                       (1L << PG_mlocked) |
1279                                       (1L << PG_uptodate)));
1280                 page_tail->flags |= (1L << PG_dirty);
1281
1282                 /* clear PageTail before overwriting first_page */
1283                 smp_wmb();
1284
1285                 /*
1286                  * __split_huge_page_splitting() already set the
1287                  * splitting bit in all pmd that could map this
1288                  * hugepage, that will ensure no CPU can alter the
1289                  * mapcount on the head page. The mapcount is only
1290                  * accounted in the head page and it has to be
1291                  * transferred to all tail pages in the below code. So
1292                  * for this code to be safe, the split the mapcount
1293                  * can't change. But that doesn't mean userland can't
1294                  * keep changing and reading the page contents while
1295                  * we transfer the mapcount, so the pmd splitting
1296                  * status is achieved setting a reserved bit in the
1297                  * pmd, not by clearing the present bit.
1298                 */
1299                 page_tail->_mapcount = page->_mapcount;
1300
1301                 BUG_ON(page_tail->mapping);
1302                 page_tail->mapping = page->mapping;
1303
1304                 page_tail->index = page->index + i;
1305
1306                 BUG_ON(!PageAnon(page_tail));
1307                 BUG_ON(!PageUptodate(page_tail));
1308                 BUG_ON(!PageDirty(page_tail));
1309                 BUG_ON(!PageSwapBacked(page_tail));
1310
1311
1312                 lru_add_page_tail(zone, page, page_tail);
1313         }
1314         atomic_sub(tail_count, &page->_count);
1315         BUG_ON(atomic_read(&page->_count) <= 0);
1316
1317         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1318         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1319
1320         /*
1321          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1322          * so adjust those appropriately if this page is on the LRU.
1323          */
1324         if (PageLRU(page)) {
1325                 zonestat = NR_LRU_BASE + page_lru(page);
1326                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1327         }
1328
1329         ClearPageCompound(page);
1330         compound_unlock(page);
1331         spin_unlock_irq(&zone->lru_lock);
1332
1333         for (i = 1; i < HPAGE_PMD_NR; i++) {
1334                 struct page *page_tail = page + i;
1335                 BUG_ON(page_count(page_tail) <= 0);
1336                 /*
1337                  * Tail pages may be freed if there wasn't any mapping
1338                  * like if add_to_swap() is running on a lru page that
1339                  * had its mapping zapped. And freeing these pages
1340                  * requires taking the lru_lock so we do the put_page
1341                  * of the tail pages after the split is complete.
1342                  */
1343                 put_page(page_tail);
1344         }
1345
1346         /*
1347          * Only the head page (now become a regular page) is required
1348          * to be pinned by the caller.
1349          */
1350         BUG_ON(page_count(page) <= 0);
1351 }
1352
1353 static int __split_huge_page_map(struct page *page,
1354                                  struct vm_area_struct *vma,
1355                                  unsigned long address)
1356 {
1357         struct mm_struct *mm = vma->vm_mm;
1358         pmd_t *pmd, _pmd;
1359         int ret = 0, i;
1360         pgtable_t pgtable;
1361         unsigned long haddr;
1362
1363         spin_lock(&mm->page_table_lock);
1364         pmd = page_check_address_pmd(page, mm, address,
1365                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1366         if (pmd) {
1367                 pgtable = get_pmd_huge_pte(mm);
1368                 pmd_populate(mm, &_pmd, pgtable);
1369
1370                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1371                      i++, haddr += PAGE_SIZE) {
1372                         pte_t *pte, entry;
1373                         BUG_ON(PageCompound(page+i));
1374                         entry = mk_pte(page + i, vma->vm_page_prot);
1375                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1376                         if (!pmd_write(*pmd))
1377                                 entry = pte_wrprotect(entry);
1378                         else
1379                                 BUG_ON(page_mapcount(page) != 1);
1380                         if (!pmd_young(*pmd))
1381                                 entry = pte_mkold(entry);
1382                         pte = pte_offset_map(&_pmd, haddr);
1383                         BUG_ON(!pte_none(*pte));
1384                         set_pte_at(mm, haddr, pte, entry);
1385                         pte_unmap(pte);
1386                 }
1387
1388                 mm->nr_ptes++;
1389                 smp_wmb(); /* make pte visible before pmd */
1390                 /*
1391                  * Up to this point the pmd is present and huge and
1392                  * userland has the whole access to the hugepage
1393                  * during the split (which happens in place). If we
1394                  * overwrite the pmd with the not-huge version
1395                  * pointing to the pte here (which of course we could
1396                  * if all CPUs were bug free), userland could trigger
1397                  * a small page size TLB miss on the small sized TLB
1398                  * while the hugepage TLB entry is still established
1399                  * in the huge TLB. Some CPU doesn't like that. See
1400                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1401                  * Erratum 383 on page 93. Intel should be safe but is
1402                  * also warns that it's only safe if the permission
1403                  * and cache attributes of the two entries loaded in
1404                  * the two TLB is identical (which should be the case
1405                  * here). But it is generally safer to never allow
1406                  * small and huge TLB entries for the same virtual
1407                  * address to be loaded simultaneously. So instead of
1408                  * doing "pmd_populate(); flush_tlb_range();" we first
1409                  * mark the current pmd notpresent (atomically because
1410                  * here the pmd_trans_huge and pmd_trans_splitting
1411                  * must remain set at all times on the pmd until the
1412                  * split is complete for this pmd), then we flush the
1413                  * SMP TLB and finally we write the non-huge version
1414                  * of the pmd entry with pmd_populate.
1415                  */
1416                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1417                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1418                 pmd_populate(mm, pmd, pgtable);
1419                 ret = 1;
1420         }
1421         spin_unlock(&mm->page_table_lock);
1422
1423         return ret;
1424 }
1425
1426 /* must be called with anon_vma->root->mutex hold */
1427 static void __split_huge_page(struct page *page,
1428                               struct anon_vma *anon_vma)
1429 {
1430         int mapcount, mapcount2;
1431         struct anon_vma_chain *avc;
1432
1433         BUG_ON(!PageHead(page));
1434         BUG_ON(PageTail(page));
1435
1436         mapcount = 0;
1437         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1438                 struct vm_area_struct *vma = avc->vma;
1439                 unsigned long addr = vma_address(page, vma);
1440                 BUG_ON(is_vma_temporary_stack(vma));
1441                 if (addr == -EFAULT)
1442                         continue;
1443                 mapcount += __split_huge_page_splitting(page, vma, addr);
1444         }
1445         /*
1446          * It is critical that new vmas are added to the tail of the
1447          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1448          * and establishes a child pmd before
1449          * __split_huge_page_splitting() freezes the parent pmd (so if
1450          * we fail to prevent copy_huge_pmd() from running until the
1451          * whole __split_huge_page() is complete), we will still see
1452          * the newly established pmd of the child later during the
1453          * walk, to be able to set it as pmd_trans_splitting too.
1454          */
1455         if (mapcount != page_mapcount(page))
1456                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1457                        mapcount, page_mapcount(page));
1458         BUG_ON(mapcount != page_mapcount(page));
1459
1460         __split_huge_page_refcount(page);
1461
1462         mapcount2 = 0;
1463         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1464                 struct vm_area_struct *vma = avc->vma;
1465                 unsigned long addr = vma_address(page, vma);
1466                 BUG_ON(is_vma_temporary_stack(vma));
1467                 if (addr == -EFAULT)
1468                         continue;
1469                 mapcount2 += __split_huge_page_map(page, vma, addr);
1470         }
1471         if (mapcount != mapcount2)
1472                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1473                        mapcount, mapcount2, page_mapcount(page));
1474         BUG_ON(mapcount != mapcount2);
1475 }
1476
1477 int split_huge_page(struct page *page)
1478 {
1479         struct anon_vma *anon_vma;
1480         int ret = 1;
1481
1482         BUG_ON(!PageAnon(page));
1483         anon_vma = page_lock_anon_vma(page);
1484         if (!anon_vma)
1485                 goto out;
1486         ret = 0;
1487         if (!PageCompound(page))
1488                 goto out_unlock;
1489
1490         BUG_ON(!PageSwapBacked(page));
1491         __split_huge_page(page, anon_vma);
1492         count_vm_event(THP_SPLIT);
1493
1494         BUG_ON(PageCompound(page));
1495 out_unlock:
1496         page_unlock_anon_vma(anon_vma);
1497 out:
1498         return ret;
1499 }
1500
1501 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1502                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1503
1504 int hugepage_madvise(struct vm_area_struct *vma,
1505                      unsigned long *vm_flags, int advice)
1506 {
1507         switch (advice) {
1508         case MADV_HUGEPAGE:
1509                 /*
1510                  * Be somewhat over-protective like KSM for now!
1511                  */
1512                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1513                         return -EINVAL;
1514                 *vm_flags &= ~VM_NOHUGEPAGE;
1515                 *vm_flags |= VM_HUGEPAGE;
1516                 /*
1517                  * If the vma become good for khugepaged to scan,
1518                  * register it here without waiting a page fault that
1519                  * may not happen any time soon.
1520                  */
1521                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1522                         return -ENOMEM;
1523                 break;
1524         case MADV_NOHUGEPAGE:
1525                 /*
1526                  * Be somewhat over-protective like KSM for now!
1527                  */
1528                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1529                         return -EINVAL;
1530                 *vm_flags &= ~VM_HUGEPAGE;
1531                 *vm_flags |= VM_NOHUGEPAGE;
1532                 /*
1533                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1534                  * this vma even if we leave the mm registered in khugepaged if
1535                  * it got registered before VM_NOHUGEPAGE was set.
1536                  */
1537                 break;
1538         }
1539
1540         return 0;
1541 }
1542
1543 static int __init khugepaged_slab_init(void)
1544 {
1545         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1546                                           sizeof(struct mm_slot),
1547                                           __alignof__(struct mm_slot), 0, NULL);
1548         if (!mm_slot_cache)
1549                 return -ENOMEM;
1550
1551         return 0;
1552 }
1553
1554 static void __init khugepaged_slab_free(void)
1555 {
1556         kmem_cache_destroy(mm_slot_cache);
1557         mm_slot_cache = NULL;
1558 }
1559
1560 static inline struct mm_slot *alloc_mm_slot(void)
1561 {
1562         if (!mm_slot_cache)     /* initialization failed */
1563                 return NULL;
1564         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1565 }
1566
1567 static inline void free_mm_slot(struct mm_slot *mm_slot)
1568 {
1569         kmem_cache_free(mm_slot_cache, mm_slot);
1570 }
1571
1572 static int __init mm_slots_hash_init(void)
1573 {
1574         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1575                                 GFP_KERNEL);
1576         if (!mm_slots_hash)
1577                 return -ENOMEM;
1578         return 0;
1579 }
1580
1581 #if 0
1582 static void __init mm_slots_hash_free(void)
1583 {
1584         kfree(mm_slots_hash);
1585         mm_slots_hash = NULL;
1586 }
1587 #endif
1588
1589 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1590 {
1591         struct mm_slot *mm_slot;
1592         struct hlist_head *bucket;
1593         struct hlist_node *node;
1594
1595         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1596                                 % MM_SLOTS_HASH_HEADS];
1597         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1598                 if (mm == mm_slot->mm)
1599                         return mm_slot;
1600         }
1601         return NULL;
1602 }
1603
1604 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1605                                     struct mm_slot *mm_slot)
1606 {
1607         struct hlist_head *bucket;
1608
1609         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1610                                 % MM_SLOTS_HASH_HEADS];
1611         mm_slot->mm = mm;
1612         hlist_add_head(&mm_slot->hash, bucket);
1613 }
1614
1615 static inline int khugepaged_test_exit(struct mm_struct *mm)
1616 {
1617         return atomic_read(&mm->mm_users) == 0;
1618 }
1619
1620 int __khugepaged_enter(struct mm_struct *mm)
1621 {
1622         struct mm_slot *mm_slot;
1623         int wakeup;
1624
1625         mm_slot = alloc_mm_slot();
1626         if (!mm_slot)
1627                 return -ENOMEM;
1628
1629         /* __khugepaged_exit() must not run from under us */
1630         VM_BUG_ON(khugepaged_test_exit(mm));
1631         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1632                 free_mm_slot(mm_slot);
1633                 return 0;
1634         }
1635
1636         spin_lock(&khugepaged_mm_lock);
1637         insert_to_mm_slots_hash(mm, mm_slot);
1638         /*
1639          * Insert just behind the scanning cursor, to let the area settle
1640          * down a little.
1641          */
1642         wakeup = list_empty(&khugepaged_scan.mm_head);
1643         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1644         spin_unlock(&khugepaged_mm_lock);
1645
1646         atomic_inc(&mm->mm_count);
1647         if (wakeup)
1648                 wake_up_interruptible(&khugepaged_wait);
1649
1650         return 0;
1651 }
1652
1653 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1654 {
1655         unsigned long hstart, hend;
1656         if (!vma->anon_vma)
1657                 /*
1658                  * Not yet faulted in so we will register later in the
1659                  * page fault if needed.
1660                  */
1661                 return 0;
1662         if (vma->vm_ops)
1663                 /* khugepaged not yet working on file or special mappings */
1664                 return 0;
1665         /*
1666          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1667          * true too, verify it here.
1668          */
1669         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1670         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1671         hend = vma->vm_end & HPAGE_PMD_MASK;
1672         if (hstart < hend)
1673                 return khugepaged_enter(vma);
1674         return 0;
1675 }
1676
1677 void __khugepaged_exit(struct mm_struct *mm)
1678 {
1679         struct mm_slot *mm_slot;
1680         int free = 0;
1681
1682         spin_lock(&khugepaged_mm_lock);
1683         mm_slot = get_mm_slot(mm);
1684         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1685                 hlist_del(&mm_slot->hash);
1686                 list_del(&mm_slot->mm_node);
1687                 free = 1;
1688         }
1689         spin_unlock(&khugepaged_mm_lock);
1690
1691         if (free) {
1692                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1693                 free_mm_slot(mm_slot);
1694                 mmdrop(mm);
1695         } else if (mm_slot) {
1696                 /*
1697                  * This is required to serialize against
1698                  * khugepaged_test_exit() (which is guaranteed to run
1699                  * under mmap sem read mode). Stop here (after we
1700                  * return all pagetables will be destroyed) until
1701                  * khugepaged has finished working on the pagetables
1702                  * under the mmap_sem.
1703                  */
1704                 down_write(&mm->mmap_sem);
1705                 up_write(&mm->mmap_sem);
1706         }
1707 }
1708
1709 static void release_pte_page(struct page *page)
1710 {
1711         /* 0 stands for page_is_file_cache(page) == false */
1712         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1713         unlock_page(page);
1714         putback_lru_page(page);
1715 }
1716
1717 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1718 {
1719         while (--_pte >= pte) {
1720                 pte_t pteval = *_pte;
1721                 if (!pte_none(pteval))
1722                         release_pte_page(pte_page(pteval));
1723         }
1724 }
1725
1726 static void release_all_pte_pages(pte_t *pte)
1727 {
1728         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1729 }
1730
1731 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1732                                         unsigned long address,
1733                                         pte_t *pte)
1734 {
1735         struct page *page;
1736         pte_t *_pte;
1737         int referenced = 0, isolated = 0, none = 0;
1738         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1739              _pte++, address += PAGE_SIZE) {
1740                 pte_t pteval = *_pte;
1741                 if (pte_none(pteval)) {
1742                         if (++none <= khugepaged_max_ptes_none)
1743                                 continue;
1744                         else {
1745                                 release_pte_pages(pte, _pte);
1746                                 goto out;
1747                         }
1748                 }
1749                 if (!pte_present(pteval) || !pte_write(pteval)) {
1750                         release_pte_pages(pte, _pte);
1751                         goto out;
1752                 }
1753                 page = vm_normal_page(vma, address, pteval);
1754                 if (unlikely(!page)) {
1755                         release_pte_pages(pte, _pte);
1756                         goto out;
1757                 }
1758                 VM_BUG_ON(PageCompound(page));
1759                 BUG_ON(!PageAnon(page));
1760                 VM_BUG_ON(!PageSwapBacked(page));
1761
1762                 /* cannot use mapcount: can't collapse if there's a gup pin */
1763                 if (page_count(page) != 1) {
1764                         release_pte_pages(pte, _pte);
1765                         goto out;
1766                 }
1767                 /*
1768                  * We can do it before isolate_lru_page because the
1769                  * page can't be freed from under us. NOTE: PG_lock
1770                  * is needed to serialize against split_huge_page
1771                  * when invoked from the VM.
1772                  */
1773                 if (!trylock_page(page)) {
1774                         release_pte_pages(pte, _pte);
1775                         goto out;
1776                 }
1777                 /*
1778                  * Isolate the page to avoid collapsing an hugepage
1779                  * currently in use by the VM.
1780                  */
1781                 if (isolate_lru_page(page)) {
1782                         unlock_page(page);
1783                         release_pte_pages(pte, _pte);
1784                         goto out;
1785                 }
1786                 /* 0 stands for page_is_file_cache(page) == false */
1787                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1788                 VM_BUG_ON(!PageLocked(page));
1789                 VM_BUG_ON(PageLRU(page));
1790
1791                 /* If there is no mapped pte young don't collapse the page */
1792                 if (pte_young(pteval) || PageReferenced(page) ||
1793                     mmu_notifier_test_young(vma->vm_mm, address))
1794                         referenced = 1;
1795         }
1796         if (unlikely(!referenced))
1797                 release_all_pte_pages(pte);
1798         else
1799                 isolated = 1;
1800 out:
1801         return isolated;
1802 }
1803
1804 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1805                                       struct vm_area_struct *vma,
1806                                       unsigned long address,
1807                                       spinlock_t *ptl)
1808 {
1809         pte_t *_pte;
1810         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1811                 pte_t pteval = *_pte;
1812                 struct page *src_page;
1813
1814                 if (pte_none(pteval)) {
1815                         clear_user_highpage(page, address);
1816                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1817                 } else {
1818                         src_page = pte_page(pteval);
1819                         copy_user_highpage(page, src_page, address, vma);
1820                         VM_BUG_ON(page_mapcount(src_page) != 1);
1821                         VM_BUG_ON(page_count(src_page) != 2);
1822                         release_pte_page(src_page);
1823                         /*
1824                          * ptl mostly unnecessary, but preempt has to
1825                          * be disabled to update the per-cpu stats
1826                          * inside page_remove_rmap().
1827                          */
1828                         spin_lock(ptl);
1829                         /*
1830                          * paravirt calls inside pte_clear here are
1831                          * superfluous.
1832                          */
1833                         pte_clear(vma->vm_mm, address, _pte);
1834                         page_remove_rmap(src_page);
1835                         spin_unlock(ptl);
1836                         free_page_and_swap_cache(src_page);
1837                 }
1838
1839                 address += PAGE_SIZE;
1840                 page++;
1841         }
1842 }
1843
1844 static void collapse_huge_page(struct mm_struct *mm,
1845                                unsigned long address,
1846                                struct page **hpage,
1847                                struct vm_area_struct *vma,
1848                                int node)
1849 {
1850         pgd_t *pgd;
1851         pud_t *pud;
1852         pmd_t *pmd, _pmd;
1853         pte_t *pte;
1854         pgtable_t pgtable;
1855         struct page *new_page;
1856         spinlock_t *ptl;
1857         int isolated;
1858         unsigned long hstart, hend;
1859
1860         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1861 #ifndef CONFIG_NUMA
1862         up_read(&mm->mmap_sem);
1863         VM_BUG_ON(!*hpage);
1864         new_page = *hpage;
1865 #else
1866         VM_BUG_ON(*hpage);
1867         /*
1868          * Allocate the page while the vma is still valid and under
1869          * the mmap_sem read mode so there is no memory allocation
1870          * later when we take the mmap_sem in write mode. This is more
1871          * friendly behavior (OTOH it may actually hide bugs) to
1872          * filesystems in userland with daemons allocating memory in
1873          * the userland I/O paths.  Allocating memory with the
1874          * mmap_sem in read mode is good idea also to allow greater
1875          * scalability.
1876          */
1877         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1878                                       node, __GFP_OTHER_NODE);
1879
1880         /*
1881          * After allocating the hugepage, release the mmap_sem read lock in
1882          * preparation for taking it in write mode.
1883          */
1884         up_read(&mm->mmap_sem);
1885         if (unlikely(!new_page)) {
1886                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1887                 *hpage = ERR_PTR(-ENOMEM);
1888                 return;
1889         }
1890 #endif
1891
1892         count_vm_event(THP_COLLAPSE_ALLOC);
1893         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1894 #ifdef CONFIG_NUMA
1895                 put_page(new_page);
1896 #endif
1897                 return;
1898         }
1899
1900         /*
1901          * Prevent all access to pagetables with the exception of
1902          * gup_fast later hanlded by the ptep_clear_flush and the VM
1903          * handled by the anon_vma lock + PG_lock.
1904          */
1905         down_write(&mm->mmap_sem);
1906         if (unlikely(khugepaged_test_exit(mm)))
1907                 goto out;
1908
1909         vma = find_vma(mm, address);
1910         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1911         hend = vma->vm_end & HPAGE_PMD_MASK;
1912         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1913                 goto out;
1914
1915         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1916             (vma->vm_flags & VM_NOHUGEPAGE))
1917                 goto out;
1918
1919         if (!vma->anon_vma || vma->vm_ops)
1920                 goto out;
1921         if (is_vma_temporary_stack(vma))
1922                 goto out;
1923         /*
1924          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1925          * true too, verify it here.
1926          */
1927         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1928
1929         pgd = pgd_offset(mm, address);
1930         if (!pgd_present(*pgd))
1931                 goto out;
1932
1933         pud = pud_offset(pgd, address);
1934         if (!pud_present(*pud))
1935                 goto out;
1936
1937         pmd = pmd_offset(pud, address);
1938         /* pmd can't go away or become huge under us */
1939         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1940                 goto out;
1941
1942         anon_vma_lock(vma->anon_vma);
1943
1944         pte = pte_offset_map(pmd, address);
1945         ptl = pte_lockptr(mm, pmd);
1946
1947         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1948         /*
1949          * After this gup_fast can't run anymore. This also removes
1950          * any huge TLB entry from the CPU so we won't allow
1951          * huge and small TLB entries for the same virtual address
1952          * to avoid the risk of CPU bugs in that area.
1953          */
1954         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1955         spin_unlock(&mm->page_table_lock);
1956
1957         spin_lock(ptl);
1958         isolated = __collapse_huge_page_isolate(vma, address, pte);
1959         spin_unlock(ptl);
1960
1961         if (unlikely(!isolated)) {
1962                 pte_unmap(pte);
1963                 spin_lock(&mm->page_table_lock);
1964                 BUG_ON(!pmd_none(*pmd));
1965                 set_pmd_at(mm, address, pmd, _pmd);
1966                 spin_unlock(&mm->page_table_lock);
1967                 anon_vma_unlock(vma->anon_vma);
1968                 goto out;
1969         }
1970
1971         /*
1972          * All pages are isolated and locked so anon_vma rmap
1973          * can't run anymore.
1974          */
1975         anon_vma_unlock(vma->anon_vma);
1976
1977         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1978         pte_unmap(pte);
1979         __SetPageUptodate(new_page);
1980         pgtable = pmd_pgtable(_pmd);
1981         VM_BUG_ON(page_count(pgtable) != 1);
1982         VM_BUG_ON(page_mapcount(pgtable) != 0);
1983
1984         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1985         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1986         _pmd = pmd_mkhuge(_pmd);
1987
1988         /*
1989          * spin_lock() below is not the equivalent of smp_wmb(), so
1990          * this is needed to avoid the copy_huge_page writes to become
1991          * visible after the set_pmd_at() write.
1992          */
1993         smp_wmb();
1994
1995         spin_lock(&mm->page_table_lock);
1996         BUG_ON(!pmd_none(*pmd));
1997         page_add_new_anon_rmap(new_page, vma, address);
1998         set_pmd_at(mm, address, pmd, _pmd);
1999         update_mmu_cache(vma, address, _pmd);
2000         prepare_pmd_huge_pte(pgtable, mm);
2001         mm->nr_ptes--;
2002         spin_unlock(&mm->page_table_lock);
2003
2004 #ifndef CONFIG_NUMA
2005         *hpage = NULL;
2006 #endif
2007         khugepaged_pages_collapsed++;
2008 out_up_write:
2009         up_write(&mm->mmap_sem);
2010         return;
2011
2012 out:
2013         mem_cgroup_uncharge_page(new_page);
2014 #ifdef CONFIG_NUMA
2015         put_page(new_page);
2016 #endif
2017         goto out_up_write;
2018 }
2019
2020 static int khugepaged_scan_pmd(struct mm_struct *mm,
2021                                struct vm_area_struct *vma,
2022                                unsigned long address,
2023                                struct page **hpage)
2024 {
2025         pgd_t *pgd;
2026         pud_t *pud;
2027         pmd_t *pmd;
2028         pte_t *pte, *_pte;
2029         int ret = 0, referenced = 0, none = 0;
2030         struct page *page;
2031         unsigned long _address;
2032         spinlock_t *ptl;
2033         int node = -1;
2034
2035         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2036
2037         pgd = pgd_offset(mm, address);
2038         if (!pgd_present(*pgd))
2039                 goto out;
2040
2041         pud = pud_offset(pgd, address);
2042         if (!pud_present(*pud))
2043                 goto out;
2044
2045         pmd = pmd_offset(pud, address);
2046         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2047                 goto out;
2048
2049         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2050         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2051              _pte++, _address += PAGE_SIZE) {
2052                 pte_t pteval = *_pte;
2053                 if (pte_none(pteval)) {
2054                         if (++none <= khugepaged_max_ptes_none)
2055                                 continue;
2056                         else
2057                                 goto out_unmap;
2058                 }
2059                 if (!pte_present(pteval) || !pte_write(pteval))
2060                         goto out_unmap;
2061                 page = vm_normal_page(vma, _address, pteval);
2062                 if (unlikely(!page))
2063                         goto out_unmap;
2064                 /*
2065                  * Chose the node of the first page. This could
2066                  * be more sophisticated and look at more pages,
2067                  * but isn't for now.
2068                  */
2069                 if (node == -1)
2070                         node = page_to_nid(page);
2071                 VM_BUG_ON(PageCompound(page));
2072                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2073                         goto out_unmap;
2074                 /* cannot use mapcount: can't collapse if there's a gup pin */
2075                 if (page_count(page) != 1)
2076                         goto out_unmap;
2077                 if (pte_young(pteval) || PageReferenced(page) ||
2078                     mmu_notifier_test_young(vma->vm_mm, address))
2079                         referenced = 1;
2080         }
2081         if (referenced)
2082                 ret = 1;
2083 out_unmap:
2084         pte_unmap_unlock(pte, ptl);
2085         if (ret)
2086                 /* collapse_huge_page will return with the mmap_sem released */
2087                 collapse_huge_page(mm, address, hpage, vma, node);
2088 out:
2089         return ret;
2090 }
2091
2092 static void collect_mm_slot(struct mm_slot *mm_slot)
2093 {
2094         struct mm_struct *mm = mm_slot->mm;
2095
2096         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2097
2098         if (khugepaged_test_exit(mm)) {
2099                 /* free mm_slot */
2100                 hlist_del(&mm_slot->hash);
2101                 list_del(&mm_slot->mm_node);
2102
2103                 /*
2104                  * Not strictly needed because the mm exited already.
2105                  *
2106                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2107                  */
2108
2109                 /* khugepaged_mm_lock actually not necessary for the below */
2110                 free_mm_slot(mm_slot);
2111                 mmdrop(mm);
2112         }
2113 }
2114
2115 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2116                                             struct page **hpage)
2117         __releases(&khugepaged_mm_lock)
2118         __acquires(&khugepaged_mm_lock)
2119 {
2120         struct mm_slot *mm_slot;
2121         struct mm_struct *mm;
2122         struct vm_area_struct *vma;
2123         int progress = 0;
2124
2125         VM_BUG_ON(!pages);
2126         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2127
2128         if (khugepaged_scan.mm_slot)
2129                 mm_slot = khugepaged_scan.mm_slot;
2130         else {
2131                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2132                                      struct mm_slot, mm_node);
2133                 khugepaged_scan.address = 0;
2134                 khugepaged_scan.mm_slot = mm_slot;
2135         }
2136         spin_unlock(&khugepaged_mm_lock);
2137
2138         mm = mm_slot->mm;
2139         down_read(&mm->mmap_sem);
2140         if (unlikely(khugepaged_test_exit(mm)))
2141                 vma = NULL;
2142         else
2143                 vma = find_vma(mm, khugepaged_scan.address);
2144
2145         progress++;
2146         for (; vma; vma = vma->vm_next) {
2147                 unsigned long hstart, hend;
2148
2149                 cond_resched();
2150                 if (unlikely(khugepaged_test_exit(mm))) {
2151                         progress++;
2152                         break;
2153                 }
2154
2155                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2156                      !khugepaged_always()) ||
2157                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2158                 skip:
2159                         progress++;
2160                         continue;
2161                 }
2162                 if (!vma->anon_vma || vma->vm_ops)
2163                         goto skip;
2164                 if (is_vma_temporary_stack(vma))
2165                         goto skip;
2166                 /*
2167                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2168                  * must be true too, verify it here.
2169                  */
2170                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2171                           vma->vm_flags & VM_NO_THP);
2172
2173                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2174                 hend = vma->vm_end & HPAGE_PMD_MASK;
2175                 if (hstart >= hend)
2176                         goto skip;
2177                 if (khugepaged_scan.address > hend)
2178                         goto skip;
2179                 if (khugepaged_scan.address < hstart)
2180                         khugepaged_scan.address = hstart;
2181                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2182
2183                 while (khugepaged_scan.address < hend) {
2184                         int ret;
2185                         cond_resched();
2186                         if (unlikely(khugepaged_test_exit(mm)))
2187                                 goto breakouterloop;
2188
2189                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2190                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2191                                   hend);
2192                         ret = khugepaged_scan_pmd(mm, vma,
2193                                                   khugepaged_scan.address,
2194                                                   hpage);
2195                         /* move to next address */
2196                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2197                         progress += HPAGE_PMD_NR;
2198                         if (ret)
2199                                 /* we released mmap_sem so break loop */
2200                                 goto breakouterloop_mmap_sem;
2201                         if (progress >= pages)
2202                                 goto breakouterloop;
2203                 }
2204         }
2205 breakouterloop:
2206         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2207 breakouterloop_mmap_sem:
2208
2209         spin_lock(&khugepaged_mm_lock);
2210         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2211         /*
2212          * Release the current mm_slot if this mm is about to die, or
2213          * if we scanned all vmas of this mm.
2214          */
2215         if (khugepaged_test_exit(mm) || !vma) {
2216                 /*
2217                  * Make sure that if mm_users is reaching zero while
2218                  * khugepaged runs here, khugepaged_exit will find
2219                  * mm_slot not pointing to the exiting mm.
2220                  */
2221                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2222                         khugepaged_scan.mm_slot = list_entry(
2223                                 mm_slot->mm_node.next,
2224                                 struct mm_slot, mm_node);
2225                         khugepaged_scan.address = 0;
2226                 } else {
2227                         khugepaged_scan.mm_slot = NULL;
2228                         khugepaged_full_scans++;
2229                 }
2230
2231                 collect_mm_slot(mm_slot);
2232         }
2233
2234         return progress;
2235 }
2236
2237 static int khugepaged_has_work(void)
2238 {
2239         return !list_empty(&khugepaged_scan.mm_head) &&
2240                 khugepaged_enabled();
2241 }
2242
2243 static int khugepaged_wait_event(void)
2244 {
2245         return !list_empty(&khugepaged_scan.mm_head) ||
2246                 !khugepaged_enabled();
2247 }
2248
2249 static void khugepaged_do_scan(struct page **hpage)
2250 {
2251         unsigned int progress = 0, pass_through_head = 0;
2252         unsigned int pages = khugepaged_pages_to_scan;
2253
2254         barrier(); /* write khugepaged_pages_to_scan to local stack */
2255
2256         while (progress < pages) {
2257                 cond_resched();
2258
2259 #ifndef CONFIG_NUMA
2260                 if (!*hpage) {
2261                         *hpage = alloc_hugepage(khugepaged_defrag());
2262                         if (unlikely(!*hpage)) {
2263                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2264                                 break;
2265                         }
2266                         count_vm_event(THP_COLLAPSE_ALLOC);
2267                 }
2268 #else
2269                 if (IS_ERR(*hpage))
2270                         break;
2271 #endif
2272
2273                 if (unlikely(kthread_should_stop() || freezing(current)))
2274                         break;
2275
2276                 spin_lock(&khugepaged_mm_lock);
2277                 if (!khugepaged_scan.mm_slot)
2278                         pass_through_head++;
2279                 if (khugepaged_has_work() &&
2280                     pass_through_head < 2)
2281                         progress += khugepaged_scan_mm_slot(pages - progress,
2282                                                             hpage);
2283                 else
2284                         progress = pages;
2285                 spin_unlock(&khugepaged_mm_lock);
2286         }
2287 }
2288
2289 static void khugepaged_alloc_sleep(void)
2290 {
2291         wait_event_freezable_timeout(khugepaged_wait, false,
2292                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2293 }
2294
2295 #ifndef CONFIG_NUMA
2296 static struct page *khugepaged_alloc_hugepage(void)
2297 {
2298         struct page *hpage;
2299
2300         do {
2301                 hpage = alloc_hugepage(khugepaged_defrag());
2302                 if (!hpage) {
2303                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2304                         khugepaged_alloc_sleep();
2305                 } else
2306                         count_vm_event(THP_COLLAPSE_ALLOC);
2307         } while (unlikely(!hpage) &&
2308                  likely(khugepaged_enabled()));
2309         return hpage;
2310 }
2311 #endif
2312
2313 static void khugepaged_loop(void)
2314 {
2315         struct page *hpage;
2316
2317 #ifdef CONFIG_NUMA
2318         hpage = NULL;
2319 #endif
2320         while (likely(khugepaged_enabled())) {
2321 #ifndef CONFIG_NUMA
2322                 hpage = khugepaged_alloc_hugepage();
2323                 if (unlikely(!hpage))
2324                         break;
2325 #else
2326                 if (IS_ERR(hpage)) {
2327                         khugepaged_alloc_sleep();
2328                         hpage = NULL;
2329                 }
2330 #endif
2331
2332                 khugepaged_do_scan(&hpage);
2333 #ifndef CONFIG_NUMA
2334                 if (hpage)
2335                         put_page(hpage);
2336 #endif
2337                 try_to_freeze();
2338                 if (unlikely(kthread_should_stop()))
2339                         break;
2340                 if (khugepaged_has_work()) {
2341                         if (!khugepaged_scan_sleep_millisecs)
2342                                 continue;
2343                         wait_event_freezable_timeout(khugepaged_wait, false,
2344                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2345                 } else if (khugepaged_enabled())
2346                         wait_event_freezable(khugepaged_wait,
2347                                              khugepaged_wait_event());
2348         }
2349 }
2350
2351 static int khugepaged(void *none)
2352 {
2353         struct mm_slot *mm_slot;
2354
2355         set_freezable();
2356         set_user_nice(current, 19);
2357
2358         /* serialize with start_khugepaged() */
2359         mutex_lock(&khugepaged_mutex);
2360
2361         for (;;) {
2362                 mutex_unlock(&khugepaged_mutex);
2363                 VM_BUG_ON(khugepaged_thread != current);
2364                 khugepaged_loop();
2365                 VM_BUG_ON(khugepaged_thread != current);
2366
2367                 mutex_lock(&khugepaged_mutex);
2368                 if (!khugepaged_enabled())
2369                         break;
2370                 if (unlikely(kthread_should_stop()))
2371                         break;
2372         }
2373
2374         spin_lock(&khugepaged_mm_lock);
2375         mm_slot = khugepaged_scan.mm_slot;
2376         khugepaged_scan.mm_slot = NULL;
2377         if (mm_slot)
2378                 collect_mm_slot(mm_slot);
2379         spin_unlock(&khugepaged_mm_lock);
2380
2381         khugepaged_thread = NULL;
2382         mutex_unlock(&khugepaged_mutex);
2383
2384         return 0;
2385 }
2386
2387 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2388 {
2389         struct page *page;
2390
2391         spin_lock(&mm->page_table_lock);
2392         if (unlikely(!pmd_trans_huge(*pmd))) {
2393                 spin_unlock(&mm->page_table_lock);
2394                 return;
2395         }
2396         page = pmd_page(*pmd);
2397         VM_BUG_ON(!page_count(page));
2398         get_page(page);
2399         spin_unlock(&mm->page_table_lock);
2400
2401         split_huge_page(page);
2402
2403         put_page(page);
2404         BUG_ON(pmd_trans_huge(*pmd));
2405 }
2406
2407 static void split_huge_page_address(struct mm_struct *mm,
2408                                     unsigned long address)
2409 {
2410         pgd_t *pgd;
2411         pud_t *pud;
2412         pmd_t *pmd;
2413
2414         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2415
2416         pgd = pgd_offset(mm, address);
2417         if (!pgd_present(*pgd))
2418                 return;
2419
2420         pud = pud_offset(pgd, address);
2421         if (!pud_present(*pud))
2422                 return;
2423
2424         pmd = pmd_offset(pud, address);
2425         if (!pmd_present(*pmd))
2426                 return;
2427         /*
2428          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2429          * materialize from under us.
2430          */
2431         split_huge_page_pmd(mm, pmd);
2432 }
2433
2434 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2435                              unsigned long start,
2436                              unsigned long end,
2437                              long adjust_next)
2438 {
2439         /*
2440          * If the new start address isn't hpage aligned and it could
2441          * previously contain an hugepage: check if we need to split
2442          * an huge pmd.
2443          */
2444         if (start & ~HPAGE_PMD_MASK &&
2445             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2446             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2447                 split_huge_page_address(vma->vm_mm, start);
2448
2449         /*
2450          * If the new end address isn't hpage aligned and it could
2451          * previously contain an hugepage: check if we need to split
2452          * an huge pmd.
2453          */
2454         if (end & ~HPAGE_PMD_MASK &&
2455             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2456             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2457                 split_huge_page_address(vma->vm_mm, end);
2458
2459         /*
2460          * If we're also updating the vma->vm_next->vm_start, if the new
2461          * vm_next->vm_start isn't page aligned and it could previously
2462          * contain an hugepage: check if we need to split an huge pmd.
2463          */
2464         if (adjust_next > 0) {
2465                 struct vm_area_struct *next = vma->vm_next;
2466                 unsigned long nstart = next->vm_start;
2467                 nstart += adjust_next << PAGE_SHIFT;
2468                 if (nstart & ~HPAGE_PMD_MASK &&
2469                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2470                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2471                         split_huge_page_address(next->vm_mm, nstart);
2472         }
2473 }